Process Reward Agents Advance Reasoning While OpenKedge Enhances Safety

New research introduces Process Reward Agents (PRA) for dynamic, step-wise rewards in knowledge-intensive reasoning, improving accuracy by up to 25.7% on medical benchmarks without policy updates (arXiv:2604.09482). OpenKedge protocol redefines API-centric agent mutations as governed processes, ensuring safety and auditability through an Intent-to-Execution Evidence Chain (arXiv:2604.08601). LOM-action equips enterprise AI with event-driven ontology simulation for grounded, auditable decisions, achieving 93.82% accuracy and outperforming baselines in tool-chain F1 by a four-fold margin (arXiv:2604.08603). In marketing, autonomous agents sustained a positive lift in engagement metrics over an 11-month period, suggesting a symbiotic model where human intervention initializes and agents preserve performance gains (arXiv:2604.08621).

Advancements in reinforcement learning for LLMs include Sequence-Level PPO (SPPO), a scalable algorithm harmonizing PPO's sample efficiency with outcome-based stability for long-horizon reasoning tasks (arXiv:2604.08865), and Stability-Augmented Reinforcement Policy Optimization (StaRPO), which incorporates reasoning stability metrics like Autocorrelation Function and Path Efficiency to enhance both accuracy and logical coherence (arXiv:2604.08905). A tutor-student multi-agent framework (PETITE) enhances LLM problem-solving by structuring interactions, achieving similar or higher accuracy with significantly fewer tokens on coding benchmarks (arXiv:2604.08931).

New benchmarks and evaluation methods are emerging: PilotBench evaluates LLMs on safety-critical flight trajectory prediction, revealing a precision-controllability dichotomy and a dynamic complexity gap in high-workload phases (arXiv:2604.08987). DRBENCHER generates synthetic benchmarks for questions requiring both web browsing and multi-step computation, highlighting limitations in systems reasoning over evolving data (arXiv:2604.09251). SAGE, a multi-agent benchmark, formalizes Standard Operating Procedures into Dynamic Dialogue Graphs for assessing service agents, revealing an 'Execution Gap' where models fail to derive correct actions despite accurate intent classification (arXiv:2604.09285). Spatial-Gym evaluates spatial reasoning as a sequential decision task, showing models struggle with scaling reasoning effort and are hindered by visual input (arXiv:2604.09338). HiL-Bench measures selective escalation skills, revealing a universal judgment gap in frontier models regarding when to ask for help (arXiv:2604.09408). SEA-Eval evaluates self-evolving agents beyond episodic assessment, identifying significant evolutionary bottlenecks and token consumption inefficiencies (arXiv:2604.08988).

Research also explores foundational aspects of agentic systems and reasoning. Artifacts in the environment can functionally serve as an agent's memory, reducing the information needed to represent history (arXiv:2604.08756). Visual-to-symbolic analytical solution inference (ViSA) models can recover analytical solutions from field visualizations, outperforming baselines with a physicist-like reasoning pipeline (arXiv:2604.08863). Parameterized complexity results show models of MSO2 formulas can be represented with decision diagrams whose size is parameterized linear in treewidth (arXiv:2604.08707). Humans exhibit a dual transition in physical planning under resource pressure, shifting both prediction mechanisms and planning strategies (arXiv:2604.09072). Hypergraph Neural Networks accelerate Minimal Unsatisfiable Subset enumeration by minimizing satisfiability checks (arXiv:2604.09001). Advantage-Guided Diffusion for MBRL steers diffusion processes using advantage estimates to improve long-term return (arXiv:2604.09035). Camera Artist, a multi-agent framework, generates narrative videos with explicit cinematic language, improving shot-to-shot continuity and filmic quality (arXiv:2604.09195). Constraint-Aware Corrective Memory (CACM) improves drug discovery agents by localizing protocol violations and biasing actions toward correction (arXiv:2604.09308). A single point-based multi-objective search framework (SPMO) focuses on finding a single high-quality solution rather than approximating the entire Pareto front (arXiv:2604.09417). LLMs exhibit both primary and strategic algorithmic monoculture in coordination games, regulating action similarity in response to incentives (arXiv:2604.09502). Enhanced Experience Exploitation (E3-TIR) improves tool-integrated reasoning by dynamically integrating expert prefixes, expert guidance, and self-exploration (arXiv:2604.09455). RAMP is a strategy for online learning of numeric planning action models via interaction (arXiv:2604.08685). Model space reasoning as search in feedback space aids planning domain generation from natural language (arXiv:2604.08712).