AI BriefingImproving LLM Decision-Making Through Diverse Reasoning Paths
Key point
Amazon researchers unveiled SSFT and GFPO, techniques that train LLMs to generate diverse reasoning paths and select the optimal strategy.
Details
Amazon researchers introduced SSFT (Set-Supervised Fine-Tuning) and GFPO (Global Forking Policy Optimization), techniques that enable LLMs to generate diverse reasoning approaches instead of relying on a single reasoning path.
Existing supervised fine-tuning (SFT) methods learn only one correct path, which limited the model's reasoning capability. To address this, the researchers introduced global forking tokens to allow the model to have distinct reasoning modes.
The key techniques are as follows:
- SSFT (Set-Supervised Fine-Tuning): Models reasoning not as a single path but as a set of completed solution paths. By assigning each path to a dedicated token through bipartite matching, it prevents the 'mode collapse' phenomenon where the model converges to a specific pattern.
- GFPO (Global Forking Policy Optimization): Through a reinforcement learning paradigm, it trains the model to decide for itself which reasoning strategy to adopt in a given problem situation.
Through this approach, they achieved a 5% to 7% improvement in single-shot accuracy on standard benchmarks such as AIME 2025 and LiveCodeBench.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Bridging the Intent-Execution Gap in Agent SystemsAgentic AI that heals systems that can't be replaced