Reproduce “DeepPrune: Parallel Scaling without Inter-trace Redundancy” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shangqing Tu, Yaxuan Li, Yushi Bai, Lei Hou, Juanzi Li
Researchers from Tsinghua University and ShanghaiTech University developed DeepPrune, a two-stage framework that identifies and prunes redundant reasoning paths in parallel Large Language Model inference. The method achieved token consumption reductions ranging from 52.5% to 91.6% while generally maintaining or improving reasoning accuracy across various models and benchmarks.
Key insights
- Inter-trace redundancy is a pervasive issue in parallel LLM reasoning, with empirical analysis showing over 80% of parallel traces yielding identical final answers across different models.
- Shallow semantic similarity or zero-shot LLM classifiers are insufficient for accurately predicting final answer equivalence from partial traces, highlighting the need for specialized training.
- A fine-tuned judge model, optimized with reasoning-step aligned features and robust class imbalance handling (focal loss + oversampling), can accurately generalize to unseen reasoning models and benchmarks for redundancy prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.