LaneRoPE: Enhancing LLM Test-Time Scaling with Collaborative Parallel Reasoning
A new arXiv preprint (2605.27570) introduces LaneRoPE, a method for improving parallel LLM test-time scaling. Traditional techniques like best-of-N generate N sequences independently, missing opportunities for reuse. LaneRoPE enables coordination among sequences via an inter-sequence attention mask and a RoPE extension that captures relative token positions within and across sequences. Evaluated on mathematical reasoning tasks, LaneRoPE shows promising results in fostering collaboration among sequences, potentially boosting accuracy and computational efficiency.
Key facts
- arXiv preprint 2605.27570 introduces LaneRoPE
- LaneRoPE enables coordination among N>1 sequences at generation time
- Uses inter-sequence attention mask to make sampling dependent
- RoPE extension injects positional information across sequences
- Evaluated on mathematical reasoning tasks
- Promising results for collaborative parallel reasoning
Entities
Institutions
- arXiv