# QA Published articles for QA. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## From Raw Sensor Signals to Regulatory-Grade Data: How Air Quality Monitoring Networks Are Engineered DevFeed: [From Raw Sensor Signals to Regulatory-Grade Data: How Air Quality Monitoring Networks Are Engineered]() Original publisher: [Read original article]() Author: Kevin John Published: 2026-11-09T09:52:38Z Content type: article Language: en Sources: [HackerNoon]() Topics: [Data Quality](), [Data pipelines]() Tags: [air-quality](), [air-quality-data-apis](), [air-quality-monitoring](), [air-sensor-co-location](), [alerts](), [api](), [cents-17660](), [cities](), [data-pipelines](), [kunak-air-cloud](), [kunak-air-pro](), [networks](), [pm2-5-monitoring](), [qa](), [qc](), [regulatory](), [sensor](), [sensor-drift-correction](), [standards]() ### AI overview Raw air-sensor readings are affected by drift, cross-sensitivity, temperature and humidity, so reliable monitoring depends on managing the whole system. The article describes factory calibration, co-location with certified reference analysers to validate performance in real conditions, ongoing baseline and span adjustments, and automated QA/QC that flags or corrects suspect data. Reference stations provide an accuracy anchor, while denser, lower-cost networks add local coverage and faster alerts. A typical pipeline buffers readings during outages, centrally compensates and validates them, supports alerts, and integrates data through APIs or protocols such as Modbus. Performance should be assessed against the intended use and relevant evidence, including CEN/TS 17660, MCERTS, EPA targets, and independent evaluations. Deployment planning should specify pollutants and averaging periods, calibration upkeep, traceability, data integration, siting, and operational ownership. The excerpt gives Kunak AIR Pro and AIR Cloud as product examples; it presents sensor networks as complementary to reference stations. ### Source excerpt Learn how air quality networks use sensor calibration, co-location, QA/QC, and data pipelines to deliver reliable measurements under recognised standards. ## Scaling an autonomous agent from one team to fifty: A serverless-to-container decision framework DevFeed: [Scaling an autonomous agent from one team to fifty: A serverless-to-container decision framework]() Original publisher: [Read original article]() Author: Rashmi Erande Published: 2026-11-08T16:30:00Z Content type: comparison Language: en Sources: [The New Stack]() Topics: [Cloud Architecture](), [AWS Lambda](), [container native](), [Serverless](), [agentic workflows](), [Agentic AI](), [Artificial Intelligence](), [Multi Agent Systems](), [observability]() Tags: [agent](), [agent-delegation](), [agentic](), [agentic-workloads](), [agents](), [aggregation](), [ai-agents](), [autonomous](), [aws-marketplace](), [bug](), [complex](), [container](), [containers](), [failure](), [framework](), [post-contributed](), [qa](), [scaling](), [serverless](), [sponsor-aws-marketplace](), [sponsored-post-contributed](), [stack](), [team](), [teams](), [test](), [ticket]() ### AI overview A Lambda based QA agent polls test results every 15 minutes, checks for duplicates and regressions, and files JIRA bugs; the author says it costs under $5 monthly for one team and saves 8-10 hours a week. The proposed decision rule is workload characteristics, not a team-count threshold: Lambda suits isolated, short event-driven jobs, while Durable Functions can handle longer workflows. For persistent shared services, cross-team analysis, centralized API rate limiting, GPU inference, and namespace isolation needed together, the author favors EKS Auto Mode. The proposed platform separates shared services, tenant namespaces, and shared ML inference, using quotas, network policies, and scoped IAM roles. Lambda remains the suggested starting point; the article estimates 50 teams on EKS Auto Mode at $200-250 monthly, but presents this as an architecture and cost projection. ### Source excerpt QA teams shipping complex products generate test failures around the clock. Every failure needs a bug ticket with steps, screenshots, The post Scaling an autonomous agent from one team to fifty: A serverless-to-container decision framework appeared first on The New Stack. ## The Manager Who Doesn't Use AI Has Already Lost: Hard Skills DevFeed: [The Manager Who Doesn't Use AI Has Already Lost: Hard Skills]() Original publisher: [Read original article]() Author: Maxi Contieri Published: 2026-11-07T03:28:32Z Content type: opinion Language: en Sources: [Maximiliano Contieri - Software Design]() Topics: [AI-assisted coding](), [Test automation](), [CI/CD](), [ai-governance](), [Self-organizing Team](), [issue tracking](), [Documentation](), [workflow automation](), [Machine learning]() Tags: [2026](), [3](), [about](), [about-you](), [acceptance](), [acceptance-tests](), [agile](), [ai](), [ai-assistant](), [code](), [designers](), [developers](), [doesn-t](), [limit](), [manager](), [mock](), [qa](), [skills](), [up](), [with]() ### AI overview The article argues that AI can expand managers' capacity to plan, monitor, and respond, while teams and people retain key decisions. It contrasts long-range estimation, manual reporting, reactive quality checks, and siloed knowledge with short-cycle replanning, small verifiable task slices, evidence-cited status reports, a source-citing company knowledge base, automated testing and incident support, and release monitoring. It names specific assistant skills for tasks such as bottleneck analysis, prioritization, task slicing, defect triage, and anomaly detection. Human review remains necessary for plans, prototypes, fixes, risk choices, incident escalation, and knowledge redistribution. The author also cautions that employee monitoring requires notice and consent, and that the company knowledge base needs organizational agreement and source validation. The excerpt says detailed functional requirements should be captured in automated and acceptance tests; it leaves several technical topics for future coverage. ### Source excerpt Developers code with AI. Designers mock up with AI. QA automates with AI. How about you? TL;DR: AI removes the manager's cognitive limit. You should use it... ## StreamGaze: Gaze-Guided Temporal Reasoning and Proactive Understanding in Streaming Videos DevFeed: [StreamGaze: Gaze-Guided Temporal Reasoning and Proactive Understanding in Streaming Videos]() Original publisher: [Read original article]() Author: Daeun Lee, Subhojyoti Mukherjee, Branislav Kveton, Ryan A. Rossi, Viet Dac Lai, Seunghyun Yoon, Trung Bui, Franck Dernoncourt, Mohit Bansal Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [llm-reasoning](), [Artificial Intelligence](), [Real-time AI interactions]() Tags: [ar](), [ar-glasses](), [augmented-reality](), [benchmark](), [benchmarks](), [cs-ai](), [cs-cl](), [cs-cv](), [develop](), [language-models](), [multimodal](), [pairs](), [pipeline](), [prediction](), [proactive](), [process](), [prompting](), [qa](), [reasoning](), [streaming](), [temporal](), [temporal-reasoning](), [videos]() ### AI overview StreamGaze benchmarks whether multimodal large language models can use human gaze to reason about streaming video and anticipate user intent. It separates past and present tasks, answered from observations available at query time, from proactive tasks that require continuous frame processing and a response at the appropriate future moment. Its QA pipeline aligns egocentric video with raw gaze trajectories using fixation extraction, region-specific visual prompting, and scanpath construction to create spatiotemporally grounded questions. The authors report substantial performance gaps between state-of-the-art models and humans across tasks, pointing to limitations in gaze-based temporal reasoning, intent modeling, and proactive prediction. The abstract says the work also analyzes gaze prompting strategies, model reasoning, and task-specific failure modes; it provides no detailed comparative results. ### Source excerpt arXiv:2512.01707v4 Announce Type: replace-cross Abstract: Streaming video understanding requires models not only to process temporally incoming frames, but also to anticipate user intent for realistic applications such as Augmented Reality (AR) glasses. While prior streaming benchmarks evaluate temporal reasoning, none measure whether Multimodal Large Language Models (MLLMs) can interpret or leverage human gaze signals within a streaming setting. To fill this gap, we introduce StreamGaze, the first benchmark designed to evaluate how effectively MLLMs utilize gaze for temporal and proactive reasoning in streaming videos. StreamGaze introduces gaze-guided past, present, and proactive tasks that comprehensively assess streaming video understanding. Past and present tasks require immediate responses based on observations available at query time, whereas proactive tasks require models to continuously process incoming frames and respond at the appropriate future moment. To build StreamGaze, we develop a gaze-video Question Answering (QA) generation pipeline that aligns egocentric videos with raw gaze trajectories through fixation extraction, region-specific visual prompting, and scanpath construction. This pipeline produces spatio-temporally grounded QA pairs that reflect human perceptual dynamics. Across all StreamGaze tasks, we observe substantial performance gaps between state-of-the-art MLLMs and human performance, highlighting key limitations in gaze-based temporal reasoning, intent modeling, and proactive prediction. We further provide detailed analyses of gaze prompting strategies, reasoning behaviors, and task-specific failure modes, offering insights into current limitations and directions for future research. All data and code are publicly available to support research in gaze-guided streaming video understanding. ## CacheGraft: Enhancing Retrieval-Augmented Generation via Latent-Space Document Enrichment DevFeed: [CacheGraft: Enhancing Retrieval-Augmented Generation via Latent-Space Document Enrichment]() Original publisher: [Read original article]() Author: Vihan Lakshman, Peter Baile Chen, Zhuohan Gu, Michael Cafarella Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Retrieval Augmented Generation (RAG)](), [Enhancing AI model performance](), [Large Language Model](), [search-engines]() Tags: [agent-communication](), [agents](), [benchmarks](), [cs-ir](), [cs-lg](), [enrichment](), [framework](), [fusion](), [generation](), [inference](), [kv-cache](), [language-model](), [machine-learning](), [model](), [offline](), [performance](), [qa](), [rag](), [rag-systems](), [receiver](), [retrieval](), [retrieval-augmented-generation]() ### AI overview CacheGraft enriches retrieval-augmented generation by encoding corpus passages offline as fused key-value cache representations between a 4B-parameter sharer model and a 0.6B-parameter receiver. At query time, the receiver can use the cache without deploying the larger model. Across five QA benchmarks and two locally run judges, it improved answer quality over a text-only receiver in 29 of 30 comparisons; 17 improvements were statistically significant. The reported gains were zero-shot for the fuser, which was not trained on retrieval data. These initial results suggest potential for settings where inference must use a small model. ### Source excerpt arXiv:2611.07393v1 Announce Type: cross Abstract: Recent work in machine learning has demonstrated that transformer-based language model agents can communicate state not only through text, but also directly in latent space via key-value (KV) caches. In this paper, we argue that such cache-to-cache translation can provide utility beyond the original motivating application of real-time agent communication and can instead be utilized for retrieval-augmented generation (RAG). Building on the Cache-to-Cache (C2C) fusion framework, our proposed system CacheGraft encodes each corpus passage offline into a fused KV cache representation between a larger sharer model (4B parameters in our experiments) and a smaller receiver model (0.6B parameters) such that this cache can be injected into the receiver at query time with no need for the larger model. Across five QA benchmarks and two independent locally-run judges, CacheGraft improves answer quality over the text-only receiver baseline in 29 out of 30 comparisons, of which 17 of the 30 are statistically significant. Notably, these gains are zero-shot with respect to the fuser, which was never trained on retrieval data. These initial results suggest that latent-space document enrichment may be a promising avenue for enhancing the performance of RAG systems, particularly in compute-constrained settings where it is only feasible to deploy a small model at inference time. ## FinMultiHopQA: A KG-Grounded Multi-Hop Financial QA Benchmark for Evidence-Driven Reasoning DevFeed: [FinMultiHopQA: A KG-Grounded Multi-Hop Financial QA Benchmark for Evidence-Driven Reasoning]() Original publisher: [Read original article]() Author: Abhinav Arun, Reetu Raj Harsh, Bhaskarjit Sarmah, Stefano Pasquali Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [llm-reasoning](), [Finance](), [база знаний]() Tags: [benchmark](), [cs-ai](), [financial](), [knowledge-graph](), [q-fin-cp](), [qa](), [reasoning](), [reproducibility]() ### AI overview FinMultiHopQA provides 15,600 qualitative and quantitative financial QA questions grounded in FinReflectKG, a temporal knowledge graph built from SEC 10-K filings with more than 17.5 million triples. Its questions instantiate reproducible two- and three-hop motifs across filing sections, fiscal years, and companies, with linked provenance. In a controlled comparison, the benchmark keeps evidence constant while varying its presentation across six formats. The reported results show lower judged correctness with less-structured presentations, especially for quantitative cross-company questions and three-hop branch motifs; KG-linked presentation also uses fewer input tokens by design. The release includes provenance annotations, model outputs, evaluation records, and technical documentation. ### Source excerpt arXiv:2510.02906v2 Announce Type: replace-cross Abstract: Multi-hop financial question answering often requires composing evidence distributed across filing sections, fiscal years, and peer companies. Prior financial QA benchmarks address parts of this problem individually, but none jointly support controlled multi-hop taxonomies, document topologies, motif-level reproducibility, and provenance-linked auditability at scale. We introduce FinMultiHopQA, a benchmark of 15,600 qualitative and quantitative questions, grounded in FinReflectKG, a temporally indexed financial knowledge graph derived from SEC 10-K filings with 17.5M+ triples. The pipeline instantiates statistically salient 2-3 hop typed KG motifs across intra-document, inter-year, and cross-company regimes. Beyond closing this gap, the same construction lets us hold each question's underlying evidence constant while varying only its presentation - embedding it in noisy surrounding context or changing its order - giving a controlled instrument to study a secondary question at scale: using six such presentations, we test whether reasoning LLMs' performance depends on evidence structure. Results confirm this: despite their reasoning capacity, judged correctness remains consistently lower under less-structured presentations of the same evidence - most sharply for quantitative cross-company reasoning and 3-hop branch motifs - indicating that reasoning alone does not fully offset how evidence is structured; KG-linked presentation also uses fewer input tokens by design. We release the benchmark, provenance annotations, model outputs, evaluation records, and supporting technical documentation. ## Where Search Agents Go Off the Rails: Graph-Grounded Step-Level Verification DevFeed: [Where Search Agents Go Off the Rails: Graph-Grounded Step-Level Verification]() Original publisher: [Read original article]() Author: Hyeseo Jeon, Seonho Yoo, Jay-Yoon Lee Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [llm-reasoning](), [Chain-of-thought](), [Multi-step AI processes]() Tags: [agents](), [auditing](), [benchmarks](), [cs-ai](), [error](), [errors](), [frameworks](), [generative](), [graph](), [qa](), [reasoning](), [search](), [structured](), [verification]() ### AI overview GraphRail audits search-augmented reasoning at the step level by representing question requirements, evidence, and reasoning as structured claims, then checking whether the evidence covers the question, the reasoning addresses it, and each step is grounded in the evidence. It does this without a generative judge or gold answer. Across three multi-hop QA benchmarks and three reasoning frameworks, the abstract reports higher error recall than generative verifiers and stability under stochastic extraction. Its diagnoses can also guide online detection and correction of intermediate errors. ### Source excerpt arXiv:2611.06662v1 Announce Type: new Abstract: Search-augmented reasoning systems can reach correct answers while making errors in intermediate steps. Reliable auditing must check whether evidence $D$ covers the question $Q$, reasoning $T$ addresses $Q$, and $T$ is grounded in $D$. We introduce GraphRail, which represents requirements, evidence, and reasoning as structured claims and compares them without a generative judge or gold answer. Across three multi-hop QA benchmarks and three reasoning frameworks, GraphRail improves error recall over generative verifiers while remaining stable under stochastic extraction. Its diagnoses can also guide online reasoning, providing a traceable way to detect and correct intermediate errors. ## EchoKV: Self-Corrective KV Cache Eviction via Iterative Query Feedback DevFeed: [EchoKV: Self-Corrective KV Cache Eviction via Iterative Query Feedback]() Original publisher: [Read original article]() Author: JianKun Zou, Tongtong Yuan Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [AI Chat](), [Cache](), [context window]() Tags: [cache](), [context](), [cs-ai](), [error](), [feedback](), [improvements](), [kv-cache](), [performance](), [qa](), [query](), [repair]() ### AI overview EchoKV addresses a feedback failure in query-driven KV cache eviction: an incorrect retained workspace can distort later eviction queries and reinforce a poor decision, especially when relevant evidence is spread across a long context. It builds a feedback-compatible bootstrap workspace from proposals at future decode positions, then uses early decode-generated queries to repair the workspace before later lookahead steps. Compared with LAQ, it uses the same number of workspace forward passes but adds tensor re-scoring for repair. On 16 LongBench tasks with Mistral-7B-Instruct-v0.2 and Qwen2.5-7B-Instruct, the authors report improvements over LAQ on several evidence-intensive tasks, including multi-hop cross-document QA and long single-document comprehension, while performance remains comparable on other subsets across budgets and backbones. ### Source excerpt arXiv:2611.05903v1 Announce Type: new Abstract: KV cache eviction is not merely a token-ranking problem: the workspace retained for generating eviction queries also determines the quality of those queries. Consequently, an early workspace error can distort later query observations and reinforce an incorrect final eviction decision, particularly when evidence is dispersed across a long context. We present EchoKV, which discovers and addresses this failure mode with iterative correction. It initializes a feedback-compatible bootstrap workspace using proposals at future decode positions, then uses early decode-generated queries to recover the workspace before the remaining lookahead steps. It uses the same number of workspace forward passes as LAQ while adding a tensor re-scoring operation for workspace repair. On LongBench (16 tasks, Mistral-7B-Instruct-v0.2 and Qwen2.5-7B-Instruct), EchoKV shows notable improvements over LAQ on several tasks requiring dispersed evidence, including multi-hop cross-document QA and long single-document comprehension. Across budgets and backbones, EchoKV selectively improves evidence-intensive tasks while maintaining comparable performance on the remaining LongBench subsets. These results identify query--workspace coupling as a practical failure mode of query-driven KV eviction. ## Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs DevFeed: [Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs]() Original publisher: [Read original article]() Author: Aizierjiang Aiersilan Published: 2026-11-05T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Deep neural networks](), [Large Language Model](), [quantization](), [Open Source Models & Datasets](), [llama](), [qwen]() Tags: [benchmarks](), [consistency](), [cs-ai](), [cs-cl](), [cs-et](), [cs-lg](), [cs-ne](), [detection](), [entropy](), [hidden-states](), [inside](), [layer](), [linear](), [llama](), [llama-3-1](), [llms](), [mistral](), [mlp](), [qa](), [quantization](), [qwen](), [reproducibility](), [sampling](), [synthetic]() ### AI overview Across three 7B-8B instruction-tuned models in 4-bit NF4 quantization, a linear probe on a single mid-network hidden-state layer achieved 0.904-1.000 AUROC on held-out hallucination benchmarks. MLP probes rarely improved on linear probes by more than 0.01 AUROC, while sampling-based detectors reached at most 0.541 under the same protocol. Peak layers clustered around blocks 13-18 of 32 for Llama and Mistral, and 19-25 of 28 for Qwen. First-block attention entropy also provided a complementary signal on HaluEval-QA (0.866-0.941 AUROC) without additional inference cost. The authors attribute sampling methods' low scores to a mismatch between paired-label evaluation and the information those methods access, rather than an inherent limitation. The abstract reports code and data for reproduction on a single 8 GB GPU. ### Source excerpt arXiv:2606.02628v1 Announce Type: cross Abstract: We investigate whether open-source LLMs encode a linearly separable truthfulness signal in their hidden states, and at which network depth this signal is strongest. Across three $7$B--$8$B instruction-tuned models (Llama-3.1-8B, Mistral-7B, Qwen2.5-7B) loaded in $4$-bit NF4 quantization, we extract per-layer hidden states on four hallucination benchmarks (TruthfulQA, HaluEval-QA, FEVER, and a controlled synthetic set) and compare four detection approaches: linear and MLP probes, INSIDE EigenScore, self-consistency, and attention entropy. A linear probe on a single mid-network layer achieves $0.904$--$1.000$ AUROC on held-out splits, while sampling-based detectors do not exceed $0.541$ AUROC under the same protocol. The truthfulness signal is approximately linear: MLP probes rarely surpass linear probes by more than $0.01$ AUROC. Peak probing layers fall in a consistent band across model families on natural-language benchmarks -- blocks~$13$--$18$ of~$32$ for Llama and Mistral, and blocks~$19$--$25$ of~$28$ for Qwen. First-block attention entropy provides a complementary signal in knowledge-grounded settings ($0.866$--$0.941$ AUROC on HaluEval-QA) at no additional inference cost. The low discriminability of sampling methods under this protocol reflects a structural mismatch between paired-label evaluation and the information these methods access, rather than an inherent limitation of those methods. Code and data are released for full reproducibility on a single $8$\,GB GPU. ## Your GitHub Actions Cron Job Can Disappear Without Failing DevFeed: [Your GitHub Actions Cron Job Can Disappear Without Failing]() Original publisher: [Read original article]() Author: Malek Damiri Published: 2026-11-04T10:00:02Z Content type: tutorial Language: en Sources: [HackerNoon]() Topics: [GitHub Actions](), [Slack](), [GitHub]() Tags: [actions](), [airflow](), [alert](), [alerts](), [automated](), [bot](), [branch](), [bridge](), [cron](), [data-engineering](), [github](), [github-actions](), [github-actions-cron](), [github-actions-schedule](), [github-actions-yaml](), [how-to](), [monitoring](), [qa](), [reporting](), [scheduled](), [scheduled-workflows](), [slack](), [workflow-monitoring](), [workflow-reliability](), [workflows]() ### AI overview GitHub Actions scheduled workflows can be delayed or silently dropped, so a daily report should check its data before posting and alert when no successful run appears. The excerpt recommends scheduling off the busy hour with an IANA timezone, testing QA rules, enabling manual dispatch, and using concurrency to prevent overlapping runs. Its report checks cover data freshness, whether driver totals reconcile, the size of the "other" category, and row counts; failed checks route the report to its owner. A Slack history check avoids duplicate posts after retries. A later watchdog can look for a successful run, but because it uses the same scheduler, it is only a second chance; the excerpt suggests an external heartbeat for more important reports. It recommends a more dependable scheduler when exact timing or frequent missed runs matter. ### Source excerpt Learn how to make GitHub Actions scheduled workflows more reliable with automated QA, Slack reporting, duplicate prevention, and watchdog monitoring. ## Balancing Chain-of-Thought and Answer Supervision for Small Language Models on Medical Question Answering DevFeed: [Balancing Chain-of-Thought and Answer Supervision for Small Language Models on Medical Question Answering]() Original publisher: [Read original article]() Author: Le Zhou, Xiao Luo, Chinmay Chandra, Zhan Zhang Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [LLM Techniques](), [Chain-of-thought](), [llm-reasoning](), [MONAI]() Tags: [chain-of-thought](), [cs-ai](), [cs-cl](), [fine-tuning](), [language-models](), [medical](), [qa](), [small-language-models](), [supervision]() ### AI overview For small medical QA models, the better chain-of-thought fine-tuning method depends on the task format and base model. Segment-weighted supervision (SwS), which weights reasoning and answer losses separately, generally performs better on multiple-choice MedQA and EMS-MCQA, especially for general-purpose models; Qwen3.5-4B improves up to 44.8% over its MedQA zero-shot baseline. Medical-domain models show smaller or sometimes negative MCQ gains. Uniform-across-Sequence supervision (UaS), which weights the full reasoning-and-answer sequence equally, works best for nearly all models on free-text Med-O1 and improves both model types; 3B Llama-3.2 approaches larger 7B medical models. Excessively long reasoning is associated with lower performance for some models, so limiting its length may help selected cases. These findings come from six models across the stated benchmarks; neither method is best in every setting. ### Source excerpt arXiv:2611.03948v1 Announce Type: cross Abstract: Small language models (<=9B parameters) can support local medical question answering (QA) while reducing privacy and latency concerns, but effective fine-tuning remains challenging. This paper compares two chain-of-thought (CoT) supervised fine-tuning methods: Uniform-across-Sequence (UaS), which applies equal loss weighting to the full CoT-and-answer sequence, and Segment-weighted (SwS), which separately weights CoT and answer losses. We evaluate six general- and medical-domain models on multiple-choice QA (MedQA and EMS-MCQA) and free-text QA (Med-O1). Results show that neither method is universally best: on multiple-choice QA (MedQA and EMS-MCQA), SwS generally performs better, with general-purpose models benefiting most; notably, Qwen3.5-4B improves by up to 44.8% over its zero-shot baseline on MedQA. In contrast, medical-domain models show smaller or even negative gains on MCQs. On free-text QA (Med-O1), UaS performs best for nearly all models and provides consistent improvements across both model types, with the 3B Llama-3.2 model approaching the performance of much larger 7B medical-domain models. We also find that excessively long CoT generation is associated with lower performance in some models on either multiple-choice or free-text QA tasks, suggesting that limiting the CoT length during training may benefit selected models. Overall, the results show that CoT loss weighting should be matched to the task format and the knowledge already encoded in the base model, rather than using a single fine-tuning strategy for all clinical QA tasks. ## Feeling of Knowing in Large Language Models DevFeed: [Feeling of Knowing in Large Language Models]() Original publisher: [Read original article]() Author: Zichuan Fu, Xian Wu, Jingtong Gao, Wenlin Zhang, Jiaxuan Li, Binhao Wang, Yimin Deng, Guojing Li, Xiaopeng Li, Derong Xu, Yefeng Zheng, Xiangyu Zhao Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [post-training](), [Fine-tuning](), [LLM Techniques](), [LLM evaluation / benchmarking](), [Model Routing]() Tags: [alignment](), [cs-ai](), [cs-cl](), [cs-lg](), [evaluation](), [fine-tuning](), [knowledge](), [language-models](), [large-language-models](), [llms](), [model-routing](), [optimization](), [policy](), [post-training](), [qa](), [retrieval](), [supervised-fine-tuning](), [train]() ### AI overview Post-training can change an LLM's ability to answer questions without keeping its answerability judgments in sync. The paper proposes functional feeling of knowing (FoK) to align those judgments with the model's current knowledge, supporting selective generation, retrieval triggering, and model routing without claiming subjective experience. Its three stages update knowledge from corrected on-policy answers (OP-SFT), adapt FoK judgments across intermediate knowledge states (KΔ-FT), then use FoK-guided policy optimization (FGPO) to train answers or explanations. The abstract reports improved judgment quality on updated knowledge in factual-QA experiments, cross-dataset evaluation, and three sequential updates; MMLU and ARC-Easy are used separately to assess capability preservation. ### Source excerpt arXiv:2611.04834v1 Announce Type: cross Abstract: Post-training updates change what large language models (LLMs) can answer, but may leave their answerability judgments misaligned. We study this knowledge-judgment desynchronization through functional feeling of knowing (FoK), which supports selective generation, retrieval triggering, and model routing without implying subjective experience. We propose two trainable channels: a knowledge channel that changes answerability and a FoK channel that judges the current knowledge state. Three stages train them: on-policy supervised fine-tuning (OP-SFT) updates knowledge from corrected on-policy answers; knowledge-delta FoK training (K$\Delta$-FT) adapts judgments across intermediate knowledge states; and FoK-guided policy optimization (FGPO) trains the subsequent answer or explanation. Factual-QA experiments, cross-dataset evaluation, and three sequential updates show that explicit FoK alignment improves judgment quality on updated knowledge. MMLU and ARC-Easy separately assess capability preservation. ## SurgStream: A Workflow-Anchored Proactive Streaming Framework and Benchmark for Real-Time Surgical Video Understanding DevFeed: [SurgStream: A Workflow-Anchored Proactive Streaming Framework and Benchmark for Real-Time Surgical Video Understanding]() Original publisher: [Read original article]() Author: Ruipeng Ma, Qing Xu, Jie Liang, Zhen Chen Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Video Analytics](), [Real-time AI interactions](), [real-time]() Tags: [alerts](), [arxiv](), [benchmark](), [code](), [cs-ai](), [cs-cv](), [documentation](), [experiments](), [framework](), [learning](), [memory](), [model](), [monitoring](), [proactive](), [qa](), [queries](), [real-time](), [responses](), [streaming](), [video](), [workflow]() ### AI overview SurgStream targets real-time surgical video understanding by combining scene perception, memory of past events, and timely responses. Workflow-Anchored Memory records structured instrument-action-target observations and consolidates them into phase-level summaries for compact, queryable recall. Deadline-aware Transition Learning supervises response states within clinical windows to help the system respond before procedural deadlines. The accompanying SurgStream-Bench contains 11.7K streaming interactions spanning real-time perception, backward memory queries, forward proactive QA, proactive alerts, and caption memory. The paper reports gains over state-of-the-art methods, especially on memory and forward proactive QA; the supplied abstract gives no quantitative results or evaluation details. ### Source excerpt arXiv:2611.01982v1 Announce Type: cross Abstract: Surgical video understanding is essential for intraoperative assistance such as workflow monitoring, instrument preparation and procedural documentation. Delivering such assistance in real time requires a model to jointly perceive the ongoing scene, remember past surgical events, and respond before clinically relevant moments pass. Existing approaches provide limited support for jointly retaining procedural evidence and deciding when to respond in streaming surgical video. In this work, we propose SurgStream, a workflow-anchored proactive streaming framework for real-time surgical video understanding. We first introduce Workflow-Anchored Memory (WAM) to record structured instrument-action-target observations and fold them into phase-level summaries, providing compact and queryable surgical memory. Deadline-aware Transition Learning (DTL) is then devised to supervise response states within clinical windows, enabling timely responses before procedural deadlines. We further introduce SurgStream-Bench, comprising 11.7K streaming interactions across various tasks covering real-time perception, backward memory queries, forward proactive QA, proactive alerts and caption memory. Experiments on SurgStream-Bench show gains over state-of-the-art methods, especially in memory and forward proactive QA. The source code will be available. ## LF-VideoQA: Evaluating Long-Video Understanding with Long-Form Theory-of-Mind Questions DevFeed: [LF-VideoQA: Evaluating Long-Video Understanding with Long-Form Theory-of-Mind Questions]() Original publisher: [Read original article]() Author: Shaden Shaar, Bradon Thymes, Majd Aldaye, Claire Cardie, Bharath Hariharan Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Machine Perception](), [AI Chat](), [generative-ai-tools]() Tags: [cs-ai](), [cs-cl](), [cs-cv](), [dataset](), [qa](), [reasoning](), [theory](), [video]() ### AI overview LF-VideoQA evaluates long-video understanding through open-ended theory-of-mind questions that require combining evidence across scenes. It contains more than 6,300 questions about 94 films; each spans scenes averaging 23 minutes and has a question-specific rubric for recall and precision. The dataset is generated from screenplay annotations of scene events and characters' beliefs, with LLM-generated and verified question-answer-rubric triples and human quality studies. In evaluations of four open-source VLM families, even a screenplay-access oracle fell short of complete answers, while the best non-oracle model reached barely half the oracle's recall. Captions and subtitles underperformed the video, suggesting visual evidence matters for these questions; missing required claims made recall the main bottleneck. ### Source excerpt arXiv:2611.00383v1 Announce Type: cross Abstract: The hardest part of understanding a long video is not pinpointing any one moment but synthesizing many, e.g., answering a question that requires tracking what develops across hours of content and articulating how the pieces fit together. To date, most long-video QA datasets require only a short answer or a multiple-choice selection, oversimplifying the long-video understanding task to retrieval. We introduce LF-VideoQA, a dataset for long-form, open-ended question answering over long narrative videos built on theory-of-mind questions: a character's belief depends on everything they have seen or been told, so each question demands assembling and reasoning from evidence distributed across the video. The dataset contains more than 6,300 questions on 94 recent films; each requires reasoning over scenes spanning an average of 23 minutes and comes with a question-specific rubric for recall and precision scoring. The dataset is created entirely automatically by annotating screenplays with scene events and each character's first-order beliefs and meta-beliefs, then generating (question, answer, rubric) triples through LLM-based question generation and verification; human studies validate their quality. We evaluate four families of state-of-the-art open-source VLMs across three settings and find the task highly challenging: (i) even an oracle with access to the screenplay falls well short of complete answers; (ii) the best non-oracle model attains barely half of the recall level of the oracle; (iii) unlike on prior long-video benchmarks, captions and subtitles underperform the video itself, indicating that the questions need visual evidence; and (iv) recall is the predominant bottleneck, with answers omitting many required rubric claims. ## Labadain-MedQA: A Tetun Medical QA Dataset for Fine-Tuning and Evaluating Large Language Models DevFeed: [Labadain-MedQA: A Tetun Medical QA Dataset for Fine-Tuning and Evaluating Large Language Models]() Original publisher: [Read original article]() Author: Gabriel de Jesus, S\\'ergio Nunes Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [peft](), [Fine-tuning](), [LLM evaluation / benchmarking](), [lora]() Tags: [cross](), [cs-ai](), [cs-cl](), [dataset](), [datasets](), [fine-tuning](), [gemini](), [gemma](), [healthcare](), [language-models](), [large-language-models](), [llama](), [llms](), [lora](), [machine-translation](), [medical](), [peft](), [performance](), [qa](), [quality](), [questions](), [translation]() ### AI overview Labadain-MedQA provides 39,077 Tetun medical question-answer pairs translated from MedAlpaca's Medical Flashcards and WikiDoc Patient Information using Gemini 2.5 Flash. In an evaluation of 50 sampled QA pairs, two Tetun-speaking healthcare professionals rated human translations higher than model-generated translations for medical accuracy and fluency (4.43/5 versus 4.02/5; p < 0.001). Fine-tuning Gemma and Llama variants with parameter-efficient methods was assessed on 325 Tetun-translated USMLE questions. Gemma 4 12B with QLoRA improved mean accuracy by 10.79% using NF4 inference and 10.42% with LoRA adapters merged into the BF16 base model, relative to corresponding base models. The abstract reports these results as evidence of the dataset's potential utility, while its translation assessment covers only a small sample. ### Source excerpt arXiv:2611.00935v1 Announce Type: cross Abstract: Advances in large language models (LLMs) have created opportunities for cross-lingual transfer of medical knowledge from high-resource to low-resource languages to support access to healthcare information. However, medical question-answering (QA) datasets for fine-tuning and evaluating LLMs remain scarce for Tetun. To address this gap, we translated MedAlpaca's Medical Flashcards and WikiDoc Patient Information collections into Tetun using Gemini 2.5 Flash, producing 39,077 medical QA pairs. Translation quality was assessed on 50 randomly sampled QA pairs (100 documents) using automatic metrics against human reference translations and evaluations by two native Tetun-speaking healthcare professionals, focusing on medical accuracy and Tetun fluency. Human translations achieved higher mean ratings than LLM-generated translations for both criteria (4.43/5 vs. 4.02/5), with statistically significant differences (Wilcoxon signed-rank test, p < 0.001). To assess downstream utility, we fine-tuned Gemma and Llama variants on the translated MedAlpaca dataset using parameter-efficient fine-tuning (PEFT) and evaluated them on 325 Tetun-translated United States Medical Licensing Examination multiple-choice questions. Gemma 4 12B fine-tuned using QLoRA achieved relative improvements in mean accuracy of 10.79% with 4-bit NormalFloat (NF4) inference and 10.42% after merging the LoRA adapters into the BF16 base model, compared with the corresponding base models. These results highlight the viability of automatic dataset translation for Tetun medical QA and the effectiveness of PEFT in enhancing model performance. The datasets, models, and evaluation resources are accessible at https://github.com/labadain/labadain-medqa to support reproducible research on Tetun medical QA and facilitate machine translation, cross lingual, and multilingual research. ## Fewer Tokens, More Answers: Verifiable-Reward Tuning of an Ornith-Style VLM for Multi-View Driving QA DevFeed: [Fewer Tokens, More Answers: Verifiable-Reward Tuning of an Ornith-Style VLM for Multi-View Driving QA]() Original publisher: [Read original article]() Author: T N Rickesh, Pranav Vinodh, Sowmya Kamath S Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [grpo](), [rlvr](), [llm-reasoning](), [post-training]() Tags: [adapter](), [budget](), [cs-cl](), [cs-cv](), [cs-lg](), [delivery](), [driving](), [grpo](), [limits](), [mechanism](), [provenance](), [qa](), [recovery](), [reward](), [vlm]() ### AI overview On a selected 400-question screen capped at 256 tokens, GRPO tuning of an Ornith checkpoint for multi-camera driving QA raised answer extraction from 78.25% to 90.25% and accuracy from 55.50% to 63.75%. Of 39 wrong-to-correct changes, 33 followed unextractable base responses; all 87 unextractable base responses hit the token cap. At 512 tokens, checkpoint-verified reconstruction produced only a 0.75-point first-pass accuracy gain, compared with 8.25 points at 256. A separately analyzed answer-cued continuation helped the base more and narrowed the adapter gap. Across the full 2,511-question development set, a recovery-enabled pipeline cut mean generated tokens from 108.6 to 47.1, but the paired accuracy interval included zero gain. Results concern Ornith alone and are subject to selection and provenance limits; they do not demonstrate transfer to Bonsai or a shared architecture mechanism. ### Source excerpt arXiv:2611.00833v1 Announce Type: cross Abstract: Reasoning-enabled model derivatives must turn intermediate deliberation into a usable answer within a deployment budget. We study that transition in an Ornith checkpoint adapted to multi-camera driving question answering with verifiable rewards. Paired archived outputs show that, on a selected 400-question development screen with a 256-token cap, GRPO increases answer extraction from 78.25% to 90.25% and accuracy from 55.50% to 63.75%. Of 39 wrong-to-correct transitions, 33 begin with an unextractable base response. The base's 87 unextractable responses all reach the token cap. Checkpoint-verified reconstruction on those same questions gives a much smaller first-pass accuracy gain at 512 tokens: +0.75 percentage points, versus +8.25 at 256. A later answer-cued continuation is analyzed separately; it helps the base more and narrows the adapter difference. On the full 2,511-question development set, this recovery-enabled pipeline reduces mean generated tokens from 108.6 to 47.1, while the paired accuracy interval includes zero gain. We frame these findings as adaptation of answer delivery under finite budgets, with explicit selection and provenance limits. Ornith's self-scaffolding lineage and newer reasoning-enabled derivatives such as Bonsai motivate the problem, but our measured results concern Ornith alone and do not establish a shared architectural mechanism or cross-model transfer. ## Answer Correction with Verifiable Rewards: A Gemma-Family Study of Multi-View Driving QA DevFeed: [Answer Correction with Verifiable Rewards: A Gemma-Family Study of Multi-View Driving QA]() Original publisher: [Read original article]() Author: T N Rickesh, Pranav Vinodh, Sowmya Kamath S Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [grpo](), [gemma](), [rlvr](), [post-training](), [autonomous vehicles]() Tags: [cs-cv](), [cs-lg](), [driving](), [gemma](), [grpo](), [language-models](), [qa](), [reward](), [safety](), [study]() ### AI overview In the matched-harness Gemma-4-12B base/GRPO comparison, accuracy rose from 73.78% to 75.43% (1.65 percentage points; scene-cluster bootstrap interval +0.84 to +2.41). Of 215 improvements, 213 began with an extractable but incorrect base answer, while two recovered previously unextractable responses, supporting an answer-selection gain for this recorded tuning configuration. Category results show a safety tradeoff: routine perception improved by 5.93 points, while safety-critical hazard prediction declined by 1.82 points. The study cautions that unweighted outcome rewards can favor aggregate accuracy over rare edge cases. A separate archived Gemma-31B thinking comparison found no detected average improvement, but uncertainty does not establish equivalence. The results do not identify a specific grounding mechanism or establish superiority over matched supervised tuning. ### Source excerpt arXiv:2611.00828v1 Announce Type: cross Abstract: Multi-camera driving question answering provides a discrete test of vision-language models, but post-training gains can reflect changed selected answers or improved answer extraction. We investigate this distinction within the Gemma model family using archived, task-linked WaymoQA evaluations. A zero-shot diagnostic on 5,519 test questions spans effective-small, mixture-of-experts, unified, and dense checkpoints, with serving differences disclosed rather than interpreted as architecture effects. The strongest controlled evidence is a matched-harness Gemma-4-12B base/GRPO pair: accuracy rises from 73.78% to 75.43%, a gain of 1.65 percentage points with a scene-cluster bootstrap interval of [+0.84, +2.41]. There are 215 improvements and 124 regressions; 213 improvements start from an extractable but incorrect base answer, and only two recover previously unextractable responses. Crossing correctness with extraction state shows why correct noncompliant outputs cannot be counted as prediction errors. Category counts reveal a critical safety asymmetry: routine perception gains (+5.93 pp) mask a regression in safety-critical hazard prediction (-1.82 pp), demonstrating that unweighted outcome rewards can trade off rare edge cases for aggregate accuracy. Reward saturation describes sampled training dynamics rather than an architectural ceiling. A separate archived Gemma-31B thinking comparison yields no detected average improvement, with uncertainty that does not establish equivalence. These results support a modest answer-selection gain for the recorded tuning configuration and a reproducible diagnostic accounting, without attributing corrected labels to specific grounding mechanisms or claiming superiority over matched supervised tuning. ## Adaptive-Policy Distillation: Knowledge Injection under Information Asymmetry DevFeed: [Adaptive-Policy Distillation: Knowledge Injection under Information Asymmetry]() Original publisher: [Read original article]() Author: Hojun Cho, Kang Eun Jeon, Jaegul Choo Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Post-training optimization](), [Retrieval Augmented Generation (RAG)](), [Neural Network](), [AI-driven information retrieval]() Tags: [adaptive](), [arxiv](), [benchmarks](), [best](), [book](), [coverage](), [cs-cl](), [cs-lg](), [distillation](), [distribution](), [gate](), [gradient](), [inference](), [information](), [knowledge](), [policy](), [qa](), [signal](), [student](), [target](), [token](), [tokens]() ### AI overview On two document-QA benchmarks, on-policy distillation (OPD) injects document knowledge less efficiently than simple off-policy distillation. The abstract attributes this to two sparse-position failures: reverse KL gives little gradient to document-specific tokens the student considers unlikely, and student rollouts reach later target positions only when preceding tokens are generated correctly, reducing coverage as targets lengthen. Adaptive-Policy Distillation (APD) uses a per-token gate to shift rollouts toward the teacher and the objective toward forward KL where needed, then withdraws both interventions as the student learns. Without labels or a verifier, APD improves over OPD by up to 54.6%, with reported best scores of 0.558 F1 on QASPER and 0.486 MRR on BioASQ. ### Source excerpt arXiv:2611.00852v1 Announce Type: cross Abstract: On-policy distillation (OPD), which uses reverse-KL supervision on student rollouts to transfer a context-conditioned model's behavior into student model parameters for closed-book inference, has shown strong results. However, we find that this success does not translate to knowledge injection: across two document-QA benchmarks, OPD injects document knowledge far less efficiently than simple off-policy distillation. We trace this reversal to two failures that are structural, yet sparse across token positions. Reverse KL weights its gradient by the student distribution and therefore barely lifts document-specific tokens the student considers unlikely; after forward KL restores this signal, student rollouts expose later positions on the target trajectory only after generating the preceding tokens correctly, so coverage deteriorates with target length. Motivated by the concentration of teacher-student divergence at these hard positions, we propose Adaptive-Policy Distillation (APD), whose per-token gate jointly shifts the rollout toward the teacher and the objective toward forward KL where needed, withdrawing both interventions as the student learns. Without labels or a verifier, APD improves over OPD by up to 54.6%, achieving the best scores of 0.558 F1 on QASPER and 0.486 MRR on BioASQ. ## From Single-Shot QA to Continuous Perception: Revisiting the Energy Composition and Optimization Boundary of Edge VLMs DevFeed: [From Single-Shot QA to Continuous Perception: Revisiting the Energy Composition and Optimization Boundary of Edge VLMs]() Original publisher: [Read original article]() Author: Yiting Duan, Miaoxin Wang, Aoqing Guo, Bo Peng Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [vlm](), [Sustainable Computing](), [LLM optimization](), [Jetson AGX Thor](), [Jetson Orin](), [Image]() Tags: [abstract](), [adaptive](), [architecture](), [arxiv](), [boundary](), [composition](), [continuous](), [cost](), [cs-ai](), [cs-cv](), [decode](), [deployment](), [edge](), [efficiency](), [encoding](), [energy](), [image](), [independent](), [jetson](), [optimization](), [qa](), [vlms]() ### AI overview Profiling nine edge VLMs across three architecture families on Jetson AGX Thor and Orin, the study finds that energy costs follow two distinct patterns: decode energy per output token rises with model scale, while visual-prefill energy depends mainly on visual-encoder efficiency. In continuous perception with N=10 and sparse output, visual prefill accounts for 92-97% of total energy, shifting the optimization focus toward visual frame reuse, adaptive output truncation, task-aware routing and encoder throughput. The findings are based on 4,508 profiling records calibrated with a pure-text task. ### Source excerpt arXiv:2611.00036v1 Announce Type: cross Abstract: Visual-prefill energy conclusions for edge vision-language models (VLMs) are commonly drawn under single-shot workloads; we revisit them under continuous perception. Profiling nine VLMs across three architecture families on Jetson AGX Thor and Orin (4,508 records, calibrated with a pure-text task), we find two mutually independent energy laws: the per-output-token decode cost grows monotonically with model scale, whereas the visual-prefill (image-encoding) cost is dominated by visual-encoder architectural efficiency, not scale. Under continuous perception ($N{=}10$) with sparse output, visual-prefill energy reaches 92--97\% of total. We derive optimization strategies---visual frame reuse, adaptive output truncation, task-aware routing, and encoder throughput optimization---for edge VLM deployment. ## Safety4DM: A Vision-Language Framework for Risk-Aware Dynamic Maps in Vehicle-Road-Cloud Collaborative Driving DevFeed: [Safety4DM: A Vision-Language Framework for Risk-Aware Dynamic Maps in Vehicle-Road-Cloud Collaborative Driving]() Original publisher: [Read original article]() Author: Zhengshu Zhou, Shaoyu Wang, Lu Tao, Xinhui Hu, Jiaqi Gu, Tingting Zhao Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [autonomous vehicles](), [vlm](), [Fine-tuning](), [Large Language Model](), [Simulation]() Tags: [advice](), [autonomous-driving](), [benchmark](), [cs-ai](), [dataset](), [driving](), [driving-safety](), [fine-tuning](), [framework](), [image](), [interface](), [language](), [language-model](), [llm](), [maps](), [natural-language](), [natural-language-interface](), [prompting](), [qa](), [risk](), [vision]() ### AI overview Safety4DM adds two safety capabilities to vehicle-road-cloud dynamic maps: turning road images into fine-grained risk entities and providing a natural-language interface for querying risks. Its RoadRisk dataset covers four road-condition types and four risk levels; a fine-tuned vision-language model produces structured representations that dynamic maps can ingest. The SafeStep prompting mechanism answers risk queries and generates driving advice. The authors report 89.83% risk-perception accuracy for fine-tuned Qwen3-VL-4B, and over 93% query accuracy for medium and large (20B+) models across five LLM families. Evaluation uses the SafeVRC-Sim simulation framework and SafeDM-QA benchmark. These results are limited to the reported datasets and evaluation settings in the abstract. ### Source excerpt arXiv:2611.01996v1 Announce Type: new Abstract: Dynamic map (DM) is a core information infrastructure for vehicle-road-cloud (VRC) cooperative autonomous driving. It fuses perception data from multiple sources to overcome the inherent line-of-sight and field-of-view limits of single-vehicle perception. However, DM still falls short on driving safety in two ways. First, DM data specifications already define road conditions semantically, but no pipeline yet turns real-world road observations into fine-grained risk entities in the DM. Second, existing DM systems lack a natural language interface for driving risks. To close these gaps, we propose Safety4DM, a safety-oriented language system for VRC dynamic maps. We first build RoadRisk, a real-world road image dataset that covers four types of road conditions and four risk levels. On this dataset, we fine-tune a vision-language model (VLM) for road safety risk perception. The model outputs structured risk representations that the DM can ingest directly. We then introduce a Safety-aware Step-wise Prompting (SafeStep) mechanism as the natural language query interface of the DM. SafeStep answers queries accurately and also generates actionable driving advice. To build and evaluate this interface, we develop SafeVRC-Sim, a VRC cooperative simulation framework, and SafeDM-QA, a safety question-answering benchmark. Experiments show that even the compact Qwen3-VL-4B reaches 89.83% accuracy on risk perception after fine-tuning. SafeStep also plugs easily into multiple LLM families. With medium and large models (20B and above) from five families, it achieves over 93% natural language query accuracy in every case, which shows strong generalization. ## You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model DevFeed: [You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model]() Original publisher: [Read original article]() Author: Ziyang Luo, Zhongyao Chu, Xinjie He, Youting Wang, Xukui Qin, Runxiong Wu, Yan-Syuan Chen Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Large Language Model](), [Uncertainty quantification LLMs](), [Neural Network](), [Fine-tuning](), [LLM optimization]() Tags: [activation](), [context](), [cs-cl](), [cs-lg](), [hidden](), [hidden-states](), [language-model](), [lora](), [map](), [module](), [network](), [one](), [qa](), [steering](), [stream](), [vector](), [weight]() ### AI overview YOPO combines answering and abstention in one forward pass of a frozen language model, without changing its weights. A trained probe edits residual states to improve answer accuracy, while a fixed direction scores answerability and triggers abstention at low scores. The probe edit interferes with that score; a small map trained only to reconstruct unedited states removes 63% of the interference on the 1.5B model, and the same repair recipe also works with LoFiT and LoReFT steering modules. On αNLI, YOPO raises three-way accuracy from 0.38 to about 0.75 on 1.5B and stays within 0.02 of a two-pass reference across 1.5B, 3B, and 7B at half the cost. The abstract also reports improvements on four human-labeled QA datasets and near-parity with the two-pass reference across nine backbones in six model families, without backbone-specific hyperparameter tuning. For answerability, transferring the fixed direction to an unfitted dataset costs under 3 points of decision accuracy on average; classifier-based alternatives lose 16 and 22 points and perform below asking the model whether its context suffices. ### Source excerpt arXiv:2608.14465v2 Announce Type: replace-cross Abstract: A frozen language model under-uses the evidence in its hidden states and answers confidently when the context contains no answer. We fix both on the residual stream without changing any weight. For accuracy, a small trained network (the probe) adds an input-dependent vector to the residual at several middle layers; it raises Qwen2.5-1.5B on $\alpha$NLI from 51% to 73% accuracy. For abstention, a fixed vector (the direction), the mean hidden state of contexts that contain the answer minus that of contexts that do not, scores the hidden state; a low score abstains. On four QA datasets with human answerability labels (Qwen2.5-7B), moving the direction to a dataset it was not fit on costs under 3 points of decision accuracy (from 0.83 in domain) on average and still beats asking the model whether its context suffices; the same judgment trained as a classifier into the probe's activation edit or into LoRA weights loses 16 and 22 points and ends below asking. We set out to put both functions on a single forward pass, and found that the probe's edit changes the very hidden state the direction scores, degrading it. A small map -- trained without labels, only to reconstruct the unedited hidden state from the edited one -- removes 63% of that interference on 1.5B. Refit on the hidden-state pairs of LoFiT or LoReFT, the same recipe repairs those methods too: the steering module is replaceable. The resulting system, YOPO, answers and abstains in one forward pass. On $\alpha$NLI it raises three-way accuracy from 0.38 to about 0.75 on 1.5B, and stays within 0.02 of a two-pass reference on 1.5B, 3B and 7B, at half the cost. We also observe improvements on the four QA datasets, where answerability is human-labeled and answers are generated as free-form text. Near-parity with the two-pass reference holds across nine backbones spanning six model families without backbone-specific hyperparameter tuning. ## SUREON: A Benchmark and Vision-Language Model for Surgical Reasoning DevFeed: [SUREON: A Benchmark and Vision-Language Model for Surgical Reasoning]() Original publisher: [Read original article]() Author: Alejandra Perez, Anita Rau, Busisiwe Mlambo, Lee White, Aina Brunet, Juan Verde, Juan Caicedo, Jo\\"el L. Lavanchy, Adam Schmidt, Chinedu Nwoye, Omid Mohareri, Muhammad Abdullah Jamal Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [vlm](), [LLM evaluation / benchmarking]() Tags: [academic](), [adversarial](), [ai](), [benchmark](), [computer-vision](), [cs-ai](), [cs-cv](), [dataset](), [grounding](), [multimodal](), [open-source](), [qa](), [reasoning](), [video](), [vision](), [vision-language-model]() ### AI overview SUREON tests whether surgical video QA models can reason from visual evidence rather than infer answers from language or procedural priors. Its pipeline creates questions from narrated academic surgical videos, refines distractors to reduce linguistic bias, and uses adversarial guessers to filter out questions answerable without seeing the video. The dataset has 76K training pairs spanning 168 procedures and 10 specialties; its 340-question, 10-category benchmark was verified by expert surgeons. No evaluated model exceeded 53% accuracy, and expert surgeons reached 64.9%. The SUREON-VLM baseline outperformed existing open-source models on SUREON and matched larger specialist models on multiple standard surgical benchmarks. Ablations found that distractor refinement and blind curation raised the multimodal gain from 8.7 to 29.1 percentage points while reducing the training set from 176.6K to 76K filtered samples. ### Source excerpt arXiv:2603.06570v2 Announce Type: replace-cross Abstract: A capable surgical AI model should reason from visual evidence; yet, many existing surgical benchmarks can be partially solved without visual input, as models may exploit procedural priors or common anatomy-instrument interactions to infer answers from language alone. To address this, we introduce SUREON, a video question-answering (QA) dataset and benchmark designed to evaluate visually grounded surgical reasoning. Our pipeline generates QA pairs from narrated surgical academic videos, refines incorrect options using multiple models to reduce linguistic biases, and uses trained adversarial guessers to filter out questions answerable without visual evidence. The resulting dataset contains 76K training pairs covering 168 procedures and 10 specialties, alongside a benchmark of 340 questions across 10 categories verified by expert surgeons. Benchmarking leading commercial and open-source models shows that surgical visual reasoning remains largely unsolved--no model exceeds 53% accuracy and even expert surgeons reach only 64.9%. We also introduce SUREON-VLM, a baseline trained on our curated data that avoids language shortcuts, outperforms existing open-source models on SUREON, and matches larger specialist models on multiple standard surgical benchmarks. Our training ablations show that distractor refinement and blind curation strengthen visual grounding, increasing the multimodal gain from 8.7 to 29.1 percentage points while reducing the training set from 176.6K to 76K filtered samples. SUREON thus offers both a verified testbed for surgical visual reasoning and training data that strengthens visual grounding. ## From Bug Report to Production: How I Fix Bugs with AI DevFeed: [From Bug Report to Production: How I Fix Bugs with AI]() Original publisher: [Read original article]() Author: inDrive.Tech Published: 2026-11-01T21:40:15Z Content type: article Language: en Sources: [HackerNoon]() Topics: [bugfixes](), [Monitoring](), [Static code analysis](), [engineering-culture](), [Risk]() Tags: [agent](), [ai](), [ai-agent](), [ai-coding-agents](), [api-testing](), [bugs](), [claude](), [clusters](), [code](), [code-review](), [good-company](), [hackernoon-top-story](), [indrive](), [logs](), [production](), [production-monitoring](), [qa](), [qa-engineer](), [qa-engineering](), [quality-assurance](), [ship](), [software-testing]() ### AI overview The article describes a three-level workflow in which a QA engineer uses Claude to investigate code, verify hypotheses with production logs and metrics, then prepare and ship small fixes through normal review and release checks. Bug reports include a proposed fix, regression criteria, risks, user impact, and code references, with confirmed findings distinguished from hypotheses. Production analysis can reveal misleading error counts: one week of data across 13 clusters showed about 1.86 million 500 errors, roughly 96.6% tied to client cancellations; the proposed handling change was an estimate, not a deployed result. A shipped idempotency-key fix validated requests before reserving keys and released them after failures; duplicate-key errors fell about 30%, with no noticeable response-time or other-error change and an assessed cache load of about 0.01%. The engineer retains responsibility for evidence, scope, tests, review, release, and post-release checks. ### Source excerpt How a QA engineer at inDrive uses AI to investigate bugs, analyze production logs, write fixes, and ship changes to production. ## DuQUAD: A Dual-View Framework with Quality-Aware Evidence Pruning for Multi-Document Question Answering DevFeed: [DuQUAD: A Dual-View Framework with Quality-Aware Evidence Pruning for Multi-Document Question Answering]() Original publisher: [Read original article]() Author: Yeonjun Choi, Seyeon Jeong, Jong Wook Kim, Beakcheol Jang Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [llm-reasoning](), [Vectors & LLMs](), [Enhancing AI model performance](), [Parallel reasoning]() Tags: [agent](), [arxiv](), [attention](), [bias](), [cs-ai](), [cs-cl](), [cs-ir](), [framework](), [language](), [language-models](), [large-language-models](), [local](), [mitigation](), [multi-agent](), [noise](), [qa](), [quality](), [reasoning](), [relevance](), [reranking](), [retrieval](), [retrieval-augmented-generation](), [score]() ### AI overview DuQUAD addresses "Lost in the Middle" in multi-document question answering by pairing a Structural Fusion Score-guided Local Agent for targeted evidence recovery with a Global Agent that reasons over the full document set. It scores candidate evidence for quality and refines it at the sentence level to reduce noise from inaccurate relevance estimates and spurious documents. The abstract reports 3.0% higher accuracy on average than the strongest competing baseline and 9.4% higher gold evidence recall, with substantially lower computational cost than existing multi-agent frameworks. It does not specify evaluation datasets or quantify the cost reduction. ### Source excerpt arXiv:2610.37701v1 Announce Type: cross Abstract: Large language models with retrieval-augmented generation systematically underutilize evidence in the middle of multi-document contexts, a phenomenon known as Lost in the Middle. Existing mitigation strategies, such as document reordering and attention steering, can reduce boundary bias but remain brittle in noisy settings, where inaccurate relevance estimation and spurious documents still obscure useful evidence in central positions. We propose DuQUAD, a multi-document QA framework that mitigates positional bias through dual-view reasoning and quality-aware evidence pruning. DuQUAD combines a Local Agent guided by a Structural Fusion Score with a Global Agent that reasons over the full document set, providing both targeted evidence recovery and broad contextual understanding. The framework further filters candidate evidence using explicit quality scoring and refines it at the sentence level to reduce noise. DuQUAD consistently outperforms strong Lost-in-the-Middle mitigation baselines, including hybrid retrieval and reranking, multi-agent reasoning, and context optimization methods, improving accuracy by 3.0% on average over the strongest competing baseline and gold evidence recall by 9.4%, while substantially reducing computational cost compared to existing multi-agent frameworks. [Next page]()