Researchers have made significant advancements in various fields, including reinforcement learning, natural language processing, and computer vision. One of the key findings is the development of PivoARL, a self-feedback retry framework for experience exploitation in large language model agents. PivoARL achieves significant improvements on Pass@2/3 across all tasks, with an average gain of about 11.5% over MetaRL. Additionally, the framework consistently improves Pass@1 on over 80% of the tasks. Another notable development is the introduction of Oyster-II, a reinforcement learning-based constructive safety alignment framework that adopts a Zero-RL paradigm combined with a multi-stage reinforcement learning strategy. Oyster-II comprehensively surpasses both Qwen3-14B and its predecessor Oyster-I on safety dimensions, achieving cross-scale performance comparable to Qwen3-Max and Qwen3.5-397B. Furthermore, researchers have proposed a novel bias mitigation approach called CO-ALIGN, which operates on the model's internal concept ontology and achieves substantial bias reduction while preserving generative integrity. CO-ALIGN outperforms the state of the art, improving fairness by 30%, ΔFID=11.4 in image quality, 2.8% in image fidelity, and reducing semantically incoherent outputs by 88%.
In the field of natural language processing, researchers have made significant progress in developing large language models that can generate coherent and context-specific text. One of the key findings is the development of iFLYTEK-Embodied-Omni, a unified multimodal foundation model that jointly models vision, language, and action within a single Omni framework. iFLYTEK-Embodied-Omni achieves IsoDDE-level co-folding accuracy within a reproducible and openly accessible framework. Additionally, researchers have proposed a novel framework called MedCalc-Pro, which covers three progressively challenging task settings: single-calculator, multi-calculator, and nested-calculator calculation settings. MedCalc-Pro contains 2,268 real-world clinical cases, covering 77 medical calculators across 14 clinical departments. The framework achieves the best performance across all three task settings.
In the field of computer vision, researchers have made significant progress in developing models that can accurately detect and classify objects in images. One of the key findings is the development of FM-ChangeNet, a pathwise-supervised framework for change detection that reformulates bi-temporal reasoning as continuous transport in feature space rather than static endpoint comparison. FM-ChangeNet produces more structured and robust change representations while achieving state-of-the-art performance. Additionally, researchers have proposed a novel framework called Graph Sparse Sampling (GSS), which shares sampled futures across many candidate decisions, rather than sampling separate successors for each candidate action. GSS substantially outperforms tree-based planners on long horizons or achieves near-optimal performance, supporting no-branching graph planning as a complementary design principle for online control.
Key Takeaways
- PivoARL achieves significant improvements on Pass@2/3 across all tasks, with an average gain of about 11.5% over MetaRL.
- Oyster-II comprehensively surpasses both Qwen3-14B and its predecessor Oyster-I on safety dimensions, achieving cross-scale performance comparable to Qwen3-Max and Qwen3.5-397B.
- CO-ALIGN achieves substantial bias reduction while preserving generative integrity, outperforming the state of the art by 30%, ΔFID=11.4 in image quality, 2.8% in image fidelity, and reducing semantically incoherent outputs by 88%.
- iFLYTEK-Embodied-Omni achieves IsoDDE-level co-folding accuracy within a reproducible and openly accessible framework.
- MedCalc-Pro achieves the best performance across all three task settings, covering single-calculator, multi-calculator, and nested-calculator calculation settings.
- FM-ChangeNet produces more structured and robust change representations while achieving state-of-the-art performance.
- Graph Sparse Sampling (GSS) substantially outperforms tree-based planners on long horizons or achieves near-optimal performance, supporting no-branching graph planning as a complementary design principle for online control.
- LLM-as-a-Verifier achieves state-of-the-art performance on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%).
- SovereignPA-Bench evaluates user-owned personal agents under evolving intent, platform mediation, privacy boundaries, consent constraints, evidence requirements, and burden tradeoffs.
- ClassicLogic is a new benchmark suite designed to evaluate an agent's ability to learn and compose problem-solving strategies, with a hierarchical, explicit knowledge base for each game.
Sources
- Agent Reinforcement Learning via Pivotal-Aware Self-Feedback Retry
- Explainable Reinforcement Learning for Adaptive Traffic Signal Control
- ASK in the Dark: Uncertainty-Gated LLM Assistance under Partial Observability
- iFLYTEK-Embodied-Omni Technical Report
- MedCalc-Pro: Solving Complex Medical Calculations with LLM Agents
- SwarmResearch: Orchestrating Coding Agents for Open-Ended Discovery
- Efficient bias mitigation in T2I diffusion models using Concept Graphs
- Reflective Dialogue or Prompt Refinement? Effects of Tutor Scaffolding on Students' Independent LLM Use for Programming
- APeB: Benchmarking Personalization Ability of Large Language Model Agents
- Oyster-II: Reinforcement Learning for Constructive Safety Alignment in Large Language Models
- Object-Centric Environment Modeling for Agentic Tasks
- Automated Data Readiness for Scientific AI
- Human-Centric Reflective Architecture for Human-AI Collaborative Decision-Making
- Internal Pluralism and the Limits of Pairwise Comparisons
- MoP-JEPA: Hard-Assigned Predictor Mixtures for Stochastic JEPA World Models
- Rethinking On-Policy Self-Distillation for Thinking Models
- AgentGym2: Benchmarking Large Language Model Agents in De-Idealized Real-World Environments
- MechMath Agent Team: LLM Driven Agents for Mathematical Research
- Evaluating Generative Agents with Actions Grounded in Socially Distributed Task Environments using Incognita
- A Sliding-Window-Based Reinforcement Learning for Dynamic Assembly Flow Shop Scheduling with Multi-Product Delivery
- VERITAS: Towards a General-Purpose Replication Tool for Scientific Research
- Beyond Forecasting: The Belief-to-Trade Layer in Prediction-Market Agents
- Reinforcement Learning for Evidence-Seeking Diagnostic Reasoning with Large Language Models
- A Clustering-Based Framework for Identifying Suspicious Trading Patterns in Capital Market
- Language models guide symbolic equation discovery by controlling search
- PLACEMEM: Toward a Compute-Aware Memory Plane for Lifelong Agents
- Explainable AI for Screening Abuse-Related Trauma in Bangladeshi Children: A Training-Free Multimodal Framework Evaluated on Noise-Aware Synthetic Data
- Online Linear Programming for Multi-Objective Routing in LLM Serving
- Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine
- MentalThink: Shaping Thoughts in Mental SVG World
- Graph Sparse Sampling: Breaking the Curse of the Horizon in Continuous MDP Planning
- Reason, Reward, Refine: Step-Level Errors Corrections with Structured Feedback for Physics Reasoning in Small Language Models
- DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
- TacReasoner: A Dynamic Tactile-Language Framework for Interactive Reasoning in Real-World Scenarios
- STAPO: Selective Trajectory-Aware Policy Optimization for LLM Agent Training
- VLA Grounder: Language-Conditioning Space Optimization for Black-Box VLA Models
- Decentralized Aggregation of LLM Predictions via Wagering Mechanisms
- Quantum-Inspired Harmonic Decision Models: A Computational Framework for Music Generation
- FM-ChangeNet: Learning Change through Pathwise Feature Transport
- Measuring Harness-Induced Belief Divergence in Multi-Step LLM Agents
- Progress- and Reliability-Oriented Group Policy Optimization for Agentic Reinforcement Learning
- Agentic IoT: Architectures, Applications, and Challenges Toward the Internet of Agents
- Demonstrating Generalization Failures via Mixtures of Conditional Policies
- Silicon Sampling via Cross-Survey Transfer
- Embodied Operators and Benchmarking: Toward Reusable and Deployable Embodied Intelligence Systems
- Organizational Memory for Agentic Business Process Execution
- When Aggregate Alignment Misleads: Auditing Policy Repair Without Per-State Expert Actions
- Personalized Causal Recourse: A Human-In-The-Loop Approach
- From Mobile Data to Business Insights: An End-to-End Analytics Framework for Large-Scale Urban Mobility Analysis and Decision Support
- Robust Feasible Route Construction through Collaborative Partition Optimization
- The Role of Rigor in Artificial Intelligence
- Applying Answer Set Programming with Fuzzy Membership Functions: a Case Study
- How to Avoid Debate: Scalable AI Safety via Doubly-Efficient Interactive Proofs
- Harness-Aware Self-Evolving: Co-Evolving Model Weights, Harness, and Task Solutions
- Can Conversational Temporal Dynamics Improve Depression Detection in Dyads? A Preliminary Investigation in Multi-Modality Perspectives
- Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth
- Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process
- Forethought: Verifiable Reasoning from Neurosymbolic Primitive Programming
- What is Left for Us? Second Scholarship Against the Degradation of Research by AI
- Unsupervised Features Mining via Activation Geometry
- Agentic SABRE: An Uncertainty-Aware Neuro-Symbolic Multi-Agent Framework for Adaptive Ransomware Detection
- Shortcut Learning in Legal Judgment Prediction: Empirical Evidence from the UK Employment Tribunal
- Biological Motifs for Agentic Control
- Nemotron-Labs-3-Puzzle-75B-A9B: Compressing Hybrid MoE LLMs
- Do GUI Agents Believe Their Eyes? Diagnosing State-Belief Reliance on Pixels versus Structure
- HAS-Bench: Evaluating LLM-Based Human-Agent Systems under Configurable Human Participation
- LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL
- Server-side Anti-cheat in FPS games for Aimbot detection using Deep learning and Machine learning
- Compressing the Validation Bottleneck: An Agentic Self-Driving Lab for Scientific Discovery
- ResearchStudio-Idea: An Evidence-Grounded Research-Ideation Skill Suite from ML Conference Outcomes
- Agent Step Value: State-Transition Measurement with State-Grounded LLM Evaluators
- Why Pure Reasoning is Not Enough: Nature as the Source of Mathematical Innovation
- Detecting Answer-Driven Reasoning in LLM-Based Educational Tutors via Truncated Chain-of-Thought Auditing
- Attention Limited Reward Learning
- Governed Individuation: Cryptographically Decoupling an Agent's Learning from Its Authority
- MRMS: A Multi-Resolution Memory Substrate for Long-Lived AI Agents
- Heaviside Continuity of Rolling Coefficients for Eliminating Epistemic Entropy in Large Language Models
- Formal Disco: Scalable Open-Ended Generation of Formally Verified Programs
- Integrated Altruistic and Fairness Preference Induces Advanced Mutual Cooperation in Sequential Social Dilemmas
- FORGE: Research-Trajectory Hijacking Attacks on Deep Research Agents
- Medi-Gemma: A Hybrid Clinical Decision Support System Integrating Deterministic EMR Analytics and Retrieval-Augmented Generation
- CARL: Constraint-Aware Reinforcement Learning for Planning with LLMs
- AgenticPD: A Stage-Aware Agentic Framework for Physical Design QoR Optimization
- ASSEMCAD: Production-Ready CAD Assembly Generation from Natural Language
- Diffusion-Guided Uncertainty-Aware Delayed Policy Optimization
- Toward Trustworthy Large Language Model Agents in Healthcare
- CP-WSP: A Declarative CP-SAT Framework for Configurable Multi-Constraint Workforce Scheduling
- The Changing Role of Symbolic Methods in Artificial Intelligence
- EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer
- ClassicLogic: A Knowledge-Driven Benchmark of Classic Puzzle Games for Evaluating Compositional Generalization
- OptiAgent: End-to-End Optimization Modeling via Multi-Agent Iterative Refinement
- MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution
- Evaluating and Understanding Model Editing for Medical Vision Language Models
- LLM-as-a-Verifier: A General-Purpose Verification Framework
- SovereignPA-Bench: Evaluating User-Owned Personal Agents under Evolving Intent, Platform Mediation, and Consent Constraints
Comments
Please log in to post a comment.