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
Same Outcome, Different Evidence: Intent Recovery in LLM Safety Evaluation Pairs attack success rate with operative understanding rate to assess whether models recognize harmful intent during safety evaluations. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Examines intent recovery in safety evaluation without focusing on coding assistants.
Same Outcome, Different Evidence: Intent Recovery in LLM Safety Evaluation Pairs attack success rate with operative understanding rate to assess whether models recognize harmful intent during safety evaluations. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Analyzes safety evaluation metrics rather than model compression or inference efficiency.
Same Outcome, Different Evidence: Intent Recovery in LLM Safety Evaluation Pairs attack success rate with operative understanding rate to assess whether models recognize harmful intent during safety evaluations. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Addresses text-based intent recovery rather than multimodal vision-language architectures.
Narrow and Deep: An Ontology Tower as the Knowledge of an LLM Agent for an Industrial Equipment System Proposes a narrow-and-deep ontology tower to structure equipment knowledge for language model agents operating industrial air-handling systems. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Applies models to equipment tasks rather than developing general LLM evaluation benchmarks.
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing Proposes an interpretability framework using sparse autoencoders to explain mixture of experts routing based on superposed fine-grained features. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Analyzes internal MoE expert routing rather than external retrieval systems.
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing Proposes an interpretability framework using sparse autoencoders to explain mixture of experts routing based on superposed fine-grained features. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Focuses on foundational MoE routing interpretability rather than code LLMs or software tools.
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing Proposes an interpretability framework using sparse autoencoders to explain mixture of experts routing based on superposed fine-grained features. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Provides a model interpretability technique rather than an evaluation benchmark.
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing Proposes an interpretability framework using sparse autoencoders to explain mixture of experts routing based on superposed fine-grained features. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Interprets MoE routing rather than measuring safety, robustness, or deployment guardrails.
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing Proposes an interpretability framework using sparse autoencoders to explain mixture of experts routing based on superposed fine-grained features. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Studies sparse MoE routing mechanisms in language models, not multimodal architectures.
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing Proposes an interpretability framework using sparse autoencoders to explain mixture of experts routing based on superposed fine-grained features. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Presents an MoE routing interpretability method, not multi-step tool-using agents.
AuraLuxMuse: Adaptive Fusion Modeling for Aesthetic Stage Lighting Design with Music and Expert Guidance Presents a system combining representation learning and music alignment to generate editable aesthetic stage lighting cues for live performances. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents a lighting synthesis system rather than general LLM evaluation benchmarks.
AuraLuxMuse: Adaptive Fusion Modeling for Aesthetic Stage Lighting Design with Music and Expert Guidance Presents a system combining representation learning and music alignment to generate editable aesthetic stage lighting cues for live performances. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Addresses aesthetic lighting design rather than AI safety, guardrails, or model alignment.
AuraLuxMuse: Adaptive Fusion Modeling for Aesthetic Stage Lighting Design with Music and Expert Guidance Presents a system combining representation learning and music alignment to generate editable aesthetic stage lighting cues for live performances. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on stage lighting design generation rather than autonomous LLM agents.
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks Proposes an external memory framework allowing frozen language model agents to learn and iteratively refine natural-language world models compiled into executable code. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Focuses on agentic world modeling rather than model compression or efficient serving.
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks Proposes an external memory framework allowing frozen language model agents to learn and iteratively refine natural-language world models compiled into executable code. Preprint	Computer vision for industrial applications via arxiv · 12 Oct, 16:28	Not relevant Evaluates simulated agent dynamics rather than industrial computer vision inspection.
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks Proposes an external memory framework allowing frozen language model agents to learn and iteratively refine natural-language world models compiled into executable code. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Uses an internal rulebook memory rather than standard retrieval-augmented generation systems.
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks Proposes an external memory framework allowing frozen language model agents to learn and iteratively refine natural-language world models compiled into executable code. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Compiles rules to code for internal agent simulation rather than assisting software developers.
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks Proposes an external memory framework allowing frozen language model agents to learn and iteratively refine natural-language world models compiled into executable code. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Introduces an agent self-improvement architecture rather than a benchmarking framework.
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks Proposes an external memory framework allowing frozen language model agents to learn and iteratively refine natural-language world models compiled into executable code. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies self-improving world models rather than safety, guardrails, or alignment in deployment.
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks Proposes an external memory framework allowing frozen language model agents to learn and iteratively refine natural-language world models compiled into executable code. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Focuses on agent reasoning and state exploration rather than multimodal model design.
From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction Evaluates eight lightweight open-source vision-language models on optical character recognition and schema-compliant JSON extraction across heritage document collections. Peer-reviewed	Computer vision for industrial applications via arxiv · 12 Oct, 16:28	Not relevant Evaluates heritage document understanding rather than industrial factory vision or inspection.
From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction Evaluates eight lightweight open-source vision-language models on optical character recognition and schema-compliant JSON extraction across heritage document collections. Peer-reviewed	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on extraction accuracy of heritage documents rather than safety or guardrails.
From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction Evaluates eight lightweight open-source vision-language models on optical character recognition and schema-compliant JSON extraction across heritage document collections. Peer-reviewed	AI agents via arxiv · 12 Oct, 16:28	Not relevant Studies document OCR and JSON extraction, not multi-step autonomous planning agents.
From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction Evaluates eight lightweight open-source vision-language models on optical character recognition and schema-compliant JSON extraction across heritage document collections. Peer-reviewed	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Comparative study on document OCR without introducing a general LLM benchmark methodology.
From Suppression to Repair: Mitigating Object Hallucination in Large Vision-Language Models via Localized Distribution Alignment Introduces a training-free inference method using Gaussian optimal transport to reduce object hallucination in vision-language models without suppressing valid representations. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Proposes a hallucination repair algorithm rather than a new evaluation suite or benchmark.
From Suppression to Repair: Mitigating Object Hallucination in Large Vision-Language Models via Localized Distribution Alignment Introduces a training-free inference method using Gaussian optimal transport to reduce object hallucination in vision-language models without suppressing valid representations. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Investigates internal representation repairs in vision-language models rather than autonomous agents.
From Suppression to Repair: Mitigating Object Hallucination in Large Vision-Language Models via Localized Distribution Alignment Introduces a training-free inference method using Gaussian optimal transport to reduce object hallucination in vision-language models without suppressing valid representations. Preprint	Computer vision for industrial applications via arxiv · 12 Oct, 16:28	Not relevant Addresses general object hallucination in VLMs, not industrial inspection or defect detection.
Probability-Signature Dynamics: Unpacking Modular Addition Learning Within Two-Layer Networks Analyzes modular addition learning in two-layer neural networks using probability signatures to explain how gradient training selects Fourier-structured representations. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Examines theoretical learning dynamics in two-layer networks, not LLM agents.
Probability-Signature Dynamics: Unpacking Modular Addition Learning Within Two-Layer Networks Analyzes modular addition learning in two-layer neural networks using probability signatures to explain how gradient training selects Fourier-structured representations. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies learning mechanics of toy networks, not deployment safety, robustness, or alignment.
Probability-Signature Dynamics: Unpacking Modular Addition Learning Within Two-Layer Networks Analyzes modular addition learning in two-layer neural networks using probability signatures to explain how gradient training selects Fourier-structured representations. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Theoretical analysis of modular arithmetic training rather than LLM benchmarking.
GRPODropout: Less is More for Online Reinforcement Learning Rollouts Presents a rollout selection strategy that drops high-probability positive-advantage trajectories to prevent entropy collapse during language model reinforcement learning. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Evaluates language reasoning models rather than multimodal or vision-language models.
GRPODropout: Less is More for Online Reinforcement Learning Rollouts Presents a rollout selection strategy that drops high-probability positive-advantage trajectories to prevent entropy collapse during language model reinforcement learning. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Pertains to post-training RL optimization, not retrieval-augmented generation systems.
GRPODropout: Less is More for Online Reinforcement Learning Rollouts Presents a rollout selection strategy that drops high-probability positive-advantage trajectories to prevent entropy collapse during language model reinforcement learning. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Proposes an algorithmic training intervention rather than an evaluation framework or benchmark.
GRPODropout: Less is More for Online Reinforcement Learning Rollouts Presents a rollout selection strategy that drops high-probability positive-advantage trajectories to prevent entropy collapse during language model reinforcement learning. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Addresses sampling diversity in RL updates, not model quantization, pruning, or small models.
GRPODropout: Less is More for Online Reinforcement Learning Rollouts Presents a rollout selection strategy that drops high-probability positive-advantage trajectories to prevent entropy collapse during language model reinforcement learning. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Presents a general RL algorithm without a primary focus on code development or SE tasks.
GRPODropout: Less is More for Online Reinforcement Learning Rollouts Presents a rollout selection strategy that drops high-probability positive-advantage trajectories to prevent entropy collapse during language model reinforcement learning. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies optimization stability in RL rollouts rather than deployment safety or guardrails.
GRPODropout: Less is More for Online Reinforcement Learning Rollouts Presents a rollout selection strategy that drops high-probability positive-advantage trajectories to prevent entropy collapse during language model reinforcement learning. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on core reinforcement learning algorithms for reasoning, not multi-step agent architectures.
Trajectory-Guided Fault Localization for Agent Skill Evolution Proposes a method linking execution traces to specific skill segments to guide automated revision of software engineering agent skills. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents a skill refinement method rather than a general LLM evaluation benchmark study.
Trajectory-Guided Fault Localization for Agent Skill Evolution Proposes a method linking execution traces to specific skill segments to guide automated revision of software engineering agent skills. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Deals with execution accuracy in coding skills rather than safety, red teaming, or jailbreaks.
LEVER: Adaptive Cost-Aware Proof Search Over AND/OR Graphs Introduces an AND/OR graph proof search algorithm optimizing customizable objectives like length and compute cost during automated formal theorem proving. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Investigates cost-aware theorem proving rather than AI safety, guardrails, or alignment risks.
LEVER: Adaptive Cost-Aware Proof Search Over AND/OR Graphs Introduces an AND/OR graph proof search algorithm optimizing customizable objectives like length and compute cost during automated formal theorem proving. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Applies a theorem proving algorithm on existing benchmarks rather than proposing evaluation methods.
How Is Automated Research Evaluated? A Survey of Benchmarks and Evaluation Practices Surveys benchmarks, metrics, and experimental practices used to evaluate automated research systems across ideation, coding, writing, and peer review. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on research benchmark design rather than deployment safety, security, or alignment.
Evaluating Exact Output and Checkpoint-State Prediction in Real Programs Evaluates language models on predicting program outputs and intermediate loop checkpoint states directly from source code and inputs. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Measures pure code-execution prediction without tools or multi-step agent actions.
Evaluating Exact Output and Checkpoint-State Prediction in Real Programs Evaluates language models on predicting program outputs and intermediate loop checkpoint states directly from source code and inputs. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Evaluates code tracing accuracy rather than safety risks, jailbreaks, or guardrails.
Forms of LLM-Integrated Applications from LLM-Chats to Autonomous AI Agent System Surveys and taxonomizes recurring architectural patterns across language-model-integrated software applications, ranging from chatbots to autonomous agent systems. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Focuses on architectural patterns across systems, not model compression, distillation, or quantization.
Forms of LLM-Integrated Applications from LLM-Chats to Autonomous AI Agent System Surveys and taxonomizes recurring architectural patterns across language-model-integrated software applications, ranging from chatbots to autonomous agent systems. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Surveys system integration patterns rather than safety guardrails, attacks, or alignment failures.
Forms of LLM-Integrated Applications from LLM-Chats to Autonomous AI Agent System Surveys and taxonomizes recurring architectural patterns across language-model-integrated software applications, ranging from chatbots to autonomous agent systems. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Analyzes text and tool architectures rather than multimodal vision-language representations.
Forms of LLM-Integrated Applications from LLM-Chats to Autonomous AI Agent System Surveys and taxonomizes recurring architectural patterns across language-model-integrated software applications, ranging from chatbots to autonomous agent systems. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Taxonomy survey of application architectures, not a benchmark or evaluation methodology paper.
Beyond Visual Enhancement: Adaptive Multi-Context Steering to Mitigate LVLM Hallucinations Proposes an adaptive multi-source steering decoding method to mitigate hallucinations in vision-language models by balancing visual and textual context contributions. Preprint	Computer vision for industrial applications via arxiv · 12 Oct, 16:28	Not relevant Addresses general LVLM hallucinations without focusing on industrial inspection or factory vision systems.
Beyond Visual Enhancement: Adaptive Multi-Context Steering to Mitigate LVLM Hallucinations Proposes an adaptive multi-source steering decoding method to mitigate hallucinations in vision-language models by balancing visual and textual context contributions. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Proposes an inference-time steering method rather than introducing LLM evaluation benchmarks or protocols.
Beyond Visual Enhancement: Adaptive Multi-Context Steering to Mitigate LVLM Hallucinations Proposes an adaptive multi-source steering decoding method to mitigate hallucinations in vision-language models by balancing visual and textual context contributions. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on multimodal decoding steering rather than agent planning, tool execution, or agentic workflows.
From Surface to Depth: Towards Cognitive Appraisal Reasoning in Multimodal Emotion Understanding Introduces a dataset and benchmark to model multimodal emotion understanding through cognitive appraisal reasoning rather than superficial affective cues. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Proposes an emotion-understanding model and dataset rather than multi-step autonomous planning agents.
From Surface to Depth: Towards Cognitive Appraisal Reasoning in Multimodal Emotion Understanding Introduces a dataset and benchmark to model multimodal emotion understanding through cognitive appraisal reasoning rather than superficial affective cues. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Deals with emotion understanding rather than deployment safety, jailbreaking, or red-teaming.
Neural Decoding as Cognitive Inference Formulates neural decoding as Bayesian cognitive inference to extract stable meta-neural semantic representations across varying neural recording modalities. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Examines cognitive inference in brain data, unrelated to AI safety, guardrails, or model hallucinations.
Neural Decoding as Cognitive Inference Formulates neural decoding as Bayesian cognitive inference to extract stable meta-neural semantic representations across varying neural recording modalities. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on computational neuroscience and biological neural decoding, not LLM agents or tool use.
Neural Decoding as Cognitive Inference Formulates neural decoding as Bayesian cognitive inference to extract stable meta-neural semantic representations across varying neural recording modalities. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents biological neural decoding experiments, not benchmarks or evaluations for large language models.
EliteTable: Self-Improving Program Search for Executable Control Policies Presents a program-search method using a curated elite solution table to guide language models in synthesizing executable control policies. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Evaluates a search algorithm on standard control tasks rather than contributing new LLM benchmark methodology.
EliteTable: Self-Improving Program Search for Executable Control Policies Presents a program-search method using a curated elite solution table to guide language models in synthesizing executable control policies. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies program search performance in control tasks, not AI safety, alignment, or deployment reliability.
Stochastic Grouping Conformal Prediction for Effective Subgroup Reliability Proposes a conformal prediction framework that learns stochastic groupings to improve subgroup coverage reliability without inflating overall prediction sets. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Does not address model compression, small language models, or efficient serving.
Stochastic Grouping Conformal Prediction for Effective Subgroup Reliability Proposes a conformal prediction framework that learns stochastic groupings to improve subgroup coverage reliability without inflating overall prediction sets. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Introduces a statistical uncertainty quantification method, not an LLM benchmark or evaluation protocol.

12,901–12,960 of 13,754← Previous Next →