Recent research advances AI efficiency, reasoning, and safety with significant performance gains. LRMoE reduces training time by 21.2–44.8%, while SkillMem cuts storage needs by 38×. Reasoning capabilities improve via ThinkJev, achieving 81.6% accuracy with 72% token savings, and WISD, which boosts accuracy by 4.0–5.5 points. Safety enhancements include LLMON, which achieves near-zero prompt injection rates, and Disparity laundering, reducing bias gaps by 67–94%.
Benchmarks expose persistent limitations in agent reliability and domain-specific tasks. OpenFrontiers finds closed-book rediscovery remains rare, while LibStructBench shows agents struggle with molecular reconstruction. Financial hallucinations persist at rates under 1%, and causal reasoning benchmarks like CLadder prove brittle. New methods address these gaps: GARDEN increases rewards by 52–58%, CHORD detects hallucinations, and ATP secures agent communication.
Verification tools such as JevAct and VerifiedLab enhance system reliability, though studies note LLMs often exploit superficial patterns rather than true logic. Domain-specific applications show promise, with MoLRSI achieving a 15.9% antibiotic hit rate and ProACT enabling VR training adaptation. CALM introduces conflict-aware memory mechanisms to further refine agent behavior in complex environments.
Key Takeaways
- LRMoE cuts training time by 21.2–44.8%
- SkillMem reduces storage requirements by 38×
- ThinkJev achieves 81.6% reasoning accuracy
- WISD boosts accuracy by 4.0–5.5 points
- LLMON reaches near-zero prompt injection rates
- Disparity laundering reduces bias gaps by 67–94%
- GARDEN increases rewards by 52–58%
- Financial hallucinations remain under 1%
- Closed-book rediscovery is rarely observed
- Verification tools improve overall reliability
Sources
- LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes
- OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery
- ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions
- Evaluating and improving AI agents for sequencing protocol understanding
- Measuring Where Intent Gets Lost in Agentic Smart Contract Synthesis
- The Anatomy of Socio-Demographics in Pluralistic Offensiveness Modeling: When and Where It Helps, and the Role of Reasoning
- What Constitutes Food Similarity? Visual Representations and Retrieval in Indian Cuisine
- GARDEN: Generalist Autonomous Reward Design Through Evidence Navigation for Cooperative Multi-Agent Reinforcement Learning
- SciGame: Let Agents Play the Scientific Discovery Games in the Real and Counterfactual World
- Beyond Low-Pass Dynamics: Frequency-Selective Spiking Reservoirs with Resonant Neurons
- Algorithmic Bias from the Perspectives of Healthcare Professionals
- SkillMem: Compact External Skill Memory for Latent Multi-Agent Reasoning
- Evaluating Causal Reasoning in LLMs: A Case Study in the Challenges of Procedural Generation
- PromptSense: Sentiment-Aware Mental Health Chatbot with Dynamic Prompting
- Too Hard, Too Easy, Just Right: Performance-Driven Challenge Adaptation in Virtual Reality Firefighter Training
- AI-Driven Digital Twin Framework for Warehouse Logistics Automation Optimization
- SkillCT: A Multidimensional Benchmark and Execution-Free Baseline for Agent Skill Assessment
- Decompose, Don't Expand: Capability-Independent AI Systems by Construction
- Post-Training Large Language Models with LLMON
- Disparity Laundering: Consensus Without Fairness in Clinical Multi-Agent Systems
- From Natural Language to Real Schematics: An Automated Workflow for Power Electronics Schematic Synthesis
- Prototype-Guided Dynamic Domain Modulation and a Large-Scale Multi-SKU Benchmark for Shampoo Bottle Surface Defect Detection
- Investor's Last Exam: Can AI Agents Do Equity Investment Without Seeing the Future?
- Can LLMs Faithfully Solve Ontology Engineering Tasks? A Mechanistic Interpretability Study
- Game-Theoretic Foundations for Undiscounted Bisimulation Metrics
- CHORD: Cross-Model Hallucination Detection via Relational Graph Discrimination
- SCOPE: Selective Critical-Step On-Policy Self-Distillation for Agentic Reinforcement Learning
- Compute Has Shape: Self-Morphing the Internal Computation of Frozen LLMs
- Encode Once, Answer Many: Internalizing Context via Parametric Flow
- DES-JEPA: Learning a Shared Vocabulary for High-Dimensional Discrete Event Sequences
- Predict the Reward, Not the World: World Reward Models for ML Agents
- Agent Transfer Protocol: Server-Mediated Agent-to-Agent Communication Architecture for the Agentic Web
- From Leniency to Suspicion: How an Open Post-Training Recipe Changes Whom LLM Fact-Checkers Believe
- Value Probes: Learned Value Functions for Beam-Search Red Teaming
- CALM: Conflict-Aware Lifecycle Memory Governance for Long-Horizon Language Agents
- A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement
- Co-Evolving Policy and Rubric via Capability Awareness for Open-Ended Generation
- PT$^3$: Prototype-Guided Self-Supervised Pretraining for Generalizable mmWave Human Point Cloud Representation
- Weaken to Strengthen: Weakening-Induced On-Policy Self-Distillation for Large Reasoning Models
- Curriculum-Based Deep Reinforcement Learning for Robust Autonomous Rocket Landing under Atmospheric Uncertainty
- "Close enough?" An investigation into multiplication errors in LLMs
- Right Numbers, Wrong Questions: Evaluating LLMs on Underspecified Financial Queries
- Leaving, Staying, or Living There? Duration Commonsense Reasoning with Human-LLM Co-annotations
- Where You Inject Matters: The Trust Region Effect in LLM Agents
- FIRME: Registry-Anchored Memory Graphs for Evolving, Heterogeneous Investor Knowledge
- Aligning Autoresearch: Co-Improvement between Agents and Humans
- Auditable Belief Updates as a Missing Safety Layer for LLM-Powered Multi-Agent Systems
- Efficient Agentic Audio-Visual Question Answering through Toolset Pruning and Modality Consolidation
- Syntactic Segmentation and Collage Generation for Videogame Levels with Perceptual Distance
- Dungeons & Dragons Encounter Analysis with Semi-Random Simulation in Foundry VTT
- Beyond LLM Wrappers: Context Engineering and the Three-Engine Architecture for Reliable Financial AI
- Reasoning or Pattern Exploitation? Mechanistic Insights from RL-Trained Language Models
- Certified Access: Test-Time Compute on Conservative Physics-Structured AI Memory
- When the Same Prompt Gives Different Answers (and the Same Answer Isn't Always Right): A Taxonomy of Failure Mechanisms, Testing, and Mitigation Strategies for LLM-Assisted Evidence Synthesis Read-and-Evaluate Tasks
- When Should AI Decide, Defer, and Deviate? Responsible Automation for Deceased Donor Kidney Allocation
- Formal Specifications with Data-Driven Signal Analysis
- Proving and Improving the Platonic Representation Hypothesis with JEPAs
- What Does "Now" Mean to Your Computer-Use Agent? Stale Screenshots, Page Loads, and Which Check to Run Before a Click
- MemSteer: Turning User Histories into Supervision for Agent Memory Optimization
- The Role of Self-Concept In Language Models
- AgentDecept: Benchmarking Deceptive Behavior in Tool-Using LLM Agents under Pressure
- CoToan: Executable Rules and a Representation Contract for Reproducible Game AI Evaluation
- Does Federated Training Degrade Explainability? A Comparative Analysis of Grad-CAM Faithfulness in Centralized vs. Federated Deepfake Detection
- Kairos: Semantics-Aware Orchestration of Planning Engines across Classical, Numeric, Temporal and Hybrid PDDL
- Timbre: Reliability-aware harmonic consensus for label-free multimodal test-time rexinforcement learning
- Verifiable LLM-Guided Focal SMT Solving for Quantified Arrays
- Category-Anchored Large Language Models for Next POI Recommendation
- Compile Away Uncertainty: Self-Evolving Agent Skills via Programmatic Compilation
- Emergence of the transformer architecture as the score of entropic laws of random measures
- From Intent to Improvement: Agentic Dataset Search for Better Model Training
- MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery
- Letting AI Formulate the Research Question: Autonomous Research Agents
- Finding Is Not Returning, Returning Is Not Reconstructing: A Multi-Dimensional Evaluation of Agent Social Re-entry
- From the Transformer to a Formal Semantics of Coordinated Transformations
- EvoTeam: Multi-Agent Collaboration via Evolving Teams
- Human--AI Interactions as First-Class Data: A Formal Interaction Model and Runtime for Explainable AI Collaboration
- Efficient Learning of Mechanism Designs and Compositionality
- KARMA: Training-Free Knowledge-Atomic Harness Evolution for Industrial Advertising Agents
- The generative paradox: the Anti-Prompt paradigm for human capability
- Reward-Adaptive Co-Evolution of Policy and Rubric
- Agentic process evidence
- Bridging Semantics and Structure: A Theoretically Grounded Framework for Link Prediction on TAGs
- A Negotiation-Aware Hybrid LLM Framework for Intent Interpretation and Service Discovery
- Beyond Selected Breakthroughs: Evaluating Language Models on Randomly Sampled Mathematical Conjectures
- Multistep and Multivariate Time Series Explainability with GradientSHAP: A Wind Power Case Study
- Neighborhood Elites: Quality-Diversity Search for Uncertainty Quantification in Question Answering
- Training language models in simulation for psychiatric intake
- Prefix Executability: Evaluating Tool-Using Agents Beyond Final Success
- Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
- Advancing Clinical Genomic Analysis with a Skill-Centred Agent
- Retrieval as Traversal: Value-Indexed Memory for Stateful Long-Context Agents
- On the Fundamental Limits of Prefill-Based Decoding Length Prediction
- Efficient Scalable Oversight via Rewinding
- Algebraic World Model: From Exploration to Proofs
- Agentic University: Collaborative Agents for Personalized Learning
- Pre-Crime for Agents, at the Last Second: Probing When a Recruited Agent Gives In
- Comparing Human Role-Players and LLM-Simulated Clients in Online Counselling Training: An Analysis of Counselling Patterns
- Hybrid AI Patterns: Thirteen Ways to Couple Language Models with Rules, and How to Choose Among Them
- Cheap Retrieval and Observation Credit Assignment from One Quantity: The Value of Information
- AegisFlow: Training-Free Non-Myopic Path-Safe Guided Flow Matching
- Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective
- Long-Term Agent Memory through Complementary Curation and Consolidation
- Non-Zero Radius: The Missing Component of Neural Networks
- DEG: Diffusion-Enhanced GFlowNet for Solving Vehicle Routing Problems
- What Breaks When Facts Become Stories? The Distortions Source-Agreement Metrics Miss When Models Retell Real Events
- Integrated Feature Selection for classification and Representation Learning with Group-Lasso Siamese Networks
- Learning to Reason from Reference Samples Alone: Reinforced Adversarial Learning for Modern LLM Post-Training
- Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?
- Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live IOS Apps
- Where to Learn and How Far to Roll Out: Efficient Adaptive Training for GUI Agents
- SocialArena: Evaluating Social Intelligence in Large Language Models
- Goal-Aware Dynamic Multi-Agent Collaboration for Long-Horizon Tasks
- Reaction precedents extend retrosynthetic prediction beyond training coverage
- From Grind to Grasp
- Beyond fabrication: Misattributed numbers in papers written by AI research agents
- Deogen: Automatic Benchmark Generation for Reasoning over Policies
- Think Before You Act: Decision Gate for LLM Agents
- Governance-Aware Federated Learning for Multi-Agent AI Systems: A Policy-Driven Framework for Trustworthy Collaborative AI
- Collaborative AI agents for autonomous experimental research with solid-state quantum spins
- VeriCUA: Verification as Agentic Exploration for Computer-Using Agents
- FARMS: A Verifiable Foundation for Automated Computational Science
- Ramanujan Challenge Results: AI-Assisted Evaluation Pipeline for Multi-Format Proofs
- EnvOS: Auditing Computer-Use Agents When the World Moves On : Phenomenon-Based Task Construction, Certification, and Evidence-Based Evaluation
- EpiAgent: Epistemic Self-Play for LLM Agents
- Locating and Closing Information Gaps in Multi-Agent Systems
- DynaRev: Benchmarking Web Agents under Dynamic Goal Revision
- ATLAS-Bench: An Agentic Tool-Use Benchmark for Local Area Search
- Retaining Repair Hypotheses for Cost-Aware Execution
- Before the Next Rollout: Diagnosing Skill Utility Shifts from Policy Updates
- LLM-Assisted Optimization for Simulation-Based Redesign
- An Alignment Benchmark for Enterprise Agents on Era by Eon
- Accelerating Nonlinear MPC by Behavior Cloning with Separated Timescales: Application to Nuclear Reactor Load-Following
- Verification-Guided Retrosynthesis: Reaction Feedback and Evidence Timing
- FACTS-MAS: A Multi-Agent Architecture for Adaptive Forecasting from Heterogeneous Evidence
- Rubricon: Automated Rubric Generation and Refinement for LLM Evaluation and Training
- GAAC: Lossless KV Compression for Bandwidth-Efficient LLM Inference on Personal Devices
- DecompGym: Benchmarking LLM Agents on Designing Decomposition Algorithms
- BridgeRouter: Gain-Aware Instance Routing over Cross-Modal Bridge Portfolios
- Understanding Large Language Models as Numerical Optimizers
- VerifiedLab: Harnessing Knowns for Unknowns
- FilingLoop: Evaluating and Evolving LLM Agents for Corporate Disclosure Review and Revision
- Verification Learning Theory: A Process Model of Checking and Retention in AI-Assisted Judgment
- Seq@K: Evaluating Language Models Under Sequential Feedback
- TreasuryQA Bench: Can AI Agents Work an Archive of Scanned Financial Records?
- Masked Erosion: How Acceptance Gates Lose Redundant Safeguards in Self-Improving Agents
- Assessing Language Models for Scientific Applications: Inertial Confinement Fusion as a Use Case
- GAUSS: A Workload-Aware Analytical and Monte-Carlo Simulator for LLM Inference Serving
- Detecting Covert Collusion between AI Agents through Partner-Outsider Positive Listening
- What Does an Ontology Catch? Curated, LLM-Generated and Data-Derived Ontologies as Checkers of LLM Agents over Knowledge Graphs
- PAPERBENCHX: Can AI Reproduce Science Across Disciplines?
- Bank-World: A Benchmark for Evaluating Agents on Banking Workflows
- The Real Cost of Context: Predicting Compute Time and Energy Consumption in LLM Prefill and Decode Phases
- Inner-Learning: From Hand-Coded Beliefs to Declarative Knowledge in Trick-Taking Games
- WebAnything: From the Wild Web to Scalable Agent Environments
- Amortized Localized Posterior Inference for Model Calibration with Application to an Ocean Turbulence Closure
- Distributional Reward Calibration and Budget-Neutral Credit Assignment for Multimodal RLVR
- Agentic AI Reasoning and Learned Sampling for Task-Aware Projection Selection in Sparse-View CT: Predicting Hemodynamics
- Context Compaction for World Models
Comments
Please log in to post a comment.