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
Trust Hinders Bidirectional Human-AI Alignment Critiques trust calibration in human-AI decision making for open-ended problems and outlines dimensions for higher-order system alignment. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Discusses human-AI decision-making philosophy rather than multi-step tool-using agents.
Trust Hinders Bidirectional Human-AI Alignment Critiques trust calibration in human-AI decision making for open-ended problems and outlines dimensions for higher-order system alignment. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Presents conceptual perspectives on trust rather than benchmark creation or validation.
Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback Trains chemical retrosynthesis models via reinforcement learning using forward reaction model consistency to credit plausible alternative pathways. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Covers chemical retrosynthesis modeling, not autonomous tool-using agents.
Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback Trains chemical retrosynthesis models via reinforcement learning using forward reaction model consistency to credit plausible alternative pathways. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on chemical synthesis accuracy, not deployment safety or alignment.
Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback Trains chemical retrosynthesis models via reinforcement learning using forward reaction model consistency to credit plausible alternative pathways. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes a domain-specific model rather than general LLM evaluation frameworks.
Harmony Harness: Multi-Objective Self-Evolution of Agent Harnesses Uses multi-objective evolutionary algorithms to optimize agent execution harnesses across competing metrics like accuracy and operating cost. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Does not address safety, guardrails, jailbreaks, or deployment reliability risks.
Harmony Harness: Multi-Objective Self-Evolution of Agent Harnesses Uses multi-objective evolutionary algorithms to optimize agent execution harnesses across competing metrics like accuracy and operating cost. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Focuses on optimizing agent harnesses rather than benchmarking methodology or dataset evaluation.
Can agents whose sole duty is to keep humans alive be hijacked to kill? Delayed multi-turn derived-action trap attacks in simulated high-stakes environments Tests delayed prompt injection attacks on life-support control agents in a simulated lunar habitat to see if safety duties persist. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Focuses on safety and injection attacks rather than evaluating benchmark methodologies.
UDHRBench: Benchmarking Human Rights in Large Language Models Evaluates commercial language models against Universal Declaration of Human Rights principles, revealing misalignments driven by efficiency arguments. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Focuses entirely on textual queries, not multimodal vision-language tasks.
UDHRBench: Benchmarking Human Rights in Large Language Models Evaluates commercial language models against Universal Declaration of Human Rights principles, revealing misalignments driven by efficiency arguments. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Evaluates model ethics, not inference optimization or model compression.
UDHRBench: Benchmarking Human Rights in Large Language Models Evaluates commercial language models against Universal Declaration of Human Rights principles, revealing misalignments driven by efficiency arguments. Preprint	LLMs for code via arxiv · 15 Oct, 06:17	Not relevant Assesses legal and ethical norms, not code synthesis or repair.
UDHRBench: Benchmarking Human Rights in Large Language Models Evaluates commercial language models against Universal Declaration of Human Rights principles, revealing misalignments driven by efficiency arguments. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on alignment benchmarking rather than autonomous agent planning.
UDHRBench: Benchmarking Human Rights in Large Language Models Evaluates commercial language models against Universal Declaration of Human Rights principles, revealing misalignments driven by efficiency arguments. Preprint	Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17	Not relevant Studies human rights alignment, not retrieval-augmented generation.
Continuous Risk-Based Scheduling for Food Safety Inspection: Integrating Historical Decay with Event-Driven Consumer Review Signals Combines historical violation decay with language-model triage of online consumer reviews to dynamically schedule restaurant health inspections. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Focuses on risk scoring algorithms rather than LLM evaluation benchmarks.
Continuous Risk-Based Scheduling for Food Safety Inspection: Integrating Historical Decay with Event-Driven Consumer Review Signals Combines historical violation decay with language-model triage of online consumer reviews to dynamically schedule restaurant health inspections. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Applies review classification to municipal inspections, not AI model safety.
Continuous Risk-Based Scheduling for Food Safety Inspection: Integrating Historical Decay with Event-Driven Consumer Review Signals Combines historical violation decay with language-model triage of online consumer reviews to dynamically schedule restaurant health inspections. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Proposes a scheduling pipeline using classifiers, not autonomous planning agents.
SocraticRepair: Planning-Guided Socratic Tutoring for Program Debugging and Misconception Repair Fine-tunes a coding model using synthetic tutoring dialogues to provide step-by-step guidance without revealing program fixes prematurely. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Addresses tutoring and debugging pedagogy, not deployment safety or attacks.
SocraticRepair: Planning-Guided Socratic Tutoring for Program Debugging and Misconception Repair Fine-tunes a coding model using synthetic tutoring dialogues to provide step-by-step guidance without revealing program fixes prematurely. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes an educational tutoring system rather than an evaluation benchmark.
SocraticRepair: Planning-Guided Socratic Tutoring for Program Debugging and Misconception Repair Fine-tunes a coding model using synthetic tutoring dialogues to provide step-by-step guidance without revealing program fixes prematurely. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on conversational tutoring dialogue rather than autonomous multi-step agents.
SocraticRepair: Planning-Guided Socratic Tutoring for Program Debugging and Misconception Repair Fine-tunes a coding model using synthetic tutoring dialogues to provide step-by-step guidance without revealing program fixes prematurely. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Deals with Socratic tutoring strategies, not model compression or efficiency.
Exploring Jailbreak Vulnerability Across Depth-Based Dynamic Inference in Large Language Models Studies how depth-based early-exit inference affects language model safety, finding shallow exits resist jailbreaks better than deep exits. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Analyzes vulnerability properties rather than presenting an LLM benchmark.
Exploring Jailbreak Vulnerability Across Depth-Based Dynamic Inference in Large Language Models Studies how depth-based early-exit inference affects language model safety, finding shallow exits resist jailbreaks better than deep exits. Preprint	Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17	Not relevant Focuses on inference depth and jailbreaks, not retrieval systems.
Exploring Jailbreak Vulnerability Across Depth-Based Dynamic Inference in Large Language Models Studies how depth-based early-exit inference affects language model safety, finding shallow exits resist jailbreaks better than deep exits. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Studies dynamic inference and jailbreak attacks, not autonomous agents.
Exploring Jailbreak Vulnerability Across Depth-Based Dynamic Inference in Large Language Models Studies how depth-based early-exit inference affects language model safety, finding shallow exits resist jailbreaks better than deep exits. Preprint	LLMs for code via arxiv · 15 Oct, 06:17	Not relevant Addresses model safety and inference depth, not code generation.
Exploring Jailbreak Vulnerability Across Depth-Based Dynamic Inference in Large Language Models Studies how depth-based early-exit inference affects language model safety, finding shallow exits resist jailbreaks better than deep exits. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Focuses on text language model inference, not multimodal architectures.
PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces Presents a multi-agent framework that decouples research planning, subquestion investigation, and factual verification via a shared semantic workspace. Preprint	LLMs for code via arxiv · 15 Oct, 06:17	Not relevant Addresses information search and research agents, not code generation or repair.
PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces Presents a multi-agent framework that decouples research planning, subquestion investigation, and factual verification via a shared semantic workspace. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Presents a deep-research multi-agent framework rather than an evaluation benchmark.
PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces Presents a multi-agent framework that decouples research planning, subquestion investigation, and factual verification via a shared semantic workspace. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Deals with research workflows, not small models or quantization techniques.
PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces Presents a multi-agent framework that decouples research planning, subquestion investigation, and factual verification via a shared semantic workspace. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on research coordination rather than deployment safety guardrails or jailbreaks.
PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces Presents a multi-agent framework that decouples research planning, subquestion investigation, and factual verification via a shared semantic workspace. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Deals with text search and reasoning corpora, not multimodal models.
Running Fewer Loops Removes Refusal Behaviour in Recurrent-Depth Language Models Shows that running fewer layer loops in recurrent-depth models significantly reduces safety refusals while preserving most reasoning performance. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Investigates internal refusal mechanisms rather than proposing an LLM benchmark.
Running Fewer Loops Removes Refusal Behaviour in Recurrent-Depth Language Models Shows that running fewer layer loops in recurrent-depth models significantly reduces safety refusals while preserving most reasoning performance. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Studies depth reduction in language models, not autonomous agent systems.
HarnessNet: Feedback-driven Hierarchical Joint Harness Evolution for Multi-agent Systems Proposes a hierarchical evolution framework that allows individual agent software harnesses in multi-agent systems to adapt using feedback. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on multi-agent performance optimization rather than deployment safety guardrails.
HarnessNet: Feedback-driven Hierarchical Joint Harness Evolution for Multi-agent Systems Proposes a hierarchical evolution framework that allows individual agent software harnesses in multi-agent systems to adapt using feedback. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes a multi-agent system optimization framework, not an evaluation benchmark.
Understanding Unified VLMs: From Attention Probing to Guidance Tuning for Multimodal Generation Probes attention heads in unified vision-language models to study condition processing and introduces an inference-time guidance tuning method. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Investigates internal attention mechanisms, not multi-step autonomous planning agents.
Understanding Unified VLMs: From Attention Probing to Guidance Tuning for Multimodal Generation Probes attention heads in unified vision-language models to study condition processing and introduces an inference-time guidance tuning method. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Presents an inference guidance technique rather than an evaluation benchmark.
Understanding Unified VLMs: From Attention Probing to Guidance Tuning for Multimodal Generation Probes attention heads in unified vision-language models to study condition processing and introduces an inference-time guidance tuning method. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Studies classifier-free guidance mechanisms, not deployment safety or alignment.
Words to World: Benchmarking LLMs in Interactive Physical Environments Introduces a closed-loop benchmark testing language models controlling physical robotic hardware and reacting to real sensor feedback. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on physical action capabilities rather than safety alignment or attacks.
Using Evidence-Centered Design to Guide Prompt Engineering in Measurement: An Illustration with Automated Essay Scoring Applies an educational measurement framework to prompt engineering to improve the validity and reproducibility of automated essay scoring systems. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Covers prompt engineering methodologies for scoring rather than LLM evaluation benchmarks.
Using Evidence-Centered Design to Guide Prompt Engineering in Measurement: An Illustration with Automated Essay Scoring Applies an educational measurement framework to prompt engineering to improve the validity and reproducibility of automated essay scoring systems. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on prompt engineering principles, not multi-step agent workflows.
Using Evidence-Centered Design to Guide Prompt Engineering in Measurement: An Illustration with Automated Essay Scoring Applies an educational measurement framework to prompt engineering to improve the validity and reproducibility of automated essay scoring systems. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Addresses educational prompt design validity, not safety guardrails or red-teaming.
MoMaTune: Model, Match, Then Tune Language Models for Numerical Records Presents a two-stage post-training strategy that improves language model handling of tabular numerical records using auxiliary pretraining and token matching. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on tabular pretraining objectives, not multi-step agent planning.
MoMaTune: Model, Match, Then Tune Language Models for Numerical Records Presents a two-stage post-training strategy that improves language model handling of tabular numerical records using auxiliary pretraining and token matching. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Addresses numerical representation accuracy rather than safety risks or alignment.
MoMaTune: Model, Match, Then Tune Language Models for Numerical Records Presents a two-stage post-training strategy that improves language model handling of tabular numerical records using auxiliary pretraining and token matching. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Presents a fine-tuning methodology rather than evaluation methodologies or benchmarks.
Framework for an AI-Ready ARPES Community Database Proposes a data-sharing ecosystem and framework to make angle-resolved photoemission spectroscopy data organized and accessible for AI research. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Proposes a scientific data infrastructure, not an AI agent system.
Framework for an AI-Ready ARPES Community Database Proposes a data-sharing ecosystem and framework to make angle-resolved photoemission spectroscopy data organized and accessible for AI research. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Discusses experimental data sharing rather than LLM benchmarking.
Framework for an AI-Ready ARPES Community Database Proposes a data-sharing ecosystem and framework to make angle-resolved photoemission spectroscopy data organized and accessible for AI research. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Covers spectroscopy data curation, not deployment safety or jailbreak defense.
ClinEdge: An Expert-Graded Benchmark of Real Patient Edge Cases Where Frontier Multimodal Models Fail Introduces a medical benchmark of real de-identified patient edge cases pairing imaging with clinical vignettes to test multimodal language models. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Evaluates medical diagnostic models rather than tool-using autonomous agents.
ClinEdge: An Expert-Graded Benchmark of Real Patient Edge Cases Where Frontier Multimodal Models Fail Introduces a medical benchmark of real de-identified patient edge cases pairing imaging with clinical vignettes to test multimodal language models. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on diagnostic benchmark failure rates rather than safety guardrails.
EvolvAttack: Backdooring self Evolving VLM Agents The paper introduces EvolvAttack, demonstrating a backdoor mechanism that activates only after a vision-language agent evolves its memory over time. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Studies backdoor attack vectors rather than general model evaluation methods.
EvolvAttack: Backdooring self Evolving VLM Agents The paper introduces EvolvAttack, demonstrating a backdoor mechanism that activates only after a vision-language agent evolves its memory over time. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Focuses on backdoor vulnerabilities rather than model compression or inference efficiency.
Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning The paper introduces Sequent-Prover, training language model agents to perform multi-turn formal reasoning using an external SMT solver scratchpad. Preprint	Multimodal models via arxiv · 15 Oct, 06:17	Not relevant Addresses formal language and text reasoning, not multimodal inputs.
Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning The paper introduces Sequent-Prover, training language model agents to perform multi-turn formal reasoning using an external SMT solver scratchpad. Preprint	Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17	Not relevant Focuses on SMT solver interactions rather than document retrieval or RAG.
Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning The paper introduces Sequent-Prover, training language model agents to perform multi-turn formal reasoning using an external SMT solver scratchpad. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Focuses on formal mathematical reasoning, not safety guardrails or alignment.
Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning The paper introduces Sequent-Prover, training language model agents to perform multi-turn formal reasoning using an external SMT solver scratchpad. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Introduces an agent framework and training method, not a general benchmark.
Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning The paper introduces Sequent-Prover, training language model agents to perform multi-turn formal reasoning using an external SMT solver scratchpad. Preprint	Small and efficient models via arxiv · 15 Oct, 06:17	Not relevant Focuses on agent fine-tuning and reasoning accuracy, not efficiency techniques.
RECAP: Calibrating Low-Bit Reasoning Models on the History They Read The paper presents RECAP, a post-training quantization method that calibrates reasoning models using dynamic prefixes to reduce error accumulation. Preprint	AI safety and reliability in deployment via arxiv · 15 Oct, 06:17	Not relevant Addresses quantization accuracy degradation rather than safety or jailbreak risks.
RECAP: Calibrating Low-Bit Reasoning Models on the History They Read The paper presents RECAP, a post-training quantization method that calibrates reasoning models using dynamic prefixes to reduce error accumulation. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Focuses on quantizing reasoning models, not agent planning or tools.
RECAP: Calibrating Low-Bit Reasoning Models on the History They Read The paper presents RECAP, a post-training quantization method that calibrates reasoning models using dynamic prefixes to reduce error accumulation. Preprint	LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17	Not relevant Proposes a post-training quantization method rather than benchmark evaluation.
LeHami: Symplectic Latent Dynamics for World Models The paper proposes LeHami, incorporating symplectic Hamiltonian dynamics into latent predictive architectures to prevent representation collapse over long rollouts. Preprint	AI agents via arxiv · 15 Oct, 06:17	Not relevant Studies dynamical systems in world models, not LLM-based agent planning.

12,241–12,300 of 17,313← Previous Next →