Every paper checked — nothing hidden

All papers

Every paper checked for brief #4 of 12–19 Oct 2026 — nothing is hidden. Brief papers first.

All17,313 In brief24 More (on-topic)3,333 Not relevant13,754 Not assessed87 Older352 Could not confirm93 Blogs with no paper266 All blogs284

Paper	Topic	Relevance
ECMR: Evidence-Aware Deliberation Control for Multi-Agent LLM Reasoning Presents ECMR, a dynamic evidence-aware controller that guides multi-agent deliberation steps to prevent false consensus and reduce redundant reasoning. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes a multi-agent control mechanism rather than an evaluation benchmark.
Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Studies persona activation vectors, not multi-step agent planning.
Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages. Preprint	LLMs for code via arxiv · 15 Oct, 06:17	Not relevant Does not address code generation or software engineering tasks.
Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Explores activation steering, not model compression or efficiency.
Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Focuses on activation vector steering rather than evaluation benchmark design.
Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages. Preprint	Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17	Not relevant Focuses on residual stream steering, not retrieval-augmented generation.
Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Focuses on text representations across languages, not vision or audio.
Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards Proposes CoSIR to train multi-agent collaboration policies using information-theoretic surprisal rewards that assess the helpfulness of communication. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes a multi-agent training framework, not an evaluation benchmark.
Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards Proposes CoSIR to train multi-agent collaboration policies using information-theoretic surprisal rewards that assess the helpfulness of communication. Preprint	LLMs for code via arxiv · 15 Oct, 06:17	Not relevant Focuses on multi-agent collaboration, not code generation or software engineering.
Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards Proposes CoSIR to train multi-agent collaboration policies using information-theoretic surprisal rewards that assess the helpfulness of communication. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Investigates text-based collaboration, not multimodal inputs.
Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards Proposes CoSIR to train multi-agent collaboration policies using information-theoretic surprisal rewards that assess the helpfulness of communication. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on task collaboration strategies rather than deployment safety or jailbreaks.
Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards Proposes CoSIR to train multi-agent collaboration policies using information-theoretic surprisal rewards that assess the helpfulness of communication. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Focuses on cooperative agent training, not inference acceleration or quantization.
Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards Proposes CoSIR to train multi-agent collaboration policies using information-theoretic surprisal rewards that assess the helpfulness of communication. Preprint	Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17	Not relevant Deals with inter-agent communication, not document database retrieval.
Interaction-as-a-Judge Presents Interaction-as-a-Judge, using computer-use agents to evaluate interactive web and desktop software applications through direct dynamic interaction. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Focuses on agentic evaluation methods, not model compression or efficiency.
Interaction-as-a-Judge Presents Interaction-as-a-Judge, using computer-use agents to evaluate interactive web and desktop software applications through direct dynamic interaction. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on preference evaluation of software rather than safety vulnerabilities.
Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Studies feature interpretability rather than model compression or efficient serving.
Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models. Preprint	LLMs for code via arxiv · 15 Oct, 06:17	Not relevant Does not address code generation or software engineering.
Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models. Preprint	Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17	Not relevant Does not involve external document retrieval or grounding.
Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Evaluates language model internal representations, not multimodal models.
Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on mechanistic interpretability of SAEs, not autonomous agents.
PaceReasoner: Label-Free Learning to Answer Briefly from Long Reasoning Introduces PaceReasoner, using long reasoning outputs during test-time reinforcement learning to boost the accuracy of brief, low-latency model answers. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Deals with reasoning token efficiency, not safety or guardrails.
PaceReasoner: Label-Free Learning to Answer Briefly from Long Reasoning Introduces PaceReasoner, using long reasoning outputs during test-time reinforcement learning to boost the accuracy of brief, low-latency model answers. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on reasoning distillation and length control, not tool use or agents.
PaceReasoner: Label-Free Learning to Answer Briefly from Long Reasoning Introduces PaceReasoner, using long reasoning outputs during test-time reinforcement learning to boost the accuracy of brief, low-latency model answers. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Introduces a test-time RL method, not a benchmark or evaluation protocol.
Visual Dependence Does Not Imply Answerability: A Systematic Analysis of Answerability Signals in Frozen InstructBLIP Analyzes answerability signals in frozen visual question answering models, finding standard confidence and alignment metrics fail to reliably detect unanswerable inputs. Preprint	Computer vision for industrial applications via arxiv · 15 Oct, 06:17	Not relevant Studies visual question answerability, not industrial defect detection.
Visual Dependence Does Not Imply Answerability: A Systematic Analysis of Answerability Signals in Frozen InstructBLIP Analyzes answerability signals in frozen visual question answering models, finding standard confidence and alignment metrics fail to reliably detect unanswerable inputs. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Analyzes question-answering behavior in VLMs, not agentic planning systems.
DriveWLA: Unifying Future Dynamics and Scene Semantics through World-Language-Action Modeling for Autonomous Driving Introduces DriveWLA, a mixture-of-transformers architecture unifying scene semantics, future dynamic prediction, and vehicle trajectory planning. Preprint	Computer vision for industrial applications via arxiv · 15 Oct, 06:17	Not relevant Applies to autonomous vehicles rather than manufacturing quality control or industrial inspection.
DriveWLA: Unifying Future Dynamics and Scene Semantics through World-Language-Action Modeling for Autonomous Driving Introduces DriveWLA, a mixture-of-transformers architecture unifying scene semantics, future dynamic prediction, and vehicle trajectory planning. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes an autonomous driving architecture, not an evaluation benchmark.
DriveWLA: Unifying Future Dynamics and Scene Semantics through World-Language-Action Modeling for Autonomous Driving Introduces DriveWLA, a mixture-of-transformers architecture unifying scene semantics, future dynamic prediction, and vehicle trajectory planning. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Deals with autonomous driving trajectory planning, not general LLM tool-use agents.
DriveWLA: Unifying Future Dynamics and Scene Semantics through World-Language-Action Modeling for Autonomous Driving Introduces DriveWLA, a mixture-of-transformers architecture unifying scene semantics, future dynamic prediction, and vehicle trajectory planning. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on trajectory prediction performance, not safety guardrails or alignment.
SpecCloak: Probing Harmful Code Generation through Engineering Specifications Finds that framing malicious coding prompts as engineering specifications systematically reduces language model refusal rates across multiple models. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Studies refusal mechanisms in code generation, not agent planning workflows.
SpecCloak: Probing Harmful Code Generation through Engineering Specifications Finds that framing malicious coding prompts as engineering specifications systematically reduces language model refusal rates across multiple models. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Focuses on safety probing mechanisms rather than benchmark design.
LinMorph: Learning to Morph Weights for Efficient Video DiT Linearization Proposes LinMorph, a learnable weight transformation method to initialize linear attention models from pretrained video diffusion transformers efficiently. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Addresses model acceleration, not safety, hallucination, or red teaming.
LinMorph: Learning to Morph Weights for Efficient Video DiT Linearization Proposes LinMorph, a learnable weight transformation method to initialize linear attention models from pretrained video diffusion transformers efficiently. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Introduces a weight morphing method, not a general evaluation framework.
LinMorph: Learning to Morph Weights for Efficient Video DiT Linearization Proposes LinMorph, a learnable weight transformation method to initialize linear attention models from pretrained video diffusion transformers efficiently. Preprint	Computer vision for industrial applications via arxiv · 15 Oct, 06:17	Not relevant Focuses on generative video transformer architecture, not industrial inspection.
LinMorph: Learning to Morph Weights for Efficient Video DiT Linearization Proposes LinMorph, a learnable weight transformation method to initialize linear attention models from pretrained video diffusion transformers efficiently. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Covers diffusion transformer linearization, not tool-using or planning agents.
CineMem: Attention-Aligned Fine-grained Memory Retrieval for Multi-Shot Long Video Generation Introduces CineMem, an attention-aligned memory retrieval framework using clustered KV caches for autoregressive multi-shot long video generation. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on video generation cache retrieval, not autonomous planning agents.
CineMem: Attention-Aligned Fine-grained Memory Retrieval for Multi-Shot Long Video Generation Introduces CineMem, an attention-aligned memory retrieval framework using clustered KV caches for autoregressive multi-shot long video generation. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on video synthesis efficiency, not AI safety or deployment guardrails.
CineMem: Attention-Aligned Fine-grained Memory Retrieval for Multi-Shot Long Video Generation Introduces CineMem, an attention-aligned memory retrieval framework using clustered KV caches for autoregressive multi-shot long video generation. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes a generation architecture, not an evaluation benchmark or methodology.
CineMem: Attention-Aligned Fine-grained Memory Retrieval for Multi-Shot Long Video Generation Introduces CineMem, an attention-aligned memory retrieval framework using clustered KV caches for autoregressive multi-shot long video generation. Preprint	Computer vision for industrial applications via arxiv · 15 Oct, 06:17	Not relevant Deals with generative video synthesis, not industrial inspection or defect detection.
GLST: Global-Local Spatio-Temporal Memory Decomposition Framework for Online Chained-Goal Navigation Presents a memory decomposition framework separating spatio-temporal beliefs to support autonomous robot navigation under evolving online goals. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Focuses on robot navigation frameworks, not LLM evaluation benchmarks.
GLST: Global-Local Spatio-Temporal Memory Decomposition Framework for Online Chained-Goal Navigation Presents a memory decomposition framework separating spatio-temporal beliefs to support autonomous robot navigation under evolving online goals. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Deals with robotic navigation beliefs, not AI safety or guardrails.
GLST: Global-Local Spatio-Temporal Memory Decomposition Framework for Online Chained-Goal Navigation Presents a memory decomposition framework separating spatio-temporal beliefs to support autonomous robot navigation under evolving online goals. Preprint	Computer vision for industrial applications via arxiv · 15 Oct, 06:17	Not relevant Concerns quadruped robot navigation, not industrial defect inspection.
GLST: Global-Local Spatio-Temporal Memory Decomposition Framework for Online Chained-Goal Navigation Presents a memory decomposition framework separating spatio-temporal beliefs to support autonomous robot navigation under evolving online goals. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on robotic spatial navigation, not LLM planning or tool-use agents.
Learning How to Remember: Memory Modality Selection for GUI Agents Introduces SwitchMem, a routing system that dynamically selects between text and visual memory representations for vision-language GUI agents. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on agent memory routing, not deployment safety or guardrails.
Learning How to Remember: Memory Modality Selection for GUI Agents Introduces SwitchMem, a routing system that dynamically selects between text and visual memory representations for vision-language GUI agents. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes an agent memory method, not general evaluation benchmarks.
MERIT: Multimodal Evidence Reasoning over Image Tables Proposes MERIT, a framework using visual table reasoning and adapted encoders to extract information directly from table images. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on visual table understanding, not multi-step agent planning.
MERIT: Multimodal Evidence Reasoning over Image Tables Proposes MERIT, a framework using visual table reasoning and adapted encoders to extract information directly from table images. Preprint	Computer vision for industrial applications via arxiv · 15 Oct, 06:17	Not relevant Focuses on table images, not industrial inspection or defect detection.
MERIT: Multimodal Evidence Reasoning over Image Tables Proposes MERIT, a framework using visual table reasoning and adapted encoders to extract information directly from table images. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes a training framework for table reasoning, not an evaluation benchmark.
MERIT: Multimodal Evidence Reasoning over Image Tables Proposes MERIT, a framework using visual table reasoning and adapted encoders to extract information directly from table images. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on visual table understanding, not deployment safety or red-teaming.
AI-Enhanced Hyperspectral Imaging for Intraoperative Detection of Low-Grade Glioma: A Clinical Study The clinical study compares hyperspectral and RGB imaging for intraoperative detection of low-grade gliomas during brain tumor surgery. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Focuses on clinical glioma classification, not LLM benchmarks.
AI-Enhanced Hyperspectral Imaging for Intraoperative Detection of Low-Grade Glioma: A Clinical Study The clinical study compares hyperspectral and RGB imaging for intraoperative detection of low-grade gliomas during brain tumor surgery. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Addresses surgical tumor classification, not autonomous language agents.
AI-Enhanced Hyperspectral Imaging for Intraoperative Detection of Low-Grade Glioma: A Clinical Study The clinical study compares hyperspectral and RGB imaging for intraoperative detection of low-grade gliomas during brain tumor surgery. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Focuses on hyperspectral medical imaging, not multimodal language models.
AI-Enhanced Hyperspectral Imaging for Intraoperative Detection of Low-Grade Glioma: A Clinical Study The clinical study compares hyperspectral and RGB imaging for intraoperative detection of low-grade gliomas during brain tumor surgery. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Pertains to surgical image classification, not AI deployment safety.
AI-Enhanced Hyperspectral Imaging for Intraoperative Detection of Low-Grade Glioma: A Clinical Study The clinical study compares hyperspectral and RGB imaging for intraoperative detection of low-grade gliomas during brain tumor surgery. Preprint	Computer vision for industrial applications via arxiv · 15 Oct, 06:17	Not relevant Deals with clinical brain surgery imaging, not industrial manufacturing.
HaloBlock: Faster and Better Decoding for Diffusion Large Language Models HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding. Preprint	LLMs for code via arxiv · 15 Oct, 06:17	Not relevant Evaluates on reasoning/code but core contribution is general diffusion LLM decoding.
HaloBlock: Faster and Better Decoding for Diffusion Large Language Models HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on decoding algorithms for diffusion LLMs, not planning agents.
HaloBlock: Faster and Better Decoding for Diffusion Large Language Models HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Focuses on text diffusion LLMs, not multimodal architectures.
HaloBlock: Faster and Better Decoding for Diffusion Large Language Models HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding. Preprint	Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17	Not relevant Concerns diffusion language model decoding, not document retrieval.
HaloBlock: Faster and Better Decoding for Diffusion Large Language Models HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes an inference decoding mechanism, not an evaluation framework.
HaloBlock: Faster and Better Decoding for Diffusion Large Language Models HaloBlock allocates adaptive context halo regions across blocks to smooth context transitions and improve diffusion language model decoding. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on decoding acceleration, not deployment safety or guardrails.

10,201–10,260 of 13,754← Previous Next →