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 |
|---|---|---|---|
| Behavioral Forgetting Is Not Privacy in Machine Unlearning Demonstrates that machine unlearning often fails to prevent targeted information recovery across alternative multilingual access paths. Ranked #146New way to see a problemPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17 | On-topic Exposes fundamental flaws in existing unlearning benchmarks and evaluation metrics. | |
| A Self-Improving Agent that Grows its Own DSL: Oracle-Free Operator Induction for Exact Question Answering The paper presents an agent that expands its domain-specific language for exact answering without external rewards or judges. Ranked #146New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 17 Oct, 06:17 | On-topic Develops a self-improving agent that invents domain-specific language operators. | |
| The Price of Precision: A Calibrated Cost Model for Mixed-Precision LLM Serving The paper presents a hardware-calibrated cost model to optimize mixed-precision serving configurations across varying batch sizes and context lengths. Ranked #146New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Introduces a cost model for mixed-precision large language model serving to optimize inference efficiency. | |
| ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills Proposes a framework that encodes visual trajectories into reusable skill cards to reinforce vision-language agent planning and rewards. Ranked #146New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 12 Oct, 16:28 | On-topic Integrates visual spatial layouts and image features directly into vision-language model agents. | |
| Forgetting is Not Always Bad: A Neuro-Inspired Memory Repair Mechanism for Poisoned LLM Agents Introduces DREAM, an agent-intrinsic defense that triggers functional forgetting of poisoned memory entries through implicit reactivation and topological diagnosis. Ranked #146New way to see a problemPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Proposes defense mechanisms against memory-poisoning and backdoor attacks in deployed systems. | |
| OPRD: On-Policy Reminder Distillation for Long-Horizon Coding Agents Uses multi-scope reminders and on-policy distillation to help coding agents maintain constraints over long software engineering interactions. Ranked #147New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 17 Oct, 06:17 | On-topic Proposes reminder distillation to improve multi-turn tool interaction and planning in coding agents. | |
| When Does "Post-Call" Begin? An Availability Contract for Earnings-Call NLP Benchmarks Proposes an availability framework to identify and resolve within-event timing mismatches in earnings-call financial NLP benchmarks. Ranked #147New testPeer-reviewed7/10 suggested | LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17 | On-topic Identifies timing lookahead contamination flaws in financial NLP benchmark designs. | |
| Fast and Accurate: Improving Personalized LLM Performance by Semantic Concentration Inspired Workload Split between RAG and Fine-Tuning This paper splits personalized language model workloads between parameter-efficient fine-tuning and retrieval based on historical data semantic concentration. Ranked #147New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Proposes splitting workloads between PEFT and RAG to lower inference latency. | |
| EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams Develops a framework for egocentric multimodal assistants to decide both when to speak and what guidance to provide. Ranked #147New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 12 Oct, 16:28 | On-topic Processes continuous first-person video and audio streams jointly with language. | |
| 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. Ranked #147Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Reduces reasoning failures and mitigates hallucinated steps by redirecting reasoning paths to improve reliability. | |
| Auditing Escalation Reliability in Healthcare LLM Agents: A Perturbation-Controlled Benchmark of Workflow-State Failures Introduces an episode-level healthcare benchmark to audit how reliably language model agents escalate actions when facing workflow disruptions. Ranked #148New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 17 Oct, 06:17 | On-topic Introduces a perturbation-controlled benchmark to evaluate agent decision-making under uncertainty. | |
| BARTER: Compact Dynamic Sparse Training with Standard Dense Kernels This paper presents a dynamic sparse training method that reduces compute overhead by maintaining compact tensors with standard dense kernels. Ranked #148New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Introduces dynamic sparse training using standard dense kernels to reduce model compute and cost. | |
| Output Discipline Sets the Floor: A Falsification-Guided, Model-Agnostic Agent Protocol for Auditable Hypothesis-Driven Decision Making The paper introduces a falsification-guided agent protocol requiring systematic hypothesis elimination to improve decision-making consistency on complex diagnostic tasks. Ranked #148Surprising resultPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Proposes a model-agnostic agent protocol using evidence falsification for auditable decisions. | |
| DiffuPlex: Accelerating Full-Duplex Spoken Dialog Models via Rolling Masked Diffusion Presents a rolling masked diffusion method that accelerates full-duplex spoken dialog models by forecasting multiple future interaction frames. Ranked #148New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 12 Oct, 16:28 | On-topic Operates on multimodal spoken dialog streaming audio and conversational text. | |
| 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. Ranked #148Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Detects and suppresses subliminal transfer of undesirable or misaligned behaviors inherited during model distillation. | |
| SelectTTRL: Representation-Guided Data Selection for Efficient Test-Time Reinforcement Learning This paper presents a method using internal representation similarity to identify which test problems benefit from test-time reinforcement learning. Ranked #149New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Proposes a data selection method to make test-time reinforcement learning inference and adaptation more efficient. | |
| Encrypted Skills Can Become New Security Exploits If AI Agents Know How to Decipher The authors demonstrate that encoding malicious instructions in agent skills bypasses scanners and causes models to execute harmful actions upon decoding. Ranked #149Surprising resultPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Demonstrates a vulnerability where LLM agents decode and execute malicious injected skills. | |
| SolBehave: Evaluating the Behavioural Correctness of LLM-Generated Smart Contracts The paper introduces a benchmark using fuzzing and transaction replays to evaluate the behavioral correctness of generated smart contracts. Ranked #149New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces SolBehave benchmark to test the behavioral correctness of LLM-generated code. | |
| Can machines reason non-verbally? Introduces a generate-verify-optimize pipeline to enable non-verbal visual reasoning in machine learning systems. Ranked #149New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 12 Oct, 16:28 | On-topic Addresses non-verbal visual reasoning in multimodal foundation models via generation and verification. | |
| VEPA: Verified Evolving Formal Policy for Lifelong LLM Agent Guardrails Develops a framework for evolving deterministic formal policies to serve as runtime guardrails for large language model agents. Ranked #149New way to see a problemPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Introduces verified evolving formal runtime guardrails to enforce safety and prevent unwanted actions. | |
| Swarm Telepathy: Prefill-Free KV-Cache Transfer Across Families Enables prefill-free multi-agent communication across different language model families using cross-model key-value cache transfer. Ranked #150New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Enables prefill-free cross-family KV-cache transfer to speed up multi-model inference. | |
| Physics-Aware Deepfake Detection via Distance-Speech Consistency Detects talking-head deepfakes in dynamic videos by matching speaker distance estimates with distance-dependent acoustic energy variations. Ranked #150New way to see a problemPeer-reviewed6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Analyzes multimodal audio-visual consistency for detecting manipulated videos. | |
| Fan Out or Relay? Ownership-Aware Routing for LLM Agents Querying Siloed Tables This paper shows data ownership determines whether agents should broadcast or relay table queries, and proposes ownership-aware routing. Ranked #150Surprising resultPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Explores coordination and query routing strategies among distributed LLM agents. | |
| Does self-distillation improve reasoning? The paper evaluates whether self-distillation using privileged information improves reasoning capabilities or reflects overfitting to training tasks. Ranked #150Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Critically evaluates whether self-distillation benchmarks reflect true reasoning gains or overfitting. | |
| Unsupervised Multimodal Reward Modeling Trains an unsupervised vision-language reward model using intermediate layer representations and caption perturbations without human preference labels. Ranked #150Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Addresses vision-language preference modeling and demonstrates significant reductions in caption hallucination. | |
| Representation Drift in Continual Learning: LoRA versus N-gram Memory Measures representation drift across continual learning methods, finding n-gram memory updates cause less internal drift than LoRA. Ranked #151Surprising resultPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Investigates parameter-efficient continual learning comparing N-gram memory against LoRA adapters. | |
| RepProbe: Can Video Language Models Count Repeated Actions? A Diagnostic Benchmark This paper introduces a benchmark to diagnose the ability of video language models to count repeated action cycles. Ranked #151Surprising resultPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Introduces a diagnostic benchmark analyzing repetition counting failures in video language models. | |
| Fresh at Decision, Stale at Action: Routing New Information to LLM Decisions Already Made This paper introduces a timestamp theory and persistent ledger to route updated information to pending LLM decisions before action. Ranked #151New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Designs a framework for routing updated information to earlier agent decisions. | |
| How Well Do LLMs Perform in Peer Review? Evidence from Finance and Accounting The authors evaluate language models on finance peer review tasks, finding predictive ability for publication outcomes through correctness rather than originality assessments. Ranked #151Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Investigates the validity and limitations of LLM-as-a-judge for academic peer review. | |
| When Faster Decisions Do Not Make Better Agents: An Empirical Study of Jev in Agent Workflows Studies how accelerated decision-making models affect overall task completion, latency, and costs across embodied and mobile agents. Ranked #151Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Analyzes agent failure modes, premature completion, and reliability issues during autonomous execution. | |
| TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models Proposes a two-phase token pruning strategy for vision-language models and identifies evaluation flaws in common decode-phase benchmarks. Ranked #152Surprising resultPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Proposes token pruning to improve efficiency in vision-language models. | |
| EdgeLLM-Nano: Auditing Polar-Guided Key-Cache Compression and Consistency-Anchored Pruning on a 33.6 MB On-Device Language Model This paper audits key-cache compression and weight pruning techniques on a small edge-deployed language model testbed. Ranked #152Surprising resultPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Audits KV-cache compression and pruning on an ultra-compact on-device model. | |
| AM-Bench: Understanding Agents under Task Underspecification across Models and Harnesses This paper introduces a benchmark to measure how language model agents handle ambiguous or missing goals, context, and constraints. Ranked #152Surprising resultPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Studies how language model agents handle ambiguous or incomplete task specifications across harnesses. | |
| Consider, Then Choose: Order-Invariant Decisions in System One Models Runoff eliminates option-order bias in multiple-choice model predictions by scoring items independently before running balanced candidate comparisons. Ranked #152New way to see a problemPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Exposes option-order sensitivity in multiple-choice evaluations and proposes an invariant scoring approach. | |
| 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. Ranked #152Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Develops internal probes to detect agent sabotage, unverbalized deception, and model lying in deployment. | |
| ObsSpec: Accelerating Agentic Reinforcement Learning with Observation Speculation This paper introduces an observation speculation framework using small world models to accelerate reinforcement learning rollouts for tool-using agents. Ranked #153New way to see a problemPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Uses speculative execution with a small world model to speed up agentic model rollouts. | |
| What Makes a Good Visual Representation for Spatial Understanding in MLLMs? This paper analyzes how different visual representation types affect the 3D spatial reasoning capabilities of multimodal large language models. Ranked #153Surprising resultPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Investigates what visual representations best enhance 3D spatial understanding in multimodal large language models. | |
| ResearchFlow-Eval: Tracing the Execution Gap in End-to-End Multi-Claim Autoresearch This paper introduces an evaluation framework grounded in published papers to assess continuous multi-stage execution in automated research agents. Ranked #153Surprising resultPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Evaluates end-to-end scientific autoresearch agent systems executing multi-stage research workflows. | |
| Claim Framing as a Driver of LLM Sycophancy This paper studies how different framings of user claims induce sycophancy in large language models on factual questions. Ranked #153Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Establishes a systematic benchmark and evaluation methodology for probing LLM sycophancy. | |
| Searching for "Harmful Refusal": A Psychometric Audit of an AI Safety Benchmark Applies psychometric auditing to HELM Safety benchmarks to test whether harmful refusal functions as a single measurable attribute. Ranked #153Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Investigates the measurement and reliability of AI refusal behaviors for dangerous and policy-violating prompts. | |
| Beyond Adjacent Layers: Graph-Guided Layer Fusion for Compressing Large Language Models This paper introduces a graph-guided layer fusion method that identifies and merges functionally redundant non-adjacent layers to compress language models. Ranked #154Surprising resultPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Presents a non-adjacent layer fusion method to compress LLMs for efficiency. | |
| MHVF: Training-Free GUI Grounding via Verification-Guided Multi-Hypothesis Search This paper introduces a training-free framework combining multi-hypothesis search and local visual verification to ground interface instructions in graphical user interfaces. Ranked #154New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Uses vision-language models for visual grounding and verification across graphical user interface layouts. | |
| Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation This paper examines how single-agent long-term memory accumulates repeated historical claims and introduces a benchmark and framework to mitigate unverified persistence. Ranked #154New testPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Investigates long-term memory accumulation and retrieval errors in single-agent architectures. | |
| AM-Bench: Understanding Agents under Task Underspecification across Models and Harnesses This paper introduces a benchmark to measure how language model agents handle ambiguous or missing goals, context, and constraints. Ranked #154Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Presents a benchmark testing agent performance under goal, context, and constraint underspecification. | |
| Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict Proposes an evaluation framework to measure how LLM agents handle uncertainty and resolve conflicting knowledge during execution. Ranked #154New way to see a problemPeer-reviewed7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Analyzes agent reliability and failure to report uncertainty when encountering conflicting knowledge. | |
| Mind the Tail: Data Selection Through Sparse Concepts Uses sparse autoencoders to select diverse training samples and prevent underrepresenting tail concepts across vision and multimodal models. Ranked #155New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 17 Oct, 06:18 | On-topic Evaluates concept-based data selection directly on multimodal pretraining and vision models. | |
| Adaptive-Depth Is a Finishing Stage The paper proposes training looped transformers at variable depths only during a short finishing stage to reduce compute costs. Ranked #155Surprising resultPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Reduces training compute and enables adaptive-depth inference compute control in looped transformers. | |
| Instructions Are Evidence, Not Objectives: Measuring Outcome Alignment of Agents This paper introduces an approach and benchmark for evaluating whether language model agents can infer unstated true requirements instead of following incomplete instructions. Ranked #155New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Studies how agents infer implicit user intent and complete tasks despite incomplete instructions. | |
| ResearchFlow-Eval: Tracing the Execution Gap in End-to-End Multi-Claim Autoresearch This paper introduces an evaluation framework grounded in published papers to assess continuous multi-stage execution in automated research agents. Ranked #155Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces an evaluation benchmark measuring process-level execution in automated research systems. | |
| An Investigation of Model Coherence: Narrow Finetunes Contradict Themselves Under Resampling Evaluates model coherence using 175 specific questions to measure internal contradictions and identity conflation under answer resampling. Ranked #155Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Investigates self-contradictions and pathologies in finetuned models that undermine reliable and aligned behavior. | |
| From Transfer Probes to Expert Grouping: Evidence from LoRA Post-Training The paper examines using transfer probes to group business tasks when sharing Low-Rank Adaptation experts under fixed parameter budgets. Ranked #156Surprising resultPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Evaluates resource-constrained LoRA expert grouping strategies for efficient multi-line model serving. | |
| EgoAlign: Situated 3D Understanding without Situation Descriptions Develops an agent-centric 3D language model that grounds spatial understanding in physical poses without requiring text situation descriptions. Ranked #156Surprising resultPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Presents a 3D multimodal language model for situated spatial reasoning and question answering. | |
| Same Attack, Different Judge: A First Controlled Adversarial Comparison of Jev, LLM Judges, and Purpose-Built Safety Classifiers This paper conducts an adversarial evaluation comparing a probability-based safety judge against large language model judges and safety classifiers across multiple attack types. Ranked #156Surprising resultPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Provides a controlled adversarial comparison and evaluation methodology for safety classifiers and LLM judges. | |
| Interaction-as-a-Judge Presents Interaction-as-a-Judge, using computer-use agents to evaluate interactive web and desktop software applications through direct dynamic interaction. Ranked #156New testPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Uses computer-use agents interacting with applications over multiple steps to conduct evaluation. | |
| Is Memorization Context-Sensitive? Prefix-Based Extraction Beyond Isolated Prefixes Analyzes training data extraction risks in language models when prompts include contextual conditioning such as retrieved documents. Ranked #156Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Examines memorization and data extraction vulnerabilities in deployed models under RAG and contextual conditioning. | |
| When More Memory Fails: Hidden Feasibility Holes in Dynamic Tensor Rematerialization The paper investigates dynamic tensor rematerialization, identifying cases where increasing memory budgets causes training to unexpectedly run out of memory. Ranked #157Surprising resultPreprint6/10 suggested | Small and efficient models via arxiv · 15 Oct, 06:18 | On-topic Analyzes memory-budget eviction anomalies in Dynamic Tensor Rematerialization for efficient model training. | |
| A Concept-informed Interpretable Artificial Intelligence Framework for Breast Ultrasound Diagnosis Develops an interpretable AI framework for breast ultrasound diagnosis incorporating clinical standard descriptors and assessment workflows. Ranked #157New way to see a problemPreprint6/10 suggested | Multimodal models via arxiv · 15 Oct, 06:18 | On-topic Integrates structured BI-RADS clinical concepts with ultrasound image modeling for interpretability. | |
| Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation This paper examines how single-agent long-term memory accumulates repeated historical claims and introduces a benchmark and framework to mitigate unverified persistence. Ranked #157New testPreprint7/10 suggested | LLM evaluation and benchmarks via arxiv · 15 Oct, 06:17 | On-topic Introduces MEchoBench to evaluate long-term memory persistence and correction in agent systems. | |
| ECMR: Evidence-Aware Deliberation Control for Multi-Agent LLM Reasoning Presents ECMR, a dynamic evidence-aware controller that guides multi-agent deliberation steps to prevent false consensus and reduce redundant reasoning. Ranked #157New way to see a problemPreprint7/10 suggested | AI agents via arxiv · 15 Oct, 06:17 | On-topic Introduces an evidence-aware controller to manage deliberation, verification, and roles in multi-agent LLM systems. | |
| Measuring Artifact Reliance in Machine-Generated Text Detectors The paper demonstrates that machine-generated text detectors rely heavily on document formatting and PDF extraction artifacts rather than text quality. Ranked #157Surprising resultPreprint7/10 suggested | AI safety and reliability in deployment via arxiv · 12 Oct, 16:28 | On-topic Examines reliability, brittleness, and failure modes of machine-generated text detectors in deployment. |
1,021–1,080 of 3,333← PreviousNext →