Every paper checked — nothing hidden
All papers
Every paper checked for brief #4 of 12–19 Oct 2026 — nothing is hidden. Brief papers first.
All17,312In brief24More (on-topic)3,333Not relevant13,753Not assessed87Older351Could not confirm93Blogs with no paper266All blogs284
| Paper | Topic | Relevance | Feedback |
|---|---|---|---|
| Scaling Agentic Data for Long-Horizon Terminal Intelligence Introduces a synthetic data generation framework to produce long-horizon task trajectories for terminal agents. Ranked #61New way to see a problemPreprint6/10 suggested | LLMs for code via arxiv · 12 Oct, 16:28 | On-topic Focuses on terminal commands and code-level execution in software environments. | |
| When Judges Agree and Are Wrong: Consistency-Accuracy Dissociation as a Blind Spot for Label-Free Evaluation Examines failure modes in label-free evaluation where models reach high consistency across perturbations while generating identical incorrect answers. Ranked #62Surprising resultPreprint8/10 suggested | LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17 | On-topic Critiques label-free evaluation and consistency-based scoring, uncovering blind spots where models agree incorrectly. | |
| FDE-Gym: The Last Mile of AI Delivery The authors introduce FDE-Gym, a benchmark of 95 customer delivery tasks evaluating AI agents on ambiguous requirements and live software environments. Ranked #62New testPreprint7/10 suggested | AI agents via arxiv · 17 Oct, 06:17 | On-topic Evaluates AI agents operating in complex environments to deliver end-to-end customer projects. | |
| Are We Really Measuring Forgetting? Pitfalls in Evaluating Machine Unlearning for Multimodal LLMs The paper identifies evaluation pitfalls in multimodal machine unlearning and introduces a benchmark grounded in real-world knowledge. Ranked #62Surprising resultPreprint7/10 suggested | Multimodal models via arxiv · 15 Oct, 15:06 | On-topic Investigates unlearning capabilities and benchmark protocols specifically for multimodal LLMs. | |
| Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure Mode of the LLM Judge Documents safety alignment methods, negative preference tuning outcomes, and automated judge limitations for a Portuguese language model. Ranked #62Surprising resultPreprint7/10 suggested | Small and efficient models via arxiv · 15 Oct, 15:06 | On-topic Presents training and alignment findings for Manacá-1B, a 1.72B parameter small language model. | |
| Computed, not Memorized: Bit-Factorized Identifiers for Efficient Generative Retrieval The paper presents a bit-factorized document identifier design that allows generative retrieval models to generalize to unseen documents without retraining. Ranked #62New way to see a problemPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:18 | On-topic Introduces a factorized identifier design to improve generative document retrieval for unseen documents. | |
| Many Steps into One: Reasoning Contraction and the Limits of Safe Memory Reuse for LLM Agents The paper analyzes the failure modes of agent memory reuse and evaluates certified memory update strategies when facts change. Ranked #62Surprising resultPreprint8/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Addresses deployment reliability by preventing agents from using stale memories to generate false answers. | |
| ReTerm: Recovery Training for Self-Improving Terminal Agents The authors present a training framework and dataset that enables terminal coding agents to self-improve by recovering from execution errors. Ranked #62New way to see a problemPreprint6/10 suggested | LLMs for code via arxiv · 15 Oct, 06:17 | On-topic Focuses on terminal agents performing complex coding tasks and software execution. | |
| INFUSE: A Unified Surface Benchmark for Indirect Prompt Injection in Agentic Systems The paper introduces an end-to-end benchmark to evaluate indirect prompt injection risks across long-horizon agent workflows. Ranked #63New testPreprint8/10 suggested | LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17 | On-topic Introduces a standardized benchmark of 40 tasks to evaluate agent vulnerabilities across frontier models. | |
| What Does LALM Post-training Learn? Rethinking Audio Understanding and Reasoning Capabilities The paper analyzes post-training in audio-language models to separate genuine audio reasoning gains from question-format memorization. Ranked #63Surprising resultPreprint7/10 suggested | Multimodal models via arxiv · 15 Oct, 15:06 | On-topic Investigates what multimodal large audio-language models learn during post-training. | |
| Embedding Geometry Shapes LLM Routing: Tracing Failure Modes and Improving Generalization The paper uses intermediate transformer hidden states instead of semantic embeddings to improve classifier-based prompt routing across unseen tasks. Ranked #63Surprising resultPreprint7/10 suggested | Small and efficient models via arxiv · 15 Oct, 15:06 | On-topic Improves LLM routing techniques to optimize multi-model selection and inference efficiency. | |
| Faithful Query Resolution for Conversational Search The paper evaluates query rewriting strategies to balance query fidelity and recall in conversational search systems. Ranked #63Surprising resultPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:18 | On-topic Examines conversational query rewriting and fusion to balance retrieval breadth and fidelity in conversational search. | |
| AutoEnvScaling: Automating the Data Flywheel with Terminal Agents The paper presents AutoEnvScaling, using terminal agents to automatically generate reinforcement learning environments from model rollouts. Ranked #63New way to see a problemPreprint6/10 suggested | LLMs for code via arxiv · 15 Oct, 06:18 | On-topic Terminal agents execute commands to solve software engineering and coding tasks. | |
| Agent-Reflex: Teaching Language Agents to Act Reflectively The paper introduces Agent-Reflex, a training framework designed to teach language agents how to reflect and recover from noisy environment feedback. Ranked #63New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Trains language agents to handle noisy tool feedback and reflect on failures. | |
| CWE-bench: Refocusing Agentic Security from Exploitation to Defense The paper presents CWE-bench, a benchmark of 120 held-out software tasks to evaluate AI agents on finding and repairing security vulnerabilities. Ranked #63New testPreprint8/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Addresses software security vulnerabilities, defensive agent capabilities, and AI deployment safety. | |
| FDE-Gym: The Last Mile of AI Delivery The authors introduce FDE-Gym, a benchmark of 95 customer delivery tasks evaluating AI agents on ambiguous requirements and live software environments. Ranked #64New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17 | On-topic Introduces a benchmark evaluating language agents on real-world engineering delivery tasks. | |
| LED-BENCH: A Benchmark for Large-format Engineering Drawing Understanding The paper presents LED-BENCH, a multidisciplinary benchmark evaluating multimodal model understanding and complex reasoning across large-format engineering drawings. Ranked #64New testPreprint7/10 suggested | Multimodal models via arxiv · 15 Oct, 15:06 | On-topic Presents a benchmark for testing multimodal LLMs on complex document reasoning over engineering diagrams. | |
| Why Generate a Click? Plug-in Decision Heads for Web Agents This paper introduces plug-in decision heads that read action choices directly from visual agent hidden states rather than generating text tokens. Ranked #64Surprising resultPreprint7/10 suggested | Small and efficient models via arxiv · 15 Oct, 15:06 | On-topic Reduces decoding latency and compute for web agents via plug-in decision heads. | |
| Data Curation and Model Souping: Unlocking Qwen3-0.6B for MS MARCO Passage Ranking The paper shows that data curation and model merging boost a small model's passage retrieval and ranking performance without architectural changes. Ranked #64Surprising resultPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:18 | On-topic Investigates passage retrieval and re-ranking performance using lightweight models via data curation and model souping. | |
| Cortex: Measurement-Gated Evolution and Pruning of Coding-Agent Skills and Rules The authors develop a system that systematically tests, evolves, and prunes skill and rule files used by coding agents. Ranked #64New way to see a problemPreprint6/10 suggested | LLMs for code via arxiv · 15 Oct, 06:17 | On-topic Automates the evolution and pruning of coding-agent skills and rules across software tasks. | |
| When LLM Agents See a Conflict but Still Proceed: Cross-Domain Dependencies in Network Slicing Shows that language agents often ignore identified cross-domain dependencies, causing unintended disruptive actions in network management. Ranked #64Surprising resultPreprint8/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Demonstrates failure modes where agents ignore critical dependencies in deployment. | |
| Self-Evolving Agent Harness The paper formalizes a framework and theoretical generalization bounds for evolving an agent's software harness without updating model weights. Ranked #64New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Studies self-evolving agent scaffolding including prompts, memory, workflows, and tool execution. | |
| TEM: Temporal State Tracking Memory for Large Language Model Agents Presents a training-free framework that tracks changing conversational facts as versioned records to improve agent memory retrieval. Ranked #65New way to see a problemPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 17 Oct, 06:17 | On-topic Structures and retrieves versioned conversational facts, directly outperforming standard RAG. | |
| MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments The paper provides an open platform and environment suite for evaluating and training GUI agents with online reinforcement learning. Ranked #65New testPreprint7/10 suggested | AI agents via arxiv · 17 Oct, 06:17 | On-topic Introduces a platform and training system for interactive GUI agents across environments. | |
| How Far Can Coding Agents Go in Visual Generation? The study evaluates coding agents generating images via executable code, analyzing their strengths in geometric constraints over photorealistic drawing. Ranked #65Surprising resultPreprint7/10 suggested | Multimodal models via arxiv · 15 Oct, 15:06 | On-topic Examines visual generation and visual question answering using code-based representations. | |
| To Reason but Not to Stop: Understanding and Recovering Pruned Large Reasoning Models The paper shows pruned reasoning models fail by looping, proposing an activation steering method to restore proper termination. Ranked #65Surprising resultPreprint7/10 suggested | Small and efficient models via arxiv · 15 Oct, 15:06 | On-topic Studies and recovers termination failure in pruned large reasoning models using activation steering. | |
| Same Operation, Different Permission: Evaluating Whether Guardrails Honor Authorization The paper introduces a benchmark to evaluate whether language-model guardrails properly distinguish authorized from unauthorized actions across identical operations. Ranked #65New testPreprint8/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Investigates guardrail reliability, blocking unauthorized actions while permitting valid permission. | |
| Returning to Development States: When Full Sandbox Snapshots Help Coding Agents Evaluates full sandbox snapshot exploration for coding agents, finding benefits primarily when recovery requires restoring runtime dependencies outside Git repositories. Ranked #65Surprising resultPreprint6/10 suggested | LLMs for code via arxiv · 12 Oct, 16:29 | On-topic Directly focuses on state restoration and sandbox snapshots to improve autonomous coding agents on software tasks. | |
| Thinking Inertia: LLMs Keep Thinking When Told Not To The paper defines evaluation metrics to show that language models continue generating reasoning traces even when instructed not to think. Ranked #65Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 12 Oct, 16:28 | On-topic Proposes new metrics and evaluation frameworks to measure LLM thinking mode compliance. | |
| Following Delegation Traces to Study Human Oversight in Interaction with Multimodal AI Investigates how user trust influences delegation and verification choices during human interaction with multimodal video question-answering systems. Ranked #66New way to see a problemPreprint7/10 suggested | Multimodal models via arxiv · 17 Oct, 06:17 | On-topic Examines human delegation and verification behaviors when interacting with multimodal AI video systems. | |
| Beyond State Capacity: Why Retrieval Breaks in Mamba The paper analyzes retrieval failures in Mamba architectures under query-first settings and explores token representations and selection gates. Ranked #66Surprising resultPreprint7/10 suggested | Small and efficient models via arxiv · 15 Oct, 15:06 | On-topic Analyzes internal limitations of Mamba, a key efficient linear-attention architecture candidate. | |
| Harness^2: Recursive Agent Harnessing for an Open World The authors propose a two-level recursive self-improvement framework to evolve and refine execution harnesses for open-world agents. Ranked #66New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Presents recursive self-improvement frameworks to evolve LLM agent harnesses for open-world tasks. | |
| AuthGateBench: Benchmarking LLMs as Authorization Gate Introduces a policy-guided benchmark evaluating language models acting as authorization gates for tool-using coding agents. Ranked #66New testPreprint8/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Studies authorization failure modes and preventing scope escalation in deployment. | |
| RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment Introduces RoTA-Bench to evaluate vision-language models on 3D spatial reasoning through interactive object rotation to match target 2D silhouettes. Ranked #66Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces RoTA-Bench to evaluate vision-language models on 3D spatial reasoning. | |
| Graph-free Retrieval with Graph-Modified Embeddings The authors evaluate offline graph-modified embeddings for vector retrieval, improving multi-hop evidence recovery without requiring query-time graph access. Ranked #66New way to see a problemPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17 | On-topic Improves knowledge retrieval for RAG using graph-modified embeddings without graph queries. | |
| Long Text to Predictive Features: LLM-Guided Blockwise Feature Engineering via Executable Program Search Presents a framework that uses language models offline to generate executable feature-engineering code for tabular risk models. Ranked #66New way to see a problemPreprint6/10 suggested | LLMs for code via arxiv · 12 Oct, 16:28 | On-topic Proposes an LLM-guided system that writes and executes programs for automated feature engineering. | |
| Cross-Prompt Inference Mitigates Rare-Class Suppression in Imbalanced VLM Finetuning Introduces cross-prompt inference to recover rare-class recall drops caused by class-imbalanced finetuning in vision-language models. Ranked #67Surprising resultPreprint7/10 suggested | Multimodal models via arxiv · 17 Oct, 06:18 | On-topic Addresses rare-class recall and prompt interactions during vision-language model finetuning. | |
| MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments The paper provides an open platform and environment suite for evaluating and training GUI agents with online reinforcement learning. Ranked #67New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17 | On-topic Presents a unified evaluation platform and benchmark suite for GUI agents. | |
| Energy Consumption Analysis of Discrete Diffusion and Autoregressive Language Models Benchmarks the inference energy efficiency of diffusion language models compared to standard autoregressive architectures. Ranked #67Surprising resultPeer-reviewed7/10 suggested | Small and efficient models via arxiv · 15 Oct, 15:06 | On-topic Compares inference energy efficiency and Joules-per-token across diffusion and autoregressive LLMs. | |
| LLM-based User-Product Relevance Evaluation for Dynamic Product Advertising The paper introduces a large language model-based metric for offline evaluation of user-product relevance in dynamic product advertising retrieval. Ranked #67New way to see a problemPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17 | On-topic Evaluates user-product relevance and quality within dynamic product retrieval systems. | |
| When Agent Evaluation Becomes Unreliable: An Empirical Study of Oracle Failures and Configuration-Dependent Measurement in an LLM Agent Benchmark The paper audits agent evaluation benchmarks for software defects, showing that broken reward functions and configuration shifts substantially distort reported success rates. Ranked #67New testPreprint8/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Investigates testing reliability defects and state-integrity failures in deployed agent systems. | |
| Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards The paper presents a dataset and framework with dense progressive rewards to train language agents on long-horizon tasks. Ranked #67New testPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Introduces environments and dense rewards for training autonomous long-horizon LLM agents. | |
| Leakage-Resistant Evaluation of AI-Generated Database Specifications in Specification-Driven Migration Evaluates LLM judges assessing AI-generated database specifications against trivial baselines using a leakage-resistant benchmark. Ranked #67Surprising resultPreprint6/10 suggested | LLMs for code via arxiv · 12 Oct, 16:29 | On-topic Benchmarks LLM judges evaluating AI-generated database specifications in software migration. | |
| Loud and Clear: Making Visual Evidence Count in Multimodal Document RAG The paper examines how increasing visual token budgets and cropping key evidence improves document retrieval-augmented generation in vision-language models. Ranked #68Surprising resultPreprint7/10 suggested | Multimodal models via arxiv · 17 Oct, 06:17 | On-topic Examines vision-language model comprehension of complex multimodal document pages. | |
| RoleQ: Role-Aligned Scales for Converting Language Models to Ternary Proposes a structural scaling method based on transformer unit roles to convert language models to ternary weights effectively. Ranked #68Surprising resultPreprint7/10 suggested | Small and efficient models via arxiv · 17 Oct, 06:17 | On-topic Introduces a method to quantize language models to ternary weights for efficient deployment. | |
| Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents The paper introduces a benchmark evaluating how well language models engineer isolated execution harnesses for coding agents using specification documents. Ranked #68New testPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 15:06 | On-topic Benchmarks how well models build harnesses that govern tool-using coding agents. | |
| Self-improving data synthesis can scale embedding models across domains and tasks This paper introduces a self-improving synthetic data generation pipeline to train embedding models across diverse retrieval domains without extensive manual annotation. Ranked #68New way to see a problemPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 06:17 | On-topic Improves embedding models used for information retrieval and RAG through synthetic data. | |
| When Agents Talk, Secrets Travel: Topology-Driven Leakage and Recipient-Conditioned Sanitization in Multi-Agent LLM Systems The paper investigates privacy leakage caused by multi-agent network topologies and proposes an execution sanitizer conditioned on recipient identity. Ranked #68Surprising resultPreprint8/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Develops recipient-conditioned sanitization to stop privacy leakage across multi-agent system topologies. | |
| Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards The paper presents a dataset and framework with dense progressive rewards to train language agents on long-horizon tasks. Ranked #68New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Presents Tev-Gym, a new benchmark dataset for evaluating long-horizon agents. | |
| Cadence: Strategic Guidance for Coding Agents Introduces a dynamic monitoring framework that adaptively schedules runtime inspections and intervention guidance for coding agents. Ranked #68New way to see a problemPreprint6/10 suggested | LLMs for code via arxiv · 12 Oct, 16:29 | On-topic Proposes a runtime monitoring and intervention framework specifically for LLM coding agents solving SWE tasks. | |
| VELD-QA: A Large-Scale Dataset for Visual Evidence Grounding and Question Answering in Long Documents Presents a benchmark with human-annotated visual evidence to evaluate retrieval and question answering across long documents. Ranked #69New testPreprint7/10 suggested | Multimodal models via arxiv · 17 Oct, 06:17 | On-topic Focuses on multimodal vision-language understanding of PDF charts and layouts. | |
| TRACER: Trajectory-Aware Tool Retrieval for LLM Agents Proposes a trajectory-aware retrieval method that indexes tool sequences in a trie to cut token costs and improve selection accuracy. Ranked #69New way to see a problemPreprint7/10 suggested | Small and efficient models via arxiv · 17 Oct, 06:17 | On-topic Significantly reduces prompt token overhead and inference compute for large tool registries. | |
| 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. Ranked #69Surprising resultPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 17 Oct, 06:17 | On-topic Addresses dynamic context admission and evidence-grounded knowledge reuse. | |
| Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents The paper introduces a benchmark evaluating how well language models engineer isolated execution harnesses for coding agents using specification documents. Ranked #69New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 15:06 | On-topic Introduces a two-stage benchmark measuring long-horizon harness engineering for agents. | |
| Vela: User Behavior as Supervision Sparks LLM Intelligence This paper presents a framework that converts long-term real user behavior records into supervisory tasks to train personalization-focused language models. Ranked #69New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Uses long-term user interaction trajectories to train personal-assistant agents. | |
| TRAP: Tool-result Replay Attack for Probing Misbinding in LLM Agents Presents a benchmark and attack method to evaluate failures caused by tool results misbinding to incorrect execution contexts in LLM agents. Ranked #69Surprising resultPreprint8/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Investigates safety vulnerabilities and execution integrity risks during tool-use deployment. | |
| SEAR: Self-Evolving Agentic Robotics The paper presents a framework and benchmark to evaluate whether self-evolving robotic agents genuinely learn reusable skills across tasks. Ranked #69New testPreprint6/10 suggested | LLMs for code via arxiv · 15 Oct, 06:17 | On-topic Examines LLMs generating, testing, and retaining robot control code. | |
| REKI: Registering Knowledge in VLAs The paper introduces a compact vision-language-action model using register tokens as persistent memory to reduce control oscillations in robotics. Ranked #70New way to see a problemPreprint7/10 suggested | Multimodal models via arxiv · 17 Oct, 06:18 | On-topic Integrates vision-language models with action prediction experts for multimodal robotic control. | |
| System1-Arena: Which Decisions Can Be Left to System 1? Introduces a benchmark evaluating single-forward-pass discrete decision models versus generative reasoning models across diverse closed-loop tasks. Ranked #70New testPreprint7/10 suggested | Small and efficient models via arxiv · 17 Oct, 06:17 | On-topic Evaluates low-latency, lower-cost single-forward-pass alternatives to expensive reasoning models. | |
| Cascaded Retrieval with Cross-Modal Query Refinement for Efficient Visual Document Search This paper proposes SCOUT, a cascaded two-track framework combining visual candidate filtering, offline semantic summaries, and query refinement for document retrieval. Ranked #70New way to see a problemPreprint6/10 suggested | Retrieval-augmented generation (RAG) via arxiv · 15 Oct, 15:06 | On-topic Proposes an efficient cascaded retrieval and reranking framework for visually rich documents. |
481–540 of 17,312← PreviousNext →