Long Live The Balance: Information Bottleneck Driven Tree-based Policy Optimization
- Published
- Source
- arXiv
- Paper number
- 264
- Field
- Machine Learning
- arXiv ID
- 2605.28109
Key points
- It improves efficiency by reusing prefixes and taking advantage of KV caching, which lets it generate more trajectories, or a higher G, within the same token budget than independent sampling.
- By branching according to the IB-Score, the model focuses exploration on the 'critical' steps where both high diversity and a strong signal toward the correct answer are present.
Paper links
External research summaries. These are not HDATF publications or measured product results.