# Post-training optimization

Machine-learning techniques applied to trained models, such as quantization, pruning, and distillation, to improve deployment efficiency or model behavior.

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

## FlowMap-OPD: On-Policy Distillation for Few-Step Flow-Map Generators

DevFeed: [FlowMap-OPD: On-Policy Distillation for Few-Step Flow-Map Generators](<https://devfeed.tech/articles/flowmap-opd-on-policy-distillation-for-few-step-flow-map-generators-144946.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.37851>)

Author: Zhiqi Li, Cheng Wan, Bo Zhu

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

Content type: article

Language: en

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

Topics: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [flux](<https://devfeed.tech/topics/flux.md>)

Tags: [consistency](<https://devfeed.tech/tags/consistency.md>), [consistency-models](<https://devfeed.tech/tags/consistency-models.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [flow](<https://devfeed.tech/tags/flow.md>), [flux](<https://devfeed.tech/tags/flux.md>), [generators](<https://devfeed.tech/tags/generators.md>), [policy](<https://devfeed.tech/tags/policy.md>), [sd3-5](<https://devfeed.tech/tags/sd3-5.md>), [supervision](<https://devfeed.tech/tags/supervision.md>), [velocity](<https://devfeed.tech/tags/velocity.md>)

### AI overview

FlowMap-OPD trains few-step flow-map students using states collected by the student's native sampler, while separating that collection from teacher supervision. The study finds that supervising instantaneous velocity with a small consistency penalty transfers teacher gains best; directly supervising the deployed flow map performs worst when teacher and student use different pretrained initializations. The authors summarize the approach as "sample with the flow map, learn the velocity." It also supports teachers that define a velocity, including multi-step teachers, and students initialized from multi-step models. On FLUX and SD3.5, one few-step student consolidated three specialist teachers within 300 updates, matched or exceeded each on its own benchmark, and outperformed multi-reward reinforcement learning and LoRA merging. The supplied text is an abstract, so it does not provide full methodological or evaluation details.

### Source excerpt

arXiv:2609.37851v2 Announce Type: replace-cross Abstract: Few-step flow-map generators, such as MeanFlow and consistency models, sample in a few long-range jumps. On-policy distillation (OPD) trains a student on states drawn from its own sampler, but existing methods compare teacher and student through the local stochastic transitions of a multi-step sampler, which a flow-map student never deploys. We present FlowMap-OPD, the first systematic study of OPD for flow-map students. It separates state collection from supervision: the student collects states with its native few-step sampler, while the teacher supervises the flow map, velocities induced from it, or the instantaneous velocity linked to the map by a separately weighted consistency penalty. We find that instantaneous-velocity supervision with a small consistency weight, our chosen form, transfers the teacher's gains best, whereas supervising the deployed map, although the natural choice, transfers them worst when teacher and student start from different pretrained models. This gives a counter-intuitive rule: \emph{sample with the flow map, learn the velocity}. As a further benefit, this form applies to any teacher that defines a velocity, including multi-step ones, and to students initialized from multi-step models. On FLUX and SD3.5, a single few-step student consolidates three specialist teachers within 300 updates, matches or exceeds each on its own benchmark (e.g., OCR 0.883 versus 0.850 on FLUX), and outperforms both multi-reward reinforcement learning and LoRA merging.

## GLiDE: A Fast, Calibrated Decision Model Distilled from Reasoning Teachers

DevFeed: [GLiDE: A Fast, Calibrated Decision Model Distilled from Reasoning Teachers](<https://devfeed.tech/articles/glide-a-fast-calibrated-decision-model-distilled-from-reasoning-teachers-145678.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26509>)

Author: Sahibzada Allahyar, Julia White, Dhruv Atreja, George Hurn-Maloney, Urchade Zaratiana

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>), [Training AI Models](<https://devfeed.tech/topics/training-ai-models.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [calibrated-decision-model](<https://devfeed.tech/tags/calibrated-decision-model.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [context](<https://devfeed.tech/tags/context.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.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>), [di](<https://devfeed.tech/tags/di.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [fast](<https://devfeed.tech/tags/fast.md>), [inference](<https://devfeed.tech/tags/inference.md>), [language](<https://devfeed.tech/tags/language.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [latency](<https://devfeed.tech/tags/latency.md>), [model](<https://devfeed.tech/tags/model.md>), [performance](<https://devfeed.tech/tags/performance.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reasoning-model](<https://devfeed.tech/tags/reasoning-model.md>)

### AI overview

GLiDE transfers reasoning capability into single-pass decision models that score candidate answers without generating text or intermediate reasoning. The 0.4B-27B family uses Qwen hybrid language models with bidirectional interactions among context, question, and candidates, plus soft-target distillation from multiple teachers--including a reasoning model--and progressive post-training on varied decision tasks. Parameter-free recurrent depth adds inference-time computation without token generation. The abstract reports that each of five sizes outperforms evaluated alternatives of equal or smaller size on the Hugging Face benchmark. The 27B model scores 71.52 on a hidden set with 2.0% expected calibration error and 92.9 ms median latency on one GPU; checkpoint averaging and recurrent depth raise the score to 73.37, with 2.1% calibration error.

### Source excerpt

arXiv:2611.26509v1 Announce Type: new Abstract: Decision models enable fast, confidence-aware decisions by scoring candidate answers in a single forward pass, without generating text or intermediate reasoning. Yet reasoning can substantially improve decision quality, raising the question of whether its benefits can be retained without the cost of generating it. We investigate this question with \textbf{GLiDE}, a family of decision models spanning 0.4B to 27B parameters. Building on Qwen hybrid language models, GLiDE enables bidirectional interactions between the context, question, and candidate answers by removing causal masking from full-attention layers. To transfer reasoning capabilities into single-pass inference, we combine soft-target distillation from multiple teachers, including a reasoning model, with progressive post-training on diverse decision tasks. We further explore parameter-free recurrent depth to increase inference-time computation without generating tokens. Across five model sizes, GLiDE lies on the performance--parameter Pareto frontier of the Hugging Face \DI{}, with each model outperforming evaluated alternatives of equal or smaller size. \dmtwo{} (27B) achieves a hidden-set score of 71.52, with an expected calibration error of 2.0\% and a median latency of 92.9\,ms on a single GPU. Further checkpoint averaging and parameter-free recurrent depth yield GLiDE~27B, which reaches 73.37 with an expected calibration error of 2.1\%.

## Model Compression vs. Adversarial Robustness: An Empirical Study on Language Models for Code

DevFeed: [Model Compression vs. Adversarial Robustness: An Empirical Study on Language Models for Code](<https://devfeed.tech/articles/model-compression-vs-adversarial-robustness-an-empirical-study-on-language-models-for-code-145758.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2508.03949>)

Author: Md. Abdul Awal, Mrigank Rochan, Chanchal K. Roy

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

Content type: article

Language: en

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

Topics: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [code](<https://devfeed.tech/tags/code.md>), [compression](<https://devfeed.tech/tags/compression.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [impact](<https://devfeed.tech/tags/impact.md>), [inference](<https://devfeed.tech/tags/inference.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [performance](<https://devfeed.tech/tags/performance.md>), [pruning](<https://devfeed.tech/tags/pruning.md>), [quantization](<https://devfeed.tech/tags/quantization.md>), [real-world](<https://devfeed.tech/tags/real-world.md>), [robustness](<https://devfeed.tech/tags/robustness.md>), [security](<https://devfeed.tech/tags/security.md>), [size](<https://devfeed.tech/tags/size.md>), [software](<https://devfeed.tech/tags/software.md>), [speeds](<https://devfeed.tech/tags/speeds.md>), [strategies](<https://devfeed.tech/tags/strategies.md>)

### AI overview

The study finds that compression can preserve language models' usual performance on software analytics tasks while significantly reducing their robustness to adversarial attacks. It evaluates pruning, quantization, and knowledge distillation across compressed versions of three code language models, three tasks, six metrics, and four classical attacks. The results point to a trade-off between reducing model size and adversarial robustness, a concern for security-critical deployments. The abstract does not specify which compression methods or attacks caused the largest effects.

### Source excerpt

arXiv:2508.03949v3 Announce Type: replace Abstract: Transformer-based language models for code have shown remarkable performance in various software analytics tasks, but their adoption is hindered by high computational costs, slow inference speeds, and substantial environmental impact. Model compression techniques such as pruning, quantization, and knowledge distillation have gained traction in addressing these challenges. However, the impact of these strategies on the robustness of compressed language models for code in adversarial scenarios remains poorly understood. Understanding how these compressed models behave under adversarial attacks is essential for their safe and effective deployment in real-world applications. To bridge this knowledge gap, we conduct a comprehensive evaluation of how common compression strategies affect the adversarial robustness of compressed models. We assess the robustness of compressed versions of three widely used language models for code across three software analytics tasks, using six evaluation metrics and four commonly used classical adversarial attacks. Our findings indicate that compressed models generally maintain comparable performance to their uncompressed counterparts. However, when subjected to adversarial attacks, compressed models exhibit significantly reduced robustness. These results reveal a trade-off between model size reduction and adversarial robustness, underscoring the need for careful consideration when deploying compressed models in security-critical software applications. Our study highlights the need for further research into compression strategies that strike a balance between computational efficiency and adversarial robustness, which is essential for deploying reliable language models for code in real-world software applications.

## VisAlloc: Allocating Visual Supervision for On-Policy Self-Distillation

DevFeed: [VisAlloc: Allocating Visual Supervision for On-Policy Self-Distillation](<https://devfeed.tech/articles/visalloc-allocating-visual-supervision-for-on-policy-self-distillation-145461.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25069>)

Author: Haoran Xu, Tian Chen, Yuhe Wu, Yanbo Zhao, Junchi Liao, Hongyu Wang, Jiaze Li, Wentao Wang, Yanfeng Li, Rui Qian

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

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>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [smolvlm](<https://devfeed.tech/topics/smolvlm.md>)

Tags: [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [filter](<https://devfeed.tech/tags/filter.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [supervision](<https://devfeed.tech/tags/supervision.md>), [token](<https://devfeed.tech/tags/token.md>)

### AI overview

VisAlloc allocates visual supervision in on-policy self-distillation using two filters: a paired-teacher sample filter keeps cases answered correctly from an evidence-centered crop but incorrectly from a spatially displaced crop; then token weights reflect view sensitivity or disagreement between EMA teachers scoring the same student response. Frozen-prior, template-bigram, and student-teacher divergence gates constrain added supervision, whose total weight is redistributed by inverse frequency within each micro-batch. The positive-view teacher alone supplies distillation targets, and inference uses only the full-image student. Across eight benchmarks, the 4B and 9B models report macro-averages of 81.44 and 84.92, exceeding Vision-OPD reproductions by 0.97 and 0.52 points, respectively.

### Source excerpt

arXiv:2611.25069v1 Announce Type: cross Abstract: On-policy self-distillation (OPSD) offers a promising approach to improving visual reasoning in multimodal large language models (MLLMs). However, not all samples benefit equally from privileged visual guidance, and uniform token supervision can dilute visually grounded signals. We propose VisAlloc, a dual-level framework combining sample-level visual filtering with token-level counterfactual reweighting. A fixed paired-teacher filter retains samples answered correctly from an evidence-centered positive crop but incorrectly from a spatially displaced negative crop. At the token level, both EMA teachers score the same student-generated response, and their absolute log-probability differences measure view sensitivity. Additional supervision targets tokens admitted by sensitivity or teacher-distribution disagreement, with shared frozen-prior, template-bigram, and student--teacher divergence gates. Inverse-frequency redistribution preserves the total additional weight within each micro-batch, while the positive teacher remains the sole distillation target. Inference requires only the full-image student. Across eight benchmarks, our 4B and 9B models achieve macro-averages of 81.44 and 84.92, outperforming Vision-OPD reproductions by 0.97 and 0.52 points, respectively. Code on GitHub: https://github.com/Kaelyn01/VisAlloc

## AVSD: Adaptive-View Self-Distillation by Balancing Consensus and Teacher-Specific Privileged Signals

DevFeed: [AVSD: Adaptive-View Self-Distillation by Balancing Consensus and Teacher-Specific Privileged Signals](<https://devfeed.tech/articles/avsd-adaptive-view-self-distillation-by-balancing-consensus-and-teacher-specific-privileged-signals-144896.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.20643>)

Author: Duy Nguyen, Hanqi Xiao, Archiki Prasad, Zaid Khan, Anirban Das, Shi-Xiong Zhang, Sambit Sahu, Hyunji Lee, Elias Stengel-Eskin, Mohit Bansal

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: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [grpo](<https://devfeed.tech/topics/grpo.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [balancing](<https://devfeed.tech/tags/balancing.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [code](<https://devfeed.tech/tags/code.md>), [code-generation](<https://devfeed.tech/tags/code-generation.md>), [competition](<https://devfeed.tech/tags/competition.md>), [consensus](<https://devfeed.tech/tags/consensus.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [grpo](<https://devfeed.tech/tags/grpo.md>), [language](<https://devfeed.tech/tags/language.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [multiple](<https://devfeed.tech/tags/multiple.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [residual](<https://devfeed.tech/tags/residual.md>), [signal](<https://devfeed.tech/tags/signal.md>), [solutions](<https://devfeed.tech/tags/solutions.md>), [stable](<https://devfeed.tech/tags/stable.md>), [student](<https://devfeed.tech/tags/student.md>), [view](<https://devfeed.tech/tags/view.md>)

### AI overview

AVSD combines multiple privileged teacher views for language-model self-distillation by separating their shared consensus signal from view-specific residuals. It uses consensus as the update direction and adds a residual only when it aligns with that direction and is proportionate to the consensus, addressing both inaccessible teacher information and task-dependent view choice. On AIME24, AIME25, and HMMT25, it outperformed single-view self-distillation and GRPO, with average Avg@8 gains over the strongest baselines of 3.1% for Qwen3-8B and 2.2% for Qwen3-4B. On Codeforces and LiveCodeBench v6, Qwen3-8B AVSD exceeded the single-view baseline by 2.4% on average.

### Source excerpt

arXiv:2605.20643v2 Announce Type: replace Abstract: Self-distillation enables language models to learn on-policy from their own trajectories by using the same model as both student and teacher, with the teacher being conditioned on privileged information unavailable to the student. Such information can come in different types or views, such as solutions, demonstrations, feedback, or final answers. This setup provides dense token-level feedback without relying on a separate external model, but creates a fundamental asymmetry: the teacher may rely on view-specific information that the student cannot access at inference time. Moreover, the best type of privileged information is often task-dependent, making it difficult to choose a single teacher view. In this work, we address both these challenges jointly by introducing AVSD (Adaptive-View Self-Distillation), a novel method of self-distillation with multiple privileged-information views, which reconstructs token-level supervision by separating stable cross-view consensus from view-specific residual signals. AVSD identifies the consensus signal shared across views, which provides a reliable update direction, and then selectively adds the view-specific residual signal to adjust the update magnitude when it both aligns with the consensus direction and remains proportionate to the consensus signal. Experiments on math competition benchmarks (AIME24, AIME25, and HMMT25) show that AVSD consistently outperforms both single-view self-distillation baselines and GRPO, achieving average Avg@8 gains of 3.1% and 2.2% over the strongest baselines on Qwen3-8B and Qwen3-4B, respectively. Moreover, on code-generation benchmarks (Codeforces, LiveCodeBench v6) using Qwen3-8B, AVSD outperforms the single-view self-distillation baseline by 2.4% on average.

## ICARE: cross-domain text classification with incremental class-aware representation and distillation learning

DevFeed: [ICARE: cross-domain text classification with incremental class-aware representation and distillation learning](<https://devfeed.tech/articles/icare-cross-domain-text-classification-with-incremental-class-aware-representation-and-distillation-learning-144992.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24959>)

Author: Son Duong, Truong X. Tran, Hai-Anh Tran

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

Content type: article

Language: en

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

Topics: [text-classification](<https://devfeed.tech/topics/text-classification.md>), [Multitask learning](<https://devfeed.tech/topics/multitask-learning.md>), [Representation learning](<https://devfeed.tech/topics/representation-learning.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [classification](<https://devfeed.tech/tags/classification.md>), [cross](<https://devfeed.tech/tags/cross.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [domain](<https://devfeed.tech/tags/domain.md>), [framework](<https://devfeed.tech/tags/framework.md>), [incremental](<https://devfeed.tech/tags/incremental.md>), [inference](<https://devfeed.tech/tags/inference.md>), [knowledge-transfer](<https://devfeed.tech/tags/knowledge-transfer.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [learning](<https://devfeed.tech/tags/learning.md>), [llms](<https://devfeed.tech/tags/llms.md>), [news](<https://devfeed.tech/tags/news.md>), [overfitting](<https://devfeed.tech/tags/overfitting.md>), [pruning](<https://devfeed.tech/tags/pruning.md>), [representation-learning](<https://devfeed.tech/tags/representation-learning.md>), [sentiment-analysis](<https://devfeed.tech/tags/sentiment-analysis.md>), [text-classification](<https://devfeed.tech/tags/text-classification.md>)

### AI overview

ICARE aims to reduce catastrophic forgetting in continual text classification by combining representation-based instance selection with knowledge distillation across tasks. Its distillation transfers feature-level and logit-level knowledge using pruning techniques and optimized objectives. The abstract reports improved target accuracy and a lower forgetting metric on benchmark text classification datasets, but provides no dataset names, numerical results, or detailed comparison baselines.

### Source excerpt

arXiv:2611.24959v1 Announce Type: cross Abstract: The rise of powerful, fine-tuned large language models (LLMs) necessitates robust natural language inference (NLI) systems. These NLI systems ideally train across various tasks, encompassing diverse concepts like sentiment analysis, news classification, and coreference resolution. However, this continual, multi-tasking approach can lead to catastrophic forgetting, where previously learned tasks are lost, especially in domains lacking specific labels. Rehearsing past examples with labels can further exacerbate this issue by causing overfitting, hindering the system's ability to classify new categories. To tackle these challenges, this research proposes ICARE, a novel framework with two key components: an instance selection module and a knowledge distillation method. The former is based on representation learning, which meticulously selects training examples to present to the NLI system. The latter, knowledge distillation with feature and logit-level transfer, facilitates knowledge transfer between tasks. This leverages pruning techniques and optimized objective functions to transfer both essential feature-level and decision-making (logit-level) knowledge, mitigating catastrophic forgetting while adapting to new patterns. Extensive evaluations on benchmark text classification datasets have demonstrated ICARE's effectiveness. The framework achieves a significant improvement in target accuracy and reduction in the ``forgetting metric''. These results solidify ICARE's ability to preserve and leverage learned knowledge across diverse tasks, ultimately enhancing the adaptability and performance of NLI systems. The source code is available at https://github.com/CongSon01/ICARE_NLP.git

## NariQ: Sub-3-Bit Quantization for Long-Horizon Agentic Tasks

DevFeed: [NariQ: Sub-3-Bit Quantization for Long-Horizon Agentic Tasks](<https://devfeed.tech/articles/nariq-sub-3-bit-quantization-for-long-horizon-agentic-tasks-145146.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26157>)

Author: Doyeob Kim

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

Content type: article

Language: en

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

Topics: [quantization](<https://devfeed.tech/topics/quantization.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>)

Tags: [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-benchmark](<https://devfeed.tech/tags/agentic-benchmark.md>), [agentic-tasks](<https://devfeed.tech/tags/agentic-tasks.md>), [aime](<https://devfeed.tech/tags/aime.md>), [bf16](<https://devfeed.tech/tags/bf16.md>), [bits](<https://devfeed.tech/tags/bits.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [deepswe](<https://devfeed.tech/tags/deepswe.md>), [gguf](<https://devfeed.tech/tags/gguf.md>), [inference](<https://devfeed.tech/tags/inference.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [quantization](<https://devfeed.tech/tags/quantization.md>), [qwen3](<https://devfeed.tech/tags/qwen3.md>), [qwen3-8-27b](<https://devfeed.tech/tags/qwen3-8-27b.md>), [terminal-bench](<https://devfeed.tech/tags/terminal-bench.md>), [trellis](<https://devfeed.tech/tags/trellis.md>)

### AI overview

NariQ targets a gap in low-bit quantization for long-horizon agent tasks: methods calibrated on short web text can retain performance on AIME while doing substantially worse on DeepSWE. Its trellis-coded post-training method calibrates on tokens generated by the BF16 model during agent sessions. For Qwen3.8-27B, the 3.0- and 2.5-bit-per-weight versions are reported to match BF16 on DeepSWE and Terminal Bench, as well as AIME and RealWorldQA; the 2.0-bit checkpoint trails that performance. The authors describe these as the first quantizations at or below 3 bits shown to match full precision on long-horizon agent tasks. They provide GGUF files and inference through a llama.cpp fork.

### Source excerpt

arXiv:2611.26157v1 Announce Type: new Abstract: Long-horizon agentic tasks require a model to call tools over many steps at contexts of 100K tokens or more, yet many low-bit quantization methods are calibrated on short web text and judged on perplexity. This mismatch is costly: existing quantizations of Qwen3.8-27B at or below 3 bits per weight match BF16 on shorter tasks like AIME, yet in published results solve at most 72% as many tasks as BF16 on DeepSWE, a long-horizon agentic benchmark. We propose NariQ, a trellis-coded post-training quantization method that calibrates on tokens generated by the BF16 model in agent sessions. On Qwen3.8-27B, NariQ at 3.0 and 2.5 bits per weight stays level with BF16 on two long-horizon agentic benchmarks, DeepSWE and Terminal Bench, as well as on AIME and RealWorldQA. Our checkpoint at 2.0 bits per weight trails this performance. To our knowledge, these are the first quantizations at or below 3 bits per weight shown to match their full-precision version on long-horizon agentic tasks. We release the files as Qwen3.8-27B-NariQ-GGUF (https://huggingface.co/nari-labs/Qwen3.8-27B-NariQ-GGUF), with inference available through our llama.cpp fork at https://github.com/nari-labs/nari-llama.cpp.

## Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization

DevFeed: [Multi-Objective Structured Pruning of LLMs for Latency and Model Size Optimization](<https://devfeed.tech/articles/multi-objective-structured-pruning-of-llms-for-latency-and-model-size-optimization-145356.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2607.22583>)

Author: Muhammad Junaid Ali, Smail Niar, El-Ghazali Talbi

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

Content type: article

Language: en

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

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Hyperparameter optimization](<https://devfeed.tech/topics/hyperparameter-optimization.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [edge-computing](<https://devfeed.tech/tags/edge-computing.md>), [latency](<https://devfeed.tech/tags/latency.md>), [llms](<https://devfeed.tech/tags/llms.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [pruning](<https://devfeed.tech/tags/pruning.md>), [size](<https://devfeed.tech/tags/size.md>), [structured](<https://devfeed.tech/tags/structured.md>)

### AI overview

The paper proposes a hardware-aware, two-stage structured pruning framework to balance LLM latency and model size for edge deployment. First, multi-objective depth pruning removes entire attention and MLP blocks; then Parallel Bayesian Optimization selects layer-wise pruning ratios under latency constraints, while importance-based strategies choose components within each layer's budget. The abstract reports that across multiple LLMs pruned by 37.5% and 50%, the approach outperformed existing methods on commonsense reasoning while reducing inference cost, with minimal reported impact on commonsense reasoning and zero-shot performance. The excerpt gives no specific model names, latency measurements, or detailed evaluation results.

### Source excerpt

arXiv:2607.22583v4 Announce Type: replace Abstract: Large Language Models (LLMs) have achieved widespread adoption because of their strong reasoning and query-response capabilities. However, deploying them in embedded and edge computing environments remains challenging because of strict latency, memory, and energy constraints. Their large parameter counts and computational demands hinder efficient execution on resource-constrained platforms. Although model pruning has emerged as a viable solution for reducing scale while preserving performance, jointly optimizing layers, attention heads, and Multi-Layer Perceptron (MLP) dimensions remains highly complex. Exhaustively exploring this combined design space is computationally expensive and often leads to local optima or unstable configurations. To address these limitations, we propose a hardware-aware, multi-objective structured pruning framework. The proposed two-stage method explicitly targets latency and model size for efficient deployment on edge devices. In the coarse-grained stage, multi-objective depth pruning removes entire attention and MLP blocks to reduce computational load and memory usage. In the subsequent fine-grained stage, Parallel Bayesian Optimization (PBO) searches for the optimal layer-wise pruning ratios for pruning under latency constraints, while importance-based strategies rank the specific components to be pruned within each layer's allocated budget. Experimental results show that our approach reduces model complexity with minimal impact on commonsense reasoning tasks and zero-shot performance. Our method achieves a favorable trade-off among accuracy, latency, and model size, making it suitable for edge deployment. Across multiple LLMs at 37.5% and 50% pruning ratios, the proposed approach achieves better performance on commonsense reasoning tasks than existing methods while significantly reducing inference cost.

## Reinforcing Few-Step Diffusion Models via Shadow Posterior Matching

DevFeed: [Reinforcing Few-Step Diffusion Models via Shadow Posterior Matching](<https://devfeed.tech/articles/reinforcing-few-step-diffusion-models-via-shadow-posterior-matching-145126.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25998>)

Author: Jingqiao Xiu, Yicong Li

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

Content type: article

Language: en

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

Topics: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>)

Tags: [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [diffusion-models](<https://devfeed.tech/tags/diffusion-models.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [generators](<https://devfeed.tech/tags/generators.md>), [hacking](<https://devfeed.tech/tags/hacking.md>), [matching](<https://devfeed.tech/tags/matching.md>), [regression](<https://devfeed.tech/tags/regression.md>), [reward](<https://devfeed.tech/tags/reward.md>), [rl](<https://devfeed.tech/tags/rl.md>)

### AI overview

Few-step diffusion generators are samplers rather than score models, and they share no transitions with their undistilled teacher, complicating reward-based reinforcement. Shadow Posterior Matching (SPM) renoises generated samples into "shadow inputs" and uses their shadow posterior to derive reward credit. Its regression loss turns reward values into a reward direction while anchoring outputs to the teacher and aligning shadow inputs with deployed inputs. On 4-step generators, the authors report higher rewards and less reward hacking than prior few-step RL methods.

### Source excerpt

arXiv:2611.25998v1 Announce Type: new Abstract: Few-step diffusion models are widely deployed for their efficiency. However, reinforcing them with black-box rewards, as is routine for multi-step models, faces two structural challenges. First, a few-step generator is a sampler rather than a score model, so its velocity is not the forward velocity of the generated distribution. Second, it must be anchored to the undistilled teacher, which shares no transitions with it. Existing few-step RL methods handle the latter with a marginal KL to the teacher, but each pays for the former in its reward credit. We propose Shadow Posterior Matching (SPM), which renoises generated samples into shadow inputs and obtains the credit from their shadow posterior. From the KL-regularized reward objective, we derive a regression loss that turns reward values into a reward direction while anchoring the generator's outputs to the teacher and aligning the shadow inputs with the deployed inputs. Applied to 4-step generators, SPM achieves higher rewards with less reward hacking than prior few-step RL methods.

## STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization

DevFeed: [STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization](<https://devfeed.tech/articles/stepquant-when-and-where-errors-matter-in-delta-rule-recurrent-state-quantization-144947.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.38169>)

Author: Bingchen Yao, Haobo Xu, Haokun Lin, Yichen Wu, Ziyu Guo, Renrui Zhang, Zhichao Lu, Zhenan Sun, Ying Wei

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: [quantization](<https://devfeed.tech/topics/quantization.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Error Propagation](<https://devfeed.tech/topics/error-propagation.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [bottleneck](<https://devfeed.tech/tags/bottleneck.md>), [budget](<https://devfeed.tech/tags/budget.md>), [compression](<https://devfeed.tech/tags/compression.md>), [concurrent](<https://devfeed.tech/tags/concurrent.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [delta](<https://devfeed.tech/tags/delta.md>), [dimensions](<https://devfeed.tech/tags/dimensions.md>), [distributions](<https://devfeed.tech/tags/distributions.md>), [error](<https://devfeed.tech/tags/error.md>), [errors](<https://devfeed.tech/tags/errors.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [framework](<https://devfeed.tech/tags/framework.md>), [generation](<https://devfeed.tech/tags/generation.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [quantization](<https://devfeed.tech/tags/quantization.md>), [recurrent](<https://devfeed.tech/tags/recurrent.md>), [state](<https://devfeed.tech/tags/state.md>)

### AI overview

STEPQuant addresses accuracy loss when low-precision quantization errors propagate through Delta-rule recurrent state updates. It accounts for error magnitude and memory lifetime, and jointly fits key-row and value-column scales using state distributions and key-row effects on output error. On Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct across short- and long-generation benchmarks, it reportedly stays close to FP32-state accuracy at a nominal 6-bit budget and outperforms uniform INT8 in its 4-bit configuration. With optimized GPU kernels in SGLang, its 6-bit version achieves over 5x recurrent-state compression and reduces total serving memory by up to 68.7%.

### Source excerpt

arXiv:2609.38169v2 Announce Type: replace-cross Abstract: Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often leads to severe accuracy degradation, as quantization errors propagate through successive state updates. We discover that the impact of these errors depends on two complementary dimensions: temporally, errors in long-lived memory can persist across many decoding steps; spatially, errors in different key rows affect model outputs differently, while state magnitudes vary substantially along both rows and columns. Motivated by these observations, we propose STEPQuant, a spatial-temporal post-training quantization framework for Delta-rule recurrent states. STEPQuant allocates precision according to error magnitude and memory lifetime, and jointly fits key-row and value-column scales based on state distributions and key-row impact on output error. Experiments on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct across both long- and short-generation benchmarks show that STEPQuant closely matches FP32-state accuracy under a nominal 6-bit budget and outperforms uniform INT8 in its 4-bit configuration. Integrated into SGLang with optimized GPU kernels, 6-bit STEPQuant achieves over 5x recurrent-state compression and reduces total serving memory by up to 68.7%. Our code is available at https://github.com/Dreamer-Toby/STEPQuant.

## Lightweight Low-Frequency Non-Intrusive Load Monitoring via Karhunen-Lo\\\`eve Feature Distillation and Temporal Convolutional Networks

DevFeed: [Lightweight Low-Frequency Non-Intrusive Load Monitoring via Karhunen-Lo\\\`eve Feature Distillation and Temporal Convolutional Networks](<https://devfeed.tech/articles/lightweight-low-frequency-non-intrusive-load-monitoring-via-karhunen-lo-eve-feature-distillation-and-temporal-convolutional-networks-145089.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25678>)

Author: Sanka Peeris, Pandula Thennakoon, Praveen Sumanasekara, Roshan Godaliyadda, Parakrama Ekanayake, Janaka Ekanayake

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

Content type: article

Language: en

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

Topics: [Image processing](<https://devfeed.tech/topics/image-processing.md>), [Feature Engineering](<https://devfeed.tech/topics/feature-engineering.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Time-sequences](<https://devfeed.tech/topics/time-sequences.md>)

Tags: [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-sy](<https://devfeed.tech/tags/cs-sy.md>), [data-augmentation](<https://devfeed.tech/tags/data-augmentation.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [disaggregation](<https://devfeed.tech/tags/disaggregation.md>), [edge](<https://devfeed.tech/tags/edge.md>), [eess-sp](<https://devfeed.tech/tags/eess-sp.md>), [eess-sy](<https://devfeed.tech/tags/eess-sy.md>), [energy](<https://devfeed.tech/tags/energy.md>), [features](<https://devfeed.tech/tags/features.md>), [filter](<https://devfeed.tech/tags/filter.md>), [framework](<https://devfeed.tech/tags/framework.md>), [grid](<https://devfeed.tech/tags/grid.md>), [lightweight](<https://devfeed.tech/tags/lightweight.md>), [load](<https://devfeed.tech/tags/load.md>), [models](<https://devfeed.tech/tags/models.md>), [monitoring](<https://devfeed.tech/tags/monitoring.md>), [network](<https://devfeed.tech/tags/network.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [smart](<https://devfeed.tech/tags/smart.md>), [smart-grid](<https://devfeed.tech/tags/smart-grid.md>), [solution](<https://devfeed.tech/tags/solution.md>), [temporal](<https://devfeed.tech/tags/temporal.md>)

### AI overview

The proposed edge-oriented NILM framework pairs activity-boundary-aware augmentation--which recombines complete or near-complete appliance activation segments without requiring new fully labeled household aggregates--with KLT-derived temporal features and a compact TCN. On UK-DALE and REDD at 10-second sampling with leave-one-house-out testing, it reports disaggregation accuracy comparable to or better than NILMFormer, BERT4NILM, BiLSTM, and BiGRU. It uses 31,009 parameters, the fewest among the five compared models, and takes 2.78 ms per 1,000 samples on one CPU thread, reported as 84-95% less time than the baselines.

### Source excerpt

arXiv:2611.25678v1 Announce Type: cross Abstract: Non-Intrusive Load Monitoring (NILM) is a critical enabler for demand-side management which allows for smart grid optimization. A deployment-ready NILM solution necessitates three key features: accurate disaggregation, low-frequency sampling, and low computational cost. This paper proposes a lightweight, edge-oriented NILM framework that combines activity-boundary-aware synthetic data augmentation with Karhunen-Lo\`eve Transform (KLT) feature extraction and a Temporal Convolutional Network (TCN). The augmentation stage extracts and recombines complete or near-complete appliance activation segments so that the model can be trained without new fully labeled household aggregates. The KLT filter bank extracts decorrelated, appliance-informed temporal features which are then fed into a compact TCN along with the original aggregate for energy disaggregation. Evaluations on two standard NILM datasets, UK-DALE and REDD, at 10-s sampling and with leave-one-house-out testing, demonstrate that the proposed architecture achieves comparable or superior disaggregation accuracy relative to NILMFormer, BERT4NILM, BiLSTM, and BiGRU. The computational footprint of the framework is 31,009 parameters, the fewest of the five models compared, and on a single CPU thread it needs 2.78 ms per 1000 samples, 84-95% less than the baselines. These results demonstrate a favorable accuracy-efficiency tradeoff for low-frequency smart-meter NILM.

## Efficient Latent Reasoning for Vision-Language Model via Dynamics-Aware Quantization and Distillation

DevFeed: [Efficient Latent Reasoning for Vision-Language Model via Dynamics-Aware Quantization and Distillation](<https://devfeed.tech/articles/efficient-latent-reasoning-for-vision-language-model-via-dynamics-aware-quantization-and-distillation-145567.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25670>)

Author: Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Yawei 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: [quantization](<https://devfeed.tech/topics/quantization.md>), [vlm](<https://devfeed.tech/topics/vlm.md>), [Chain-of-thought](<https://devfeed.tech/topics/chain-of-thought.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [quantization](<https://devfeed.tech/tags/quantization.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [vision](<https://devfeed.tech/tags/vision.md>), [vision-language-model](<https://devfeed.tech/tags/vision-language-model.md>)

### AI overview

QL-VLM aims to make latent chain-of-thought reasoning in vision-language models cheaper without losing predictive accuracy under 4-bit quantization. Its Dynamics-Guided Quantization uses nonuniform codebooks to preserve rotary-key structure for information retrieval; Dynamics-Preserving Distillation aligns continuous thoughts, latent-query readouts, and predictions with a full-precision teacher to support later low-precision reasoning. On ScienceQA-IMG with Qwen2-VL-2B-Instruct, the abstract reports 90.43% accuracy, a 40.33% improvement over baseline, 16.64x fewer bit-operations, and 6.40x lower measured latency than explicit CoT.

### Source excerpt

arXiv:2611.25670v1 Announce Type: new Abstract: Efficient chain-of-thought (CoT) reasoning in vision-language models (VLMs) requires reducing both intermediate reasoning steps and their execution cost. Latent CoT methods compress textual rationales into continuous thoughts, yet their full-precision (FP) models retain substantial storage and computation costs. Lowering numerical precision could reduce these remaining costs, yet direct quantization severely degrades reasoning accuracy, particularly with 4-bit weights and activations. Operator and perturbation analyses reveal key sensitivity and direction-dependent predictive effects of continuous-thought errors, motivating preservation of information retrieval and subsequent reasoning beyond reconstruction accuracy. We therefore introduce QL-VLM, a dynamics-aware Quantization framework for efficient Latent reasoning in Vision-Language Models (VLMs). First, Dynamics-Guided Quantization (DGQ) uses nonuniform codebooks that preserve rotary-key structure, mitigating quantization-induced distortion in vision-language information retrieval. Second, Dynamics-Preserving Distillation (DPD) aligns continuous thoughts, latent-query readouts, and predictions with a full-precision teacher, maintaining predictive consistency through subsequent low-precision reasoning. On ScienceQA-IMG with Qwen2-VL-2B-Instruct, 4-bit QL-VLM achieves 90.43% accuracy and improves baseline by 40.33%, with 16.64x fewer bit-operations and 6.40x lower measured latency than explicit CoT.

## Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost

DevFeed: [Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost](<https://devfeed.tech/articles/introducing-harness-adapters-adapt-your-favorite-model-to-your-harness-to-improve-performance-and-cost-146558.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/harness-adapters-at-scale>)

Author: Younes B; ben burtenshaw; Nathan Habib; Lysandre; Amine Dirhoussi; Adithya S K; shaun smith; Sergio Paniego; Joel Niklaus; Lewis Tunstall

Published: 2026-11-20T00:00:00Z

Content type: tutorial

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [ai evaluation](<https://devfeed.tech/topics/ai-evaluation.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>)

Tags: [adapters](<https://devfeed.tech/tags/adapters.md>), [agent-harness](<https://devfeed.tech/tags/agent-harness.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-coding](<https://devfeed.tech/tags/agentic-coding.md>), [agents](<https://devfeed.tech/tags/agents.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [cost](<https://devfeed.tech/tags/cost.md>), [harness](<https://devfeed.tech/tags/harness.md>), [lora](<https://devfeed.tech/tags/lora.md>), [model](<https://devfeed.tech/tags/model.md>), [open](<https://devfeed.tech/tags/open.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [openenv](<https://devfeed.tech/tags/openenv.md>), [performance](<https://devfeed.tech/tags/performance.md>), [rl](<https://devfeed.tech/tags/rl.md>), [source](<https://devfeed.tech/tags/source.md>), [trl](<https://devfeed.tech/tags/trl.md>)

### AI overview

The article reports that reinforcement learning inside each target agent harness improved Qwen3.8-27B's coding benchmark results and reduced tool use. Separate LoRA adapters trained for Pi, Hermes Agent, and OpenCode each beat the base model under their own harness on Terminal-Bench 2.1 and hard SWE-bench Pro tasks. On tasks both models solved, specialists used 53-72% fewer tool calls. A merged adapter also beat the base under all three harnesses; with thinking off it averaged 39.5% to 49.0% pass@1, while cutting tool calls by 53-73%. The recipe uses validated Harbor tasks, difficulty filtering so GRPO receives tasks the base model solves inconsistently, and TRL's AsyncGRPOTrainer with the harness running as shipped. Results are not directly comparable across harnesses because their defaults differ. On 300 AutomationBench tasks outside the training domain and harnesses, adapters stayed near the base model, so transfer to that setting was not clear. The authors also note possible reasoning leakage in non-thinking outputs and call for further study.

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## Model Compression Beyond Quantization

DevFeed: [Model Compression Beyond Quantization](<https://devfeed.tech/articles/model-compression-beyond-quantization-143867.md>)

Original publisher: [Read original article](<https://machinelearningmastery.com/model-compression-beyond-quantization/>)

Author: Adrian Tam

Published: 2026-11-19T14:06:50Z

Content type: article

Language: en

Sources: [MachineLearningMastery.com](<https://devfeed.tech/sources/machinelearningmastery-com.md>)

Topics: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [Training AI Models](<https://devfeed.tech/topics/training-ai-models.md>)

Tags: [compression](<https://devfeed.tech/tags/compression.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [inference-from-transformer-models](<https://devfeed.tech/tags/inference-from-transformer-models.md>), [knowledge-distillation](<https://devfeed.tech/tags/knowledge-distillation.md>), [model-compression](<https://devfeed.tech/tags/model-compression.md>), [pruning](<https://devfeed.tech/tags/pruning.md>), [quantization](<https://devfeed.tech/tags/quantization.md>), [sparsity](<https://devfeed.tech/tags/sparsity.md>), [structured](<https://devfeed.tech/tags/structured.md>)

### AI overview

Compression methods can reduce model size by removing weights or structures, or by training a smaller model to imitate a larger one. Unstructured magnitude pruning zeros low-magnitude weights, but speed or storage gains depend on sparse representation, matrix shape, sparsity, dtype, hardware, and kernel support. Structured pruning removes coordinated units such as MLP channels; matching dimensions must be edited together, and hardware-friendly sizes matter. Knowledge distillation trains a student from teacher signals such as output distributions or intermediate representations. It requires teacher targets, compatible logits and token positions, and representative training data; generating targets has a cost. The text stresses evaluating held-out quality, memory, latency, and throughput, since reduced parameter counts do not guarantee preserved behavior or faster serving.

### Source excerpt

Overview This post is divided into three parts; they are: - Pruning and Sparsity - Structured Pruning - Knowledge Distillation Pruning and Sparsity Pruning sets selected model parameters to zero or removes larger structures.

## Arbor: How Search Policy Shapes Agent Behavior in System Optimization

DevFeed: [Arbor: How Search Policy Shapes Agent Behavior in System Optimization](<https://devfeed.tech/articles/arbor-how-search-policy-shapes-agent-behavior-in-system-optimization-142530.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.12563>)

Author: Neha Prakriya, Chaojun Hou, Zheng Gong, Huasha Zhao, Yanyuan Qin, Chris Cai, Xi Zhao, Zhenyu Gu, Emad Barsoum

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

Content type: article

Language: en

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

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [policy](<https://devfeed.tech/tags/policy.md>), [search](<https://devfeed.tech/tags/search.md>), [system](<https://devfeed.tech/tags/system.md>)

### AI overview

Arbor organizes accelerator optimization as an explicit tree that agents can traverse, revisit, and deepen. Its shared harness provides containerized isolation, automatic rollback, and end-to-end validation; search options include scored bandit traversal, parallel swarms, and depth-first sweeps. The abstract reports gains of 40%-193% across six production inference models on two GPU generations, 42.8% for LLM training, and 41% for recommendation training. In controlled comparisons using the same agent, tree, and tools, bandit search gained 30.1% through targeted framework fixes, sweeps 26.4% through deep kernel rewrites, and swarms 33.4% by combining independent changes. A loop that chose its own targets gained 14% and wrote four times more code per percentage point of improvement than Arbor's structured policies. The paper also introduces an evaluation framework covering efficiency, grounding, robustness, and depth.

### Source excerpt

arXiv:2606.12563v2 Announce Type: replace Abstract: Achieving peak performance on modern accelerators requires coordinated optimization across kernels, operator fusion, KV-cache management, quantization, scheduling, and communication overlap, a process that demands weeks of expert effort per model and hardware platform. We present Arbor, a system that gives autonomous agents an explicit optimization tree to reason over, backtrack through, and deepen with experience, with pluggable search policies, scored bandit traversal, parallel agent swarms, and depth-first sweep, sharing a common harness of containerized isolation, automatic rollback, and end-to-end validation. Arbor generalizes across domains: +40% to +193% on six production inference models across two GPU generations, +42.8% on LLM training, and +41% on recommendation training, where agents autonomously wrote kernel-level and framework-level optimizations spanning thousands of lines of code. In controlled ablations where every search policy receives the same agent, tree, and tools, we find that search structure shapes both the magnitude and the nature of the optimizations. The bandit achieves +30.1% through targeted framework fixes, the sweep +26.4% through deep kernel rewrites, and the swarm +33.4% by compounding independent changes. An agent-driven loop that receives the same harness, tools, and information, but selects its own targets reaches +14%, writing 4x more code per percentage point of improvement than Arbor's structured search policies. We additionally introduce an evaluation framework for autonomous loop-based agents spanning efficiency, grounding, robustness, and depth, designed to characterize agent search behavior across long-running optimization campaigns.

## FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment

DevFeed: [FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment](<https://devfeed.tech/articles/flexmoe-learning-one-for-all-multi-budget-expert-pruning-for-flexible-moe-deployment-143284.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.27866>)

Author: Fan Mo, Yuxuan Han, Geng Zhang, Wangbo Zhao, Yang You

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

Content type: article

Language: en

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

Topics: [Mixture of Experts (MoE)](<https://devfeed.tech/topics/mixture-of-experts-moe.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>)

Tags: [acceleration](<https://devfeed.tech/tags/acceleration.md>), [actions](<https://devfeed.tech/tags/actions.md>), [activation](<https://devfeed.tech/tags/activation.md>), [budget](<https://devfeed.tech/tags/budget.md>), [co-design](<https://devfeed.tech/tags/co-design.md>), [compression](<https://devfeed.tech/tags/compression.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [deployment](<https://devfeed.tech/tags/deployment.md>), [experts](<https://devfeed.tech/tags/experts.md>), [ffn](<https://devfeed.tech/tags/ffn.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [flexible](<https://devfeed.tech/tags/flexible.md>), [inference](<https://devfeed.tech/tags/inference.md>), [kernel](<https://devfeed.tech/tags/kernel.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [learning](<https://devfeed.tech/tags/learning.md>), [llms](<https://devfeed.tech/tags/llms.md>), [masks](<https://devfeed.tech/tags/masks.md>), [memory](<https://devfeed.tech/tags/memory.md>), [moe](<https://devfeed.tech/tags/moe.md>), [pruning](<https://devfeed.tech/tags/pruning.md>)

### AI overview

FlexMoE aims to produce deployable MoE subnetworks at multiple parameter budgets from one pretrained model, avoiding separate compression runs for each budget. It ranks expert FFN channels to create a shared sliceable weight space, then learns expert-wise discrete retention actions under increasing cost pressure. A single fine-tune at a mid-range budget can recover performance and transfer to unseen budgets, reducing repeated recovery and separate checkpoint storage. The abstract reports gains over tested compression baselines; on Qwen2-57B-A14B, pruning 50% of routed-expert parameters retained about 99.8% of base performance without recovery fine-tuning. It also reports memory reduction and throughput gains, with kernel co-design enabling runtime budget switching without reloading checkpoints.

### Source excerpt

arXiv:2606.27866v2 Announce Type: replace Abstract: Mixture-of-Experts (MoE) language models scale model capacity through sparse expert activation, yet deployment still requires storing and serving all experts. Moreover, deployment budgets can vary substantially across devices, users, and workloads. Existing MoE compression methods typically optimize a separate compressed model for each target budget. We introduce FlexMoE, a post-training one-for-all framework that converts a large pretrained MoE model into a multi-budget family of smaller deployable subnetworks. FlexMoE first ranks expert FFN channels by importance to construct a sliceable weight space shared across budgets, and then learns expert-wise discrete retention actions under gradually increasing cost pressure, allowing a single action-training run to produce reliable pruning masks across budgets. Moreover, only a single mid-budget fine-tune can further recover degraded performance and yields a shared recovered model that transfers to other unseen budgets, avoiding repeated per-budget recovery and separated checkpoint storage. Across tested MoE LLMs with diverse architectures and scales, FlexMoE consistently outperforms strong MoE compression baselines. In particular, on Qwen2-57B-A14B, it retains ~99.8% of base performance while pruning 50% of routed-expert parameters even without recovery fine-tuning. For deployment, pruned subnetworks deliver real memory reduction and end-to-end throughput gains, while our kernel co-design enables runtime budget switching without checkpoint reloading and delivers substantial inference acceleration.

## Distill What You Trust: Reliability-Aware Multi-Teacher On-Policy Distillation

DevFeed: [Distill What You Trust: Reliability-Aware Multi-Teacher On-Policy Distillation](<https://devfeed.tech/articles/distill-what-you-trust-reliability-aware-multi-teacher-on-policy-distillation-143313.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.23697>)

Author: Jie Sun, Mao Zheng, Mingyang Song, Zeyuan Liu, Gengsheng Li, Houcheng Jiang, Yilin Cheng, Bichuan Feng, Yuchen Cai, Junfeng Fang, Xiang Wang

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

Content type: article

Language: en

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

Topics: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>)

Tags: [code](<https://devfeed.tech/tags/code.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [policy](<https://devfeed.tech/tags/policy.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [trust](<https://devfeed.tech/tags/trust.md>)

### AI overview

TrustMOPD allocates multi-teacher distillation supervision token by token, using each specialist's calibrated deviation from a shared reference at the student's current prefix as a proxy for local reliability. This label-free approach addresses the limits of choosing one teacher per example, which requires domain labels and cannot adapt as expertise needs shift within a response. In evaluations across mathematics, code, and instruction following, it closed 91.5% and 98.0% of the overall-score gap between the initial student and oracle-routed teachers on SingleCap and MultiCap, respectively; the strongest label-free baselines closed 54.4% and 54.5%. On SingleCap, it approached label-based MOPD without domain labels.

### Source excerpt

arXiv:2609.23697v4 Announce Type: replace-cross Abstract: Multi-teacher on-policy distillation allows a student to learn from complementary specialists on its own trajectories. Domain-routed approaches, however, select one teacher per example and keep it fixed throughout the response. This design both depends on domain labels that mixed training corpora often lack and cannot adapt teacher selection when the expertise required changes within a trajectory. We observe that each specialist deviates more from a shared reference on in-domain prompts than on out-of-domain prompts, on average. Based on this observation, we propose \textbf{TrustMOPD}, which replaces example-level teacher selection with label-free, token-level supervision allocation. At each student-generated prefix, TrustMOPD measures this displacement in next-token preferences, calibrates its magnitude across teachers, and uses the resulting scores as proxies for local reliability to weight teacher-specific distillation losses. Evaluated across mathematics, code, and instruction following, TrustMOPD closes 91.5\% and 98.0\% of the overall-score gap between the initial student and oracle-routed teachers when trained on \textsc{SingleCap} and \textsc{MultiCap}, respectively, compared with 54.4\% and 54.5\% for the strongest label-free baseline in each setting. On \textsc{SingleCap}, it approaches label-based MOPD without using domain labels. Our code is available at https://github.com/sunjie279/TrustMOPD-.

## FedS2R: One-Shot Federated Domain Generalization for Synthetic-to-Real Semantic Segmentation in Autonomous Driving

DevFeed: [FedS2R: One-Shot Federated Domain Generalization for Synthetic-to-Real Semantic Segmentation in Autonomous Driving](<https://devfeed.tech/articles/feds2r-one-shot-federated-domain-generalization-for-synthetic-to-real-semantic-segmentation-in-autonomous-driving-142496.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2507.19881>)

Author: Tao Lian, Jose L. G\\'omez, Antonio M. L\\'opez

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

Content type: article

Language: en

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

Topics: [Federated Learning](<https://devfeed.tech/topics/federated-learning.md>), [autonomous vehicles](<https://devfeed.tech/topics/autonomous-vehicles.md>), [data augmentation](<https://devfeed.tech/topics/data-augmentation.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>)

Tags: [autonomous-driving](<https://devfeed.tech/tags/autonomous-driving.md>), [classification](<https://devfeed.tech/tags/classification.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [data-augmentation](<https://devfeed.tech/tags/data-augmentation.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [federated](<https://devfeed.tech/tags/federated.md>), [federated-learning](<https://devfeed.tech/tags/federated-learning.md>), [fusion](<https://devfeed.tech/tags/fusion.md>), [knowledge-distillation](<https://devfeed.tech/tags/knowledge-distillation.md>)

### AI overview

FedS2R targets synthetic-to-real semantic segmentation for autonomous driving while keeping client data local. It combines inconsistency-driven augmentation to generate images for unstable classes with multi-client knowledge distillation and feature fusion to build a global model from client models in a one-shot federated setting. On Cityscapes, BDD100K, Mapillary, IDD, and ACDC, the global model reportedly outperforms individual client models and comes within 2 mIoU points of a model trained with simultaneous access to all client data.

### Source excerpt

arXiv:2507.19881v4 Announce Type: replace-cross Abstract: Federated domain generalization has shown promising progress in image classification by enabling collaborative training across multiple clients without sharing raw data. However, its potential in the semantic segmentation of autonomous driving remains underexplored. In this paper, we propose FedS2R, the first one-shot federated domain generalization framework for synthetic-to-real semantic segmentation in autonomous driving. FedS2R comprises two components: an inconsistency-driven data augmentation strategy that generates images for unstable classes, and a multi-client knowledge distillation scheme with feature fusion that distills a global model from multiple client models. Experiments on five real-world datasets, Cityscapes, BDD100K, Mapillary, IDD, and ACDC, show that the global model significantly outperforms individual client models and is only 2 mIoU points behind the model trained with simultaneous access to all client data. These results demonstrate the effectiveness of FedS2R in synthetic-to-real semantic segmentation for autonomous driving under federated learning

## CoreQ: Mismatch-Corrected Successive Rounding for LLM Quantization

DevFeed: [CoreQ: Mismatch-Corrected Successive Rounding for LLM Quantization](<https://devfeed.tech/articles/coreq-mismatch-corrected-successive-rounding-for-llm-quantization-142504.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2602.05902>)

Author: Seohyeon Cha, Huancheng Chen, Dongjun Kim, Haoran Zhang, Kevin Chan, Gustavo de Veciana, Haris Vikalo

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: [quantization](<https://devfeed.tech/topics/quantization.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [beam](<https://devfeed.tech/tags/beam.md>), [beam-search](<https://devfeed.tech/tags/beam-search.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [framework](<https://devfeed.tech/tags/framework.md>), [github](<https://devfeed.tech/tags/github.md>), [greedy](<https://devfeed.tech/tags/greedy.md>), [layer](<https://devfeed.tech/tags/layer.md>), [layers](<https://devfeed.tech/tags/layers.md>), [learning](<https://devfeed.tech/tags/learning.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-quantization](<https://devfeed.tech/tags/llm-quantization.md>), [overfitting](<https://devfeed.tech/tags/overfitting.md>), [performance](<https://devfeed.tech/tags/performance.md>), [perplexity](<https://devfeed.tech/tags/perplexity.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [precision](<https://devfeed.tech/tags/precision.md>), [quantization](<https://devfeed.tech/tags/quantization.md>), [search](<https://devfeed.tech/tags/search.md>), [target](<https://devfeed.tech/tags/target.md>)

### AI overview

CoreQ addresses activation mismatch in sequential post-training quantization, where earlier layers' quantization errors change the inputs to later layers. It uses a geometric decomposition to derive a closed-form, layer-adaptive coefficient that limits how much empirical mismatch shifts each calibration target, aiming to reduce overfitting on small calibration sets without coefficient tuning. It then applies greedy successive rounding to a triangular least-squares objective. K-CoreQ adds bounded beam search, trading modest extra compute for reported performance gains. Across multiple LLM families, scales, bit-widths and quantization settings, the abstract reports lower perplexity and higher downstream accuracy than strong PTQ baselines.

### Source excerpt

arXiv:2602.05902v3 Announce Type: replace-cross Abstract: Post-training quantization (PTQ) enables efficient deployment of large language models by mapping pretrained weights to low-bit formats without retraining, typically using a small calibration set to minimize a layer-wise calibration objective. However, this sequential procedure induces a mismatch: errors from earlier quantized layers alter the inputs received by later layers, causing the activations to deviate from those of the full-precision model. Recent approaches introduce mismatch-aware calibration objectives to compensate for this effect, but leave open how much of the observed mismatch should shift each layer's calibration target. We show that fully applying the empirical mismatch correction can overfit finite calibration data, while using a fixed correction coefficient ignores variation in mismatch estimates across layers. To address these limitations, we propose CoreQ, a learning-free PTQ framework that uses a geometric decomposition of the mismatch to compute a closed-form, layer-adaptive correction coefficient. This coefficient controls how much of the observed mismatch enters each layer's calibration target, reducing overfitting to limited calibration data without coefficient tuning. CoreQ then applies efficient greedy successive rounding to the resulting triangular least-squares objective. Its bounded beam-search extension, K-CoreQ, trades modest additional compute for improved performance. Across multiple LLM families, scales, bit-widths, and quantization settings, CoreQ and K-CoreQ achieve lower perplexity and higher downstream accuracy than strong PTQ baselines. Code is available at GitHub.

## Think Before You Distill: Generalization Recovery in Quantized VLAs via Selective Policy Distillation

DevFeed: [Think Before You Distill: Generalization Recovery in Quantized VLAs via Selective Policy Distillation](<https://devfeed.tech/articles/think-before-you-distill-generalization-recovery-in-quantized-vlas-via-selective-policy-distillation-143325.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.05745>)

Author: Shen Ruan, Wenchang Gao, Jin Wang, Siao Liu, Zhaxizhuoma, Dongchun Ren, Xin Zheng

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

Content type: article

Language: en

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

Topics: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [quantization](<https://devfeed.tech/topics/quantization.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [anchor](<https://devfeed.tech/tags/anchor.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [budget](<https://devfeed.tech/tags/budget.md>), [but](<https://devfeed.tech/tags/but.md>), [can](<https://devfeed.tech/tags/can.md>), [changes](<https://devfeed.tech/tags/changes.md>), [code](<https://devfeed.tech/tags/code.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [distilling](<https://devfeed.tech/tags/distilling.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [effective](<https://devfeed.tech/tags/effective.md>), [effects](<https://devfeed.tech/tags/effects.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [framework](<https://devfeed.tech/tags/framework.md>), [full](<https://devfeed.tech/tags/full.md>), [guidance](<https://devfeed.tech/tags/guidance.md>), [policy](<https://devfeed.tech/tags/policy.md>), [recovery](<https://devfeed.tech/tags/recovery.md>)

### AI overview

Post-training quantization can preserve in-distribution VLA performance while making policies fragile to subtle environmental changes. The abstract reports that action discrepancies cluster in a small share of rollout states, and teacher guidance helps at high-discrepancy states but can hurt at low-discrepancy states. PIVOT-Q addresses this with on-policy distillation from a frozen full-precision teacher: it detects vulnerable states through discounted discrepancies over a short horizon, applies phase-balanced sparse supervision, and uses a Behavioral Anchor to limit unintended changes. In experiments across seven LIBERO-Plus variations, multiple VLA backbones, and quantization methods, it reportedly recovered performance with lower variance than full-state distillation using 20% of its state-level distillation budget.

### Source excerpt

arXiv:2610.05745v3 Announce Type: replace Abstract: Post-training quantization has been shown to preserve VLA performance under standard evaluation conditions, but whether it preserves the full-precision model's robustness and generalization remains underexplored. In this study, we systematically study the robustness and generalization of post-quantized VLA policies under environmental disturbances. Empirical results show that quantized policies can become fragile to subtle environmental variations despite retaining comparable in-distribution performance. We further observe that action discrepancies are concentrated in a small subset of the rollout states, while teacher guidance has opposite effects depending on discrepancy: it improves performance at high-discrepancy states but can degrade it at low-discrepancy states. These findings reveal that effective post-quantization recovery requires selective intervention on vulnerable states rather than globally distilling the student. We therefore propose Policy-Induced Vulnerability-Oriented Tuning (PIVOT-Q), a vulnerability-aware On-Policy Distillation (OPD) framework that selectively corrects vulnerable states encountered during quantized-student rollouts using the frozen full-precision policy as a teacher. PIVOT-Q identifies vulnerable states using discounted accumulated discrepancies over a short horizon, applies phase-balanced sparse supervision, and uses a Behavioral Anchor to prevent unintended changes. Experiments under seven LIBERO-Plus environmental variations demonstrate consistent recovery across multiple VLA backbones and quantization methods. Notably, PIVOT-Q outperforms full-state distillation with smaller variance across evaluated settings while using only 20\% of its state-level distillation budget. Our code is available at https://github.com/ruanruan-andy/PIVOT-Q.

## FedE2E: One-Shot Federated Learning for End-to-End Autonomous Driving with Closed-Loop Evaluation

DevFeed: [FedE2E: One-Shot Federated Learning for End-to-End Autonomous Driving with Closed-Loop Evaluation](<https://devfeed.tech/articles/fede2e-one-shot-federated-learning-for-end-to-end-autonomous-driving-with-closed-loop-evaluation-142772.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24254>)

Author: Tao Lian, Antonio M. L\\'opez

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

Content type: article

Language: en

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

Topics: [Federated Learning](<https://devfeed.tech/topics/federated-learning.md>), [autonomous vehicles](<https://devfeed.tech/topics/autonomous-vehicles.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [autonomous-driving](<https://devfeed.tech/tags/autonomous-driving.md>), [consensus](<https://devfeed.tech/tags/consensus.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [end-to-end-autonomous-driving](<https://devfeed.tech/tags/end-to-end-autonomous-driving.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [federated-learning](<https://devfeed.tech/tags/federated-learning.md>), [stat-ml](<https://devfeed.tech/tags/stat-ml.md>)

### AI overview

FedE2E targets end-to-end autonomous driving with one-shot federated learning, where clients collaborate without sharing datasets. It combines multi-client calibration, anchor-based feature alignment, and token-wise consensus distillation to transfer knowledge across heterogeneous models. In closed-loop Bench2Drive evaluation, the method reports a one-shot Driving Score of 55.85, 4.07 points above prior work. With five communication rounds, it reports 60.35, the best score among compared methods at the same round count.

### Source excerpt

arXiv:2611.24254v1 Announce Type: new Abstract: Federated learning enables collaborative model training without sharing datasets. However, existing federated learning methods for autonomous driving typically require multi-round communication and lack closed-loop evaluation, while one-shot methods mainly focus on perception tasks. To address these limitations, we propose \method, a one-shot federated learning method for end-to-end autonomous driving. It comprises multi-client calibration, anchor-based feature alignment, and token-wise consensus distillation to distill knowledge from heterogeneous models. Closed-loop evaluation on Bench2Drive demonstrates the effectiveness of \method in both one-shot and multi-round federated learning. In the one-shot setting, it achieves a Driving Score of 55.85, which is 4.07 points higher than prior work. In the multi-round setting, it reaches 60.35 after five communication rounds, the best performance among all compared methods under the same number of communication rounds.

## Which Tokens, in What Order, and How Fast? A Theoretical Analysis of On-Policy Distillation Objectives

DevFeed: [Which Tokens, in What Order, and How Fast? A Theoretical Analysis of On-Policy Distillation Objectives](<https://devfeed.tech/articles/which-tokens-in-what-order-and-how-fast-a-theoretical-analysis-of-on-policy-distillation-objectives-142651.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23397>)

Author: Hubert Strauss, Liam Fowl, Noam Razin

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: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [Chinchilla scaling law](<https://devfeed.tech/topics/chinchilla-scaling-law.md>)

Tags: [bias](<https://devfeed.tech/tags/bias.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [future](<https://devfeed.tech/tags/future.md>), [gradient](<https://devfeed.tech/tags/gradient.md>), [kl-divergence](<https://devfeed.tech/tags/kl-divergence.md>), [language](<https://devfeed.tech/tags/language.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [learning](<https://devfeed.tech/tags/learning.md>), [objectives](<https://devfeed.tech/tags/objectives.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [stat-ml](<https://devfeed.tech/tags/stat-ml.md>), [theoretical](<https://devfeed.tech/tags/theoretical.md>), [tokens](<https://devfeed.tech/tags/tokens.md>)

### AI overview

The analysis finds that sequence-level on-policy distillation can have a future agreement bias: it matches the teacher at later token positions first and may favor teacher-improbable tokens earlier. Myopic token-level objectives avoid this bias. The choice of divergence also affects optimization: forward KL has a well-behaved objective landscape, while other f-divergences such as reverse KL can have vanishing gradients even when student and teacher differ substantially. The abstract says the paper examines implications across language-model post-training settings and emphasizes choosing objectives in light of the student, teacher, and task.

### Source excerpt

arXiv:2611.23397v1 Announce Type: cross Abstract: On-policy distillation (OPD) trains a student language model to match a teacher model based on its own generations. When using OPD, one must choose whether to apply it at the sequence or token level and which divergence to minimize. However, despite the wide adoption of OPD in practice, the formal understanding of which OPD objective is appropriate for a given setting is limited. Toward addressing this gap, we theoretically analyze how the choice of OPD objective governs which tokens are prioritized, in what order, and how efficiently. Our analysis reveals that sequence-level OPD exhibits a future agreement bias, which myopic token-level objectives avoid. This bias induces a reverse learning order, in which sequence-level OPD first matches the teacher at later token positions, while at earlier positions it can favor tokens that are improbable under the teacher, and thus often undesirable. As for the choice of divergence, we prove that it strongly influences the amenability of OPD to gradient-based optimization. In particular, while the forward KL divergence enjoys a well-behaved objective landscape, other $f$-divergences, such as reverse KL, can suffer from vanishing gradients even when the student and teacher are far apart. Across language model post-training settings, we demonstrate the implications of these theoretical findings for which OPD objective is preferable, highlighting the importance of accounting for the interplay between the student, teacher, and task at hand.

## QDA: Closed-Loop LLM Quantization Design with Agents

DevFeed: [QDA: Closed-Loop LLM Quantization Design with Agents](<https://devfeed.tech/articles/qda-closed-loop-llm-quantization-design-with-agents-142650.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23388>)

Author: Xingyang Li, Tianlang Zhao, Song Han, Ligeng Zhu

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: [quantization](<https://devfeed.tech/topics/quantization.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [NVFP4](<https://devfeed.tech/topics/nvfp4.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [kernels](<https://devfeed.tech/tags/kernels.md>), [kl-divergence](<https://devfeed.tech/tags/kl-divergence.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-quantization](<https://devfeed.tech/tags/llm-quantization.md>), [moe](<https://devfeed.tech/tags/moe.md>), [nvfp4](<https://devfeed.tech/tags/nvfp4.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [precision](<https://devfeed.tech/tags/precision.md>), [quantization](<https://devfeed.tech/tags/quantization.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reasoning-model](<https://devfeed.tech/tags/reasoning-model.md>), [reasoning-models](<https://devfeed.tech/tags/reasoning-models.md>), [recipe](<https://devfeed.tech/tags/recipe.md>), [search](<https://devfeed.tech/tags/search.md>), [serving](<https://devfeed.tech/tags/serving.md>)

### AI overview

QDA uses an LLM agent to design model- and precision-specific post-training quantization recipes through iterative build-and-evaluate rounds. It scores candidates in minutes using output KL divergence on reasoning traces hidden from the agent, then checks promising recipes with long generations to catch failures such as repetition loops. A compatible library of public PTQ methods reduces manual porting across architectures. Across four reasoning models, QDA reports average accuracy gains over the strongest public recipe of 0.9, 1.4, and 2.2 percentage points at W4A16, W4A8, and W3A16, respectively, using 64 GPU-hours per search. For GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of FP8 accuracy, versus 96.9% for the best vendor checkpoint. The authors say QDA checkpoints use existing serving kernels with no inference overhead.

### Source excerpt

arXiv:2611.23388v1 Announce Type: cross Abstract: Post-training quantization is the standard way to serve large language models at low cost. However, the best quantization recipe differs across models and precisions. As a result, experts must redesign it for every new model, which is time-consuming and often sub-optimal. Fortunately, an LLM agent can take over this design by refining the recipe over rounds of building and evaluating candidates, so that more rounds lead to better recipes. However, evaluating one quantized reasoning model takes tens of GPU-hours, and combining methods from different codebases means porting each to the new architecture by hand. We present QDA, a system in which an agent designs the recipe for each model and precision. To make evaluation fast, QDA scores each candidate in minutes by its output KL divergence from the original model on reasoning traces that the agent never sees, and checks promising ones with long generations to catch failures such as repetition loops. To make building fast, QDA provides a library of public PTQ methods that are compatible with one another and with new model architectures. Across four reasoning models, QDA improves accuracy over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search. On the 320B-parameter MoE model GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of the FP8 model's accuracy, against 96.9% for the best vendor checkpoint. Every QDA checkpoint also runs on existing serving kernels with no inference overhead. We will release the library, the evaluator and the harness.

## Lumi-4B AI Model: Everything You Need to Know

DevFeed: [Lumi-4B AI Model: Everything You Need to Know](<https://devfeed.tech/articles/lumi-4b-ai-model-everything-you-need-to-know-140476.md>)

Original publisher: [Read original article](<https://hackernoon.com/lumi-4b-ai-model-everything-you-need-to-know?source=rss>)

Author: aimodels44

Published: 2026-11-18T06:07:46Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [context window](<https://devfeed.tech/topics/context-window.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [base](<https://devfeed.tech/tags/base.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [chat](<https://devfeed.tech/tags/chat.md>), [coding](<https://devfeed.tech/tags/coding.md>), [content-creation](<https://devfeed.tech/tags/content-creation.md>), [context](<https://devfeed.tech/tags/context.md>), [customer-success](<https://devfeed.tech/tags/customer-success.md>), [data-science](<https://devfeed.tech/tags/data-science.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [english](<https://devfeed.tech/tags/english.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [knowledge-distillation](<https://devfeed.tech/tags/knowledge-distillation.md>), [language](<https://devfeed.tech/tags/language.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [languages](<https://devfeed.tech/tags/languages.md>), [lumi-4b-explained](<https://devfeed.tech/tags/lumi-4b-explained.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [parameter](<https://devfeed.tech/tags/parameter.md>), [programming](<https://devfeed.tech/tags/programming.md>), [specifications](<https://devfeed.tech/tags/specifications.md>), [supervised-fine-tuning](<https://devfeed.tech/tags/supervised-fine-tuning.md>), [training](<https://devfeed.tech/tags/training.md>), [use-cases](<https://devfeed.tech/tags/use-cases.md>)

### AI overview

Lumi-4B is a 4B-parameter instruct model focused on Brazilian Portuguese, with English also intended. CEIA/UFG's NLP team started from Nemotron 3 Nano 4B and used continual pre-training with distillation (50 billion tokens, 8,192-token training context), followed by supervised fine-tuning. It targets conversational and instruction-following text; mathematics and coding capabilities were intended to be retained but should be evaluated for each use. Keep thinking disabled. This release has no additional RL/RLVR or dedicated identity training, and tool use is undocumented and unvalidated. Accuracy, bias, unsupported content, specialized-domain performance, and use beyond Brazilian Portuguese require evaluation. The text gives no numerical benchmark results or validated inference context length. Its vLLM template uses a placeholder repository ID and requires checking against the released checkpoint. Related-model descriptions identify GAIA Gemma-3 Gaia PT-BR 4B as based on Gemma 3 and Manacá-1B as a base model, but provide no comparative quality evidence. Claims about v0.2 should not be applied to this release.

### Source excerpt

Explore Lumi-4B, a 4B-parameter AI model focused on Brazilian Portuguese. Learn about its training, use cases, specifications, and limitations.

[Next page](<https://devfeed.tech/topics/post-training-optimization.md?cursor=WyIyMDI2LTExLTE4VDA2OjA3OjQ2KzAwOjAwIiwgImMxNjQyNjBlLTVmYzMtNGNlNC1hYjBmLTE2MTFlOWExZDdjMSJd>)