A new arXiv study (2608.19760v1) auditing step-level credit assignment in LLM agents against causal ground truth from executed replay in ALFWorld found that none of the credit signals used to train LL…
Researchers propose PlanPO, a group planning-aware policy optimization method that improves multi-turn agentic large language models by distinguishing advantages among successful trajectories based on…
ReAct agents, which loop through reasoning and acting by having a model call tools and feed results back into its prompt, reported an absolute success-rate gain of 34% over imitation and reinforcement…
Researchers introduce CODS, a critic-guided data selection method for reusable offline reinforcement learning, which at a 10% budget retains 96.6% of eligible-pool performance across 20 D4RL task-algo…
Researchers introduced EvoHarness-RL, a framework that trains long-horizon LLM agents to learn runtime harness policies for constructing and coordinating external state, achieving 96.9% success on ALF…
Researchers propose NeSyFS, a neuro-symbolic fast-slow thinking framework for LLM agents under partial observability, using a knowledge graph for belief state representation and a twisted sequential M…
Researchers propose CAST (Credit Assignment from Solver Teachers), a method that converts game solver state-value changes into turn-level signals to improve reinforcement learning for LLM agents in lo…
Researchers propose Progress-conditioned Group Policy Optimization (ProGPO), a method that uses first-visit observation coverage to assign higher advantages to trajectories visiting new states when al…
FlowEvo achieves 82.8% success rate on ALFWorld, 23.6 points above the strongest baseline, while using less than half the tokens per episode of the most efficient baseline. The system allows agents to…
Researchers propose Prefix-GRPO, a reinforcement learning framework that improves small language model agents by decomposing teacher trajectories into replay-aligned prefix queries and online continua…
Researchers introduce masked diffusion language models (MDLMs) as steerable text-based world models for reinforcement learning, achieving up to 47% absolute gains over baselines in zero-shot transfer …
A developer evaluating the SDAR paper's claims found that the reported +9.4% accuracy gain on ALFWorld is less interesting than SDAR's ability to avoid training instability. The developer designed a t…
Researchers introduce Branching Policy Optimization (BPO), a reinforcement learning algorithm for LLM agents that exploits deterministic, snapshottable sandboxes to reduce variance and improve trainin…
A new framework called Experience Memory Graph (EMG) offers a promising solution to the persistent challenge of AI agents failing in complex, long-horizon tasks by treating error recovery as a graph m…
Researchers have introduced normalized entropy and the STAPO framework to address trajectory neglect in reinforcement learning, where AI models lose track of goals during complex tasks. Tests on ALFWo…
Researchers introduced TRIAGE, a novel framework that refines credit assignment in agentic reinforcement learning by classifying actions based on their role, improving success rates and reducing ineff…
Researchers introduced TREK (Teacher-Routed Exploration via Forward KL), a new AI training method that enhances learning through unconventional exploration strategies. TREK significantly improved perf…
Researchers introduced TREK, a method that improves AI problem-solving by using verified output trajectories to extend model learning. TREK boosted Qwen3-8B's performance on AIME 2024 from 36.9 to 40.…
Researchers propose TurnOPD, a turn-level budgeting strategy for efficient on-policy distillation of long-horizon language agents. The method addresses inefficiencies in vanilla agent OPD by using ada…
Microsoft researchers developed SkillOpt, a method that treats AI agent skill files as trainable parameters outside frozen target models, enabling controlled optimization through bounded text edits an…