Skill Selection for Agents

How this thread developed

1. July 2026 · Alibaba, The Chinese University of Hong Kong

   Self-play with evolving skills boosts LLM benchmarks

   Proposes Skill Self-Play for dynamic skill routing to improve LLM benchmark performance.

2. 1 further paper

   Aug 2026 · Appier AI Research, National Taiwan University

   Joint training helps small language models create and use tools

   Joint training on tool creation and use enables small models to generate their own reusable tools, expanding the skill selection paradigm.

3. 1 further paper

   Aug 2026 · Google Research, Virginia Tech

   Persistent Wikis Help AI Agents Build Better Transferable Skills

   Introduces WikiSkill to separate experience, knowledge, and skills for improved transferable skill selection.

4. Sept 2026 · Jilin University

   Memory lets traffic-simulation agents build reusable skills without retraining

   Builds and selects reusable skills from memory, optimizing skill usage under budget constraints.

   released code

5. 2 further papers

   Sept 2026 · The Chinese University of Hong Kong, Shenzhen, Shenzhen Future Network of Intelligence Institute (FNii-Shenzhen)

   Enumerating Tool Choices Beats Sampling Them in Genomic Reasoning

   Replaces sampling in RL with exact enumeration over tool combinations, providing an alternative to probabilistic skill selection.

6. 3 further papers

   Sept 2026 · East China Normal University, Shanghai AI Laboratory

   Verified workflow training gives language models reusable procedural skills

   Creates executable environments from human-written skills and collects trajectories to train models on reusable procedural skills.

7. 1 further paper

   Sept 2026 · Zhejiang University, Ant Group

   Reinforcement learning improves tool selection across large repositories

   Extends budget-constrained skill and tool selection to repository-scale toolkits, with RL as the selection policy.

   released code

7 of 15 papers shown