Researchers have made significant progress in understanding the modular architecture of large language models (LLMs), with a study finding that LLMs develop a modular architecture that mirrors the human brain. This modular organization may be a fundamental principle of how intelligent systems must be built, or an evolutionary accident specific to biological brains.
A new study has proposed a framework for evaluating learning harnesses end-to-end without a labeled benchmark, grounded in the scaling hypothesis. The framework scores a harness by how much its student converges toward the teacher over time, and has shown that improvement relative to the teacher correlates with improvement relative to a held-out gold standard.
Researchers have introduced a benchmark for the anchoring effect in LLMs, which evaluates multiple anchor pathways under an explicit anchor relevance axis. The study found that anchoring is strongly pathway-dependent, and that plausible anchors usually induce larger shifts than irrelevant ones when introduced through stronger pathways.
A new study has proposed a framework for generating grounded multimodal reports, which integrates images, tables, and text with supporting references in a unified report. The framework has been evaluated on a human evaluation study, which showed that the figures' informativeness is perceived as superior to that of a recent baseline in 87% of cases.
Key Takeaways
- Large language models (LLMs) develop a modular architecture that mirrors the human brain.
- The scaling hypothesis can be used to evaluate learning harnesses end-to-end without a labeled benchmark.
- The anchoring effect in LLMs is strongly pathway-dependent and can be evaluated using a benchmark that considers multiple anchor pathways.
- A framework for generating grounded multimodal reports has been proposed, which integrates images, tables, and text with supporting references in a unified report.
- The figures' informativeness is perceived as superior to that of a recent baseline in 87% of cases.
- LLMs can be used to generate grounded multimodal reports that are more informative than those generated by recent baselines.
- The proposed framework for generating grounded multimodal reports has been evaluated on a human evaluation study.
- The study found that the framework generates reports that are more informative than those generated by recent baselines.
- The framework has been shown to be effective in generating grounded multimodal reports that are more informative than those generated by recent baselines.
- The study found that the proposed framework for generating grounded multimodal reports is a significant improvement over recent baselines.
Sources
- Modular Cognitive Architecture Emerges in Large Language Models
- Depth-Aware Sensitivity Analysis of Mixture-of-Experts Models via Magnitude-Based Expert Masking
- Inducing Reward-Free Judging Rubrics that Reduce Over-Crediting in Agent Evaluation
- Measuring Cross-Task Behavioral Consistency in Language Model Agents
- Stable Miscalibration in Large Language Models: A Practical View of High-Confidence Errors
- AI Evaluation Should Work With Humans
- No Universal Signal Predicts Sample-Level LLM Regression under Version Updates
- MobileMem: Learning from a Year of Mobile Experiences
- Active Perception for Embodied Disambiguation
- How Compliant is Sepsis Treatment? An Expert-Guided Neuro-symbolic Pipeline for Generating Clinical Compliance Insights
- Evaluating Agentic Learning Harness Capabilities Without Labels via the Scaling Hypothesis
- Exploring ESC Winners with Nested Diagrams
- A Calibrated Test of Internal Action Maps: State Signals Without Global Affine Closure
- Your Probabilistic JEPA Is Secretly a Hidden Markov Model: A State-Space Interpretation of Joint-Embedding Predictive Learning
- Reward Machines for Signal Temporal Logic
- ARC: Fair Relative Advantage Comparison in Open-Ended Real-World Interaction
- From Passive Delegates to Strategic Negotiators: Reinforcing Social Reasoning in Small Language Models with SocialRL
- Simulation-Aware In-Context Policy Improvement for LLM-Aided Analog Layout Refinement
- Explanation Multiplicity: Circuit-Level Interpretability Evidence Does Not Survive Defensible Analytic Variation
- Learning to Assemble Novel Structures with Unfamiliar Parts under Semantic Constraints
- MemoryLake on MemoryArena: A Matched Study of Agent Memory Backends
- Joint Optimization of Memory and Computing Frequency for Energy-Efficient DNN Inference
- SDO: Subspace Deconflicting Operator for Multi-Adapter Composition
- AI Research Preference Models
- The Past and Future of AI Scientists
- LLMs Don't Pay for the Jump
- Tripwire: Triggering Aligned Refusal via Statistically Certified Safety Neurons
- AgentRewind: Recoverable Execution for Long-Horizon LLM Agents
- Removing Temporal Note Redundancy Improves Multimodal Reinforcement Learning for Medicine
- Towards Efficient Multimodal and Multilingual Opinion Extraction for STI: A QLoRA-Based Fine-Tuning Approach
- Traj-LeWM: Path-Aware World-Model Planning via Latent Trajectory Cost
- Nanbeige4.2-3B on Apple Silicon: Fixing Deployment Bugs and Decreasing Looped Transformer Memory Overhead
- Algorithm Design and Physician Liability
- SemPlan: Benchmarking Structured Semantic Planning for LLM-Based Queries over Enterprise Data
- Never the Number: Structural Abstention for AI Systems Whose Answers Are Consumed as Fact
- Split the Labor: Separating Evidence Interpretation from Decision Aggregation
- A Pathway to General-Purpose Scientific AI: Multimodal Comprehension of Scientific Images
- Reinforcement Learning-Based Production Scheduling in an Industry-Based Coating Scenario Using the Digital Model Playground
- A Graph-Based Reinforcement Learning Framework for Structured Drift Diagnosis and Recovery in Autonomous LLM Agents
- Retrieval Grounding Latent Reasoning for Dense Retrieval
- QuaSAR: Quantization Compensation via Stable Activation-Aware Rank Truncation
- Can Language Models Understand mmWave Data? Benchmarking Large Language Models for mmWave Radar-Based Human Understanding
- BiasTrace: Linking Reasoning Behaviours to Biased Outputs in LLMs
- A Generalized Parallelogram Rule for Proportional Analogies on Riemannian Manifolds
- APTER: Adaptive Post-Training with Expert-Grounded Rubrics
- FreeBalance: Pre-Routing Online Moe Load Balancing via Residual Workload Prediction
- Grounding Without Corrective Control: Truth-Tracking Profiles for Large Language Models
- Attributing Preprocessing Invariance in Spectral Foundation Models
- Scaling Domain Data Repetition in LLM Pretraining
- Coverage Aware Active Evaluation for Failure Discovery with Paired Systems
- Second Thought: Reasoning in Parallel as LLM Agents Act and Observe
- Sensor-Driven Mission Synthesis for UAV/UGV Swarms: A TB-CSPN Coordination Architecture with Hardware-Enforced Safety
- Program-space Diffusion for Morphology-to-Transcriptomics Prediction
- Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages
- Designing Sustainable Federated Learning as a Service using Neural Architecture Search
- Disentangled Shared Representations Improve Morpho-Transcriptomic Integration
- ScienceFlow: A long-horizon agent for ML research, scientific discovery and beyond
- Clearing the Fog: Towards Installing and Refining Proactive Exploration Capabilities in LLM Agents
- Handover of In-Context Learning State Across Session Boundaries
- Participatory Moral AI Is Not Neutral: The Invisible Hand of Developers
- Shift Aware Transfer Learning with Adaptive Dual-Encoder Fusion for PM Forecasting in Data-Limited Environments
- Wyvern: An Agentic Framework for Generating Grounded Multimodal Reports
- AnchorBench: A Multi-Pathway Benchmark for the Anchoring Effect in LLMs
- Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
- Cross-Disciplinary Taxonomy and Modeling of Misunderstanding Generation, Amplification, and Detection, from Pragmatics to AI Agents
- Agentao: A Governed Local-First Runtime for Tool-Using LLM Agents
- A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing
- Content Depth Matters in Short-Video Recommendation: Rethinking the Attention Economy
- Implementing Computational Law in Wolfram Language for the Governance of Artificial Intelligence
- HELIX: Model-Harness Co-evolution for Recursive Self-Improvement
- Demystifying Agent Skills: Why They Work-Until They Don't
- Agent-Orchestration in Autonomous Chip Design
- Buy the Rumor, Sell the News: When Is News Priced In?
- The Dynamics of Intelligence Explosions
- Knowing When to Stop: Bayesian Optimal Stopping for LLM Evaluations
- Ontology-Grounded Project Memory for Coding Agents
- Twin: Playing an Unknown Game with a Test-Time Digital Twin
- SheetCompass: Hierarchical Relation Graphs for Agentic Spreadsheet Reasoning
- PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments
- TimeSage-EV: A Live Benchmark for Agentic Time Series Analysis in Evolving Environments
- Polaris : Multi Agentic System for Conversational Enterprise Analytics
- MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement
- Regime-Conditional Verification: Correctness Estimation for Adapting and Monitoring Safety Classifiers
- Mandato: Protocol-Level Enforcement of Digitally Signed Mandates on AI Agent Actions with Cryptographically Chained Audit Trails
- Benchmarking data-driven material models on the classic Treloar dataset
- Residual Dominance as a Structural Account of Last-Item Reliance in Causal Self-Attention Recommenders
- Simulation-Driven Vehicular Traffic Data Augmentation: Extending Sensor Coverage Through Virtual Sensing
- FLARE MCMC: Fidelity-based Layer-Adaptive REcursive proposals for MCMC
- When Personal Memory Has No Single Answer: Evaluating LLM Agents under Irreducible Conflict
Comments
Please log in to post a comment.