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
Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception Develops multi-layer linear probes on model activations to detect agent deception and sabotage across diverse scenarios. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Does not investigate model compression, distillation, or serving efficiency.
Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception Develops multi-layer linear probes on model activations to detect agent deception and sabotage across diverse scenarios. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on internal probe mechanisms to detect deception rather than agent architectures.
Bi-FORK: Generative Modeling of High-Dimensional Bifurcating Systems Proposes a generative flow matching method with repulsion-guided sampling to model one-to-many bifurcation trajectories in physical systems. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents a scientific generative framework rather than an LLM evaluation benchmark.
Bi-FORK: Generative Modeling of High-Dimensional Bifurcating Systems Proposes a generative flow matching method with repulsion-guided sampling to model one-to-many bifurcation trajectories in physical systems. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Not related to AI safety, guardrails, or alignment.
Bi-FORK: Generative Modeling of High-Dimensional Bifurcating Systems Proposes a generative flow matching method with repulsion-guided sampling to model one-to-many bifurcation trajectories in physical systems. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Does not address small language models or inference compression techniques.
Bi-FORK: Generative Modeling of High-Dimensional Bifurcating Systems Proposes a generative flow matching method with repulsion-guided sampling to model one-to-many bifurcation trajectories in physical systems. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on flow matching for physical bifurcating systems, not LLM agents.
BrickBench: Evaluating Agentic Brick Design Introduces a benchmark and environment for evaluating coding agents on text-conditioned LEGO design with physical and discrete assembly constraints. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Focuses on 3D discrete assembly and constraint validation rather than multimodal vision-language models.
BrickBench: Evaluating Agentic Brick Design Introduces a benchmark and environment for evaluating coding agents on text-conditioned LEGO design with physical and discrete assembly constraints. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Does not focus on safety risks, red-teaming, or guardrails.
BrickBench: Evaluating Agentic Brick Design Introduces a benchmark and environment for evaluating coding agents on text-conditioned LEGO design with physical and discrete assembly constraints. Preprint	Computer vision for industrial applications via arxiv · 12 Oct, 16:28	Not relevant Studies text-conditioned LEGO assembly rather than industrial vision inspection, defect detection, or robotics.
From Reactive Containment to Proactive Assurance: Lessons from OpenAI, Anthropic, and Google Agent Security Incidents Analyzes security incidents from frontier AI agent evaluations and proposes a framework for proactive boundary assurance. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents a case study and security framework, not a new evaluation benchmark.
On the estimation and validity of AI time horizons---a statistical look at the METR plot Recomputes METR task time horizons using splines and item-response theory to examine the validity of AI difficulty curves. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies benchmark metric validity rather than deployment safety, alignment, or guardrails.
On the estimation and validity of AI time horizons---a statistical look at the METR plot Recomputes METR task time horizons using splines and item-response theory to examine the validity of AI difficulty curves. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Provides a statistical analysis of an evaluation metric, not agent systems or architectures.
On the estimation and validity of AI time horizons---a statistical look at the METR plot Recomputes METR task time horizons using splines and item-response theory to examine the validity of AI difficulty curves. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Does not focus on model compression, quantization, or efficient serving.
AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits Introduces a generative model that allows flexible post-hoc text editing through learned insertion, substitution, and deletion operations. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Tests code correction only as a generic text editing application; main focus is decoding architecture.
AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits Introduces a generative model that allows flexible post-hoc text editing through learned insertion, substitution, and deletion operations. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Deals with text token sequence generation, not multimodal models.
AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits Introduces a generative model that allows flexible post-hoc text editing through learned insertion, substitution, and deletion operations. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Not focused on safety, security, jailbreaks, or guardrails.
AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits Introduces a generative model that allows flexible post-hoc text editing through learned insertion, substitution, and deletion operations. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Focuses on edit flexibility rather than model compression, quantization, or serving efficiency.
AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits Introduces a generative model that allows flexible post-hoc text editing through learned insertion, substitution, and deletion operations. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Focuses on generative decoding mechanisms, not retrieval-augmented generation.
AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits Introduces a generative model that allows flexible post-hoc text editing through learned insertion, substitution, and deletion operations. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents a generative modeling technique rather than an evaluation framework.
AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits Introduces a generative model that allows flexible post-hoc text editing through learned insertion, substitution, and deletion operations. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Proposes a decoding mechanism for sequence editing, not multi-step autonomous agents.
Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning Proposes a multi-agent framework that evolves the data generation space using learner feedback to synthesize agent training trajectories. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Introduces a data synthesis framework for training agents, not an evaluation benchmark.
Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning Proposes a multi-agent framework that evolves the data generation space using learner feedback to synthesize agent training trajectories. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Addresses training data generation rather than AI safety or guardrails.
SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire Presents a benchmark evaluating agentic large language models on long-horizon strategic reasoning in the game Slay the Spire. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Does not involve software engineering or code generation.
SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire Presents a benchmark evaluating agentic large language models on long-horizon strategic reasoning in the game Slay the Spire. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Focuses on strategic game playing, not retrieval-augmented generation.
SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire Presents a benchmark evaluating agentic large language models on long-horizon strategic reasoning in the game Slay the Spire. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Does not study model compression, distillation, or inference optimization.
SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire Presents a benchmark evaluating agentic large language models on long-horizon strategic reasoning in the game Slay the Spire. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies gameplay strategy and evaluation rather than safety, guardrails, or alignment.
SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire Presents a benchmark evaluating agentic large language models on long-horizon strategic reasoning in the game Slay the Spire. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Focuses on text-based strategic game reasoning, not multimodal perception.
Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents Introduces behavioral self-rewarding to refine reinforcement learning feedback for software engineering agents by comparing execution trajectories. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Focuses on code debugging performance rather than security vulnerabilities or deployment safety.
Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents Introduces behavioral self-rewarding to refine reinforcement learning feedback for software engineering agents by comparing execution trajectories. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Uses SWE-bench to train and evaluate agents rather than proposing a new benchmark methodology.
Self-Preservation in Language Models Investigates self-preservation in language models by using task banks to steer synthetic replay during continual instruction tuning. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Investigates continual learning and catastrophic forgetting in LLMs, not autonomous agents.
Self-Preservation in Language Models Investigates self-preservation in language models by using task banks to steer synthetic replay during continual instruction tuning. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Studies capability retention across instruction tuning rather than efficiency, quantization, or distillation.
Self-Preservation in Language Models Investigates self-preservation in language models by using task banks to steer synthetic replay during continual instruction tuning. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies catastrophic forgetting rather than safety, red-teaming, or guardrails.
Self-Preservation in Language Models Investigates self-preservation in language models by using task banks to steer synthetic replay during continual instruction tuning. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Not specifically about code generation or software engineering.
Self-Preservation in Language Models Investigates self-preservation in language models by using task banks to steer synthetic replay during continual instruction tuning. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Focuses on continual learning methods rather than an LLM evaluation benchmark or judge system.
Self-Preservation in Language Models Investigates self-preservation in language models by using task banks to steer synthetic replay during continual instruction tuning. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Examines continual instruction tuning and replay, not retrieval-augmented generation.
Self-Preservation in Language Models Investigates self-preservation in language models by using task banks to steer synthetic replay during continual instruction tuning. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Focuses on language models and instruction tuning, not multimodal systems.
Mesh-Native Tokenisation: Retrofitting Foundation Models to Unstructured Data Proposes a mesh-native tokenization approach to adapt foundation model architectures to unstructured scientific simulation data. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on tokenisation for physics PDE emulators, not language model agents.
Mesh-Native Tokenisation: Retrofitting Foundation Models to Unstructured Data Proposes a mesh-native tokenization approach to adapt foundation model architectures to unstructured scientific simulation data. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents an architecture for physics data, not an LLM benchmark or evaluation methodology.
Mesh-Native Tokenisation: Retrofitting Foundation Models to Unstructured Data Proposes a mesh-native tokenization approach to adapt foundation model architectures to unstructured scientific simulation data. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Not related to deployment safety, guardrails, or alignment.
Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building Benchmarks machine learning and pretrained foundation models for day-ahead electricity forecasting in a university building. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Compares existing forecasting models rather than developing small or efficient language models.
Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building Benchmarks machine learning and pretrained foundation models for day-ahead electricity forecasting in a university building. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Applies time-series models to electricity forecasting without agent planning or tool use.
Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building Benchmarks machine learning and pretrained foundation models for day-ahead electricity forecasting in a university building. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Not related to AI safety, hallucination, or adversarial robustness.
Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building Benchmarks machine learning and pretrained foundation models for day-ahead electricity forecasting in a university building. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant An applied forecasting case study, not a benchmark or evaluation method for LLMs.
Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller Demonstrates a backdoor attack where malicious reasoning degradation is triggered incrementally using invisible prompt tokens and stage tracking. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Studies backdoor attacks on LLMs rather than autonomous agent systems or planning.
Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller Demonstrates a backdoor attack where malicious reasoning degradation is triggered incrementally using invisible prompt tokens and stage tracking. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Does not study small models or inference efficiency.
Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller Demonstrates a backdoor attack where malicious reasoning degradation is triggered incrementally using invisible prompt tokens and stage tracking. Preprint	Multimodal models via arxiv · 12 Oct, 16:28	Not relevant Focuses on textual LLM backdoor attacks, not multimodal systems.
Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller Demonstrates a backdoor attack where malicious reasoning degradation is triggered incrementally using invisible prompt tokens and stage tracking. Preprint	Retrieval-augmented generation (RAG) via arxiv · 12 Oct, 16:28	Not relevant Not related to retrieval-augmented generation.
Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller Demonstrates a backdoor attack where malicious reasoning degradation is triggered incrementally using invisible prompt tokens and stage tracking. Preprint	LLMs for code via arxiv · 12 Oct, 16:28	Not relevant Not focused on code generation or software engineering tasks.
Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller Demonstrates a backdoor attack where malicious reasoning degradation is triggered incrementally using invisible prompt tokens and stage tracking. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Focuses on introducing and demonstrating an attack mechanism, not an evaluation benchmark.
CLEAT: Learning Value-Guided Clarification for Interactive Agents Co-trains an interactive agent's task policy with a clarification-value controller to optimize when to ask follow-up questions. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Studies interactive task execution and intent clarification rather than safety risks or guardrails.
CLEAT: Learning Value-Guided Clarification for Interactive Agents Co-trains an interactive agent's task policy with a clarification-value controller to optimize when to ask follow-up questions. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Develops a co-training framework for interactive agents rather than an evaluation benchmark.
Kronecker Product Attention: Memory-Efficient KV Compression via Structured Factorization Introduces a key-value cache compression method representing attention activations as sums of compact Kronecker products. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on attention-level KV cache compression, not LLM agents or planning.
Kronecker Product Attention: Memory-Efficient KV Compression via Structured Factorization Introduces a key-value cache compression method representing attention activations as sums of compact Kronecker products. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Does not address safety, reliability, or guardrails.
Kronecker Product Attention: Memory-Efficient KV Compression via Structured Factorization Introduces a key-value cache compression method representing attention activations as sums of compact Kronecker products. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Presents an attention architecture, not an evaluation framework or benchmark.
MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis Analyzes zero-advantage training plateaus in reinforcement learning with verifiable rewards and proposes recursive data synthesis to address them. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on training data synthesis for RLVR reasoning, not agent workflows or multi-agent systems.
MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis Analyzes zero-advantage training plateaus in reinforcement learning with verifiable rewards and proposes recursive data synthesis to address them. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Investigates RLVR training dynamics rather than deployment safety, jailbreaks, or red-teaming.
MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis Analyzes zero-advantage training plateaus in reinforcement learning with verifiable rewards and proposes recursive data synthesis to address them. Preprint	LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28	Not relevant Analyzes RLVR training data synthesis rather than proposing a general LLM evaluation benchmark.
MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis Analyzes zero-advantage training plateaus in reinforcement learning with verifiable rewards and proposes recursive data synthesis to address them. Preprint	Small and efficient models via arxiv · 12 Oct, 16:28	Not relevant Does not focus on small models, distillation, or hardware-efficient serving.
Incremental Retrieval-Augmented Generation for Human-in-the-Loop Decision Support Evaluates an incremental retrieval-augmented generation framework that progressively populates its knowledge base from human decision precedents. Preprint	AI safety and reliability in deployment via arxiv · 12 Oct, 16:28	Not relevant Addresses human-in-the-loop decision support, but main focus is RAG workflow rather than safety or guardrails.
Incremental Retrieval-Augmented Generation for Human-in-the-Loop Decision Support Evaluates an incremental retrieval-augmented generation framework that progressively populates its knowledge base from human decision precedents. Preprint	AI agents via arxiv · 12 Oct, 16:28	Not relevant Focuses on a decision-support retrieval framework rather than autonomous multi-step agents or tool use.

16,801–16,860 of 17,313← Previous Next →