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 |
|---|---|---|---|
| CT0: Cybersecurity Transformer The paper presents a compact 4B-parameter transformer specialized for cybersecurity tasks using on-policy distillation and dual-mode inference. Ranked #170New way to see a problemPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Introduces a specialized transformer and agent framework for defensive cybersecurity operations. | |
| EvoWeb: A Benchmark for Coding Agents on Evolving Web Applications Introduces a benchmark evaluating coding agents on evolving web applications across cumulative feature requests. Ranked #170New testPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Evaluates autonomous web-coding agents executing multi-step development requests. | |
| Releasing the Inner Jev: A Simple Recipe for Open Decision Models The paper demonstrates recipes to extract single-pass probability decisions directly from open instruction-tuned language models. Ranked #170Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces OpenDecisions benchmark to evaluate non-generative LLM decision heads. | |
| BELT : BEhavior-Aligned Low-Rank KV Cache Compression with Learned Top-k Rank Allocation BELT compresses key-value caches using attention-distribution alignment and learned global rank allocation across model heads. Ranked #170New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:17 | On-topic Introduces low-rank KV cache compression to improve inference throughput and memory efficiency. | |
| Silence of Commit Messages: An Empirical Study for Vulnerability Commit Message Generation using Large Language Models This paper investigates using large language models to generate structured commit messages for software security patches. Ranked #171Peer-reviewed0/10 suggestedOlder paper (2026) Older paper (2026) — discussed this week | LLMs for code via acm · 17 Oct, 06:19 | On-topic Investigates using language models to automate the generation and evaluation of vulnerability commit messages. | |
| PhySimCode: A Benchmark and Evaluation Method for Physics Video to Code Generation This paper introduces a benchmark evaluating whether vision-language models can infer underlying physical laws from motion videos and generate runnable simulation code. Ranked #171New testPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Presents a benchmark evaluating multimodal models translating video dynamics into executable code. | |
| CircuitBreak: Pruning a Sparse Safety Circuit of Neurons The paper presents a greedy neuron-pruning method targeting a sparse safety circuit to remove refusal behavior across multiple instruction-tuned language models. Ranked #171Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Investigates safety degradation via sparse neuron pruning targeting LLM refusal mechanisms. | |
| How Should a Camera-First AI Agent Work? This paper proposes a two-deadline architecture for visual agents to rapidly capture fleeting visual targets before slower semantic reasoning completes. Ranked #171New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Designs an architectural framework for real-time multimodal agents interacting with dynamic environments. | |
| Understanding Personalization Gains in Whisper The paper studies low-rank adaptation in Whisper to evaluate whether transcription gains require group-specific adapters rather than pooled fine-tuning. Ranked #171Surprising resultPeer-reviewed6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:17 | On-topic Analyzes lightweight LoRA parameter-efficient adaptation strategies for Whisper models. | |
| The Format is the Filter: Scoring Format Fixes What an LLM Judge Can Express Demonstrates that the scoring format chosen for an evaluation model strongly restricts its expressive capacity and ranking reliability. Ranked #171Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Analyzes how scoring formats bias and constrain LLM-as-a-judge evaluations. | |
| Purpose IR: Compiling User Purpose for Privacy-Preserving Agent Workflow State The paper introduces an intermediate representation that compiles user purpose into executable semantic boundaries for reusable language-model agent states. Ranked #172New way to see a problemPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Proposes an intermediate representation to constrain privacy and authorized state persistence in LLM agents. | |
| Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate Introduces an adaptive multi-agent debate framework that uses a leader-supporter structure to trigger debate only when simple aggregation is insufficient. Ranked #172Surprising resultPeer-reviewed7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Presents an adaptive multi-agent debate framework for complex reasoning tasks. | |
| SuPER-Mamba: Surprisal Prioritization for Extended-Context Routing in Mamba Improves long-context handling in Mamba state space models without retraining by using surprisal scores to guide state routing. Ranked #172New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:17 | On-topic Provides a training-free state routing framework for efficient long-context processing in Mamba. | |
| Structural Baselines Matter for Agent Failure Detection and Attribution Shows agent failure detection models often rely on trivial cues like run length, exposing benchmark flaws when attributing critical failure steps. Ranked #172Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Critically exposes flawed evaluation baselines and labeling artifacts in agent failure benchmarks. | |
| Improving Semantic Coverage in AI Test & Evaluation through Generative Augmentation Uses generative models and vision-language labeling to detect and synthesize missing semantic domains in test datasets. Ranked #172New testPeer-reviewed6/10 suggested | Multimodal models via arxiv · 12 Oct, 16:28 | On-topic Leverages vision-language models and diffusion models to perform cross-modal semantic labeling and test augmentation. | |
| CreepScope: Measuring Post-Use Authorization Exposure in Tool-Using Agents The paper presents an evaluation protocol to measure post-use authorization exposure and privilege creep across tool-using agent workflows. Ranked #173New testPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Measures post-use authorization exposure and privilege creep in deployed tool-using LLM agents. | |
| 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 #173New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:17 | On-topic Distills teacher LLM relevance judgments into an efficient student model for deployment. | |
| SLDBench v2: Active Scaling Law Discovery Expands a benchmark to evaluate whether language model agents can actively select experiments to discover reliable model scaling laws under compute budgets. Ranked #173New testPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Evaluates language-model agents making sequential decisions and taking actions under budget constraints. | |
| Does a Hosted Decision Model Respect Irrelevant Alternatives? A Pre-Registered Test of Option-Set Dependence in Jev Tests a hosted decision model to show that adding irrelevant options systematically shifts output probability distributions between top candidates. Ranked #173Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Presents a formal evaluation test checking decision models against independence of irrelevant alternatives. | |
| Mitigating Hallucinations in Large Vision-Language Models via Reasoning Uncertainty-Guided Refinement This paper introduces an uncertainty-guided refinement framework to reduce hallucinated responses in large vision-language models. Ranked #173New way to see a problemPeer-reviewed6/10 suggested | Multimodal models via ieee · 14 Oct, 06:19 | On-topic Proposes a framework to mitigate hallucinations in large vision-language models. | |
| Metric-Driven Reinforcement Learning for Object Hallucination Reduction in LVLM-Based Image Captioning The paper proposes a reinforcement learning framework using natural language metrics to reduce object hallucination in vision-language image captioning. Ranked #174New way to see a problemPeer-reviewed6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Improves large vision-language models for multimodal image captioning tasks. | |
| Attacking Only Half the Data: Two-Sided Evaluation of AI-Text Detector Robustness Evaluates AI text detectors under attacks applied to both human and machine text, revealing high false-positive rates on edited human writing. Ranked #174Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Evaluates the robustness and failure modes of AI text detectors against adversarial attacks. | |
| Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation Measures teacher entity familiarity to filter out uninformative search trajectories when distilling deep-search capabilities into smaller agents. Ranked #174New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Studies trajectory distillation and tool-use behaviors in deep-search agents. | |
| Growing and Transferring Pathways for LLM Reinforcement Learning Proposes structural growth pathways to preserve pretrained weights during reinforcement learning and enable transfer across language models. Ranked #174New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:17 | On-topic Enables efficient parameter reuse and transfer across LLMs by freezing backbones and training compact pathways. | |
| VLA-GAUGE: Systematically Evaluating Generalization and Robustness in Vision-Language-Action Models Introduces a diagnostic evaluation framework to systematically test generalization and perturbation robustness in vision-language-action robotic models. Ranked #174New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces a diagnostic evaluation framework and benchmark protocol for Vision-Language-Action models. | |
| OS-Helix: Unifying GUI Interaction and Code Execution in Native Computer-Using Agents The paper presents an end-to-end trained multimodal agent that combines visual graphical interface actions and executable code execution. Ranked #175New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 17 Oct, 06:17 | On-topic Employs multimodal vision-language models capable of perceiving desktop visual interfaces and generating code. | |
| To Cache or Not to Cache: Understanding and Mitigating Safety Incompatibility in Cross-Model KV Cache Reuse The paper studies safety behavioral deviations caused by cross-model key-value cache reuse and proposes an alignment technique to mitigate them. Ranked #175Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Analyzes jailbreak vulnerabilities and safety regressions caused by cross-model KV cache reuse during inference. | |
| ProCARE: Real-World Study Automation via Profile-Grounded Evidence Contracts This paper proposes ProCARE, a framework that uses structured contracts to reliably automate healthcare observational studies with language model agents. Ranked #175New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Proposes an agentic framework with deterministic validation for automating real-world healthcare studies. | |
| REVEAL: Reverse Engineering Agentic Latents from Frontier-Model Traces Extracts reusable local control signals from frontier model traces to guide smaller language model agents without extra target-model training. Ranked #175New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:17 | On-topic Enables transferring high-capability frontier agent behaviors to smaller, frozen open-source models. | |
| EvoWeb: A Benchmark for Coding Agents on Evolving Web Applications Introduces a benchmark evaluating coding agents on evolving web applications across cumulative feature requests. Ranked #175New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces EvoWeb, a novel benchmark dataset for iterative coding tasks. | |
| To Think or Not to Think: Pre-Decisional Reasoning Budgets for Referring Audio-Visual Segmentation This paper studies when audio-visual reasoning harms grounding and trains a probe to determine if chain-of-thought is needed before generation begins. Ranked #176Surprising resultPreprint6/10 suggested | Multimodal models via arxiv · 17 Oct, 06:18 | On-topic Studies multimodal reasoning budgets and overthinking in audio-visual segmentation models. | |
| When the Approved File Is Unchanged: Authority-Set Integrity for Deployment and Coding-Agent Approvals The paper analyzes integrity risks in deployment workflows when secondary repository files change without modifying the approved primary configuration file. Ranked #176New way to see a problemPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Examines deployment and authorization integrity to prevent agents from executing unauthorized repository changes. | |
| ELECT: Settling Identity Before Rendering for Scalable Story Visualization The paper proposes a training-free method establishing entity identity before final rendering to maintain visual consistency across long storyboards. Ranked #176New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:17 | On-topic Reduces story generation memory footprint and rendering inference costs. | |
| Pollard: Governed Semantic Execution Trees for Framework-Independent AI Agent Control and Replay This paper presents Pollard, a framework-independent agent runtime that uses semantic execution trees for deterministic tracking, governance, and replay. Ranked #176New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Proposes a runtime framework for governed execution, tracing, and replay of AI agents. | |
| SLDBench v2: Active Scaling Law Discovery Expands a benchmark to evaluate whether language model agents can actively select experiments to discover reliable model scaling laws under compute budgets. Ranked #176New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces a benchmark measuring LLM agent capabilities in active scaling law discovery. | |
| Counterfactual Neighborhood Risk Fields: Evidence-Grounded Selective Safety for Vision-Language Models The paper proposes a selective safety framework for vision-language models using counterfactual pairs and representation density estimation. Ranked #177New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 17 Oct, 06:17 | On-topic Develops representation-based risk detection specifically tailored across vision-language model architectures. | |
| MetaCascade: End-to-End Attacks on Agent Skills via Metadata-Only Poisoning The paper demonstrates an end-to-end attack framework targeting agent skill selection using metadata-only poisoning. Ranked #177Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Analyzes supply-chain metadata poisoning attacks that compromise tool selection and execution in LLM agents. | |
| Recurrent In-Context Learning for Tabular Foundation Models This paper introduces parameter-tied recurrent blocks to reduce model size in tabular foundation models without sacrificing performance. Ranked #177New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:17 | On-topic Shares layer parameters to reduce model size by 62% while enabling flexible inference compute. | |
| Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM Pipelines The paper shows that LLM review pipelines rarely outperform the stronger model alone due to confirmation bias rather than true verification. Ranked #177Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Examines benchmark evaluation protocols and reveals flaws in testing LLM review pipelines. | |
| Phantom Merge: When Your Large Language Model Agents Pick One but Tell You About Another This paper identifies and mitigates Phantom Merge, an error where language model agents bind retrieved facts to the wrong entity. Ranked #177New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Identifies and fixes entity-evidence binding errors in multi-entity LLM agents. | |
| AlignPred: Error-Controlled KV Cache Compression via Alignment and Predictive Coding This paper presents an error-controlled compression codec for key-value caches using coordinate alignment and predictive coding. Ranked #178New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Proposes an error-controlled KV cache compression and predictive coding scheme to optimize serving memory and speed. | |
| SecReason: Vulnerability Detection and In-Stream Self-Repair for LLM Code Generation The paper presents an in-stream monitoring tool that detects security vulnerabilities and prompts repairs during reasoning-model code generation. Ranked #178New way to see a problemPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Prevents LLMs from generating insecure code during production deployments. | |
| Visible but Not Avoided: State-Dependent Re-Proposal of Rejected Candidates in Catalog-Based LLM Optimization Loops The paper demonstrates that language model optimizers frequently re-propose previously rejected candidates during search trajectories. Ranked #178Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Evaluates the disparity between static probing and dynamic trajectory behavior in LLM optimizers. | |
| LLM Value Functions for Credit Assignment in Agentic RL The paper uses in-context language model evaluations as state value functions to improve credit assignment in agentic reinforcement learning. Ranked #178New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Improves credit assignment for reinforcement learning in long-horizon LLM agent tasks. | |
| World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models The paper proposes using test-time training memory to maintain consistent long-horizon visual scenes in real-time camera-controlled video world models. Ranked #178New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:17 | On-topic Develops a video world model for persistent visual memory and interactive generation. | |
| How Low Can You Go: Fine-Tuning LLMs with Bayesian Optimization (LlMBO) This paper proposes LlMBO, applying Bayesian optimization to ultra-low-dimensional fine-tuning as an alternative to gradient-based policy optimization. Ranked #179Surprising resultPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Proposes an ultra-low parameter, sample-efficient Bayesian optimization method to fine-tune LLMs with minimal compute. | |
| Towards Safe Autonomy for Computer Use This paper proposes a dual-LLM architecture to establish trust boundaries and prevent indirect prompt injections in visual computer-use agents. Ranked #179New way to see a problemPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Develops defense mechanisms to protect computer-use agents against indirect prompt injection. | |
| Big-Big-SycoBench: A Benchmark to Detect Sycophancy Through Model Responses to Valid vs. Invalid Arguments The paper introduces a benchmark measuring whether language models resist sycophancy when presented with invalid user arguments versus valid critiques. Ranked #179New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces a benchmark evaluating LLM sycophancy via valid versus invalid arguments. | |
| Visible but Not Avoided: State-Dependent Re-Proposal of Rejected Candidates in Catalog-Based LLM Optimization Loops The paper demonstrates that language model optimizers frequently re-propose previously rejected candidates during search trajectories. Ranked #179Surprising resultPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Analyzes candidate selection and re-proposal failures in LLM optimization loops for GUI agents. | |
| The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors The paper introduces an in-context learning method that selects demonstration examples by mining a vision-language model's own past prediction mistakes. Ranked #179Surprising resultPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:17 | On-topic Improves vision-language model adaptation through self-mined visual error demonstrations. | |
| FastEEP: Fast Intra-Instance Elastic Scaling for Expert-Parallel MoE Serving This paper presents FastEEP, a system that speeds up intra-instance elastic scaling for Mixture-of-Experts serving by reusing runtime state. Ranked #180Surprising resultPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Introduces elastic intra-instance scaling to make large MoE LLM serving and inference faster and more resource-efficient. | |
| Kairos: Transactional Security Enforcement at the System Boundary for LLM Agents This paper introduces a transactional isolation monitor that delays irreversible actions to stop indirect prompt injections in software agents. Ranked #180New way to see a problemPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:18 | On-topic Introduces transactional security enforcement to mitigate indirect prompt injection in LLM agents. | |
| REPRO: Reflective Evolutionary Prompt Optimizer for Automated LLM Red-Teaming This paper proposes a black-box evolutionary prompt optimization framework using a small open-weight model to discover adversarial jailbreaks against language models. Ranked #180New way to see a problemPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Presents an automated red-teaming evaluation framework tested across multiple LLMs. | |
| Recall What Is Missing: Composing Evidence from Long-Term Agent Memory The paper presents a training-free memory selection method that progressively composes non-redundant evidence sets for long-term agents. Ranked #180New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Proposes an evidence composition method for long-term agent memory. | |
| From Token Correction to Claim Admission: Training-Free Hallucination Mitigation for Large Vision-Language Models Proposes a training-free framework that mitigates vision-language model hallucinations by admitting claims only when supported relative to a visual-null input. Ranked #180New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:17 | On-topic Mitigates object hallucination in large vision-language models via visual evidence admission. | |
| MemDuet: Learning Dual-Process Memory for Long-Context Reasoning The paper introduces a dual-process memory framework coordinating small and large models for efficient long-context reasoning. Ranked #181New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 15:06 | On-topic Improves long-context reasoning inference efficiency and speed by coordinating a small and large model. | |
| Sensitivity or Criterion? Measuring and Moving When Language Models Act This paper introduces a diagnostic framework using signal detection theory to measure when language model agents decide to use tools. Ranked #181New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces a diagnostic framework and benchmark for measuring model action decisions. | |
| When Messages Go Wrong: Understanding Communication Faults in LLM-Based Multi-Agent Systems This paper provides an empirical study and taxonomy of communication faults and error propagation across multi-agent language model systems. Ranked #181Surprising resultPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Empirically studies communication faults and failure propagation across LLM multi-agent systems. | |
| Revocation Is Not Complete at Receipt: An Executable Outcome Contract for Distributed Agent-Tool Gateways The paper introduces an executable contract and monitor to verify that delegated authority revocations actually take effect across distributed agent gateways. Ranked #181Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 15 Oct, 06:17 | On-topic Addresses access control enforcement, authorization revocation, and security guarantees in agent-tool gateways. | |
| Proactive Physical AI through Continuous Observation and Timely Intervention The paper proposes a system for physical AI that continuously monitors environments to identify intervention needs and execute timely actions. Ranked #181New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:17 | On-topic Uses streaming video understanding coupled with language models to initiate physical actions. |
1,201–1,260 of 17,312← PreviousNext →