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
LoopAhead: Speculating Ahead for Faster Decoding in Looped Language Models The paper introduces a speculative decoding framework that uses early recurrent states to draft tokens in looped transformers. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Focuses on latency reduction, not deployment safety or robustness.
Decoding What iOS Location Services Send to Apple: An LLM Is an Unreliable Analyst but a Useful Toolmaker The study evaluates using language models to directly reconstruct mobile network schemas versus generating validation code tools. Preprint	AI agents via arxiv · 17 Oct, 06:17	Not relevant Evaluates code models for reverse engineering, not autonomous agents.
Decoding What iOS Location Services Send to Apple: An LLM Is an Unreliable Analyst but a Useful Toolmaker The study evaluates using language models to directly reconstruct mobile network schemas versus generating validation code tools. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Studies protocol reverse-engineering accuracy, not general LLM benchmarks.
Evaluating Inquiry Learning in LLM-Based Agents: Hypothesizing, Reasoning, and Exploration The authors present an interactive environment to evaluate how language models integrate exploration, hypothesis generation, and reasoning. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Evaluates reasoning capabilities rather than deployment safety or guardrails.
Learning Intrinsic Dynamics in Latent World Models The paper studies whether predicting actions from consecutive latent states helps world models retain essential environmental dynamics. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Addresses representation learning in RL, not safety or guardrails.
Learning Intrinsic Dynamics in Latent World Models The paper studies whether predicting actions from consecutive latent states helps world models retain essential environmental dynamics. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Studies latent representation dynamics, not LLM benchmarking frameworks.
Learning Intrinsic Dynamics in Latent World Models The paper studies whether predicting actions from consecutive latent states helps world models retain essential environmental dynamics. Preprint	AI agents via arxiv · 17 Oct, 06:17	Not relevant Focuses on world model dynamics, not interactive LLM agents.
Qwen-Audio-Jev: A Unified Interface for Audio Decisions The authors introduce an audio model with a non-autoregressive decision head for selection, judgment, and scoring tasks. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Focuses on an audio model architecture, not benchmark methodology.
Qwen-Audio-Jev: A Unified Interface for Audio Decisions The authors introduce an audio model with a non-autoregressive decision head for selection, judgment, and scoring tasks. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Focuses on latency and output formats, not deployment safety.
Qwen-Audio-Jev: A Unified Interface for Audio Decisions The authors introduce an audio model with a non-autoregressive decision head for selection, judgment, and scoring tasks. Preprint	AI agents via arxiv · 17 Oct, 06:17	Not relevant Introduces an audio decision head, not an interactive agent.
Efficient Pairwise Difference Learning: A Comparative Study of Encoding Strategies and Training Pair Selection The paper evaluates feature encoders and pair selection strategies for pairwise difference learning in low-data classification regimes. Peer-reviewed	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Addresses low-data classification, not AI deployment safety or guardrails.
Efficient Pairwise Difference Learning: A Comparative Study of Encoding Strategies and Training Pair Selection The paper evaluates feature encoders and pair selection strategies for pairwise difference learning in low-data classification regimes. Peer-reviewed	AI agents via arxiv · 17 Oct, 06:17	Not relevant Focuses on tabular classification, not LLM agents or tool use.
Efficient Pairwise Difference Learning: A Comparative Study of Encoding Strategies and Training Pair Selection The paper evaluates feature encoders and pair selection strategies for pairwise difference learning in low-data classification regimes. Peer-reviewed	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Studies pairwise classical classifiers, not small or efficient LLMs.
Efficient Pairwise Difference Learning: A Comparative Study of Encoding Strategies and Training Pair Selection The paper evaluates feature encoders and pair selection strategies for pairwise difference learning in low-data classification regimes. Peer-reviewed	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Studies classical classification encoders, not LLM evaluation benchmarks.
The Self-Improvement Problem The paper formalizes the theoretical limits of automated gates evaluating self-improvement updates in autonomous agent code. Preprint	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Studies safe update acceptance, not model compression or inference.
The Self-Improvement Problem The paper formalizes the theoretical limits of automated gates evaluating self-improvement updates in autonomous agent code. Preprint	LLMs for code via arxiv · 17 Oct, 06:17	Not relevant Addresses theoretical self-improvement limits, not software development tools.
The Self-Improvement Problem The paper formalizes the theoretical limits of automated gates evaluating self-improvement updates in autonomous agent code. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Focuses on theoretical computability rather than empirical benchmarking.
The Self-Improvement Problem The paper formalizes the theoretical limits of automated gates evaluating self-improvement updates in autonomous agent code. Preprint	Retrieval-augmented generation (RAG) via arxiv · 17 Oct, 06:17	Not relevant Examines theoretical program modification, not retrieval-augmented generation.
The Self-Improvement Problem The paper formalizes the theoretical limits of automated gates evaluating self-improvement updates in autonomous agent code. Preprint	Multimodal models via arxiv · 17 Oct, 06:17	Not relevant Deals with theoretical agent self-modification, not multimodal models.
Data World: Scaling Data Environments from Scratch The authors present an engine to synthesize environments, tasks, and rubrics for training and evaluating data analysis agents. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Focuses on benchmark creation rather than deployment safety and guardrails.
Router Drift Control: Mitigating Off-Policy Degradation in Mixture-of-Experts Reinforcement Learning The paper proposes filtering high-drift responses to mitigate off-policy degradation in mixture-of-experts reinforcement learning. Preprint	AI agents via arxiv · 17 Oct, 06:17	Not relevant Addresses mixture-of-experts training stability, not interactive AI agents.
Router Drift Control: Mitigating Off-Policy Degradation in Mixture-of-Experts Reinforcement Learning The paper proposes filtering high-drift responses to mitigate off-policy degradation in mixture-of-experts reinforcement learning. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Focuses on RL optimization dynamics rather than safety or guardrails.
Router Drift Control: Mitigating Off-Policy Degradation in Mixture-of-Experts Reinforcement Learning The paper proposes filtering high-drift responses to mitigate off-policy degradation in mixture-of-experts reinforcement learning. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Proposes an RL training algorithm rather than LLM benchmarking.
Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures The paper uses human interventions to localize and steer computations behind agent failure to decide between acting and clarifying. Preprint	Multimodal models via arxiv · 17 Oct, 06:17	Not relevant Does not involve multimodal reasoning or vision-language data.
Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures The paper uses human interventions to localize and steer computations behind agent failure to decide between acting and clarifying. Preprint	LLMs for code via arxiv · 17 Oct, 06:17	Not relevant Examines agent ask-vs-act decisions, not code generation or repair.
Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures The paper uses human interventions to localize and steer computations behind agent failure to decide between acting and clarifying. Preprint	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Investigates internal state editing, not model compression or efficiency.
Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures The paper uses human interventions to localize and steer computations behind agent failure to decide between acting and clarifying. Preprint	Retrieval-augmented generation (RAG) via arxiv · 17 Oct, 06:17	Not relevant Focuses on internal agent activation steering, not retrieval systems.
Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures The paper uses human interventions to localize and steer computations behind agent failure to decide between acting and clarifying. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Focuses on mechanistic interpretability rather than benchmark development.
Encoded but Exposed: Language Embeddings can Reveal More Than They Should in Recommender Systems The authors evaluate how easily item identities and user profiles can be recovered from language embeddings in recommender systems. Peer-reviewed	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Studies privacy leakage rather than general LLM evaluation benchmarks.
Encoded but Exposed: Language Embeddings can Reveal More Than They Should in Recommender Systems The authors evaluate how easily item identities and user profiles can be recovered from language embeddings in recommender systems. Peer-reviewed	AI agents via arxiv · 17 Oct, 06:17	Not relevant Addresses recommender embeddings, not LLM agents or tool use.
FIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning The authors introduce a Fisher-informed adapter merging method to maintain retained task performance during federated language model unlearning. Preprint	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Focuses on knowledge removal rather than model compression or efficiency.
FIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning The authors introduce a Fisher-informed adapter merging method to maintain retained task performance during federated language model unlearning. Preprint	AI agents via arxiv · 17 Oct, 06:17	Not relevant Focuses on federated parameter merging, not agentic planning.
FIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning The authors introduce a Fisher-informed adapter merging method to maintain retained task performance during federated language model unlearning. Preprint	Multimodal models via arxiv · 17 Oct, 06:17	Not relevant Deals with text unlearning rather than multimodal models.
FIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning The authors introduce a Fisher-informed adapter merging method to maintain retained task performance during federated language model unlearning. Preprint	LLMs for code via arxiv · 17 Oct, 06:17	Not relevant Studies federated unlearning, not code generation or software engineering.
FIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning The authors introduce a Fisher-informed adapter merging method to maintain retained task performance during federated language model unlearning. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Proposes an unlearning algorithm rather than general benchmark evaluation.
FIRM: Fisher-Informed Retention Merging for Federated LLM Unlearning The authors introduce a Fisher-informed adapter merging method to maintain retained task performance during federated language model unlearning. Preprint	Retrieval-augmented generation (RAG) via arxiv · 17 Oct, 06:17	Not relevant Addresses model weight unlearning rather than retrieval-augmented generation.
Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence The paper analyzes why search-augmented language models produce outdated answers even when relevant evidence is retrieved. Preprint	Multimodal models via arxiv · 17 Oct, 06:17	Not relevant Focuses entirely on text-based web search and question answering.
Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence The paper analyzes why search-augmented language models produce outdated answers even when relevant evidence is retrieved. Preprint	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Studies evidence utilization rather than model compression or inference latency.
Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence The paper analyzes why search-augmented language models produce outdated answers even when relevant evidence is retrieved. Preprint	AI agents via arxiv · 17 Oct, 06:17	Not relevant Evaluates evidence utilization in retrieval, not agent planning or tool use.
Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence The paper analyzes why search-augmented language models produce outdated answers even when relevant evidence is retrieved. Preprint	LLMs for code via arxiv · 17 Oct, 06:17	Not relevant Focuses on web search QA, not software engineering.
In-Context Preference-Conditioned Pareto Set Learning from Limited Offline Data Introduces an in-context learning approach to map specified user preferences directly to Pareto-optimal solutions using offline data. Preprint	AI agents via arxiv · 17 Oct, 06:17	Not relevant Focuses on Pareto set multi-objective optimization, not LLM agents.
In-Context Preference-Conditioned Pareto Set Learning from Limited Offline Data Introduces an in-context learning approach to map specified user preferences directly to Pareto-optimal solutions using offline data. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Deals with hypernetwork optimization methods, not LLM evaluation frameworks.
In-Context Preference-Conditioned Pareto Set Learning from Limited Offline Data Introduces an in-context learning approach to map specified user preferences directly to Pareto-optimal solutions using offline data. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Focuses on offline multi-objective optimization, not deployment safety.
Reliable Knowledge Acquisition for Continual Learning Agents Proposes a structured knowledge admission and verification system to prevent continual learning agents from storing harmful past experience entries. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Introduces a knowledge acquisition system rather than LLM evaluation methodology.
Reliable Knowledge Acquisition for Continual Learning Agents Proposes a structured knowledge admission and verification system to prevent continual learning agents from storing harmful past experience entries. Preprint	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Focuses on agent context management rather than model efficiency.
Reliable Knowledge Acquisition for Continual Learning Agents Proposes a structured knowledge admission and verification system to prevent continual learning agents from storing harmful past experience entries. Preprint	LLMs for code via arxiv · 17 Oct, 06:17	Not relevant Evaluates general task streams rather than code generation tasks.
Reliable Knowledge Acquisition for Continual Learning Agents Proposes a structured knowledge admission and verification system to prevent continual learning agents from storing harmful past experience entries. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Focuses on agent task performance rather than safety guardrails.
Reliable Knowledge Acquisition for Continual Learning Agents Proposes a structured knowledge admission and verification system to prevent continual learning agents from storing harmful past experience entries. Preprint	Multimodal models via arxiv · 17 Oct, 06:17	Not relevant Does not address multimodal reasoning across images or audio.
The Delegation Dilemma: Economies of Scale, Accountability, and Polycentric Governance Models repeated games with endogenous delegation to analyze trade-offs between economies of scale and institutional accountability. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Focuses on game-theoretic governance dynamics rather than AI deployment safety.
The Delegation Dilemma: Economies of Scale, Accountability, and Polycentric Governance Models repeated games with endogenous delegation to analyze trade-offs between economies of scale and institutional accountability. Preprint	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Studies organizational dynamics of agents, not model compression or efficiency.
The Delegation Dilemma: Economies of Scale, Accountability, and Polycentric Governance Models repeated games with endogenous delegation to analyze trade-offs between economies of scale and institutional accountability. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Focuses on multi-agent governance dynamics, not LLM benchmark evaluation.
Harmful Partial Interventions and Benign Leakage in Concept-based Models Investigates how partial interventions in concept-based models can unexpectedly reduce downstream task performance due to information leakage. Preprint	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Addresses model interpretability and interventions, not model efficiency.
Harmful Partial Interventions and Benign Leakage in Concept-based Models Investigates how partial interventions in concept-based models can unexpectedly reduce downstream task performance due to information leakage. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Studies concept intervention harm rather than LLM benchmarking methods.
Harmful Partial Interventions and Benign Leakage in Concept-based Models Investigates how partial interventions in concept-based models can unexpectedly reduce downstream task performance due to information leakage. Preprint	AI agents via arxiv · 17 Oct, 06:17	Not relevant Focuses on concept-based models, not multi-step tool-using agents.
DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents Develops a domain harness and policy learning pipeline to improve language model agent execution in computer-aided design environments. Preprint	Computer vision for industrial applications via arxiv · 17 Oct, 06:17	Not relevant Interactive CAD modeling, not industrial defect inspection on manufacturing lines.
DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents Develops a domain harness and policy learning pipeline to improve language model agent execution in computer-aided design environments. Preprint	AI safety and reliability in deployment via arxiv · 17 Oct, 06:17	Not relevant Focuses on CAD policy optimization, not deployment safety or guardrails.
DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents Develops a domain harness and policy learning pipeline to improve language model agent execution in computer-aided design environments. Preprint	Small and efficient models via arxiv · 17 Oct, 06:17	Not relevant Focuses on CAD harness policy learning, not model compression.
DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents Develops a domain harness and policy learning pipeline to improve language model agent execution in computer-aided design environments. Preprint	Multimodal models via arxiv · 17 Oct, 06:17	Not relevant Addresses procedural CAD modeling code, not multimodal vision-language models.
DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents Develops a domain harness and policy learning pipeline to improve language model agent execution in computer-aided design environments. Preprint	LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17	Not relevant Introduces a CAD agent framework, not a general LLM benchmark.
DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents Develops a domain harness and policy learning pipeline to improve language model agent execution in computer-aided design environments. Preprint	Retrieval-augmented generation (RAG) via arxiv · 17 Oct, 06:17	Not relevant Focuses on CAD action harnesses, not retrieval-augmented generation.

1,921–1,980 of 13,754← Previous Next →