Tag
本文提出Tev-Gym——首个具有可验证、密集渐进奖励的长周期智能体训练数据集（7个领域1170个任务），并基于进度-势能奖励塑形（PPRS）训练出Tev-27B，在Agents' Last Exam计算与数学科学领域匹配GPT-6 Astra和Claude Opus 5的Pass@5表现，并可泛化至Terminal-Bench 3.0/4.0。
This paper tests Reinforcement Learning with Verifiable Rewards (RLVR) on small language models using retrieval-augmented generation, achieving a 3.8-fold gain without distillation and emphasizing the need for reward design tailored to small models.
IncentRL presents a framework for preference-based reward shaping in reinforcement learning that uses KL divergence penalty to balance preference guidance with task performance, supported by theoretical analysis and experiments on MiniGrid benchmarks.
When2Think is a post-training framework that dynamically allocates computation in large reasoning models based on problem difficulty, improving accuracy-efficiency trade-offs on mathematical benchmarks.
This paper studies how reinforcement learning can lead LLM agents to learn spurious tool-use policies based on superficial cues rather than task requirements, and introduces a dense reward method to mitigate this issue.
Cliff improves reinforcement learning with verifiable rewards by using an off-the-shelf LLM to detect the first reasoning mistake and shape token-level advantages, leading to consistent performance gains over existing methods.
The paper introduces Answerable Working Memory (AWM) and AWM-GRPO to enhance the quality of terminal working memory in long-document VQA agents, improving accuracy and reducing memory issues.
The paper proposes calibrating an artificial guilt reward signal from human neural and behavioral fMRI data, then embeds it in multi-agent PPO agents. The neurally calibrated agents match human social decision rates far better than hand-tuned or selfish baselines.
After 123 failed PPO experiments on Atari Breakout, adding a simple proximity reward for tracking the ball during descent finally achieved reactive, non-scripted play that transfers to unfamiliar brick layouts.
MeRLa is a meta-learned reward shaping framework for RLHF that improves alignment by learning task-specific shaping functions, achieving state-of-the-art results on multiple benchmarks with significant reductions in training instability.
Introduces AgenticAI-Supervisor, an API and UI-driven simulation environment for scalable reinforcement learning of LLM agents, using verifiable execution outcomes and reward shaping to mitigate reward hacking.
This work presents a model that learns shaped 'process rewards' for robotic reinforcement learning, which evolves automatically as the policy improves, enhancing performance on benchmarks and in real-world settings.
This paper proposes a safe transfer reinforcement learning framework for autonomous highway lane-changing, using adaptive teacher intervention and reward shaping to improve sample efficiency and safety. Experiments show over 52% improvement in safety and 5% improvement in efficiency over baselines.
ExpRL is a new RL-based mid-training method that uses human-written reference solutions as dense reward scaffolds (never shown to the policy) to improve LLM reasoning, achieving significant gains on hard math benchmarks like AIME-2026.
SlimSearcher is a framework that improves efficiency in deep research agents by combining Pareto-efficient trajectory filtering and adaptive reward shaping, reducing tool-call rounds by 17-58% while maintaining accuracy on benchmarks like GAIA, BrowseComp, and XBenchDeepSearch.
This paper frames LLM-generated reward shaping for sparse structured RL as a debugging problem, identifying failure modes like reward flooding and semantic misunderstanding. The authors propose diagnostic-driven iterative refinement, achieving dramatic success rate improvements (e.g., DoorKey-8×8 from 2.3% to 97.6%) compared to one-shot generation.
The paper introduces OracleTSC, a method using oracle-informed reward hurdles and uncertainty regularization to stabilize reinforcement fine-tuning of LLMs for traffic signal control. It demonstrates significant improvements in traffic flow metrics on the LibSignal benchmark using LLaMA-3-8B while maintaining interpretability.
This paper proposes an empirical 'sparse-to-dense' reward principle for language model post-training, arguing that scarce labeled data should be used with sparse rewards for teacher model discovery and dense rewards for student compression via distillation. The authors demonstrate that this staged approach, bridging sparse RL and on-policy distillation, outperforms direct GRPO on deployment-sized models in math benchmarks.
This paper proposes a signal reshaping method for Group Relative Policy Optimization (GRPO) to improve weak-feedback agentic code repair, showing significant gains in compile and semantic accuracy.
This paper introduces Retention-aware Policy Optimization (RaPO) to mitigate catastrophic forgetting in visual continual learning using reinforcement fine-tuning. RaPO uses trajectory-level reward shaping and cross-task advantage normalization to close the gap between reinforcement and supervised fine-tuning in class- and domain-incremental learning.