Researchers introduce Reasoning Primitive Induction, a method that extracts reusable reasoning patterns from ReAct-style LLM agent traces and converts them into a compact library of pseudo-tools. The induced libraries consistently outperform the original agents by 22-44 percentage points across multiple reasoning tasks, suggesting a systematic path to improve LLM reasoning through learned decomposition.
The research addresses a fundamental inefficiency in how large language models perform complex reasoning. ReAct agents generate successful problem-solving traces but fail to retain or reuse the reasoning strategies they discover, forcing them to rediscover similar patterns repeatedly. By mining these traces and clustering recurring reasoning moves, the researchers created a mechanism to crystallize implicit reasoning strategies into explicit, reusable components—essentially distilling agent intelligence back into the model's toolkit.
This work builds on years of research demonstrating that LLMs benefit from structured reasoning frameworks. While Chain-of-Thought prompting showed initial promise, methods like ReAct revealed that agents perform better with environmental interaction and tool access. However, the static nature of these tools limited their ability to capture domain-specific reasoning patterns. The primitive induction approach bridges this gap by making reasoning patterns themselves discoverable and composable, treating successful problem-solving strategies as learnable abstractions.
The performance gains—particularly the 44-point improvement on RuleArena NBA tasks—suggest meaningful practical value. The method matches or exceeds expert-authored decompositions while operating at lower inference cost than alternative approaches, indicating it captures genuinely efficient reasoning patterns rather than mere task overfitting. The single fixed configuration performing consistently across diverse subtasks (narrative deduction, rule application, constraint satisfaction) demonstrates generalizability.
Future implications center on whether this approach scales to more complex domains and whether induced primitives transfer across different task families. The work also raises questions about optimal primitive library size, the relationship between trace quantity and library quality, and whether this methodology could inform better pre-training objectives for foundation models.
- →Reasoning Primitive Induction extracts recurring problem-solving patterns from successful LLM agent traces and converts them into reusable pseudo-tools.
- →Induced primitive libraries outperformed source agents by 22-44 percentage points across five benchmark tasks in reasoning and planning domains.
- →The approach matches or surpasses expert-authored decompositions while maintaining lower average inference costs than competing methods.
- →A single fixed configuration of induced primitives generalizes effectively across diverse reasoning subtasks without task-specific tuning.
- →The method provides a systematic mechanism to retain and compose learned reasoning strategies rather than discarding them in transient execution contexts.