Recent advances in LLMs and agents leverage architectural innovations like LangDiT and Co^4, specialized frameworks such as OntoAct and RSIFlow, and efficiency strategies including VISTA and LoopAhead. LangDiT achieves state-of-the-art perplexity of 21.72 GenPPL, while RAG-RST boosts nuclearity F1 to 0.846 without training, though performance saturates beyond 50 exemplars. Reasoning accuracy improves by 4.71 points via SIF integration.
Safety and reliability studies demonstrate significant gains, with VISTA delivering 5.35x speedups alongside toxicity reduction and StepSafe reaching a Youden's J of 53.0. However, RealtimeBench indicates only 44% success in live environments, and Verus auditing reveals that feedback can suppress evidence discovery. Quantum ML research matches classical counterparts but shows no consistent performance gains.
Robustness investigations expose critical vulnerabilities, such as Believed but Not Said uncovering hidden belief collapse and Diviner introducing human-governed root cause analysis. Limitations persist in benchmark transcripts failing to generalize across distribution shifts. Disagreements remain regarding harness efficacy in AlphaDiana and the criticality of cost estimation in Constrained Multi-Agent RL.
Key Takeaways
- LangDiT achieves SOTA perplexity of 21.72 GenPPL.
- RAG-RST boosts nuclearity F1 to 0.846 without training.
- SIF improves reasoning accuracy by 4.71 points.
- VISTA delivers 5.35x speedups with toxicity reduction.
- StepSafe reaches a Youden's J of 53.0.
- RealtimeBench shows only 44% success in live environments.
- Feedback suppresses evidence discovery in Verus auditing.
- Quantum ML matches classical counterparts with no consistent gains.
- Believed but Not Said exposes hidden belief collapse.
- Benchmark transcripts fail to generalize across shifts.
Sources
- Designing Transformers for Continuous Diffusion Language Models
- RAG-RST: Retrieved Exemplars as In-Context Discourse Schema for RST Parsing
- CARA: Causally Guided Retrieval Agent for Question Answering
- OntoAct: Hypothesis-Based Neuro-Symbolic Agents through Ontological Abstraction
- RSIFlow: Adaptive Self-Improvement Workflows with an Evolving Meta-Agent
- Constrained Multi-Agent Reinforcement Learning in Imagination
- AI-Driven Disaster Response and Displacement Monitoring
- Learning to Reason with Selective Internal Feedback
- ARGUS: An Agenda-Driven Framework for Reliable User Simulation
- Quantum Versus Classical Machine Learning for Neuropsychiatric Disorders: A Feature-Matched Multi-Cohort Evaluation
- Probe-and-Resume: When and How to Evaluate Agent Execution
- ARIA: Runtime Inspection and Amendment for Multi-Agent Systems
- When Reality Outpaces Cognition: A New Scaling Principle for Reliable Intelligence
- L2G-SR: Learning to Guide LLM-based Symbolic Regression
- Planning Commitment of LLM under Partial Observability
- DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents
- Functional Subnetwork Localization for Efficient Continual Learning in Large Language Models
- Decomposing Brain-LLM Alignment with Interpretable Semantic Representations
- GNN-Guided Local Refinement for Efficient Layout Decomposition with DSA-MP
- FlashAttention with Statistics: Low-Overhead Attention Features for Compact Hallucination Detectors
- The Density of Thought: Predicting and Steering Reasoning in Large Language Models
- MemScope: Evolving Agent Memory Architectures through Localized Code Repair
- X-RIC: An Agentic RAN-Intelligent Controller Benchmark for Enhancing Reinforcement Learning
- Who May Release an Agent's Promise? Standing over Commitments in Multi-Party Chat
- In-Context Preference-Conditioned Pareto Set Learning from Limited Offline Data
- Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion
- The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock
- When Correctness Feedback Erases Evidence: Phase-Separated Completeness Auditing in Verus
- AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks
- IneqAgent: Agentic Inequality Proving in Lean with Symbolic Computation
- Counterfactual Neighborhood Risk Fields: Evidence-Grounded Selective Safety for Vision-Language Models
- When Reasoning Traces Conflict with Evidence: How Answer Configuration Shapes Evidence Following
- EvolvingBrain: An Adaptive Skill System for Multi-modal Agents Self-Evolution in Open Worlds
- Open-world AI: a Tale of Three Requirements
- A Multi-Tiered AI Accessibility Model (MAIAM): Proportionate Access and Capability Limits for Comprehensive Frontier AI Safety
- Anticipatory Commonsense with Moving Objects: An ASP-Based Framework for Predictive Reasoning with Interacting Trajectories
- Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents
- StepSafe: Learning Timely Intervention with Context-Aware Safety Reasoning
- MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments
- RAC: A Relational Agent Contract for Constraint-Governed Engineering Tasks
- SkillNEO: Non-Elitist Evolutionary Optimization of Agent Skills
- Learning When to Withdraw: Curriculum Coverage and Rule-Specific Failures in Fine-Tuned Language Agents
- Failure Is the Mother of Success: Recovering Supervision with Grounding Hindsight
- Deep Reinforcement Learning for HPC Job Scheduling: A Statistical Evaluation
- MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data
- OmicsBench: a long-horizon multi-turn benchmark of AI agents on research-study-scale bioinformatics
- MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning
- Kolab: Benchmarking Human-Agent Collaboration in Long-Horizon Workflows
- All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory Agents
- CleanRM: Reward Modeling from Noisy Feedback with a Few Clean Feedback for RLHF
- GRADS: Graph-Regulated Agentic Data Synthesis with Controllable Difficulty
- Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models
- AgentEval: Evidence-Bounded Assessment of Stateful Agents
- Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally
- Condition-Contrastive Extrapolation for On-Policy Distillation of Language Models
- Auditing Escalation Reliability in Healthcare LLM Agents: A Perturbation-Controlled Benchmark of Workflow-State Failures
- AutoOnboarder: Autonomously Onboarding LLM Agents
- DocPilot: Interactive Agent Orchestration and Training for Long Visual Document Understanding
- Money-Mule Detection as a Relational Learning Problem: A Study on the RBIH NFPC Dataset
- Bias Control and Injection for Multi-Agent Aggregation
- Self-Guided LLM Reasoning for Error Diagnosis in Vehicle Routing Problems
- Hierarchical Graph Convolutional Network Approach for Detecting Low-Quality Documents
- FedSkillGit: Federated Skill Evolution via Version Control
- Jev-LiteMAS: Separating Decisions from Generation in Multi-Agent Workflows
- Watts If? A Benchmark for Counterfactual Reasoning About the AI Data Center Buildout
- Incentivizing Tool-Use Reinforcement Learning for Post-Arrival Emergency Care Prediction
- Do Agents Need a New Kind of Paper? SPRA: Structured Paper Representation for Agents
- ISW-Agent: Integrating LLM-Driven Dialogue with n8n Workflow for Internal Solitary Wave Remote Sensing Image Processing
- MetaExp: Learning to Self-Improve through Experience via Verifiable Cross-Task Environment Scaling
- CriticaLLM: Using LLMs as Autonomous Game Master for Tabletop Role-Playing Games
- SCRIBE: Scalable Contextual Representations of Interpretable Brain-Grounded Semantics
- QREF: Improving DeepSearch Agents by Aligning Search with Reasoning via Query Refinement and Evidence Filtering
- ClaimScope: Label-Blind Witness Verification for Mathematical Claim Auditing
- Reinforcement Learning-driven Diffusion Models: Language, Vision, and Beyond
- Agent Status Protocol (ASP): A Lightweight Specification for State Reporting in Multi-Agent Systems
- Learning Intrinsic Dynamics in Latent World Models
- LoopAhead: Speculating Ahead for Faster Decoding in Looped Language Models
- The Self-Improvement Problem
- From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research
- CoWorld Reasoner: Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness
- RoleQ: Role-Aligned Scales for Converting Language Models to Ternary
- CIM-Embed: Learning Instruction-Conditioned Multimodal Embeddings for Clustering
- Co2In: a Bi-level Memory Incremental Learning Framework with Knowledge Encoding, Consolidation, and Integration
- Prism-1: Position- and Rival-Independent Scoring for System One Models
- FAOPs: Normal Forms and an Algebra for AI Agent Workflows
- GraphACM: Graph-based Action Consistency Measurement for Uncertainty Quantification in LLM Agents
- CART: Compatible Routing for Multimodal Continual Instruction Tuning
- DropKV: Decoupling Residual-Output Perturbation for Near-Optimal KV-Cache Eviction
- PROBE: Learning to Audit Policy Compliance in Tool-Using LLM Agents
- Learning from Unverified Peers via Shared Residual Distillation
- Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine
- Multi-sourced AI Tutor for Non-Native English Speaker Students
- Realtime-Bench: Can Frontier Models Solve Real-time Interactive Problems?
- AutoSFT: An Agentic System for Automated Data Research in Supervised Fine-Tuning
- HTN-Orchestrate: A Hierarchical Task Network Framework for Multi-Agent Orchestration
- Disagreement-Guided Weak-to-Strong Routing
- Shortcut Sensitivity in Agent-Safety Evaluations
- Private Retrieval Augmented Generation via Random Projection
- EliteReplay: Preserving Diverse Solutions in RL Post-Training via Quality-Diversity Search
- Improving Environment-Agnostic Coding Agents: The Role of Scaffolding
- FDE-Gym: The Last Mile of AI Delivery
- Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?
- KnowGPO: Knowledge-Grouped Policy Optimization for Search Agents
- MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering
- A team of Gemini 4 Argon agents beats Minecraft
- LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents
- Are You Sure Your LLM Is Sure? Repairing and Auditing Confidence
- Believed but Not Said: Measuring Hidden Belief Collapse in Multi-Agent LLM Discussions
- When Reasoning Is Not Enough: Proactive Inquiry for Cross-Modal Safety in Multimodal Large Language Models
- Progress Reward Model for Reinforcement Learning via Large Language Models
- Data World: Scaling Data Environments from Scratch
- AgentPerturb: Evaluating LLM Agent Robustness under Controlled Interaction Perturbations
- Evaluating Inquiry Learning in LLM-Based Agents: Hypothesizing, Reasoning, and Exploration
- LOADBench: A Large Open Analog Dataset and Benchmark for Machine Learning in IC Design
- Tool-Augmented Large Language Models for Reliable Medical Calculation from Patient Notes
- Ontology Refinement with Embeddings and LLMs: A Case Study on CASTON
- Learning from Interaction Experience via Localized Context Distillation
- JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs
- MemCut: Every Secret Leaves a Trail in the Long-Term Memory of LLM Agents
- Mallea: Building, Steering, and Memorizing Worlds with Coding Agent
- Diviner: Agentic RCA with Human-Governed Adaptation for Large-Scale Infrastructure
- Rubric Judgments Lack Uncertainty: Conformal Weighting for LLM Evaluation
- Does Latent Chain-of-Thought Actually Reason? A Single, Unifying Taxonomy for Latent Reasoning
- Sci$\Pi$vot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents
- Athanor: An Agent-Native System for Autonomous Post-Training Research
- Recursive Reward Improvement via Adversarial Meta-Prompt Evolution
- LEV: Thinking Fast and Deep with Looped Decision Models
- Defining and Generating Compositional Structure
- Confidence-Calibrated Inference Expansion for Evaluator-Guided Test-Time Reasoning
- Temporal Specification Guidance for Text-to-Video Generation
- A Self-Improving Agent that Grows its Own DSL: Oracle-Free Operator Induction for Exact Question Answering
- Evaluating Model Capabilities After Benchmark Saturation
Comments
Please log in to post a comment.