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
Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation Proposes Road MOPD, using non-routed teachers as cross-domain critics to resolve gradient conflicts in multi-teacher on-policy distillation. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on multi-teacher knowledge distillation training algorithms rather than multi-step agent planning or tool use.
Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation Proposes Road MOPD, using non-routed teachers as cross-domain critics to resolve gradient conflicts in multi-teacher on-policy distillation. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Proposes a model training method rather than introducing or analyzing LLM evaluation benchmarks.
MAGMA-Bench: Evaluating Long-Horizon Embodied Agents under Dynamic Tasks and Execution Failures Introduces MAGMA-Bench to evaluate long-horizon embodied agent reasoning during dynamic task execution failures and multi-robot coordination. Preprint	Computer vision for industrial applications via arxiv · 12 Oct, 16:28	Not relevant Evaluates embodied multi-robot planning, not computer vision systems for manufacturing or industrial inspection.
Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning Introduces an unsupervised method to identify and remove subliminally transferred traits from model parameter updates during knowledge distillation. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Focuses on discovering hidden traits in weights rather than general LLM benchmarking protocols.
Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning Introduces an unsupervised method to identify and remove subliminally transferred traits from model parameter updates during knowledge distillation. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Deals with parameter-space trait extraction during distillation rather than autonomous LLM agents.
Constructing Pareto Frontiers of Optimal Reasoning Termination Methods for Overthinking in Reasoning Models Develops Step-wise Truncation Probing and termination strategies to quantify and mitigate accuracy degradation from overthinking in reasoning models. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Studies reasoning termination dynamics and token budgets rather than tool-using autonomous agents.
Constructing Pareto Frontiers of Optimal Reasoning Termination Methods for Overthinking in Reasoning Models Develops Step-wise Truncation Probing and termination strategies to quantify and mitigate accuracy degradation from overthinking in reasoning models. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on overthinking token efficiency and accuracy trade-offs rather than deployment safety or guardrails.
MENTOR: Learning from Failure with On-Policy Pedagogical Intervention Presents MENTOR, a multi-agent framework that intervenes adaptively during failed reasoning rollouts to prevent zero-reward collapse in RLVR. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Proposes a pedagogical intervention training method rather than an evaluation framework or benchmark.
MENTOR: Learning from Failure with On-Policy Pedagogical Intervention Presents MENTOR, a multi-agent framework that intervenes adaptively during failed reasoning rollouts to prevent zero-reward collapse in RLVR. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on preventing zero-reward collapse during RL training rather than deployed safety guardrails.
AudioTETRIS: Token Elimination via Transitional Importance Signals for Efficient Audio LLMs Proposes AudioTETRIS, a training-free token pruning framework that uses acoustic transition signals to compress audio representations efficiently. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Optimizes audio LLM inference efficiency rather than addressing safety, guardrails, or alignment.
AudioTETRIS: Token Elimination via Transitional Importance Signals for Efficient Audio LLMs Proposes AudioTETRIS, a training-free token pruning framework that uses acoustic transition signals to compress audio representations efficiently. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Proposes token elimination in audio encoders rather than multi-step agent planning or tools.
AudioTETRIS: Token Elimination via Transitional Importance Signals for Efficient Audio LLMs Proposes AudioTETRIS, a training-free token pruning framework that uses acoustic transition signals to compress audio representations efficiently. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents an efficient token compression method rather than benchmark design or LLM evaluation.
SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development Lifecycle Introduces SWE-DeceptionBench, a benchmark measuring active and passive deceptive model behaviors across software development lifecycle tasks. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Evaluates deceptive text outputs of LLMs rather than autonomous agent tool use or planning.
Humanize Olympiad Agents: A Multi-Session Workflow for Olympiad Problem Solving with Solve, Review, Wiki, Verify Introduces HOA, a multi-session LLM agent framework coordinating review, wiki knowledge, and verification for complex Olympiad problem solving. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on mathematical problem solving performance rather than safety, alignment, or guardrails.
Humanize Olympiad Agents: A Multi-Session Workflow for Olympiad Problem Solving with Solve, Review, Wiki, Verify Introduces HOA, a multi-session LLM agent framework coordinating review, wiki knowledge, and verification for complex Olympiad problem solving. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Proposes an agent workflow rather than introducing or critiquing evaluation benchmarks.
Model-agnostic Knowledge Augmented Machine Learning for Weakly Annotated Bioimaging Data Proposes a model-agnostic method to integrate biological domain knowledge into machine learning models trained on weakly annotated bioimaging. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Introduces a domain-knowledge augmentation technique for bioimaging rather than an LLM benchmark.
Model-agnostic Knowledge Augmented Machine Learning for Weakly Annotated Bioimaging Data Proposes a model-agnostic method to integrate biological domain knowledge into machine learning models trained on weakly annotated bioimaging. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Focuses on specialized biological imaging classification rather than multimodal vision-language models.
Model-agnostic Knowledge Augmented Machine Learning for Weakly Annotated Bioimaging Data Proposes a model-agnostic method to integrate biological domain knowledge into machine learning models trained on weakly annotated bioimaging. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Addresses biomedical image classification rather than deployed AI safety guardrails or red-teaming.
Model-agnostic Knowledge Augmented Machine Learning for Weakly Annotated Bioimaging Data Proposes a model-agnostic method to integrate biological domain knowledge into machine learning models trained on weakly annotated bioimaging. Preprint	Computer vision for industrial applications via arxiv · 12 Oct, 16:28	Not relevant Focuses on biomedical microscopy and cytomorphology rather than industrial inspection or manufacturing.
Model-agnostic Knowledge Augmented Machine Learning for Weakly Annotated Bioimaging Data Proposes a model-agnostic method to integrate biological domain knowledge into machine learning models trained on weakly annotated bioimaging. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Applies machine learning to bioimaging analysis rather than language model agents or tool execution.
Learning Size Independent Constraint Models from Input Data and Positive and Negative Example Solutions Presents a constraint acquisition tool that synthesizes size-independent constraint programs from input data and example solutions. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents a constraint acquisition tool rather than an LLM evaluation suite or benchmark.
Learning Size Independent Constraint Models from Input Data and Positive and Negative Example Solutions Presents a constraint acquisition tool that synthesizes size-independent constraint programs from input data and example solutions. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on constraint programming acquisition from examples, not language model planning or agents.
Learning Size Independent Constraint Models from Input Data and Positive and Negative Example Solutions Presents a constraint acquisition tool that synthesizes size-independent constraint programs from input data and example solutions. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Deals with inductive logic and constraint programming rather than deployed AI safety or reliability.
SelectArena: A Modern Benchmark for Feature Selection on Real-world Tabular Data Introduces SelectArena, a benchmark designed to evaluate tabular feature selection methods across real-world domain datasets and contemporary models. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on tabular feature selection accuracy rather than AI safety, red-teaming, or guardrails.
SelectArena: A Modern Benchmark for Feature Selection on Real-world Tabular Data Introduces SelectArena, a benchmark designed to evaluate tabular feature selection methods across real-world domain datasets and contemporary models. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Presents a benchmark for tabular feature selection, not language model agents or tool usage.
SelectArena: A Modern Benchmark for Feature Selection on Real-world Tabular Data Introduces SelectArena, a benchmark designed to evaluate tabular feature selection methods across real-world domain datasets and contemporary models. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Covers feature selection for tabular data rather than small language models or LLM compression.
SelectArena: A Modern Benchmark for Feature Selection on Real-world Tabular Data Introduces SelectArena, a benchmark designed to evaluate tabular feature selection methods across real-world domain datasets and contemporary models. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Benchmarks classical and tabular models on tabular datasets, not LLM capabilities or evaluation methodology.
Pretraining and Test-Time Training Budgets Matter for ARC: A Controlled Study of VARC and TRM Evaluates how pretraining and test-time training computational budgets affect ARC benchmark performance across two distinct reasoning model architectures. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Studies training and test-time budget effects on ARC rather than model compression or quantization.
Pretraining and Test-Time Training Budgets Matter for ARC: A Controlled Study of VARC and TRM Evaluates how pretraining and test-time training computational budgets affect ARC benchmark performance across two distinct reasoning model architectures. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Analyzes performance scaling on abstract reasoning tasks without addressing AI safety or deployment risks.
Pretraining and Test-Time Training Budgets Matter for ARC: A Controlled Study of VARC and TRM Evaluates how pretraining and test-time training computational budgets affect ARC benchmark performance across two distinct reasoning model architectures. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Investigates neural networks on ARC grids rather than multi-step tool-using LLM agent systems.
A Multi-Agent LLM System for Automated Tuning of Specialized Neural Network Models Presents a multi-agent LLM framework that automatically tunes specific internal algorithm components while preserving overall neural network structures. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Deals with code and numeric modeling tasks without multimodal image, video, or audio integration.
A Multi-Agent LLM System for Automated Tuning of Specialized Neural Network Models Presents a multi-agent LLM framework that automatically tunes specific internal algorithm components while preserving overall neural network structures. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Does not focus on model compression, quantization, distillation, or small language models.
A Multi-Agent LLM System for Automated Tuning of Specialized Neural Network Models Presents a multi-agent LLM framework that automatically tunes specific internal algorithm components while preserving overall neural network structures. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Does not involve external document retrieval, dense indexing, or retrieval-augmented generation systems.
A Multi-Agent LLM System for Automated Tuning of Specialized Neural Network Models Presents a multi-agent LLM framework that automatically tunes specific internal algorithm components while preserving overall neural network structures. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Focuses on an automated model-tuning framework rather than LLM evaluation or benchmark design.
A Multi-Agent LLM System for Automated Tuning of Specialized Neural Network Models Presents a multi-agent LLM framework that automatically tunes specific internal algorithm components while preserving overall neural network structures. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on automated hyperparameter and architecture coding rather than safety guardrails or alignment.
A Step Harder, Not a Step Too Far: The Benefits and Limits of Curriculum Difficulty Progression in GRPO Proposes a dynamic curriculum for GRPO that re-estimates sample difficulty to progressively transition policy training toward harder problems. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Investigates reinforcement learning curriculum design rather than LLM agents, planning, or tool use.
A Step Harder, Not a Step Too Far: The Benefits and Limits of Curriculum Difficulty Progression in GRPO Proposes a dynamic curriculum for GRPO that re-estimates sample difficulty to progressively transition policy training toward harder problems. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Explores training curriculum progression rather than benchmarks, saturation, or evaluation protocols.
A Step Harder, Not a Step Too Far: The Benefits and Limits of Curriculum Difficulty Progression in GRPO Proposes a dynamic curriculum for GRPO that re-estimates sample difficulty to progressively transition policy training toward harder problems. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies curriculum progression for GRPO optimization rather than safety guardrails, jailbreaks, or monitoring.
Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction Introduces UPTURN, a framework that trains a small classifier to inject redirection prompts as new user turns during reasoning failures. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Operates entirely on text-based reasoning tasks without multimodal inputs or outputs.
Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction Introduces UPTURN, a framework that trains a small classifier to inject redirection prompts as new user turns during reasoning failures. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Investigates test-time reasoning self-correction without using external document or database retrieval.
Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction Introduces UPTURN, a framework that trains a small classifier to inject redirection prompts as new user turns during reasoning failures. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Evaluates general reasoning on math and science benchmarks rather than coding or software repair.
Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction Introduces UPTURN, a framework that trains a small classifier to inject redirection prompts as new user turns during reasoning failures. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Focuses on reasoning accuracy via prompt injection rather than model compression or efficient serving.
Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction Introduces UPTURN, a framework that trains a small classifier to inject redirection prompts as new user turns during reasoning failures. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Proposes a self-correction inference technique rather than novel benchmarks or evaluation methodologies.
Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels Presents Outlier-Aware Attention Linearization, using quadratic approximations to retain outlier dimensions during efficient Transformer attention evaluation. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on attention kernel approximation rather than autonomous agent planning or actions.
Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels Presents Outlier-Aware Attention Linearization, using quadratic approximations to retain outlier dimensions during efficient Transformer attention evaluation. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Develops a linearized attention mechanism rather than evaluation benchmarks or LLM-as-a-judge protocols.
Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels Presents Outlier-Aware Attention Linearization, using quadratic approximations to retain outlier dimensions during efficient Transformer attention evaluation. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Addresses inference efficiency rather than deployed AI safety, guardrails, or alignment.
Variance-Reduced Policy Gradient is Nearly Minimax-Optimal Introduces Loopless Variance-Reduced Policy Gradient, proving near-optimal sample complexity under weak gradient domination using first-order updates. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Analyzes policy gradient theoretical bounds rather than efficient model inference, serving, or compression.
Variance-Reduced Policy Gradient is Nearly Minimax-Optimal Introduces Loopless Variance-Reduced Policy Gradient, proving near-optimal sample complexity under weak gradient domination using first-order updates. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Theoretical analysis of policy gradient sample complexity, not an LLM benchmark or evaluation study.
Variance-Reduced Policy Gradient is Nearly Minimax-Optimal Introduces Loopless Variance-Reduced Policy Gradient, proving near-optimal sample complexity under weak gradient domination using first-order updates. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Presents core reinforcement learning theory rather than LLM planning, multi-agent systems, or tool use.
Variance-Reduced Policy Gradient is Nearly Minimax-Optimal Introduces Loopless Variance-Reduced Policy Gradient, proving near-optimal sample complexity under weak gradient domination using first-order updates. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on generic RL theoretical convergence rates rather than AI safety or alignment risks.
Calibration in One Go: Jointly Learning Accurate and Calibrated Neural Networks Employs a bilevel Stackelberg formulation to jointly optimize classification accuracy and confidence calibration without separate calibration datasets. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Investigates neural network output calibration, not multi-step LLM agents or tool usage.
Calibration in One Go: Jointly Learning Accurate and Calibrated Neural Networks Employs a bilevel Stackelberg formulation to jointly optimize classification accuracy and confidence calibration without separate calibration datasets. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Focuses on prediction calibration rather than model compression, quantization, or inference acceleration.
Calibration in One Go: Jointly Learning Accurate and Calibrated Neural Networks Employs a bilevel Stackelberg formulation to jointly optimize classification accuracy and confidence calibration without separate calibration datasets. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Addresses model training loss formulations rather than LLM benchmarking or evaluation frameworks.
A Baseline Assessment for the Agentic Web Technology Stack Benchmarks combinations of language models, communication protocols, and agent frameworks on web interaction tasks. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Evaluates web-navigation agents on Mind2Web rather than models writing, reviewing, or fixing software.
A Baseline Assessment for the Agentic Web Technology Stack Benchmarks combinations of language models, communication protocols, and agent frameworks on web interaction tasks. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on system stack evaluation rather than safety vulnerabilities, guardrails, or jailbreaks.
Adaptive DHiSparse: Speculative Decoding over a Load-Adaptive KV Hierarchy Combines load-adaptive hierarchical key-value caching with speculative decoding to accelerate long-context language model serving. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Evaluates text LLMs on context serving without multimodal document, vision, or audio components.
Adaptive DHiSparse: Speculative Decoding over a Load-Adaptive KV Hierarchy Combines load-adaptive hierarchical key-value caching with speculative decoding to accelerate long-context language model serving. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on inference hardware optimization and KV cache serving rather than autonomous agent systems.
Adaptive DHiSparse: Speculative Decoding over a Load-Adaptive KV Hierarchy Combines load-adaptive hierarchical key-value caching with speculative decoding to accelerate long-context language model serving. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Investigates inference serving throughput rather than benchmark methodology or LLM evaluation frameworks.
Adaptive DHiSparse: Speculative Decoding over a Load-Adaptive KV Hierarchy Combines load-adaptive hierarchical key-value caching with speculative decoding to accelerate long-context language model serving. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Optimizes serving infrastructure rather than software engineering or code intelligence.
Adaptive DHiSparse: Speculative Decoding over a Load-Adaptive KV Hierarchy Combines load-adaptive hierarchical key-value caching with speculative decoding to accelerate long-context language model serving. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Addresses system performance and throughput rather than safety, alignment, or guardrails.

13,621–13,680 of 13,754← Previous Next →