cs.AINov 2, 2026

Think with Skills: Pseudocode Planning Before Execution for Advanced Reasoners

Authors: Boshen Zhou, Boxuan Yue, Xiang Li, Shanchao Yang

Organizations: Honor Device Co., Ltd., China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Language model agents often struggle with complex tasks even when equipped with expert-written skills. Such skills encode declarative knowledge, whereas successful execution requires procedural knowledge tailored to the task at hand. Natural language skills may leave conditions and dependencies implicit, and tasks involving multiple skills further require agents to coordinate intermediate results and execution order. We introduce Think with Skills, a method that trains reasoning models to turn relevant skill descriptions into task-specific pseudocode plans before execution. The pseudocode explicitly represents operations, conditions, intermediate results, and dependencies, providing a structured bridge from skill descriptions to agent actions. We construct training data that pairs tasks and skills with pseudocode plans and execution trajectories, use supervised fine tuning to initialize this behavior, and further optimize end-to-end task performance with Group Relative Policy Optimization. On SkillsBench, Think with Skills raises task success from 35.0% to 62.0%. These results show that proceduralizing expert knowledge before execution substantially improves skill use.

Figures & tables

Figure 1: Overview of the Think with Skills training pipeline. A Draft Agent constructs and validates tasks from related skills. Strong and weak Solvers then attempt each task. Tasks solved by the strong Solver but not the weak Solver are retained, and the corresponding trajectories provide pseudocode and execution supervision for SFT. GRPO continues training from the SFT checkpoint with verifier rewards.
Figure 2: Illustrative comparison of direct skill use and Think with Skills on a routing-diagnosis task with the same input files and expert skill. Direct skill use consults relevant guidance during execution. Think with Skills first expresses the skill as task-specific pseudocode, then follows its branches and checks while executing. Failed checks trigger diagnosis and revision of the affected step. Passed checks allow execution to continue or finish.
Figure 3: Verifiable and difficulty-calibrated data construction. (A) A Draft Agent constructs a task, input files, and verifier from related skills and self-validates the candidate. (B) The strong and weak Solvers independently attempt the same candidate once, and the task-specific verifier scores both outputs. (C) Their joint outcome determines whether to retain the candidate, increase its difficulty, repair or simplify it, or discard it. Revised tasks return to self-validation before another independent evaluation by both Solvers.
Agent frameworkModelSkillsBench result (%)
Claude CodeThink with Skills (Qwen3.5-35B-A3B)62.0 (+27.0 pp)
 Qwen3.5-35B-A3B35.0
 Kimi K2.562.5
 DeepSeek V4 Pro61.67
 MiniMax M2.742.76
OpenHandsGLM 5.158.4†
Table 1: SkillsBench results with expert skills.
Model / Training ConfigurationSkillsBench Score (%)
Qwen3.5-35B-A3B (base model, direct skill use)35.0
Qwen3.5-35B-A3B + pseudocode prompting45.0
Qwen3.5-35B-A3B + SFT60.0
Qwen3.5-35B-A3B + SFT + GRPO (Think with Skills)62.0
Table 2: SkillsBench task success rates for Qwen3.5-35B-A3B under direct skill use, pseudocode prompting, SFT, and SFT followed by GRPO.
Skills per Training TaskTraining ExamplesSkillsBench Success (%)
Single skill (1)90155.0
Multiple skills (>1)91360.0
Table 3: Comparison of single-skill and multi-skill training on SkillsBench.
More in this topic

LLM Agent Planning

Large language model (LLM) agent planning studies how agents decompose goals, choose actions and tools, and adapt plans from feedback.

See all 231 papers

Explore similar work

CardsList

  1. Generative Skill Composition for LLM Agents

    Jun 30, 2026Xinyu Zhao, Zhen Tan, Vaishnav Tadiparthi +5AI Coding AgentsLLM Agent Skill Learning

    University of North Carolina at Chapel Hill · Arizona State University · Honda Research Institute USA

  2. SkillReason: Reasoning-Enhanced Agent Skill Retrieval for Implicit User Requests

    Aug 9, 2026Donghong Jiang, Endian Lin, Luoping Cui +6CoT SupervisionAI Agent Benchmarks

    Beijing University of Posts and Telecommunications · Peking University · Beijing ZOYEN Technology Co., Ltd. +1

  3. SkillComposer: Learning to Evolve Agent Skills for Specification and Generalization

    Jun 4, 2026Qi Zhang, Zhaopeng Feng, Xiaonan Shi +8LLM Agent Self-ImprovementLLM Agent Skill Learning

    1Zhejiang University · 2Tongyi Lab · 3National University of Singapore