# completion

Published articles for completion.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Generalist Meets Specialists: An Evolving Multimodal Agent with Persistent Tool Orchestration for Traffic Scenario Understanding

DevFeed: [Generalist Meets Specialists: An Evolving Multimodal Agent with Persistent Tool Orchestration for Traffic Scenario Understanding](<https://devfeed.tech/articles/generalist-meets-specialists-an-evolving-multimodal-agent-with-persistent-tool-orchestration-for-traffic-scenario-understanding-145597.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25868>)

Author: Linjia Kang, Binhan Jiang, Juan Yang, Zhiqin Song, Xiangsen Wu, Qionglin Ren, Jiang Liu, Zhongjiang He, Hao Sun

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [agentic workflows](<https://devfeed.tech/topics/agentic-workflows.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [adoption](<https://devfeed.tech/tags/adoption.md>), [agent](<https://devfeed.tech/tags/agent.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [challenges](<https://devfeed.tech/tags/challenges.md>), [collaboration](<https://devfeed.tech/tags/collaboration.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cost-savings](<https://devfeed.tech/tags/cost-savings.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [error](<https://devfeed.tech/tags/error.md>), [evolution](<https://devfeed.tech/tags/evolution.md>), [execution](<https://devfeed.tech/tags/execution.md>), [fast](<https://devfeed.tech/tags/fast.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [generalist](<https://devfeed.tech/tags/generalist.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [orchestration](<https://devfeed.tech/tags/orchestration.md>), [specialists](<https://devfeed.tech/tags/specialists.md>), [tool](<https://devfeed.tech/tags/tool.md>), [traffic](<https://devfeed.tech/tags/traffic.md>)

### AI overview

StarTV-Agent combines policy learning with reusable tool workflows to balance adaptive reasoning with lower-cost handling of recurring traffic-understanding tasks. During training, Role-Sensitive Process Optimization combines process and terminal advantages, while Weakness-Driven Demonstration Learning uses validated teacher examples aimed at the student's weak questions. At inference, Lightweight Adaptive Workflow Reuse selects a suitable workflow and binds task-specific arguments and resource references; tasks without a suitable workflow use a slower ReAct-based path. Label-free feedback from execution completion, workflow adoption, and cost savings guides reuse and memory updates. Across three benchmarks, the abstract reports up to 8.0% relative performance improvement and up to 52.1% lower token consumption.

### Source excerpt

arXiv:2611.25868v1 Announce Type: new Abstract: Traffic scenario understanding requires combining fine-grained visual evidence with spatial and temporal reasoning. Tool-augmented vision-language models (VLMs) can use specialized perception tools, but effective collaboration faces two challenges: terminal rewards provide limited guidance for intermediate decisions and error recovery, while repeated planning adds unnecessary cost across similar tasks. Echoing the fast-slow dichotomy of human cognition, we propose StarTV-Agent, which combines policy learning with experience evolution to support adaptive reasoning when needed and efficient reuse of prior experience when possible. During training, Role-Sensitive Process Optimization (RSPO) combines process advantages with terminal advantages, while Weakness-Driven Demonstration Learning (WDL) provides validated teacher demonstrations targeting the student's weak questions. During inference, Lightweight Adaptive Workflow Reuse (LAWR) enables a fast path by selecting suitable workflows and binding task-specific arguments and resource references, while a slow ReAct-based path handles tasks without suitable workflows. We further design label-free feedback based on execution completion, workflow adoption, and cost savings to guide workflow reuse and memory updates. Across three challenging benchmarks, StarTV-Agent achieves up to 8.0% relative performance improvement, while reducing token consumption by up to 52.1%.

## ResearchFlow-Eval: Tracing the Execution Gap in End-to-End Multi-Claim Autoresearch

DevFeed: [ResearchFlow-Eval: Tracing the Execution Gap in End-to-End Multi-Claim Autoresearch](<https://devfeed.tech/articles/researchflow-eval-tracing-the-execution-gap-in-end-to-end-multi-claim-autoresearch-144962.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.18088>)

Author: Peng Jiang, Dongyang Li, Quanying Liu, Xiaoxuan Jia

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [hierarchical agent systems](<https://devfeed.tech/topics/hierarchical-agent-systems.md>), [Orchestration](<https://devfeed.tech/topics/orchestration.md>), [DeepFact-Bench](<https://devfeed.tech/topics/deepfact-bench.md>), [iteration](<https://devfeed.tech/topics/iteration.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-harnesses](<https://devfeed.tech/tags/agent-harnesses.md>), [anchor](<https://devfeed.tech/tags/anchor.md>), [artifacts](<https://devfeed.tech/tags/artifacts.md>), [autoresearch](<https://devfeed.tech/tags/autoresearch.md>), [capabilities](<https://devfeed.tech/tags/capabilities.md>), [completion](<https://devfeed.tech/tags/completion.md>), [continuous](<https://devfeed.tech/tags/continuous.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [design](<https://devfeed.tech/tags/design.md>), [dimensions](<https://devfeed.tech/tags/dimensions.md>), [eval](<https://devfeed.tech/tags/eval.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [execution](<https://devfeed.tech/tags/execution.md>), [experimental](<https://devfeed.tech/tags/experimental.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [framework](<https://devfeed.tech/tags/framework.md>), [gap](<https://devfeed.tech/tags/gap.md>), [improvements](<https://devfeed.tech/tags/improvements.md>), [tracing](<https://devfeed.tech/tags/tracing.md>)

### AI overview

ResearchFlow-Eval assesses whether autoresearch systems can carry multiple interdependent claims through continuous research stages and revise experiments as evidence accumulates. It derives stage-specific rubrics from published anchor papers and scores work at both stage and claim levels. Across the evaluated general-purpose, skill-augmented, and research-orchestrated systems, experimental execution scored below design and planning. Systems often struggled to produce high-quality results that met claim requirements, and repeated attempts rarely improved scores. Skill augmentation and research-specific orchestration did not reliably raise performance. Results are averaged over two runs per system and scored by two LLM judges.

### Source excerpt

arXiv:2610.18088v3 Announce Type: replace-cross Abstract: Autoresearch systems increasingly undertake entire research projects. In practice, scientific research requires systems to pursue multiple interdependent claims through a continuous, multi-stage workflow and revise experiments as evidence accumulates. Existing evaluations of autoresearch systems, however, typically score isolated tasks or final artifacts, leaving these process-level capabilities largely unmeasured. We therefore introduce ResearchFlow-Eval, a fine-grained framework for assessing continuous scientific execution and iteration at both the stage and claim levels. Each evaluation is grounded in a published anchor paper, from which we derive stage-specific rubrics that decompose each research stage into dimensions and separately scored rubric items. Using these rubrics, we score each research stage and examine experimental execution in detail, focusing on multi-claim completion and revisions across attempts. We apply the evaluation framework to general-purpose agent harnesses, skill-augmented harnesses, and systems with research-specific orchestration. Averaged over two runs per system, each scored by two LLM judges, experimental execution lags behind design and planning for every evaluated system. Within execution, systems struggle to produce high-quality results and meet claim requirements, and repeated attempts rarely yield measurable improvements. Moreover, neither skill augmentation nor research-specific orchestration reliably yields higher scores. ResearchFlow-Eval thus extends autoresearch evaluation beyond isolated, single-objective runs to continuous workflows in which systems pursue multiple claims and refine their work across iterations.

## Should Belgians Install an Advanced Energy Management System in Their Electrified Dwelling ?

DevFeed: [Should Belgians Install an Advanced Energy Management System in Their Electrified Dwelling ?](<https://devfeed.tech/articles/should-belgians-install-an-advanced-energy-management-system-in-their-electrified-dwelling-145182.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26441>)

Author: Julian Ruddick, Cedric De Cauwer, Thierry Coosemans, Maarten Messagie

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [belgium](<https://devfeed.tech/tags/belgium.md>), [benefits](<https://devfeed.tech/tags/benefits.md>), [completion](<https://devfeed.tech/tags/completion.md>), [control](<https://devfeed.tech/tags/control.md>), [costs](<https://devfeed.tech/tags/costs.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-sy](<https://devfeed.tech/tags/cs-sy.md>), [decision-trees](<https://devfeed.tech/tags/decision-trees.md>), [eess-sy](<https://devfeed.tech/tags/eess-sy.md>), [electric](<https://devfeed.tech/tags/electric.md>), [electric-vehicles](<https://devfeed.tech/tags/electric-vehicles.md>), [energy](<https://devfeed.tech/tags/energy.md>), [heat-pumps](<https://devfeed.tech/tags/heat-pumps.md>), [learning](<https://devfeed.tech/tags/learning.md>), [load](<https://devfeed.tech/tags/load.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [mpc](<https://devfeed.tech/tags/mpc.md>), [performance](<https://devfeed.tech/tags/performance.md>), [prices](<https://devfeed.tech/tags/prices.md>)

### AI overview

A simulation of 500 representative Belgian dwellings under a 2023 contract with hourly prices and a monthly peak tariff found modest savings from advanced energy management. Against a rule-based controller, annual average savings were EUR 82.50 for TreeC, EUR 116.60 for model predictive control (MPC), and EUR 177.40 for MPC with perfect forecasts, used as an upper bound. Thermal comfort was similar across systems (26-31 Kh). MPC saved more, while TreeC achieved higher EV charge completion (93.3% versus 88.8% for MPC, with a 95% maximum) and a smoother aggregate load. Estimated purchase costs (EUR 280-680) and subscriptions (EUR 0-360 per year) make the savings unattractive for widespread adoption. The findings reflect the simulated dwellings and stated pricing conditions.

### Source excerpt

arXiv:2611.26441v1 Announce Type: cross Abstract: As dwellings electrify and more complex electricity pricing spreads in Belgium, the question arises: What are the potential benefits of installing an advanced energy management systems in an electrified Belgian dwelling ? To answer this question, we simulate 500 representative Belgian dwellings in which fossil systems are replaced by controllable electric technologies: space heating via air-to-air heat pumps, domestic hot water via electric boilers, and private transport via electric vehicles with home chargers. Photovoltaic systems are assigned according to current adoption rates. We evaluate two advanced energy management systems (EMS): model predictive control (MPC) and TreeC, a machine learning method using decision trees to model the EMS strategy. These are compared to a rule-based controller (RBC) that mimics default operation and an MPC with perfect forecasts (MPC P) as an upper bound. Performance for these EMSs is assessed for 2023 under a real contract with hourly prices and a monthly peak tariff. Average annual savings versus RBC are 82.5 EUR, 116.6 EUR, and 177.4 EUR for TreeC, MPC, and MPC P, respectively. All EMSs deliver similar thermal comfort (26-31 Kh). MPC saves more, while TreeC achieves higher EV charge completion (93.3% vs 88.8% for MPC, with a max charge of 95%) and a smoother aggregate load across the 500 dwellings. Savings remain modest relative to EMS purchase (280-680 EUR) and subscription (0-360 EUR/year) costs which is not favourable towards widespread adoption. The data and code to reproduce this study are available at https://github.com/EVERGi/should_belgians_install_advanced_ems_paper.

## Context as a Resource

DevFeed: [Context as a Resource](<https://devfeed.tech/articles/context-as-a-resource-145662.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26346>)

Author: Xufei Tian, Han Hu, Wenli Du

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>), [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [scheduling](<https://devfeed.tech/topics/scheduling.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [blocks](<https://devfeed.tech/tags/blocks.md>), [bus](<https://devfeed.tech/tags/bus.md>), [cells](<https://devfeed.tech/tags/cells.md>), [completion](<https://devfeed.tech/tags/completion.md>), [concurrent](<https://devfeed.tech/tags/concurrent.md>), [constraint](<https://devfeed.tech/tags/constraint.md>), [context](<https://devfeed.tech/tags/context.md>), [continuous](<https://devfeed.tech/tags/continuous.md>), [cross](<https://devfeed.tech/tags/cross.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [events](<https://devfeed.tech/tags/events.md>), [gate](<https://devfeed.tech/tags/gate.md>), [generative](<https://devfeed.tech/tags/generative.md>), [handoffs](<https://devfeed.tech/tags/handoffs.md>), [interface](<https://devfeed.tech/tags/interface.md>), [linear](<https://devfeed.tech/tags/linear.md>), [resource](<https://devfeed.tech/tags/resource.md>)

### AI overview

ClonAct proposes routerless management of concurrent agent contexts, using a non-generative Context MMU, in-place Context Cells, a footprint-aware Synaptic Bus with an isolated trial merge gate, and a wave-flushed Shared Trunk. The abstract reports that the merge gate raises true-conflict repository merge success from a 33.3% three-run mean to 63.3%; ClonAct reaches that rate with 45.9% fewer prompt tokens and avoids router collapse on compact models. Across three multi-stream benchmarks and three model families, it also matches the best reported completion rate under continuous surges (95.0%), with the lowest median wait and no timed-out status checks. These are abstract-level claims; the supplied text gives no further benchmark methodology or results.

### Source excerpt

arXiv:2611.26346v1 Announce Type: new Abstract: Managing concurrent tasks across separate agent sessions quickly exhausts human attention, motivating always-on personal agents behind a single continuous interface. Yet when every context shares identical frozen backbone weights, funneling interleaved dialogues and external events into one linear session induces cross-topic constraint bleed and head-of-line stalling. Conversely, router-ticket hierarchies incur lossy prose handoffs, heavy routing overhead, and blind merges - collapsing true-conflict repository merges to a 33.3% three-run mean despite high isolated branch passes. Rather than scheduling anthropomorphic roles, we introduce ClonAct, a routerless architecture that virtualizes context blocks directly through a pluggable single-pass non-generative Context MMU, in-place Context Cells, a footprint-aware Synaptic Bus with an isolated trial merge gate, and a wave-flushed Shared Trunk. Across three programmatic multi-stream benchmarks and three model families, our portable trial merge gate lifts true-conflict merges to 63.3%, while ClonAct's routerless virtualization attains this ceiling with 45.9% fewer prompt tokens and prevents compact-model router collapse. Under continuous surge streams, ClonAct matches top completion (95.0%) with the lowest median wait and zero timed-out status checks.

## ScaleWire: An End-to-End Scale-Up Ethernet Simulation Framework for LLM Systems

DevFeed: [ScaleWire: An End-to-End Scale-Up Ethernet Simulation Framework for LLM Systems](<https://devfeed.tech/articles/scalewire-an-end-to-end-scale-up-ethernet-simulation-framework-for-llm-systems-146150.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26086>)

Author: Kai Ma, Quanfeng Lv, Bowei Dai, Jingguo Ge, Kefan Ruan, Ruoqi Li

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Networking](<https://devfeed.tech/sources/arxiv-networking.md>)

Topics: [Ethernet](<https://devfeed.tech/topics/ethernet.md>), [Simulation](<https://devfeed.tech/topics/simulation.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>)

Tags: [accelerators](<https://devfeed.tech/tags/accelerators.md>), [balance](<https://devfeed.tech/tags/balance.md>), [bandwidth](<https://devfeed.tech/tags/bandwidth.md>), [budget](<https://devfeed.tech/tags/budget.md>), [communication](<https://devfeed.tech/tags/communication.md>), [comparison](<https://devfeed.tech/tags/comparison.md>), [completion](<https://devfeed.tech/tags/completion.md>), [constraint](<https://devfeed.tech/tags/constraint.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cross](<https://devfeed.tech/tags/cross.md>), [cs-ni](<https://devfeed.tech/tags/cs-ni.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [deepseek-r1](<https://devfeed.tech/tags/deepseek-r1.md>), [design](<https://devfeed.tech/tags/design.md>), [difference](<https://devfeed.tech/tags/difference.md>), [ethernet](<https://devfeed.tech/tags/ethernet.md>), [framework](<https://devfeed.tech/tags/framework.md>), [llm](<https://devfeed.tech/tags/llm.md>), [scale-up](<https://devfeed.tech/tags/scale-up.md>), [simulation](<https://devfeed.tech/tags/simulation.md>), [systems](<https://devfeed.tech/tags/systems.md>)

### AI overview

ScaleWire combines packet-level simulation of Ethernet network-interface mechanisms across transport, network, and link layers with analytical estimates to explore on-chip resource sizing for LLM systems. The authors report validation against hardware measurements and nine analytical checks, and evaluate six NI designs across four communication workloads. For DeepSeek-R1 dispatch, design choices produce a 4.5x difference in per-XPU storage budget; the exploration engine uses 6-19% as many simulations as exhaustive search in grid comparisons. With vLLM prefill traces and GH200-measured kernel times, it also guides resource provisioning under an inference completion-time constraint. The abstract positions the framework as a way to balance communication performance against resource cost.

### Source excerpt

arXiv:2611.26086v1 Announce Type: new Abstract: LLM training and inference rely on scale-up networks for high-bandwidth, low-latency memory-semantic communication among accelerators (XPUs). Ethernet is emerging as an open interconnect for scale-up systems. In Ethernet systems, network-interface (NI) mechanisms across transport, network, and link layers interact, jointly shaping communication performance and on-chip resource requirements. We present ScaleWire, an end-to-end scale-up Ethernet simulation framework for LLM systems. It couples packet-level simulation of cross-layer NI mechanisms with a resource exploration engine that combines analytical estimates and simulation to size on-chip resources. We validate ScaleWire against hardware measurements and nine analytical checks, and evaluate six NI designs across four LLM communication workloads. For DeepSeek-R1 dispatch, NI design choices lead to a 4.5$\times$ difference in per-XPU on-chip storage budget. In grid-comparison experiments, the exploration engine uses 6-19% of the simulations required by exhaustive search. Using vLLM prefill traces with kernel times measured on GH200 GPUs, ScaleWire further guides resource provisioning under a completion-time constraint for inference execution. ScaleWire thus helps designers select NI mechanisms and provision on-chip resources to balance communication performance and resource cost.

## ReLay: Bidirectional Layout-Code Translation and Library Evolution for Reusable and Process-Portable Analog Design

DevFeed: [ReLay: Bidirectional Layout-Code Translation and Library Evolution for Reusable and Process-Portable Analog Design](<https://devfeed.tech/articles/relay-bidirectional-layout-code-translation-and-library-evolution-for-reusable-and-process-portable-analog-design-145497.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25328>)

Author: Minxing Chu, Quan Cheng, Anhang Li, Martins Atilla, Junyi Luo, Xinting Jiang, Ruichen Qi, Mehdi Saligane

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Chip design](<https://devfeed.tech/topics/chip-design.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [analog](<https://devfeed.tech/tags/analog.md>), [analog-design](<https://devfeed.tech/tags/analog-design.md>), [api](<https://devfeed.tech/tags/api.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [cells](<https://devfeed.tech/tags/cells.md>), [chips](<https://devfeed.tech/tags/chips.md>), [circuit](<https://devfeed.tech/tags/circuit.md>), [code](<https://devfeed.tech/tags/code.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cost](<https://devfeed.tech/tags/cost.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ar](<https://devfeed.tech/tags/cs-ar.md>), [design](<https://devfeed.tech/tags/design.md>), [evolution](<https://devfeed.tech/tags/evolution.md>), [layout](<https://devfeed.tech/tags/layout.md>), [library](<https://devfeed.tech/tags/library.md>), [portable](<https://devfeed.tech/tags/portable.md>), [process](<https://devfeed.tech/tags/process.md>), [relay](<https://devfeed.tech/tags/relay.md>), [translation](<https://devfeed.tech/tags/translation.md>)

### AI overview

ReLay recovers generator programs from existing analog layouts, enabling cross-PDK portability and parametric resizing. Repeated constructions become library functions as more layouts are processed. Its benchmark contains 1,050 hand-drawn layouts from 27 circuit families across three open-source PDKs, from primitive cells to full chips. Learned functions account for 53-66% of layout-building calls in held-out layouts. In reported experiments, frontier LLM agents using ReLay had 64-77% lower API cost per completed task than agents generating geometry directly, with comparable completion rates and electrical quality.

### Source excerpt

arXiv:2611.25328v1 Announce Type: cross Abstract: Analog layout generators provide a programmable interface for design reuse and process migration, but their coverage is limited by handcrafted libraries. The designer knowledge embedded in complete layouts remains implicit in their geometry, making it difficult to capture in reusable generator libraries. We present ReLay, a bidirectional layout-code framework that recovers generator programs from existing layouts, enabling cross-PDK portability and parametric resizing of existing designs. Constructions that recur across recovered programs are promoted to library functions, so the library evolves as layouts accumulate. We also introduce the first layout-code benchmark: 1,050 hand-drawn layouts from 27 circuit families across three open-source PDKs, ranging from primitive cells to full chips. Learned functions account for 53-66% of layout-building calls in held-out layouts, demonstrating reuse beyond the learning corpus. Frontier LLM agents use ReLay to produce new layouts at 64-77% lower API cost per completed task than direct geometry generation, with comparable completion rates and electrical quality.

## Evaluating Real-Time Voice Agents: A Survey from Component Quality to Grounded Outcomes

DevFeed: [Evaluating Real-Time Voice Agents: A Survey from Component Quality to Grounded Outcomes](<https://devfeed.tech/articles/evaluating-real-time-voice-agents-a-survey-from-component-quality-to-grounded-outcomes-145378.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.30798>)

Author: Shivam Negi, Arpit Rawat, Rashi Jain

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Human-AI evaluation](<https://devfeed.tech/topics/human-ai-evaluation.md>), [ai observability](<https://devfeed.tech/topics/ai-observability.md>), [speech-to-speech](<https://devfeed.tech/topics/speech-to-speech.md>)

Tags: [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [external](<https://devfeed.tech/tags/external.md>), [grounded](<https://devfeed.tech/tags/grounded.md>), [real-time](<https://devfeed.tech/tags/real-time.md>), [real-time-voice-agents](<https://devfeed.tech/tags/real-time-voice-agents.md>), [recovery](<https://devfeed.tech/tags/recovery.md>), [standard](<https://devfeed.tech/tags/standard.md>), [survey](<https://devfeed.tech/tags/survey.md>), [timing](<https://devfeed.tech/tags/timing.md>), [tool-calling](<https://devfeed.tech/tags/tool-calling.md>), [traces](<https://devfeed.tech/tags/traces.md>), [user-interaction](<https://devfeed.tech/tags/user-interaction.md>), [voice](<https://devfeed.tech/tags/voice.md>)

### AI overview

The survey finds that real-time voice-agent evaluation rarely verifies grounded outcomes: only 13 of 91 evaluation works check external state or action traces, and only 2 of 26 multiparty benchmarks verify state. It codes 136 works by capability (timing, recovery, grounded outcomes, and sometimes multiparty conduct) and evidence strength, then proposes TRG, a minimum reporting standard that reports these dimensions without combining them into a score. One reviewed report has tool-selection F1 of 82.5% but exact-call success of 33.0%, without external task-state evidence; none of the three reports measures task state after interruption. The abstract also notes that one-on-one evaluation cannot establish whether an agent addresses the correct person, verifies their authority, or chooses appropriate silence.

### Source excerpt

arXiv:2609.30798v2 Announce Type: replace Abstract: Real-time voice agents must decide when to speak, recover from disruption, and act only when the conversation requires it. But currently their evaluation is fragmented and is split across speech quality, turn-taking, tool calling, and task completion. We review 136 primary works published up to October 2, 2026, and code each on two axes: the capability an evaluation exercises, namely timing, recovery, grounded outcome, and, where relevant, multiparty conduct; and the strength of its evidence, namely evaluation conditions and grading instrument. The coding locates the gaps. Only 13 of 91 evaluation works check external state or action traces, only 5 of 21 grounded-outcome benchmarks use real human speech, and only 2 of 26 multiparty benchmarks verify state. Three claims follow. Architecture constrains deployment but does not determine results. It has been observed that benchmarks can verify outcomes, but few do. Evaluating an agent solely through one-on-one user interaction leaves critical gaps: it cannot confirm if the agent is speaking to the correct person, verify that person's authority, or identify the proper moments for strategic silence. From the taxonomy we derive TRG, short for Timing-Recovery-Grounded, a minimum reporting standard, and apply it to three published reports with different architectures. In one, tool-selection F1 is 82.5% while exact-call success is 33.0%, with no evidence of external task state. None of the three reports measures task state after an interruption. TRG governs what is reported and assigns no composite score. Per-source codes and tooling accompany the submission.

## System-One RSI: Building Muscle Memory for Interactive Agents through Guided Practice

DevFeed: [System-One RSI: Building Muscle Memory for Interactive Agents through Guided Practice](<https://devfeed.tech/articles/system-one-rsi-building-muscle-memory-for-interactive-agents-through-guided-practice-145621.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26007>)

Author: Qingxu Fu, Xuanfa Jin, Ziqian Chen, Zhaoyang Liu, Bolin Ding

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [few-shot](<https://devfeed.tech/topics/few-shot.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [act](<https://devfeed.tech/tags/act.md>), [actions](<https://devfeed.tech/tags/actions.md>), [agents](<https://devfeed.tech/tags/agents.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [black](<https://devfeed.tech/tags/black.md>), [building](<https://devfeed.tech/tags/building.md>), [but](<https://devfeed.tech/tags/but.md>), [can](<https://devfeed.tech/tags/can.md>), [completion](<https://devfeed.tech/tags/completion.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [decision](<https://devfeed.tech/tags/decision.md>), [decision-model](<https://devfeed.tech/tags/decision-model.md>), [decision-models](<https://devfeed.tech/tags/decision-models.md>), [interactive](<https://devfeed.tech/tags/interactive.md>), [memory](<https://devfeed.tech/tags/memory.md>), [one](<https://devfeed.tech/tags/one.md>), [rsi](<https://devfeed.tech/tags/rsi.md>), [system](<https://devfeed.tech/tags/system.md>), [system-one](<https://devfeed.tech/tags/system-one.md>)

### AI overview

System-One RSI pairs a fast online control loop with offline improvement by reasoning agents to address the latency of reasoning models in sub-second tasks. In each OODA cycle, a decision model turns visual observations and task questions into structured judgments, and a decision tree converts them into timed actions without calling a reasoning model. Between interactions, agents revise the prompt and tree, record experience, derive supervision, and train or deploy updated parameters; the system then gathers experience for another round. The approach is reported to use a small number of real episodes without expert demonstrations, environment changes, sandboxing, or parallel simulation. With a 27B decision model, reported results include Snake mean score rising from 1.30 to 7.35, an 8.4% reduction in median completion time in one Zenless Zone Zero scenario, and Black Myth: Wukong results of 11 times the stationary reference's mean survival time for an evasive policy and offensive win rate rising from 0% to 66.7% across evaluated configurations. These are results from the specified scenarios, not evidence of general performance.

### Source excerpt

arXiv:2611.26007v1 Announce Type: new Abstract: Reasoning models provide strong multimodal task understanding but struggle with sub-second control because deliberation increases decision latency. Fast decision models can meet the response window, yet prompting and few-shot examples do not readily provide the fine-grained experience required by a new task. Inspired by coached motor learning, this paper presents System-One Recursive Self-Improvement (System-One RSI), which couples two loops. The high-frequency online inner loop is an Observe, Orient, Decide, Act (OODA) cycle: a decision model maps each visual observation and task question to structured judgments, and a decision tree converts them into timed actions without invoking a reasoning model. The offline outer loop is a recursive improvement cycle carried out by a team of reasoning agents. The team revises the decision-model prompt and decision tree, captures interaction records, derives supervision from those records, and trains or deploys updated model parameters. The improved system then returns to the environment and generates the experience for the next round, consolidating learned distinctions as external muscle memory. The framework operates with a small number of real episodes, without expert demonstrations, environment modification, sandboxing, or parallel simulation. Experiments in Snake, Zenless Zone Zero, and Black Myth: Wukong use a 27B decision model. Mean Snake score increases from 1.30 to 7.35. In an evaluated scenario of Zenless Zone Zero, median completion time decreases by 8.4%. In an evaluated combat scenario of Black Myth: Wukong, the evasive policy achieves 11 times the mean survival time of a stationary reference, while the offensive policy win rate rises from 0% to 66.7% across the evaluated configurations.

## A$^2$-VLA: Adapting Vision-Language-Action Policies to New Tools through Agentic Affordance Discovery

DevFeed: [A$^2$-VLA: Adapting Vision-Language-Action Policies to New Tools through Agentic Affordance Discovery](<https://devfeed.tech/articles/a-2-vla-adapting-vision-language-action-policies-to-new-tools-through-agentic-affordance-discovery-145563.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25656>)

Author: Chenyuan Zhou, Chaoyi Huang, Zibo Wang, Zhiyuan Li, Yong-Lu Li

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Agentic AI](<https://devfeed.tech/topics/what-is-agentic-ai.md>), [smolvla](<https://devfeed.tech/topics/smolvla.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [adapting](<https://devfeed.tech/tags/adapting.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [based](<https://devfeed.tech/tags/based.md>), [but](<https://devfeed.tech/tags/but.md>), [can](<https://devfeed.tech/tags/can.md>), [challenge](<https://devfeed.tech/tags/challenge.md>), [combines](<https://devfeed.tech/tags/combines.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [discovery](<https://devfeed.tech/tags/discovery.md>), [language](<https://devfeed.tech/tags/language.md>), [policies](<https://devfeed.tech/tags/policies.md>), [tools](<https://devfeed.tech/tags/tools.md>), [vision](<https://devfeed.tech/tags/vision.md>), [vision-language-action](<https://devfeed.tech/tags/vision-language-action.md>), [vla](<https://devfeed.tech/tags/vla.md>)

### AI overview

A$^2$-VLA adapts manipulation policies to unseen tools by discovering and validating task-relevant affordances. From teleoperated demonstrations with a few source tools, an embodied agent reconstructs scenes in simulation, identifies functional regions and their geometric properties, then refines affordance features and extraction skills through autonomous experiments. Validated skills are collected for reuse on unseen tools; encoded functional-region features condition a π₀.₅ backbone during training and inference. The abstract reports 86.1% success in simulated affordance-based cross-tool trajectory retargeting and a 30.7 percentage-point task-success improvement over π₀.₅ across three real-robot tool-use tasks.

### Source excerpt

arXiv:2611.25656v1 Announce Type: cross Abstract: Tools with different shapes can have similar functions, but transferring manipulation skills across tools requires understanding how tool geometry relates to the specific task and adjusting the manipulation strategy accordingly. The key challenge is to discover the common affordances that contribute to task completion across different tools and verify whether they support new tools in performing similar work. We present A$^2$-VLA, a tool-adaptation framework that combines an embodied agent with a VLA policy. It treats tool affordances as hypotheses the agent can construct, test, and revise. Given teleoperated demonstrations with a few source tools, the agent reconstructs the scenes in simulation and autonomously determines task-relevant affordance features: functional regions and their key geometric properties. It develops affordance-extraction skills and refines both the features and skills through autonomous experiments in the reconstructed scenes. Validated extraction skills accumulate in a reusable skill library for unseen tools. Finally, A$^2$-VLA uses encoded functional-region features as conditioning inputs to a $\pi_{0.5}$ backbone during training and inference. Evaluation examines affordance reliability in reconstructed scenes and real-robot task success on unseen tools. Our method achieves 86.1% success in simulated cross-tool trajectory retargeting based on affordances. On three real-robot tool-use tasks, A$^2$-VLA improves task success by 30.7 percentage points over $\pi_{0.5}$.

## CoachSI: Self-Improvement of Robot Policies with a Language Coach

DevFeed: [CoachSI: Self-Improvement of Robot Policies with a Language Coach](<https://devfeed.tech/articles/coachsi-self-improvement-of-robot-policies-with-a-language-coach-145529.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25529>)

Author: Jiachen Tao, Gengyu Zhang, Haoxuan Wang, Bing Liu, Yan Yan

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Robot Navigation](<https://devfeed.tech/topics/robot-navigation.md>), [vlm](<https://devfeed.tech/topics/vlm.md>), [post-training](<https://devfeed.tech/topics/post-training.md>)

Tags: [camera](<https://devfeed.tech/tags/camera.md>), [coaching](<https://devfeed.tech/tags/coaching.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [horizon](<https://devfeed.tech/tags/horizon.md>), [improvement](<https://devfeed.tech/tags/improvement.md>), [instructions](<https://devfeed.tech/tags/instructions.md>), [language](<https://devfeed.tech/tags/language.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [model](<https://devfeed.tech/tags/model.md>), [naming](<https://devfeed.tech/tags/naming.md>), [next](<https://devfeed.tech/tags/next.md>), [policies](<https://devfeed.tech/tags/policies.md>), [policy](<https://devfeed.tech/tags/policy.md>), [robot](<https://devfeed.tech/tags/robot.md>), [self-improvement](<https://devfeed.tech/tags/self-improvement.md>), [vision](<https://devfeed.tech/tags/vision.md>), [vision-language-action](<https://devfeed.tech/tags/vision-language-action.md>), [vision-language-model](<https://devfeed.tech/tags/vision-language-model.md>), [vlm](<https://devfeed.tech/tags/vlm.md>)

### AI overview

CoachSI improves long-horizon robot task performance by having a frozen vision-language model watch the robot and name the next step in the policy's atomic-instruction format; it never supplies actions. Successful coached runs are relabeled with the original task instruction and used to fine-tune the policy, which is then deployed without a coach. The action backbone remains frozen. On stalled states, next-step coaching raises subgoal completion 4.6-fold over the unchanged task instruction. Unassisted success in unseen scenes rises from 47.4% to 59.3%, exceeding human-demonstration fine-tuning by 9.8 percentage points and equal-budget self-collected-success fine-tuning by 7.3 points. With a coach limited to camera images, success reaches 56.8%, above both baselines.

### Source excerpt

arXiv:2611.25529v1 Announce Type: cross Abstract: Generalist robot policies execute atomic skills reliably but fail when the same skills are composed into long-horizon tasks. Self-improvement from the policy's own experience does not close this gap: experience collected by the policy alone stays within its own distribution and reproduces what it can already do. We show that what the policy lacks is not a skill but the next step. We present CoachSI, a self-improvement method in which a frozen, off-the-shelf vision-language model (VLM) coaches the policy in language: it watches the robot and names the next step in the format of the policy's atomic instructions, and never outputs an action. Successful coached rollouts are relabeled in hindsight with the original task instruction and used to fine-tune the policy. The fine-tuned policy is then deployed on its own. On states where a World Action Model and a Vision-Language-Action model have stalled, naming the next step raises subgoal completion by 4.6x over the unchanged task instruction, and the frozen VLM supplies it from the same camera images the policy sees. After fine-tuning, unassisted success on long-horizon tasks in unseen scenes rises from 47.4% to 59.3%, 9.8 points above fine-tuning on human demonstrations and 7.3 points above fine-tuning on self-collected successes at an equal data budget. With a coach that sees only camera images, it reaches 56.8%, above both baselines. The gain is obtained with the action backbone frozen: language coaching teaches the policy what should happen next, and the skills it already has do the rest.

## Improving Imitation Policies for Multi-Phase Tasks through Residual Learning Allocation

DevFeed: [Improving Imitation Policies for Multi-Phase Tasks through Residual Learning Allocation](<https://devfeed.tech/articles/improving-imitation-policies-for-multi-phase-tasks-through-residual-learning-allocation-145042.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25320>)

Author: Clinton Enwerem, Saleh Kemal, John S. Baras, Calin Belta

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [assembly](<https://devfeed.tech/tags/assembly.md>), [budget](<https://devfeed.tech/tags/budget.md>), [completion](<https://devfeed.tech/tags/completion.md>), [components](<https://devfeed.tech/tags/components.md>), [controller](<https://devfeed.tech/tags/controller.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [cs-sy](<https://devfeed.tech/tags/cs-sy.md>), [eess-sy](<https://devfeed.tech/tags/eess-sy.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [failure](<https://devfeed.tech/tags/failure.md>), [imitation](<https://devfeed.tech/tags/imitation.md>), [learning](<https://devfeed.tech/tags/learning.md>), [phases](<https://devfeed.tech/tags/phases.md>), [policies](<https://devfeed.tech/tags/policies.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [release](<https://devfeed.tech/tags/release.md>), [residual](<https://devfeed.tech/tags/residual.md>), [robotic](<https://devfeed.tech/tags/robotic.md>), [tasks](<https://devfeed.tech/tags/tasks.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

The paper allocates residual reinforcement learning by estimated completion gain rather than phase failure counts. Completion gain measures the change in full-task completion probability when a state-feedback controller replaces selected base-policy action components in one phase; paired episodes estimate it by factoring in both the chance of reaching that phase and the conditional improvement from there. Positive estimates guide which components to correct and where to spend training episodes. In point-mass docking, the channel phase had the largest gain despite most failures occurring during final docking; gain-based allocation outperformed failure-count allocation, and in a longer-approach variant beat uniform allocation at half the interaction budget. In simulated bimanual assembly, gains were positive only for insertion and release, with few release failures; training in release improved held-out completion by 16.9% in a second configuration.

### Source excerpt

arXiv:2611.25320v1 Announce Type: new Abstract: Residual reinforcement learning corrects an imitation-trained base policy using a limited budget of environment interaction. In a task with several phases, this budget must be divided among the phases, and the phase in which the base policy fails is not necessarily the phase in which a correction most improves task completion. We formalize this distinction through the completion gain: the change in full-task completion probability when, during a single phase, a state-feedback controller replaces selected components of the base policy's action. We decompose the completion gain into the probability of reaching the phase and the conditional improvement in completion from the start of the phase, and we estimate it from paired episodes. The positive estimates determine which action components the residual policy corrects and how training episodes are distributed across phases. In stochastic point-mass docking experiments, most failures occurred in the final docking phase, whereas the largest completion gain occurred in the preceding channel phase; gain-based allocation improved completion over allocation proportional to failure counts and, in a variant with a longer approach, over uniform allocation at half of the budget. In simulated bimanual robotic assembly, the completion-gain estimates were positive only for insertion and release, although few failures occurred during release; in a second assembly configuration, a residual policy trained in the release phase improved completion from held-out states at the start of release by 16.9%.

## Design and Evaluation of Epira: An Automated Interview System for Case Investigation and Contact Tracing

DevFeed: [Design and Evaluation of Epira: An Automated Interview System for Case Investigation and Contact Tracing](<https://devfeed.tech/articles/design-and-evaluation-of-epira-an-automated-interview-system-for-case-investigation-and-contact-tracing-145072.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25577>)

Author: Ishita Asnani, Shaghayegh Fazliani, Zachary Ryan, Kriyoshig Prasai, Chirag Kumar, Indra Gonzalez Ojeda, Marc Lipsitch, Madeleine Udell

Published: 2026-11-20T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [rule-engine](<https://devfeed.tech/topics/rule-engine.md>), [human review](<https://devfeed.tech/topics/human-review.md>), [Trustworthy AI](<https://devfeed.tech/topics/trustworthy-ai.md>), [human computer interaction](<https://devfeed.tech/topics/human-computer-interaction.md>), [trust](<https://devfeed.tech/topics/trust.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [automated](<https://devfeed.tech/tags/automated.md>), [backends](<https://devfeed.tech/tags/backends.md>), [based](<https://devfeed.tech/tags/based.md>), [case](<https://devfeed.tech/tags/case.md>), [cases](<https://devfeed.tech/tags/cases.md>), [chatbots](<https://devfeed.tech/tags/chatbots.md>), [choice](<https://devfeed.tech/tags/choice.md>), [cloud](<https://devfeed.tech/tags/cloud.md>), [completeness](<https://devfeed.tech/tags/completeness.md>), [completion](<https://devfeed.tech/tags/completion.md>), [conclusion](<https://devfeed.tech/tags/conclusion.md>), [cs-hc](<https://devfeed.tech/tags/cs-hc.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [design](<https://devfeed.tech/tags/design.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [interview](<https://devfeed.tech/tags/interview.md>), [investigation](<https://devfeed.tech/tags/investigation.md>), [system](<https://devfeed.tech/tags/system.md>), [tracing](<https://devfeed.tech/tags/tracing.md>)

### AI overview

Epira combines LLM dialogue with a rule-based controller that enforces required questions, safety handoffs, and structured case records; new diseases can be added by uploading a questionnaire. In 108 simulated measles and gonorrhea cases, the local Qwen2.5 version captured 88.9% of required fields and 60.1% of known contacts. Open-ended interviewers captured 31.1-41.9 percentage points fewer required fields than the same models under Epira, and protocol control mattered more than model choice. In a separate study of 72 volunteers, completion was 100%, with mean usability of 82.95/100 and trust of 3.85/5. Local Qwen2.5 and cloud GPT-5.4 produced similar records. The authors support local use with human review of routed interviews and missed contacts; the results do not establish that all contacts are captured.

### Source excerpt

arXiv:2611.25577v1 Announce Type: new Abstract: Objective: Case investigation and contact tracing are labor-intensive. Web forms struggle when respondents need clarification before disclosing contacts, and open-ended chatbots may deviate from protocols. We developed and evaluated Epira, a protocol-controlled large language model (LLM) interviewer for multi-disease case investigation and contact tracing. Materials and Methods: Epira pairs LLM dialogue with a rule-based controller enforcing required questions, safety handoffs, and a structured case record, and adapts to new diseases by questionnaire upload. Local (Qwen2.5) and cloud (GPT-5.4) backends were tested. Study 1 used 108 simulated cases with ground-truth fields and contacts for measles and gonorrhea under Epira and open-ended conditions (864 sessions). Study 2 enrolled 72 volunteers in a 2$\times$2 disease-by-model interview assessing completion, usability, acceptability, and trust. Results: Across Study 1, Epira with a local LLM captured 88.9% of required fields and 60.1% of ground-truth contacts. Open-ended interviewers recovered 31.1-41.9 percentage points fewer required fields than the same models in Epira. In Study 2, completion was 100%; mean adapted System Usability Scale (SUS) was 82.95 $\pm$ 16.17 on a 0-100 scale and mean trust was 3.85 $\pm$ 0.77 on a 1-5 scale. Discussion: Protocol control drove completeness more than model choice. Local and cloud deployments produced similar records, supporting privacy-preserving local use with human review of routed interviews and missed contacts. Conclusion: Epira recovers required fields and contacts under diverse disclosure conditions, works with local and cloud LLMs, and received favorable usability and trust ratings.

## Stop Going in Circles:Loop-Triggered Routing for System 1 Decision Agents

DevFeed: [Stop Going in Circles:Loop-Triggered Routing for System 1 Decision Agents](<https://devfeed.tech/articles/stop-going-in-circles-loop-triggered-routing-for-system-1-decision-agents-145503.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25363>)

Author: Zeming Yu, Huamin Ma, Yingyu Chen, Yuan Sun, Peng Hu, Yang Qin

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [rule-engine](<https://devfeed.tech/topics/rule-engine.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [actions](<https://devfeed.tech/tags/actions.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [also](<https://devfeed.tech/tags/also.md>), [api](<https://devfeed.tech/tags/api.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [based](<https://devfeed.tech/tags/based.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [but](<https://devfeed.tech/tags/but.md>), [call](<https://devfeed.tech/tags/call.md>), [can](<https://devfeed.tech/tags/can.md>), [capabilities](<https://devfeed.tech/tags/capabilities.md>), [completion](<https://devfeed.tech/tags/completion.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [decision](<https://devfeed.tech/tags/decision.md>), [routing](<https://devfeed.tech/tags/routing.md>), [stop](<https://devfeed.tech/tags/stop.md>), [system](<https://devfeed.tech/tags/system.md>)

### AI overview

LoopRoute uses recurring state-action pairs as a signal that a System 1 decision agent has stalled. The decision model selects actions by default; when a loop is detected, an LLM provides brief, feedback-driven help, then control returns to the decision model. Later loops can trigger more assistance. The rule needs no extra training or model-specific confidence calibration. Across three benchmarks and seven LLMs, it improved task success over standalone Jev by mean gains of 21.0% on ALFWorld, 60.1% on MapReader, and 30.1% on TextWorld cooking. Compared with full LLM control, it substantially reduced mean cumulative API time and cost per task. The abstract also reports higher success rates for six frozen decision models, but provides no further methodological or quantitative detail for that result.

### Source excerpt

arXiv:2611.25363v1 Announce Type: new Abstract: System 1 decision models (represented by Jev) provide fast, low-cost action selection, but existing agent integrations do not reliably translate these local advantages into task completion. At certain points during execution, an agent may repeatedly select the same actions without making progress toward its goal. LLM-based policies offer strong reasoning capabilities that support planning within an agent, but invoking them at every step incurs substantial overhead. The effective coordination should therefore preserve Jev's fast local action selection while introducing LLM reasoning when execution stalls. We observe that recurring state-action pairs can signal stalled execution. We therefore propose LoopRoute, a simple but effective action router that uses this signal to allocate control within an agent. The decision model acts by default. When a loop is detected, the LLM provides brief, feedback-driven assistance and then returns control to the decision model. Later loops can request further assistance. The rule requires neither additional training nor model-specific confidence calibration. Across three benchmarks and seven LLMs, LoopRoute improves task success over standalone Jev with only limited use of the LLM policy. The mean gains are 21.0%, 60.1%, and 30.1% on ALFWorld, MapReader, and TextWorld cooking, respectively. Compared with full LLM control, LoopRoute also substantially reduces the mean cumulative API time and API-call cost per task. The same rule also improves the success rates of six frozen decision models.

## Learning from Context: Leave-One-Out Prediction in Jump Diffusions

DevFeed: [Learning from Context: Leave-One-Out Prediction in Jump Diffusions](<https://devfeed.tech/articles/learning-from-context-leave-one-out-prediction-in-jump-diffusions-145169.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26359>)

Author: Christian Wald

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [Embeddings](<https://devfeed.tech/topics/embeddings.md>)

Tags: [completion](<https://devfeed.tech/tags/completion.md>), [context](<https://devfeed.tech/tags/context.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [drift](<https://devfeed.tech/tags/drift.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [generation](<https://devfeed.tech/tags/generation.md>), [graph](<https://devfeed.tech/tags/graph.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [prediction](<https://devfeed.tech/tags/prediction.md>)

### AI overview

The paper studies leave-one-out (LOO) prediction in jump diffusions, where each coordinate is predicted from the corrupted states of the other coordinates to encourage learning dependencies between them. Under stated assumptions, complete coordinate resets with state-independent clocks preserve prescribed marginals when reset destinations use exact LOO probabilities and drift uses the full endpoint posterior. The authors show this drift can be built from LOO probabilities through analytic likelihood completion, without extra network evaluations. On Text8, LM1B, and ZINC250k, LOO-based variants improve key generation metrics over matched unrestricted controls. Controlled interventions examine reliance on local inputs and context; the abstract does not specify the metrics or size of the improvements.

### Source excerpt

arXiv:2611.26359v1 Announce Type: new Abstract: Continuous diffusion models generate discrete data through stochastic dynamics in an embedding space. We consider jump diffusions that additionally reset coordinates to embeddings of categorical values. A coordinate's own corrupted state can make prediction easy, potentially reducing the pressure to learn dependencies between coordinates. We study leave-one-out (LOO) prediction, in which each coordinate is predicted from the corrupted states of the other coordinates. We show that, under stated assumptions, jump diffusions with complete coordinate resets and state-independent clocks preserve the prescribed marginals when reset destinations use exact LOO probabilities and the drift uses the full endpoint posterior. We further show that this drift can be constructed from LOO probabilities by analytic likelihood completion without additional network evaluations. We evaluate unrestricted, LOO, and completed predictions on Text8 and LM1B text generation and ZINC250k molecular graph generation. LOO-based variants improve key generation metrics over matched unrestricted controls on all three datasets. Controlled interventions assess reliance on local inputs and context, and we analyze the effect of analytic likelihood completion.

## BatchGen: An Architecture for Scalable and Efficient Batch Inference

DevFeed: [BatchGen: An Architecture for Scalable and Efficient Batch Inference](<https://devfeed.tech/articles/batchgen-an-architecture-for-scalable-and-efficient-batch-inference-142960.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.21712>)

Author: Tairan Xu, Leyang Xue, Zhan Lu, Jinfu Deng, Hongyang Xiao, Yinsicheng Jiang, Congjie He, Matej Sandor, Le Xu, Luo Mai

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [AI Architecture](<https://devfeed.tech/topics/ai-architecture.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [accelerators](<https://devfeed.tech/tags/accelerators.md>), [ai](<https://devfeed.tech/tags/ai.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [batch](<https://devfeed.tech/tags/batch.md>), [batches](<https://devfeed.tech/tags/batches.md>), [cluster](<https://devfeed.tech/tags/cluster.md>), [completion](<https://devfeed.tech/tags/completion.md>), [compute](<https://devfeed.tech/tags/compute.md>), [coroutine](<https://devfeed.tech/tags/coroutine.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [devices](<https://devfeed.tech/tags/devices.md>), [engines](<https://devfeed.tech/tags/engines.md>), [event](<https://devfeed.tech/tags/event.md>), [event-driven](<https://devfeed.tech/tags/event-driven.md>), [execution](<https://devfeed.tech/tags/execution.md>), [existing](<https://devfeed.tech/tags/existing.md>), [fail](<https://devfeed.tech/tags/fail.md>), [inference](<https://devfeed.tech/tags/inference.md>)

### AI overview

BatchGen uses a sequence coroutine compute model to address runtime load variation and GPU utilization challenges in large-scale batch inference. It represents each sequence as a fine-grained, event-driven coroutine, allowing the runtime to form larger expert-level batches, mitigate stragglers, move work across devices, and sustain utilization on cost-effective or memory-constrained GPUs. The abstract reports up to 2.3x shorter batch completion time on a 128-GPU cluster and up to 9.6x better performance than the strongest offloading baseline on memory-constrained accelerators.

### Source excerpt

arXiv:2606.21712v2 Announce Type: replace Abstract: Batch inference has become a central mode of AI computation, yet existing inference engines still rely on execution models designed for interactive serving. When scaled to millions of sequences, batch workloads reveal two fundamental requirements: the ability to handle extreme inter- and intra-sequence load variation that emerges only at runtime, and the ability to sustain high utilization across large fleets of GPUs. Existing systems fail to meet these requirements, losing substantial fractions of achievable throughput. We introduce a new architectural foundation for batch inference: the sequence coroutine compute model, which represents each sequence as a fine-grained, event-driven coroutine. This model exposes expressive primitives that allow the runtime to reorganize work dynamically, enabling larger expert-level batches, mitigating stragglers, reallocating work across devices, and maintaining utilization even on cost-effective or memory-constrained GPUs. Building on this abstraction, we implement BatchGen, a production-ready system that uses the coroutine model at cluster scale. On a 128-GPU cluster, BatchGen reduces batch completion time by up to $2.3\times$, and on memory-constrained accelerators it outperforms the strongest offloading baseline by up to $9.6\times$. We will open-source BatchGen at https://github.com/batchgen-project/batchgen

## Checking Retry Transparency in Agent Tools

DevFeed: [Checking Retry Transparency in Agent Tools](<https://devfeed.tech/articles/checking-retry-transparency-in-agent-tools-142168.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23307>)

Author: Jiapeng Li (Microsoft)

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [Test-driven development](<https://devfeed.tech/topics/tdd.md>), [retry](<https://devfeed.tech/topics/retry.md>), [Test automation](<https://devfeed.tech/topics/test-automation.md>)

Tags: [absorption](<https://devfeed.tech/tags/absorption.md>), [admission](<https://devfeed.tech/tags/admission.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agent-tools](<https://devfeed.tech/tags/agent-tools.md>), [annotation](<https://devfeed.tech/tags/annotation.md>), [atomic](<https://devfeed.tech/tags/atomic.md>), [boundary](<https://devfeed.tech/tags/boundary.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [browser](<https://devfeed.tech/tags/browser.md>), [completion](<https://devfeed.tech/tags/completion.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [effect](<https://devfeed.tech/tags/effect.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [file](<https://devfeed.tech/tags/file.md>), [git](<https://devfeed.tech/tags/git.md>), [graph](<https://devfeed.tech/tags/graph.md>), [identity](<https://devfeed.tech/tags/identity.md>), [mcp](<https://devfeed.tech/tags/mcp.md>), [retry](<https://devfeed.tech/tags/retry.md>), [transparency](<https://devfeed.tech/tags/transparency.md>)

### AI overview

A successful retry can still leave an earlier attempt able to overwrite a later acknowledged update, so a tool's local idempotence claim may not guarantee recovery. The prototype tests manually constructed, source-backed effect contracts at two levels: atomic retry absorption and bounded interleavings of successful-attempt microsteps. It produces model witnesses, conditional admission decisions, and explicit uncertainty when completion boundaries are unjustified. In 27 missing-reply checks, its selective rule completed six justified workflows without acknowledged goal violations; strict coalescing completed none, while blind retry completed nine with three violations. An intent-aware Porcupine specification agreed with all 156 complete-history oracles. The evaluation covers a purposive frame of six MCP components from four repositories, with one backend unavailable; the authors caution these mechanism-selected tests do not estimate defect prevalence or introduce a new linearizability algorithm.

### Source excerpt

arXiv:2611.23307v1 Announce Type: new Abstract: A successful retry does not necessarily put the earlier attempt behind it. An old write may still overwrite a later acknowledged update; a partially executed file write may even empty a file after the retry has returned successfully. We study how to test the recovery guarantee behind an agent tool's local idempotence claim. Our prototype takes manually constructed, source-backed effect contracts and checks them at two levels: atomic retry-absorption obligations and bounded interleavings of successful-attempt microsteps. It emits model witnesses, conditional admission decisions, and explicit uncertainty when a completion boundary cannot be justified. The evaluation preserves a locally frozen purposive frame of 6 MCP components from 4 repositories; selected paths in 5 execute on the study machine and one backend remains unavailable. Real tests include file, graph, Git-index, document-update and browser bindings, together with a native Qdrant development matrix. In 27 missing-reply checks, the selective rule completes six justified workflows without an acknowledged goal violation, whereas strict coalescing completes none and blind retry completes nine with three violations. An intent-aware Porcupine specification independently agrees with all 156 complete-history oracles. These are mechanism-selected tests, not defect-prevalence estimates or a new linearizability algorithm. The contribution is an executable recovery-contract workflow that makes effect granularity, logical identity and completion assumptions testable rather than inheriting them from a tool annotation.

## CriticHack: Evaluating Visual Rewards Under Robot Policy Optimization

DevFeed: [CriticHack: Evaluating Visual Rewards Under Robot Policy Optimization](<https://devfeed.tech/articles/critichack-evaluating-visual-rewards-under-robot-policy-optimization-142571.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.02527>)

Author: Jiaxuan Luo, Xingguo Xu, Shanshan Wang, Yuhan Zhou, Zhen Zhang

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Robot Manipulation](<https://devfeed.tech/topics/robot-manipulation.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>)

Tags: [camera](<https://devfeed.tech/tags/camera.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [drawer](<https://devfeed.tech/tags/drawer.md>), [execution](<https://devfeed.tech/tags/execution.md>), [exploit](<https://devfeed.tech/tags/exploit.md>), [failure](<https://devfeed.tech/tags/failure.md>), [model](<https://devfeed.tech/tags/model.md>), [object](<https://devfeed.tech/tags/object.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [policy](<https://devfeed.tech/tags/policy.md>), [progress](<https://devfeed.tech/tags/progress.md>), [repair](<https://devfeed.tech/tags/repair.md>), [reward](<https://devfeed.tech/tags/reward.md>), [reward-model](<https://devfeed.tech/tags/reward-model.md>), [robot](<https://devfeed.tech/tags/robot.md>), [simulator](<https://devfeed.tech/tags/simulator.md>), [stacking](<https://devfeed.tech/tags/stacking.md>), [training](<https://devfeed.tech/tags/training.md>), [trust](<https://devfeed.tech/tags/trust.md>)

### AI overview

Fine-tuning a robot policy against visual rewards can improve task success while also increasing wrong-object failures. In drawer opening, Robometer training raised success by 10.2 percentage points and wrong-drawer failures by 10.9; RoboReward-8B showed the same pattern, which also appeared on a physical robot. The proposed mechanism is that wrong-object actions look like progress and score better than many ordinary failures, so optimization amplifies them even when successes score higher. The authors suggest a pre-training check using initial-policy executions to identify outcomes that score unusually well relative to alternatives. They also report a label-free repair: rescore executions on frames re-rendered with unnamed objects restored, which removed the amplification in the tested tasks and robot setting. In a stacking task, the check predicted which placement would be exploited: wrong-object stacks rose from 20% to 69% only for that placement.

### Source excerpt

arXiv:2610.02527v2 Announce Type: replace-cross Abstract: Learned visual reward models score a robot's execution from camera images and an instruction, and are increasingly used to fine-tune robot policies. Two checks decide whether to trust one: the reward model separates successes from failures, and task success rises during training. We show that both checks can pass while the policy learns to act on the wrong object. When a diffusion policy is fine-tuned against the reward model Robometer to open a drawer, task success rises by 10.2 percentage points. Wrong-object failures, which open a different drawer, rise by 10.9 points, whereas the same training on the simulator's task-completion signal raises success alone. A second reward model, RoboReward-8B, behaves the same way, and the effect reappears on a physical robot. This happens because of how the reward scores the policy's own failures. Most failures make little progress and score low, while a wrong-object failure looks like progress and scores well above them. Optimization shifts a policy toward outcomes that score above the executions they compete with, the executions they can take the place of. It therefore turns low-scoring failures into both successes and wrong-object failures, even when the reward model scores successes higher. This rule needs only executions of the initial policy, so it can be tested before training. On a stacking task we predicted which of three object placements Robometer would exploit, and training raised wrong-object stacks from 20% to 69% in that placement and in neither of the others. The same rule gives a check to run before optimization and a repair that needs no labels: scoring each execution again on frames re-rendered with the objects the instruction does not name put back removes the amplification on both tasks and on the robot.

## H3Remover: Reinforcing Video Object Removal without Paired Data

DevFeed: [H3Remover: Reinforcing Video Object Removal without Paired Data](<https://devfeed.tech/articles/h3remover-reinforcing-video-object-removal-without-paired-data-142815.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24633>)

Author: Kangjie Chen, Xiangyu Li, Jinhao Deng, Dongbin Zhang, Chaoda Zheng, Shijia Chen, Yuqi Zhang, Choo Sin Wai, Boyi Sun, Puyi Wang, Yu Zhang, Xianming Liu, Boyang Wang

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Generative AI](<https://devfeed.tech/topics/generative-ai.md>), [multimodal-ai](<https://devfeed.tech/topics/multimodal-ai.md>)

Tags: [adapters](<https://devfeed.tech/tags/adapters.md>), [appearance](<https://devfeed.tech/tags/appearance.md>), [background](<https://devfeed.tech/tags/background.md>), [comparisons](<https://devfeed.tech/tags/comparisons.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [data](<https://devfeed.tech/tags/data.md>), [editors](<https://devfeed.tech/tags/editors.md>), [edits](<https://devfeed.tech/tags/edits.md>), [effects](<https://devfeed.tech/tags/effects.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [lora](<https://devfeed.tech/tags/lora.md>), [minimax](<https://devfeed.tech/tags/minimax.md>), [minimax-h3](<https://devfeed.tech/tags/minimax-h3.md>), [object](<https://devfeed.tech/tags/object.md>), [ordering](<https://devfeed.tech/tags/ordering.md>), [original](<https://devfeed.tech/tags/original.md>), [video](<https://devfeed.tech/tags/video.md>)

### AI overview

H3Remover reinforces MiniMax-H3 Base Ref2VA for video object removal without paired object-free training targets. It combines a depth-based removal reward, a background-geometry reward that preserves relative depth ordering, and an appearance reward that limits RGB changes in protected regions while exempting targets and estimated shadows or reflections. Frozen HPSv3 provides perceptual-quality feedback; DiffusionART optimizes LoRA adapters with the backbone frozen and evaluation regions fixed across candidates. The excerpt reports substantial reconstruction gains over original H3 and qualitative improvements in removal and surrounding-scene preservation.

### Source excerpt

arXiv:2611.24633v1 Announce Type: cross Abstract: Pretrained video editors possess rich removal and completion priors, yet may retain targets or alter unrelated content. We present H3Remover, which reinforces MiniMax-H3 Base Ref2VA for video object removal without paired object-free training targets. To discourage superficial edits such as recoloring, a target-removal reward favors calibrated depth changes consistent with revealing farther content. To limit unintended scene changes, a background-geometry reward encourages preservation of relative depth ordering. To preserve background appearance without retaining unwanted shadows or reflections, an appearance reward constrains RGB changes in source-derived protected regions while exempting targets and estimated effects. Frozen HPSv3 complements these constraints with perceptual-quality feedback for synthesized content. Evaluation regions remain fixed across candidates, and DiffusionART optimizes LoRA adapters with the backbone frozen. Experiments show substantial reconstruction gains over original H3, with qualitative comparisons demonstrating more effective object removal and better preservation of surrounding scene content.

## Optimal Development Path: An Empirical Evaluation of Tiered Dependency-Aware Sequencing

DevFeed: [Optimal Development Path: An Empirical Evaluation of Tiered Dependency-Aware Sequencing](<https://devfeed.tech/articles/optimal-development-path-an-empirical-evaluation-of-tiered-dependency-aware-sequencing-142164.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23303>)

Author: Azzeddine Ihsine, Sara Ihsine

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [Combinatorial optimization](<https://devfeed.tech/topics/combinatorial-optimization.md>), [jira](<https://devfeed.tech/topics/jira.md>)

Tags: [alternatives](<https://devfeed.tech/tags/alternatives.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [block](<https://devfeed.tech/tags/block.md>), [calendar](<https://devfeed.tech/tags/calendar.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [dependency](<https://devfeed.tech/tags/dependency.md>), [development](<https://devfeed.tech/tags/development.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [jira](<https://devfeed.tech/tags/jira.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [operational](<https://devfeed.tech/tags/operational.md>), [optimum](<https://devfeed.tech/tags/optimum.md>), [orders](<https://devfeed.tech/tags/orders.md>), [ratio](<https://devfeed.tech/tags/ratio.md>), [reference](<https://devfeed.tech/tags/reference.md>), [rework](<https://devfeed.tech/tags/rework.md>), [risk](<https://devfeed.tech/tags/risk.md>), [scheduling](<https://devfeed.tech/tags/scheduling.md>), [sequences](<https://devfeed.tech/tags/sequences.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>)

### AI overview

ODP (L0) sequences prerequisite-eligible feature blocks by choosing the lowest effort-and-risk-to-value ratio. On synthetic dependency graphs, unlocked-value (L2) and one-step rollout (L4) come closer to the surrogate optimum than L0, especially at larger sizes; on 373 Jira sprint proxies with sparse recorded dependencies, L0 reaches 0.998 of the surrogate optimum and is exactly optimal in 69% of sprints. Synthetic-calibrated unlock bonuses yield no significant average gain on the Jira-derived instances and can worsen the surrogate. Although L4 never worsens that surrogate relative to L0, retrospective calendar replays show surrogate gains do not reliably improve value-weighted completion. The evidence supports L0 as the default, with rollout and exact references as auditable alternatives. Effects on interruptions, rework, costs, and delivery times remain unmeasured.

### Source excerpt

arXiv:2611.23303v1 Announce Type: new Abstract: The Optimal Development Path (ODP) in D-POAF orders Feature Blocks in a Wave Candidate by choosing the eligible block with the lowest effort-and-risk-to-value ratio (ERS/BVS), while respecting declared prerequisites. We evaluate native ODP (L0) against four dependency-aware strategies: unlock count (L1), unlocked value (L2), multi-hop potential (L3), and one-step rollout (L4), alongside local search, Sidney decomposition, and an exact reference. On held-out synthetic graphs with about one prerequisite per block, L0 attains 0.978 of the surrogate optimum at 12 blocks, versus 0.993 for L2 and 0.994 for L4. At 100 blocks, L0 attains 0.942 and L2 0.983. In contrast, 373 historical Jira sprints from 21 open-source projects in TAWOS, used as scheduling proxies rather than D-POAF Waves, contain sparse recorded dependencies. On these instances, L0 attains 0.998 of the surrogate optimum and is exactly optimal in 69% of sprints. Unlock bonuses calibrated on synthetic graphs do not deliver significant average gains on these sprint-derived instances and can worsen the surrogate. L4 never worsens the declared surrogate relative to L0, but this guarantee does not extend to operational outcomes. Retrospective calendar replays using recorded Jira durations show that surrogate improvements do not reliably improve value-weighted completion. The findings support keeping L0 as the default and presenting rollout and exact references as auditable alternatives. Dependency-valid sequences expose prerequisite readiness, but their effects on interruptions, rework, costs, and delivery times remain unmeasured.

## Safe Personal RSI: Evaluating and Repairing Personal-Agent Updates

DevFeed: [Safe Personal RSI: Evaluating and Repairing Personal-Agent Updates](<https://devfeed.tech/articles/safe-personal-rsi-evaluating-and-repairing-personal-agent-updates-142190.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23651>)

Author: Jiapeng Pei, Yuhang Yao, Sitan Liu, Hanxin Chen, Hao Li, Guogeng Li, Jiapei Liu, Ruijie Chen, Hongshan Lin, Yang Mo, Tianyu Shi, Shangding Gu

Published: 2026-11-19T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [ai alignment](<https://devfeed.tech/topics/ai-alignment.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [AI Prompt Templates](<https://devfeed.tech/topics/ai-prompt-templates.md>), [gaia](<https://devfeed.tech/topics/gaia.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [ai-governance](<https://devfeed.tech/topics/ai-governance.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [actions](<https://devfeed.tech/tags/actions.md>), [adoption](<https://devfeed.tech/tags/adoption.md>), [agent](<https://devfeed.tech/tags/agent.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [authorization](<https://devfeed.tech/tags/authorization.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [calendar](<https://devfeed.tech/tags/calendar.md>), [can](<https://devfeed.tech/tags/can.md>), [cases](<https://devfeed.tech/tags/cases.md>), [changes](<https://devfeed.tech/tags/changes.md>), [code](<https://devfeed.tech/tags/code.md>), [completion](<https://devfeed.tech/tags/completion.md>), [component](<https://devfeed.tech/tags/component.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [personal](<https://devfeed.tech/tags/personal.md>), [personal-agent](<https://devfeed.tech/tags/personal-agent.md>), [rsi](<https://devfeed.tech/tags/rsi.md>), [updates](<https://devfeed.tech/tags/updates.md>)

### AI overview

Safe Personal RSI evaluates proposed changes to Kicker's reusable instructions while keeping model weights and runtime code fixed. It tests current and proposed prompts, repeats failed cases, temporarily reverses changes to guide repairs, and adopts an update only if it passes safety and task-performance checks; otherwise, the current instructions remain. In a 29-case benchmark across three models and two risky edits, it rejected 20 of 21 update attempts. A confirmed GPT-5.6 Luna repair reduced attack violations from 10.7% to 0% and raised safe task completion from 89.7% to 100%, matching a broad handwritten authority rule. A repair transferred from Qwen2.5-7B-Instruct to Qwen3-14B reduced violations under stronger attacks from 69.0% to 17.9% and raised safe completion from 19.5% to 64.4%. Separate runtime probes found missing local OAuth scope checks in the evaluated snapshot.

### Source excerpt

arXiv:2611.23651v1 Announce Type: new Abstract: Kicker is a personal health assistant. Changes to an assistant's reusable instructions can improve task handling while changing what it discloses, which requests it follows, and how it reports actions. We study Safe Personal RSI, an update-evaluation component developed for Kicker. We restrict recursive self-improvement (RSI) to prompt updates, with model weights and runtime code fixed. The procedure tests proposed and current instructions, repeats failing cases, temporarily undoes changes to guide repairs, and checks safety and task performance before adoption. If no version passes, the current instructions stay in use. Our executable benchmark contains 29 email, calendar, health, and marketplace cases, with separate probes for authorization reuse, memory, and tool-output injection. Across three models and two deliberately risky edits, the procedure rejects 20 of 21 update attempts. In repair confirmation, an admitted GPT-5.6 Luna repair reduces attack violations from 10.7% to 0% and increases safe task completion from 89.7% to 100%; a broad handwritten authority rule matches these outcomes. Under stronger attacks, a repair transferred from Qwen2.5-7B-Instruct to Qwen3-14B reduces violations from 69.0% to 17.9% and increases safe task completion from 19.5% to 64.4%. Separate runtime probes examine memory-entry deletion and identify missing local OAuth scope checks in the evaluated snapshot.

## When Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation

DevFeed: [When Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation](<https://devfeed.tech/articles/when-does-subspace-direction-matter-for-lora-regime-analysis-of-the-magnitude-principle-in-few-shot-adaptation-143392.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23428>)

Author: Nischal Subedi, Cencheng Shen, Peng Zhao

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [lora](<https://devfeed.tech/topics/lora.md>), [few-shot](<https://devfeed.tech/topics/few-shot.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>)

Tags: [activation](<https://devfeed.tech/tags/activation.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [bert](<https://devfeed.tech/tags/bert.md>), [classification](<https://devfeed.tech/tags/classification.md>), [completion](<https://devfeed.tech/tags/completion.md>), [covariance](<https://devfeed.tech/tags/covariance.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [dimension](<https://devfeed.tech/tags/dimension.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [few-shot](<https://devfeed.tech/tags/few-shot.md>), [llama](<https://devfeed.tech/tags/llama.md>), [lora](<https://devfeed.tech/tags/lora.md>), [magnitude](<https://devfeed.tech/tags/magnitude.md>), [span](<https://devfeed.tech/tags/span.md>), [stat-ml](<https://devfeed.tech/tags/stat-ml.md>), [structure](<https://devfeed.tech/tags/structure.md>), [subspace](<https://devfeed.tech/tags/subspace.md>), [theory](<https://devfeed.tech/tags/theory.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

The paper finds that LoRA initialization direction matters most when labeled data are scarce relative to representation dimension and the initialization subspace captures task structure. Its theory includes a finite-sample stability analysis and a first-step account that motivates an alignment coefficient, ρ, which tracks classification gains. PivotLoRA uses activation PCA at prediction-relevant positions, then applies an output-preserving orthogonal pivot early in training. In few-shot BERT-base classification it ranks best on average among 13 methods and beats vanilla LoRA in 29 of 30 settings despite a smaller input-factor initialization norm. Gains diminish with more labeled data; targeted experiments also report improvements on GSM8K and StrategyQA. The supplied text does not detail those evaluation setups or effect sizes.

### Source excerpt

arXiv:2611.23428v1 Announce Type: new Abstract: We show that the importance of a LoRA adapter's initial input-subspace direction is regime-dependent, and we characterize this dependence in theory and experiment. This refines the recent view that spectral initializers help through update magnitude alone. Direction matters most when labeled data are scarce relative to the representation dimension and the activation subspace used for initialization captures supervised task structure. In this regime, the adapter's row span remains sensitive to its starting point, making task-aligned initialization especially useful. A spiked-covariance perturbation analysis characterizes the finite-sample stability of the estimated subspace. A first-step analysis of LoRA training explains why the subspace's alignment with the task matters and motivates an alignment coefficient $\rho$ that closely tracks gains across classification tasks. We test this account with PivotLoRA, which initializes adapters from activation PCA at prediction-relevant positions and applies one output-preserving orthogonal pivot early in training. In few-shot BERT-base classification, PivotLoRA achieves the best average rank among 13 methods and improves over vanilla LoRA in 29 of 30 settings, even though it starts from a smaller input-factor initialization norm. Additional classification experiments extend the evaluation to 7B and 8B Llama decoders. Classification gains diminish with more labeled data, consistent with the proposed regime dependence. A targeted extension to completion-only reasoning also shows improvements on GSM8K and StrategyQA. These findings identify conditions under which initialization direction matters, with subspace-task alignment offering a low-cost diagnostic before training for classification.

## Meaning Reaches Object Choice but Not Completion: Where Is a VLA Vulnerable to Landmark Synonyms?

DevFeed: [Meaning Reaches Object Choice but Not Completion: Where Is a VLA Vulnerable to Landmark Synonyms?](<https://devfeed.tech/articles/meaning-reaches-object-choice-but-not-completion-where-is-a-vla-vulnerable-to-landmark-synonyms-143480.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24076>)

Author: Ryoichi Kurihara, Kota Sunamoto

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [vlm](<https://devfeed.tech/topics/vlm.md>)

Tags: [choice](<https://devfeed.tech/tags/choice.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [finetuning](<https://devfeed.tech/tags/finetuning.md>), [object](<https://devfeed.tech/tags/object.md>), [robotics](<https://devfeed.tech/tags/robotics.md>), [robustness](<https://devfeed.tech/tags/robustness.md>), [vla](<https://devfeed.tech/tags/vla.md>)

### AI overview

In released GR00T N1.7, landmark synonyms absent from training instructions often direct the robot to the named bowl but do not ensure task completion: one LIBERO-Spatial task reached 67.6% closed-loop success with synonyms versus 100% with the original landmark. Before the vision-language/action-generator connector, swapping landmark-token vectors nearly always changed the chosen bowl. After it, changing any one of the landmark, remaining instruction, or image tokens rarely changed choice, while changing any two did. Masking landmark tokens after the connector still reduced synonym-task completion, suggesting choice and execution can fail at different stages. Diversified-word finetuning repaired completion for tested synonyms, whereas finetuning on original instructions did not; connector-only training or transferring a diversified policy's connector also sufficed. Starting-frame open-loop scores missed repair carried by weights outside the connector. The study supports evaluating rewording in closed loop and scoring object choice separately from completion; evidence is limited to the released policy and two tasks.

### Source excerpt

arXiv:2611.24076v1 Announce Type: cross Abstract: VLA policies are fragile to rewording. Success rates alone cannot say whether a policy fails to choose the object or to complete the task, nor where in the policy to intervene. We study the released GR00T N1.7 on two LIBERO-Spatial tasks differing only in the landmark word telling which of two bowls to take. Synonyms absent from the training instructions send the policy to the bowl they name, unlike made-up words, yet complete 67.6% of one task's closed-loop episodes, against the original word's 100%. Before the connector between the vision-language model and the action generator, replacing only the landmark tokens' vectors with the other landmark's (pasting) sends nearly every episode to the other bowl. After the connector, pasting at the landmark, the remaining instruction, or the image tokens alone rarely moves the choice, but any two do. Hiding the landmark tokens from the action generator after the connector still lowers completion for the synonyms that fall short. Finetuning on diversified landmark words repairs that one task's completion for synonyms absent from finetuning; finetuning on the original instructions does not. Training only the connector on the diversified words suffices, as does copying the connector of a policy fully finetuned on them into one finetuned on the original instructions. An open-loop score on starting frames misses that weights outside the connector also carry part of the repair. For now, on this evidence, robustness to rewording should be judged in closed loop, with actions executed, intervening on both sides of the connector, scoring choice and completion separately.

## AInfer-PD: Communication-Safe In-Place Prefill-Decode Multiplexing for Distributed MoE Rollouts

DevFeed: [AInfer-PD: Communication-Safe In-Place Prefill-Decode Multiplexing for Distributed MoE Rollouts](<https://devfeed.tech/articles/ainfer-pd-communication-safe-in-place-prefill-decode-multiplexing-for-distributed-moe-rollouts-140376.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.00993>)

Author: Guowei Wang, Chaokun Yang, Zhenxuan Pan, Yipeng Wei, Yuhong Guo, Minghua Zhu, Zechuan Zhang, Shuo Wan, Xiaowei Zhu

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>)

Topics: [Distributed Systems & Parallel Computing](<https://devfeed.tech/topics/distributed-systems-parallel-computing.md>), [async](<https://devfeed.tech/topics/async.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [distributed algorithm](<https://devfeed.tech/topics/distributed-algorithm.md>)

Tags: [ablation](<https://devfeed.tech/tags/ablation.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [asynchronous](<https://devfeed.tech/tags/asynchronous.md>), [batches](<https://devfeed.tech/tags/batches.md>), [collective](<https://devfeed.tech/tags/collective.md>), [communication](<https://devfeed.tech/tags/communication.md>), [completion](<https://devfeed.tech/tags/completion.md>), [continuation](<https://devfeed.tech/tags/continuation.md>), [cross](<https://devfeed.tech/tags/cross.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [d](<https://devfeed.tech/tags/d.md>), [decode](<https://devfeed.tech/tags/decode.md>), [deepep](<https://devfeed.tech/tags/deepep.md>), [devices](<https://devfeed.tech/tags/devices.md>), [distributed](<https://devfeed.tech/tags/distributed.md>), [engine](<https://devfeed.tech/tags/engine.md>), [moe](<https://devfeed.tech/tags/moe.md>), [multiplexing](<https://devfeed.tech/tags/multiplexing.md>), [pd](<https://devfeed.tech/tags/pd.md>), [prefill](<https://devfeed.tech/tags/prefill.md>), [safe](<https://devfeed.tech/tags/safe.md>)

### AI overview

AInfer-PD lets decode iterations progress during long prefills in distributed MoE rollouts by coordinating collective ordering across ranks at safe communication boundaries and separating prefill and decode communication buffers and protocol state within one DeepEP runtime. This allows prefills to use DeepEP's normal path while decode keeps its low-latency path, with model weights and KV storage shared on the same devices. In reported prefill-intensive workloads, fixed-workload completion time fell 7.1-22.5% versus the same AInfer engine without P/D multiplexing and 24.8-32.9% versus SGLang on one node; on two nodes, reductions were 18.0-35.3% and 18.3-31.8%, respectively. A same-engine ablation found fine-grained boundaries improved completion time by a further 8.6-19.8% over whole-prefill asynchronous enqueue.

### Source excerpt

arXiv:2609.00993v3 Announce Type: replace Abstract: In agentic reinforcement learning (RL), asynchronous multi-turn trajectories repeatedly introduce continuation prefills (P) while other trajectories remain in decode (D). Efficient rollout completion therefore requires D to advance during long P executions. In-place multiplexing shares devices, model weights, and KV state, but phase-specific communication in distributed MoE introduces two constraints: cross-rank collective ordering and ownership of mutable communication state. Whole-prefill coordination can ensure safe ordering but delays D. DeepEP's normal and low-latency paths suit large P and small D batches, respectively, but safely overlapping them requires separating their mutable communication state. We present AInfer-PD, which extends in-place P/D multiplexing to distributed MoE rollouts. AInfer-PD establishes rank-consistent collective ordering at safe communication boundaries, allowing successive D iterations to advance without waiting for an entire prefill execution to be enqueued or completed. It gives P and D separate communication buffers and protocol state within one DeepEP runtime, allowing P to use the normal path while D retains the low-latency path. Together, these mechanisms preserve useful P/D overlap while sharing model weights and KV storage on the same devices. Across repeated single-node prefill-intensive workloads, AInfer-PD reduces fixed-workload rollout completion time by 7.1-22.5% relative to the same AInfer engine with P/D multiplexing disabled and by 24.8-32.9% relative to SGLang. On two nodes, the reductions are 18.0-35.3% and 18.3-31.8%, respectively. In a same-engine ablation, fine-grained boundaries reduce completion time by a further 8.6-19.8% over whole-prefill asynchronous enqueue.

## Streaming On-Policy Distillation: Decoupling Three Clocks for Stage-Wise Supervision Consumption

DevFeed: [Streaming On-Policy Distillation: Decoupling Three Clocks for Stage-Wise Supervision Consumption](<https://devfeed.tech/articles/streaming-on-policy-distillation-decoupling-three-clocks-for-stage-wise-supervision-consumption-140007.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22298>)

Author: Yun Zhu Da Wang

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Streaming](<https://devfeed.tech/topics/streaming.md>)

Tags: [budget](<https://devfeed.tech/tags/budget.md>), [checkpoint](<https://devfeed.tech/tags/checkpoint.md>), [clocks](<https://devfeed.tech/tags/clocks.md>), [completion](<https://devfeed.tech/tags/completion.md>), [computing](<https://devfeed.tech/tags/computing.md>), [continuation](<https://devfeed.tech/tags/continuation.md>), [control](<https://devfeed.tech/tags/control.md>), [cost](<https://devfeed.tech/tags/cost.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [data](<https://devfeed.tech/tags/data.md>), [decoupling](<https://devfeed.tech/tags/decoupling.md>), [delivery](<https://devfeed.tech/tags/delivery.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [distilling](<https://devfeed.tech/tags/distilling.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [gap](<https://devfeed.tech/tags/gap.md>), [policy](<https://devfeed.tech/tags/policy.md>), [stage](<https://devfeed.tech/tags/stage.md>), [streaming](<https://devfeed.tech/tags/streaming.md>), [supervision](<https://devfeed.tech/tags/supervision.md>)

### AI overview

Streaming on-policy distillation (OPD) delivers newly generated response segments while generation continues, preserving each segment's full causal prefix but computing loss only on new positions. A consumption protocol manages supervision budgets, position selection, staleness, and update timing across training stages. In GSM8K experiments distilling Qwen3-4B into Qwen3-0.6B with VERL, budget matching narrowed an early test-accuracy gap across seven common candidates from 4.55 to 2.42 points. Synchronous continuation reached 73.39% best accuracy, below synchronous training's 74.30% and above the budget-matched run's 71.95%; a 76.12% extended record was unpaired. Freshness filtering and disagreement-overlap selection improved local peaks without showing sustained advantages. A prototype reduced data-supply waiting time but did not establish equal-update speedup; end-to-end gains must account for continuation cost.

### Source excerpt

arXiv:2611.22298v1 Announce Type: new Abstract: In on-policy distillation (OPD), response completion, supervision delivery, and parameter updates need not share the same granularity. We develop streaming OPD by decoupling these three clocks: a response repeatedly delivers newly generated segments while generation continues, retaining the full causal prefix at each delivery and computing loss only on new positions. A consumption protocol then organizes supervision budgets, position selection, staleness handling, and update scheduling, allowing control over quantity, coverage, and version lag across training stages. In GSM8K experiments distilling Qwen3-4B into Qwen3-0.6B with VERL, supervised-budget matching reduces the early gap between best test accuracies over seven common candidates from 4.55 to 2.42 points. Synchronous continuation from the same checkpoint reaches 73.39% best test accuracy over seven common candidates at u=384,...,576, versus 74.30% synchronous and 71.95% budget-matched (differences -0.91 and -2.35 points); an unpaired extended record reaches 76.12%. Freshness filtering and disagreement-overlap selection improve local peaks but do not establish sustained advantages. A system prototype shows lower data-supply waiting times without establishing an equal-update speedup. Quality recovery and supply-side benefits are assessed separately; end-to-end gains must account for continuation cost.

[Next page](<https://devfeed.tech/tags/completion.md?cursor=WyIyMDI2LTExLTE4VDA1OjAwOjAwKzAwOjAwIiwgImVkMDA5M2QzLTUyNTItNGU0MS04NTJlLWE2MzIzZDc4MDkyYyJd>)