Reproduce “Segment-Level Attribution for Selective Learning of Long Reasoning Traces” 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
Siyuan Wang, Yanchen Liu, Xiang Ren
Researchers from the University of Southern California developed a framework that employs Integrated Gradients to identify crucial segments within large language model reasoning traces. This method, applied through selective supervised finetuning, improves reasoning accuracy by up to 4.7% and reduces output length by up to 18.2% across multiple models and mathematical reasoning benchmarks.
Key insights
- A significant portion of CoT segments (approximately 60-70%) is redundant and contributes minimally to the final answer prediction, regardless of CoT correctness.
- Important segments, as identified by high attribution strength and moderate direction consistency, show lower perplexity, less repetition, and lead to greater improvements in correct answer confidence.
- Selectively learning from critical CoT segments by masking loss (selective SFT) is more effective than physically pruning them, maintaining contextual flow while focusing learning on high-impact components.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.