More Accurate or More Efficient? Evaluating Locally Deployed Compact Open-Weight Language Models for Mathematical Reasoning
LinkedQwen 3.8 Max
Loading…
Radar
Which research topics are accelerating, and where earlier papers have already landed in shipped models.
Papers this week
50
Trending topic
Agents
▲ 22%papers this window vs prior window
Papers linked to models
137
linked by the desk, past 90 days
Median days paper → model
—
lower is faster
papers this window by topic · Δ vs prior window · a paper can carry more than one topic
| Topic | Papers this window | Prior window | Δ |
|---|---|---|---|
| Agents | 133 | 109 | +22% |
| Evaluation & Benchmarks | 39 | 27 | +44% |
| Multimodal | 26 | 33 | -21% |
Links are AI-inferred by the desk's LLM from tech reports, system cards and citations — each carries a confidence level and is not a claim by the authors.
Evaluation of Small Vision-Language Models on Qualitative Mechanical Problems
arXiv:2608.22143AI-inferred · highGenCoord: Skill-Path Commitments under Private Information
arXiv:2608.22055AI-inferred · highMore Accurate or More Efficient? Evaluating Locally Deployed Compact Open-Weight Language Models for Mathematical Reasoning
arXiv:2608.22048AI-inferred · highDisagree to Explore, Agree to Commit: Routing-Guided Test-Time Scaling for Software Agents
| 23 |
| 15 |
| +53% |
| Training & Optimization | 21 | 14 | +50% |
| Efficiency & Inference | 20 | 21 | -5% |
| Reasoning | 15 | 11 | +36% |
| Interpretability | 12 | 15 | -20% |
| Reinforcement Learning | 10 | 14 | -29% |
| Retrieval & RAG | 9 | 11 | -18% |
| Privacy & Security | 8 | 2 | +300% |
| Code Generation | 8 | 11 | -27% |
| Robotics & Embodied AI | 6 | 7 | -14% |
| Language Understanding | 5 | 4 | +25% |
| Generative Models | 5 | 15 | -67% |
| Speech & Audio | 5 | 6 | -17% |
| Mixture of Experts | 3 | 2 | +50% |
| Long Context | 3 | 5 | -40% |
| Computer Vision | 3 | 8 | -62% |
| Data & Synthetic Data | 1 | 4 | -75% |
| Other | 165 | 147 | +12% |
Safety & Alignment
Redteaming Leading Arabic LLMs with ASAS
arXiv:2608.21985AI-inferred · highTask-Driven 3D Printability Assistance via Geometry- and Knowledge-Grounded LLM Reasoning
arXiv:2608.22128AI-inferred · highMEMORY Wins All: Indirect Bias Injection Attacks via Social Media Feeds
arXiv:2608.22061AI-inferred · highHiMA-MDD: A Hierarchical Multi-Agent Harness for Interpretable Multimodal Depression Detection in Clinical Interviews
arXiv:2608.21868AI-inferred · highHints, Critics, and Teachers: Prior Injection for Sparse-Reward RL in Vision-Language Math Reasoning
arXiv:2608.21811AI-inferred · highHIRA: A Human-in-the-Loop Retrieval-Augmented Cascade for Document Classification in Regulated Industries
arXiv:2608.21792AI-inferred · highATHENA: Knowledge-guided agentic neural architecture search for AutoFormer-based electronic health record modeling
arXiv:2608.21712AI-inferred · high| Paper | Linked model | Relation | Confidence |
|---|---|---|---|
| Evaluation of Small Vision-Language Models on Qualitative Mechanical Problems (2608.22143) | Qwen 3.8 Max | evaluates | 98% |
| Evaluation of Small Vision-Language Models on Qualitative Mechanical Problems (2608.22143) | Gemma 4 | evaluates | 98% |
| GenCoord: Skill-Path Commitments under Private Information (2608.22055) | Qwen 3.8 Max | technique used | 98% |
| More Accurate or More Efficient? Evaluating Locally Deployed Compact Open-Weight Language Models for Mathematical Reasoning (2608.22048) | Qwen 3.8 Max | evaluates | 100% |
| Disagree to Explore, Agree to Commit: Routing-Guided Test-Time Scaling for Software Agents (2608.22191) | Qwen 3.8 Max | evaluates | 98% |
| Disagree to Explore, Agree to Commit: Routing-Guided Test-Time Scaling for Software Agents (2608.22191) | GPT-5.6 | evaluates | 98% |
| Redteaming Leading Arabic LLMs with ASAS (2608.21985) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 98% |
| Redteaming Leading Arabic LLMs with ASAS (2608.21985) | GPT-5.6 | evaluates | 98% |
| Task-Driven 3D Printability Assistance via Geometry- and Knowledge-Grounded LLM Reasoning (2608.22128) | Gemini 3.7 Flash | evaluates | 95% |
| MEMORY Wins All: Indirect Bias Injection Attacks via Social Media Feeds (2608.22061) | GPT-5.6 | evaluates | 95% |
| HiMA-MDD: A Hierarchical Multi-Agent Harness for Interpretable Multimodal Depression Detection in Clinical Interviews (2608.21868) | Qwen 3.8 Max | technique used | 99% |
| Hints, Critics, and Teachers: Prior Injection for Sparse-Reward RL in Vision-Language Math Reasoning (2608.21811) | Qwen 3.8 Max | evaluates | 98% |
| HIRA: A Human-in-the-Loop Retrieval-Augmented Cascade for Document Classification in Regulated Industries (2608.21792) | Qwen 3.8 Max | technique used | 99% |
| ATHENA: Knowledge-guided agentic neural architecture search for AutoFormer-based electronic health record modeling (2608.21712) | Athena-Brain-8B | technique used | 99% |
| Context as an Environment: Programmatic Context Management for Long-Horizon Agents (2608.21690) | Qwen 3.8 Max | evaluates | 95% |
| From Mastery Profile to Simulated Response: Stochastic Student Knowledge Graphs (SSKG) for Faithful LLM Student Simulation (2608.21668) | GPT-5.6 | evaluates | 98% |
| From Mastery Profile to Simulated Response: Stochastic Student Knowledge Graphs (SSKG) for Faithful LLM Student Simulation (2608.21668) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 98% |
| From Mastery Profile to Simulated Response: Stochastic Student Knowledge Graphs (SSKG) for Faithful LLM Student Simulation (2608.21668) | Gemini 3.7 Flash | evaluates | 98% |
| Let Credit Follow Computation: Architecture-Aware Credit Transport for Large Language Model Reinforcement Learning (2608.21501) | Qwen 3.8 Max | evaluates | 99% |
| KVBoost: Chunk-Level Key-Value Cache Reuse with Deviation-Guided Recomputation for Efficient Large Language Model Inference (2608.21362) | Qwen 3.8 Max | evaluates | 98% |
| K-Bench: measuring model performance on real scientific agent requests (2608.21601) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 100% |
| K-Bench: measuring model performance on real scientific agent requests (2608.21601) | GPT-5.6 | evaluates | 100% |
| IB-RL: Isolated Bilateral Reinforcement Learning for Strategic Dialogue Agents (2608.06735) | DeepSeek-V4-Flash-0731 | evaluates | 95% |
| NiyamAI - An Intent-Bound AI Agent with Cryptographically Verifiable Guardrails using Zero-Knowledge Proofs (2608.07167) | GPT-5.6 | evaluates | 98% |
| WebGrader: Training LLMs for Web Development with Self-Evolving Programmatic Grader (2608.06474) | GPT-5.6 | evaluates | 99% |
| Divergent Response Modes in Frontier Language Models Under Steering Pressure (2608.06578) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 99% |
| Divergent Response Modes in Frontier Language Models Under Steering Pressure (2608.06578) | GPT-5.6 | evaluates | 99% |
| Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory (2608.07169) | GPT-5.6 Luna | technique used | 98% |
| Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills (2608.07885) | GPT-5.6 Luna | evaluates | 98% |
| SodaMem: Evidence-Grounded Temporal Graph Memory for LLM Agents (2608.08055) | DeepSeek-V4-Flash-0731 | technique used | 98% |
| Who Verifies the Benchmark? Decentralizing Trust in Large Language Model Evaluation (2608.07762) | DeepSeek-V4-Flash-0731 | evaluates | 98% |
| Who Verifies the Benchmark? Decentralizing Trust in Large Language Model Evaluation (2608.07762) | GLM-5.3 | evaluates | 98% |
| Who Verifies the Benchmark? Decentralizing Trust in Large Language Model Evaluation (2608.07762) | GPT-5.6 Luna | evaluates | 98% |
| An Agentic AI Framework Overcomes Fundamental Limitations of Large Language Models for Glaucoma Detection from Fundus Photography (2608.07651) | GPT-5.6 Luna | evaluates | 99% |
| An Agentic AI Framework Overcomes Fundamental Limitations of Large Language Models for Glaucoma Detection from Fundus Photography (2608.07651) | Gemini 3.7 Flash | evaluates | 99% |
| When the Judge Should Not Decide: Evidence-Locked, Non-Compensatory Selection Bounds LLM-Judge Failure in Reasoning Pipelines (2608.07813) | DeepSeek-V4-Flash-0731 | evaluates | 95% |
| The Authority Expectancy Effect in Multi-User Conflict (2608.08026) | Grok 4.6 | evaluates | 95% |
| The Authority Expectancy Effect in Multi-User Conflict (2608.08026) | GPT-5.6 Luna | evaluates | 95% |
| The Authority Expectancy Effect in Multi-User Conflict (2608.08026) | Gemini 3.7 Flash | evaluates | 95% |
| The Authority Expectancy Effect in Multi-User Conflict (2608.08026) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 95% |
| Can Legal AI Know When It Is Wrong? And Do Students Know When It Is? (2608.21089) | GPT-5.6 Luna | evaluates | 99% |
| Calibrating Criterion Revision in LLM Agents: Failure Modes and a Trace-Anchored Protocol (2608.20729) | Qwen 3.8 Max | evaluates | 98% |
| Why2Speak: Faithful Reasoning for Abstaining Action Policies (2608.20670) | Qwen 3.8 Max | evaluates | 98% |
| No Judgment Without a Reason: Counterfactual Receipts for Versioned AI Evaluators (2608.20938) | Qwen 3.8 Max | evaluates | 99% |
| Automated Trajectory Evaluation for Mobile Agents via Step-Level Consequence Reasoning and Aggregation (2608.20797) | Qwen 3.8 Max | technique used | 98% |
| Beyond Endpoint Gains: A Weight-Delta Audit of Medical Specialization (2608.20768) | GPT-5.6 Luna | evaluates | 99% |
| Beyond Endpoint Gains: A Weight-Delta Audit of Medical Specialization (2608.20768) | Qwen 3.8 Max | evaluates | 99% |
| Beyond Endpoint Gains: A Weight-Delta Audit of Medical Specialization (2608.20768) | Gemma 4 | evaluates | 99% |
| Natural-Language-Guided Generator-Agnostic Shortlisting for Protein Binder Design (2608.20755) | GPT-5.6 Luna | technique used | 95% |
| DirEAG: Dirichlet Evidence Aggregation for Calibrating Verbalized Confidence in Mathematical Reasoning (2608.20717) | Gemma 4 | evaluates | 95% |
| DirEAG: Dirichlet Evidence Aggregation for Calibrating Verbalized Confidence in Mathematical Reasoning (2608.20717) | Mistral OCR 4 | evaluates | 95% |
| DirEAG: Dirichlet Evidence Aggregation for Calibrating Verbalized Confidence in Mathematical Reasoning (2608.20717) | Qwen 3.8 Max | evaluates | 95% |
| Nexus: Depth-Adaptive KV-Cache Splicing and Retrieval-Decoupled Tool Routing for Agentic LLMs on Unified Memory (2608.20397) | Qwen 3.8 Max | evaluates | 99% |
| Truth Lies Deep: Countering Semantic Camouflage via Latent Intent Verification (2608.20378) | Gemma 4 | evaluates | 98% |
| Truth Lies Deep: Countering Semantic Camouflage via Latent Intent Verification (2608.20378) | Qwen 3.8 Max | evaluates | 98% |
| PrimeAgentOrchestrator: Memory-Primed Agent Spawning for Personal AI Infrastructure (2608.20342) | claude-fable-5, claude-sonnet-5, claude-opus-5 | technique used | 98% |
| Personalized Privacy Control in LLMs via Attention Head Intervention (2608.21209) | Qwen 3.8 Max | evaluates | 99% |
| TreeWY: Speculative Verification for Gated DeltaNet Hybrids (2608.20961) | Qwen 3.8 Max | evaluates | 98% |
| Weighted Memory Tree: Remembering What Matters for Long-Horizon LLM Agents (2608.20631) | Gemma 4 | evaluates | 99% |
| Weighted Memory Tree: Remembering What Matters for Long-Horizon LLM Agents (2608.20631) | Qwen 3.8 Max | evaluates | 99% |
| FlavourBench: Ranking Frontier Language Models with Executable Culinary Ground Truth (2608.20574) | Grok 4.6 | evaluates | 99% |
| StateSight: Benchmarking Latent Spatial-State Reconstruction in Vision-Language Models (2608.20414) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 99% |
| StateSight: Benchmarking Latent Spatial-State Reconstruction in Vision-Language Models (2608.20414) | GPT-5.6 Luna | evaluates | 99% |
| SkillLens: Visual Skill Cards for Retrieval-Augmented GUI Action Prediction and On-Policy Distillation (2608.10775) | GPT-5.6 Luna | evaluates | 98% |
| CHORUS: Complementary Experts for High-Coverage Testbench Stimulus Generation (2608.10090) | DeepSeek-V4-Flash-0731 | evaluates | 98% |
| DSAgentBench: Can Agents Automate End-to-End Data-Science Workflows in Real Computer Environments? (2608.10366) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 98% |
| Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence (2608.11341) | GPT-5.6 Luna | evaluates | 98% |
| AutoWorldModel-Bench: A State-Centric Benchmark for Automated World-Model Research (2608.11216) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 98% |
| MBA: Multimodal Benchmark and Agents for Real-World Business Ideation (2608.11616) | GPT-5.6 Luna | technique used | 98% |
| When Self-Consistency Backfires: Majority Vote Hurts the Majority of Hard Science Problems for Small LLMs (2608.11403) | Qwen 3.8 27B | evaluates | 99% |
| XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication (2608.11676) | Qwen 3.8 27B | evaluates | 95% |
| Graph-Structured Rubrics: Compiling Rubrics into Typed Evaluation Graphs for LLM Judges (2608.12097) | GPT-5.6 Luna | evaluates | 95% |
| HUGIN: Enhancing Vision-Language Planning for Autonomous Logistics Sorting (2608.11692) | Qwen 3.8 27B | evaluates | 99% |
| Claim-Level Reliability Assessment for Efficient Test-Time Reasoning (2608.11994) | GPT-5.6 Luna | evaluates | 95% |
| From Numbers to Judgment: Specialist LLM Agents and Reinforcement Learning for European Listed Real Estate (2608.11381) | Qwen 3.8 27B | technique used | 98% |
| Retrofitting Recurrent Depth into a Pretrained Language Model: Installation, Extrapolation, Transfer, and Retention at Two Parameter Budgets (2608.11233) | Qwen 3.8 27B | technique used | 99% |
| Backtrader-Bench: Benchmarking LLM Agents on Algorithmic Trading with Self-Generated MCQs (2608.11232) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 95% |
| Backtrader-Bench: Benchmarking LLM Agents on Algorithmic Trading with Self-Generated MCQs (2608.11232) | GPT-5.6 Luna | evaluates | 95% |
| FrontierFinance: A Challenging Benchmark for Measuring Frontier Intelligence of Finance Agents (2608.11683) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 98% |
| Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence (2608.12036) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 86% |
| Harnessing agent memory to build lifelong AI partners for materials scientists (2608.11224) | GPT-5.6 Luna | evaluates | 95% |
| Can Frontier LLMs Match Natively Multimodal Embeddings? A Comparison on Hard-Negative Text-to-Image Retrieval (2608.11343) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 99% |
| Can Frontier LLMs Match Natively Multimodal Embeddings? A Comparison on Hard-Negative Text-to-Image Retrieval (2608.11343) | GPT-5.6 Luna | evaluates | 99% |
| An Agentic Workflow for Legacy HPC Modernization: Converting the Two-Electron-Integral Core of GAMESS (2608.12249) | claude-fable-5, claude-sonnet-5, claude-opus-5 | technique used | 95% |
| Distribird: Literature-Informed Prior Distribution Design for Bayesian Model Calibration (2608.11210) | Qwen 3.8 27B | evaluates | 95% |
| From Monolithic to Modular: Segment-level Automatic Prompt Optimization (2608.11219) | GPT-5.6 Luna | evaluates | 98% |
| RecSys Factory: Bounding LLM Agent Autonomy to Decision Points in the Industrial Recommender Lifecycle (2608.11241) | claude-fable-5, claude-sonnet-5, claude-opus-5 | technique used | 90% |
| Reasoning Jury: Multi-Model Consensus for Evaluating Reasoning Traces (2608.12585) | GPT-5.6 Luna | evaluates | 95% |
| Enhancing Virtual Agents through SLMs and Edge-Computing: An Exploratory Evaluation of Think and Memory Processes (2608.13420) | Qwen 3.8 27B | evaluates | 98% |
| Rethinking Normalization Placement for LLMs: Post-Norm under Curriculum Depth Growing (2608.13156) | Qwen 3.8 27B | evaluates | 95% |
| Jointly Predicting Courses and Grades Using a Transformer-Based Model (2608.13409) | trace-supervised symbolic neural CPU | technique used | 99% |
| Explanatory Engagement Under Rare Anomalous Failure: Asymptotic Rarity in Model Behavior (or: The Asymptotic AI) (2608.13063) | Qwen 3.8 27B | evaluates | 99% |
| ε-MemEvo: Adaptive Cross-Task Memory Transfer for LLM Program Evolution (2608.12522) | AlphaEvolve | related | 94% |
| ε-MemEvo: Adaptive Cross-Task Memory Transfer for LLM Program Evolution (2608.12522) | GPT-5.6 Luna | evaluates | 98% |
| Governed Persistent Memory: Source-Bound State Semantics and Fail-Closed Release for Long-Horizon Agents (2608.12476) | Qwen 3.8 27B | evaluates | 99% |
| Privacy-Preserving RAG by Concealing Sensitive Information from External LLMs (2608.12675) | Qwen 3.8 27B | evaluates | 95% |
| Polish Medical Visual Question Answering: Vision-Language Models Underutilize Visual Evidence (2608.12928) | GPT-5.6 Luna | evaluates | 99% |
| From Passive Delegates to Strategic Negotiators: Reinforcing Social Reasoning in Small Language Models with SocialRL (2608.13787) | GPT-5.6 Luna | evaluates | 95% |
| Explanation Multiplicity: Circuit-Level Interpretability Evidence Does Not Survive Defensible Analytic Variation (2608.13754) | GPT-5.6 Luna | evaluates | 99% |
| Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages (2608.14375) | GPT-5.6 Luna | evaluates | 95% |
| MemoryLake on MemoryArena: A Matched Study of Agent Memory Backends (2608.13883) | GPT-5.6 Luna | evaluates | 95% |
| Implementing Computational Law in Wolfram Language for the Governance of Artificial Intelligence (2608.13958) | GPT-5.6 Luna | evaluates | 100% |
| Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning (2608.14290) | Qwen 3.8 Max | related | 95% |
| StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (2608.15089) | DeepSeek-V4-Flash-0731 | evaluates | 99% |
| StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (2608.15089) | GPT-5.6 Luna | evaluates | 95% |
| Large Language Models Show Metacognitive Sensitivity in Medical Reasoning (2608.14552) | GPT-5.6 Luna | evaluates | 99% |
| When Agentic Executions Fail: Detecting and Localizing Runtime Faults from Telemetry (2608.14680) | DeepSeek-V4-Flash-0731 | evaluates | 98% |
| SKILL: Self-correcting Knowledge-guided Iterative Large Language Model Agent for Logic Optimization (2608.14579) | GPT-5.6 Luna | technique used | 98% |
| ACTS-SQL: Agentic and Critic-Oriented Tree-Structured SQL Correctness with Large Language Models (2608.15145) | GPT-5.6 Luna | evaluates | 95% |
| The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning (2608.14558) | GPT-5.6 Luna | evaluates | 98% |
| The Hallucination Snowball: Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines (2608.14588) | GPT-5.6 Luna | evaluates | 98% |
| LLMs Can Predict Failure Risk, But Struggle to Predict Which Collaboration Protocol Pays Off: Cost-Aware Protocol Routing Across Reasoning Tasks (2608.14927) | GPT-5.6 Luna | evaluates | 95% |
| Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability (2608.19338) | Qwen 3.8 Max | evaluates | 98% |
| Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability (2608.19338) | GPT-5.6 Luna | evaluates | 98% |
| When AI Writes, Who Gets Cited? Evidence of Citation Monoculture Across Language Models (2608.19230) | GPT-5.6 Luna | evaluates | 98% |
| MidTool: Mid-training Data Synthesis for Agentic Tool Use (2608.20314) | Qwen 3.8 Max | technique used | 98% |
| Automatic bioinformatic software named entity recognition from literature (2608.19201) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 90% |
| Automatic bioinformatic software named entity recognition from literature (2608.19201) | Grok | evaluates | 90% |
| Automatic bioinformatic software named entity recognition from literature (2608.19201) | Gemini 3.7 Flash | evaluates | 90% |
| Automatic bioinformatic software named entity recognition from literature (2608.19201) | GPT-5.6 Luna | evaluates | 90% |
| Can Conversational AI loosen Us-Versus-Them Boundaries? The Effects of Common, Dual, and Separate Identity Framings on Pro-Immigrant Intergroup Helping (2608.19220) | GPT-5.6 Luna | technique used | 99% |
| A Virtual Member of a Community of Practice for the Society of Petroleum Engineers: From Prototype to Deployment (2608.19199) | Athena-Brain-8B | evaluates | 98% |
| Phantom Gains: Auditing Self-Improvement Against a Measured Null (2608.20290) | Qwen 3.8 Max | evaluates | 98% |
| ReguSim: Evaluating LLM Agent Rule Grounding in Financial Compliance (2608.19974) | Gemini 3.7 Flash | evaluates | 95% |
| ReguSim: Evaluating LLM Agent Rule Grounding in Financial Compliance (2608.19974) | DeepSeek-V4-Flash-0731 | evaluates | 95% |
| PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents (2608.19861) | Gemini 3.7 Flash | evaluates | 99% |
| PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents (2608.19861) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 99% |
| PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents (2608.19861) | GPT-5.6 Luna | evaluates | 99% |
| SAPO: Single-Rollout Autoregressive Policy Optimization for Agentic Reinforcement Learning (2608.19842) | Qwen 3.8 Max | evaluates | 98% |
| Can Agent Memory Systems Track Evolving State? (2608.19652) | Qwen 3.8 Max | evaluates | 98% |
| Can Agent Memory Systems Track Evolving State? (2608.19652) | DeepSeek-V4-Flash-0731 | evaluates | 98% |
| From Retrieved Context to Runtime Control: Adaptive Compression for Edge-based RAG (2608.19535) | Qwen 3.8 Max | evaluates | 99% |
| Air Traffic Control Using Large Language Models: Prompt Engineering, Architecture, and Evaluation (2608.19299) | GPT-5.6 Luna | evaluates | 80% |
| When Personalization Becomes Bias: Structural and Discursive Religious Framing in AI-Generated Financial Advice (2608.16909) | Grok | evaluates | 98% |
| Different Facets of Verbalised Overconfidence: an Interpretability Study (2608.18106) | Qwen 3.8 Max | evaluates | 99% |
| DeepTCM1.0: A Multi-Expert AI Agent for Deciphering Mechanisms of Chinese Herbal Formulae Based on General Large Language Models (2608.18103) | DeepSeek V4-Flash | technique used | 99% |
| Abliteration Mitigation via Refusal Aliases (2608.18093) | Gemma 4 | evaluates | 99% |
| Nine Emotion Centroids: A Label-Free Valence Axis That Transfers Across Four Modalities (2608.18090) | Gemma 4 | evaluates | 85% |
| Nine Emotion Centroids: A Label-Free Valence Axis That Transfers Across Four Modalities (2608.18090) | Qwen 3.8 Max | evaluates | 85% |
| Nine Emotion Centroids: A Label-Free Valence Axis That Transfers Across Four Modalities (2608.18090) | Mistral OCR 4 | evaluates | 85% |
| Latent Space Refusal Anchoring for Low-Resource African Languages: Mechanistic Safety Recovery Without Retraining (2608.18089) | Qwen 3.8 Max | evaluates | 99% |
| Latent Space Refusal Anchoring for Low-Resource African Languages: Mechanistic Safety Recovery Without Retraining (2608.18089) | Mistral OCR 4 | evaluates | 99% |
| Beyond the Transcript: Detecting Covert Co ordination in Latent Multi-Agent Communication (2608.19161) | Qwen 3.8 Max | evaluates | 98% |
| Can a Lightweight Multimodal Model Estimate LLM Reasoning Performance? A Study for Compute-Optimal Document Inference (2608.18591) | Qwen 3.8 Max | technique used | 100% |
| A Jagged Frontier: Evaluating Robustness of Code Agents to Semantics-Preserving Transformations (2608.18389) | Qwen 3.8 Max | evaluates | 98% |
| A Jagged Frontier: Evaluating Robustness of Code Agents to Semantics-Preserving Transformations (2608.18389) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 98% |
| Governance Records as Supervision: Verifier-Selected Self-Training for Structured Workflow Repair (2608.18324) | Qwen 3.8 Max | evaluates | 98% |
| Efficient Adaptation of LLMs for Hate Speech Detection in Low-Resource Languages: A Comparative Study on Roman Urdu (2608.18142) | Mistral OCR 4 | evaluates | 99% |
| Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions (2608.18078) | DeepSeek V4-Flash | evaluates | 95% |
| Computational Orientalism: Measuring Structural Discourse Bias in Large Language Models Using the Middle East Cultural Sensitivity Score (MECSS) (2608.18100) | GPT-5.6 Luna | evaluates | 99% |
| Breaking the weakest link to evade vision language models (2608.18938) | Qwen 3.8 Max | evaluates | 99% |
| Training-Free Inference-Time Self-Reflection and Cost-Bounded Early Stopping for Large Language Models (2608.18884) | Qwen 3.8 Max | evaluates | 98% |
| CTIFoundry: An Agent-Native Corpus Scaffold for Cyber Threat Intelligence (2608.18613) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 80% |
| FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents (2608.18423) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 100% |
| Measuring the Partial-Credit Gap: A Strict Benchmark on Vietnam's 2025 Convex Marking Scheme (2608.18336) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 90% |
| Measuring the Partial-Credit Gap: A Strict Benchmark on Vietnam's 2025 Convex Marking Scheme (2608.18336) | Qwen 3.8 Max | evaluates | 98% |
| Solving Is Not Drawing: A Benchmark for Diagrammatic Reasoning in Olympiad Geometry (2608.18111) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 80% |
| Solving Is Not Drawing: A Benchmark for Diagrammatic Reasoning in Olympiad Geometry (2608.18111) | GPT-5.6 Luna | evaluates | 80% |
| ComponentBench: Diagnosing Component-Level Failures in Computer-Use Agents (2608.18307) | Qwen 3.8 Max | evaluates | 99% |
| ComponentBench: Diagnosing Component-Level Failures in Computer-Use Agents (2608.18307) | Gemini 3.7 Flash | evaluates | 99% |
| ComponentBench: Diagnosing Component-Level Failures in Computer-Use Agents (2608.18307) | GPT-5.6 Luna | evaluates | 99% |
| Potential of ChatGPT in predicting stock market trends based on Twitter Sentiment Analysis (2311.06273) | GPT-5.6 Luna | evaluates | 95% |
| StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents (2608.18050) | GPT-5.6 Luna | evaluates | 98% |
| StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents (2608.18050) | Gemini 3.7 Flash | evaluates | 98% |
| Auditing Self-Evolution in Financial Agents: Capability Gains, Security Drift, and Execution-Interface Mismatch (2608.17684) | Qwen 3.8 Max | evaluates | 98% |
| The Price of Thinking: Reasoning Effort as a Model-Specific API Contract (2608.16956) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 95% |
| Beyond the Trace: Coupling an Interpretable Reasoning-State Readout to Native MoE Routing (2608.17638) | GPT-5.6 Luna | evaluates | 95% |
| TRUSS: Towards Task-Reliable and User-Safe Automated Agent Skill Generation (2608.17588) | GPT-5.6 Luna | evaluates | 95% |
| Agent Lightning v1.0: Towards Harnessed Agentic RL (2608.17528) | Qwen 3.8 Max | evaluates | 99% |
| LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents (2608.17393) | Qwen 3.8 Max | evaluates | 99% |
| TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration (2608.17336) | Qwen 3.8 Max | evaluates | 98% |
| SignalReasoner: Assessing the Upper Bound of 3B Models for Signal Mathematical Reasoning (2608.17301) | Qwen 3.8 Max | evaluates | 98% |
| When Personalization Becomes Bias: Structural and Discursive Religious Framing in AI-Generated Financial Advice (2608.16909) | Grok 4.6 | evaluates | 99% |
| When Personalization Becomes Bias: Structural and Discursive Religious Framing in AI-Generated Financial Advice (2608.16909) | Gemini 3.7 Flash | evaluates | 98% |
| When Personalization Becomes Bias: Structural and Discursive Religious Framing in AI-Generated Financial Advice (2608.16909) | GPT-5.6 Luna | evaluates | 98% |
| FedPref: Federated Preference Learning for Structured Radiology Report Extraction (2608.16971) | Qwen 3.8 Max | technique used | 99% |
| Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification (2608.17247) | GPT-5.6 Luna | evaluates | 98% |
| GxP-Agent: Process-DAG Topology for Reliable Clinical Trial Programming with LLM Agents (2608.16890) | GPT-5.6 Luna | evaluates | 99% |
| GxP-Agent: Process-DAG Topology for Reliable Clinical Trial Programming with LLM Agents (2608.16890) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 99% |
| StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (2608.15089) | DeepSeek V4-Flash | evaluates | 99% |
| StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (2608.15089) | GPT-5.6 Sol | evaluates | 99% |
| Does a Tool Result Carry More Authority Than Plain Text? Three Prospective Studies of False-Claim Adoption in a Synthetic Assignment Task with Claude Opus 5 (2608.14992) | claude-fable-5, claude-sonnet-5, claude-opus-5 | evaluates | 100% |
| Small Models Scout Bottleneck Order for Large-Model Data Control (2608.14936) | Qwen 3.8 Max | evaluates | 95% |
| LLMs Can Predict Failure Risk, But Struggle to Predict Which Collaboration Protocol Pays Off: Cost-Aware Protocol Routing Across Reasoning Tasks (2608.14927) | GPT-5.6 Sol | technique used | 98% |
| The Hallucination Snowball: Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines (2608.14588) | Qwen 3.8 Max | evaluates | 98% |
| The Hallucination Snowball: Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines (2608.14588) | GPT-5.6 Sol | evaluates | 99% |
| SKILL: Self-correcting Knowledge-guided Iterative Large Language Model Agent for Logic Optimization (2608.14579) | Gemini 3.7 Flash | technique used | 98% |
| SKILL: Self-correcting Knowledge-guided Iterative Large Language Model Agent for Logic Optimization (2608.14579) | claude-fable-5, claude-sonnet-5, claude-opus-5 | technique used | 98% |
| SKILL: Self-correcting Knowledge-guided Iterative Large Language Model Agent for Logic Optimization (2608.14579) | GPT-5.6 Sol | technique used | 99% |
| Anatomy of a Quantized Agent: VRAM Stability and Forecasting in Code-Synthesis Agentic Workloads (2608.15117) | Qwen 3.8 Max | evaluates | 99% |
| When Agentic Executions Fail: Detecting and Localizing Runtime Faults from Telemetry (2608.14680) | DeepSeek V4-Flash | evaluates | 99% |
| ACTS-SQL: Agentic and Critic-Oriented Tree-Structured SQL Correctness with Large Language Models (2608.15145) | GPT-5.6 Sol | evaluates | 95% |
| ACTS-SQL: Agentic and Critic-Oriented Tree-Structured SQL Correctness with Large Language Models (2608.15145) | GPT-5.6 Sol | technique used | 95% |
| The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning (2608.14558) | Gemini 3.7 Flash | evaluates | 98% |
| The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning (2608.14558) | GPT-5.6 Sol | evaluates | 98% |
| Large Language Models Show Metacognitive Sensitivity in Medical Reasoning (2608.14552) | GPT-5.6 Sol | evaluates | 99% |
| MemoryLake on MemoryArena: A Matched Study of Agent Memory Backends (2608.13883) | GPT-5.6 Sol | technique used | 95% |
| From Passive Delegates to Strategic Negotiators: Reinforcing Social Reasoning in Small Language Models with SocialRL (2608.13787) | GPT-5.6 Sol | evaluates | 99% |
| Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning (2608.14290) | Qwen 3.8 Max | technique used | 85% |
| Implementing Computational Law in Wolfram Language for the Governance of Artificial Intelligence (2608.13958) | GPT-5.6 Sol | evaluates | 99% |