RL for Tool Agents

How this thread developed

1. Sept 2026 · NVIDIA

   Open voice model combines real-time conversation with native tool calling

   Brings native tool calling to real-time speech-to-speech models, enabling new capabilities for tool-use in voice-driven agents.

2. 2 further papers

   Sept 2026 · Zhejiang University, Ant Group

   Reinforcement learning improves tool selection across large repositories

   ToolSearcher applies RL to choose tools from repository-scale toolkits without fitting every description into context.

   released code

3. 3 further papers

   Sept 2026 · Georgia Institute of Technology, Alibaba Token Foundry, Alibaba Group

   Solved mathematical mechanisms can generate verifiable training worlds for AI agents

   Generates verifiable training worlds by solving mathematical mechanisms, enabling RL training of a 35B parameter tool agent.

4. Oct 2026 · Microsoft, tuyoogame

   Explicit diversity training gives LLM agents distinct successful strategies

   uses trajectory-guided joint policy optimization to train diverse behavioral strategies

5. 1 further paper

   Oct 2026 · Surge AI

   Reinforcement learning helps coding agents finish tasks across benchmarks

   Demonstrates RLVR post-training improves coding agent task completion across benchmarks.

6. 1 further paper

   Oct 2026 · Tsinghua University, The Chinese University of Hong Kong

   Diffusion reward models capture multimodal preference that scalar scores miss

   Introduces diffusion reward models that capture multimodal preferences, improving reward modeling for RL-based agent training.

7. 1 further paper

   Oct 2026 · Anthropic Fellows Program, University of Edinburgh

   Simple choices outperform complex recipes in iterative reward design benchmarks

   Compares iterative reward design methods under controlled conditions and finds simple choices outperform complex recipes.

8. Oct 2026 · MIT, NVIDIA

   Visual language models refuse harmful requests less often when using tools

   Provides empirical evidence that tool-use behavior in agents undermines safety alignment, a critical consideration for RL-based tool training.

9. Oct 2026

   Existing scientific software can generate verified training tasks for terminal agents

   Uses existing scientific software to self-supervise training tasks for terminal agents, eliminating the need for handwritten reference solutions.

10. 3 further papers

    Oct 2026 · ETH Zurich

    Testing transfer helps language-model tutors teach instead of simply telling

    Redesigns RL for language-model tutors using student learning outcomes, relevant to RL training of agents with tools

11. 2 further papers

    Oct 2026 · UC Berkeley, AWS AI Labs

    Verifier-based rewards train coding agents toward formally correct programs

    Trains coding agents with verifier-based rewards to achieve formally correct programs.

12. 6 further papers

    Oct 2026 · Hanyang University, Hunyuan, Tencent

    Tool graph memory rarely improves multi-turn agents by itself

    Tests whether stored tool transition graphs improve multi-turn tool selection, finding no significant gains, which informs training designs.

13. 3 further papers

    Oct 2026 · South China University of Technology, FiT, Tencent

    Separating search and answer credit improves accuracy and evidence grounding

    Shows RL on final answer correctness can hurt evidence grounding in search agents, and proposes separate credit assignment to improve grounding.

14. 1 further paper

    Oct 2026 · Renmin University of China, Key Laboratory of Data Engineering and Knowledge Engineering, MOE

    Evidence annotations improve step-level credit for deep search agents

    Trains deep search agents with RL by rewarding individual evidence-grounded retrieval steps, improving tool use for search.

14 of 37 papers shown