Reproduce “Correct Reasoning Paths Visit Shared Decision Pivots” 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
Dongkyu Cho, Amy B.Z. Zhang, Bilel Fehri, Sheng Wang, Rumi Chunara, Hengrui Cai, Rui Song
Researchers at NYU and Amazon developed ROMA, a self-training framework that identifies "decision pivots" by aggregating multiple successful reasoning paths from Large Language Models. This approach synthesizes concise "short-path reasoning" (SPR) to provide targeted supervision, resulting in improved task accuracy across commonsense, medical, and mathematical reasoning benchmarks and generating significantly more compact explanations.
Key insights
- Correct reasoning paths, despite stylistic variations, consistently converge upon a sparse and shared set of "decision pivots"—minimal, semi-verifiable checkpoints essential for a sound explanation.
- Decoupling reasoner generation from verifier-guided pivot extraction and SPR synthesis allows for targeted, efficient training that injects domain-specific priors without requiring human-annotated rationales or task-specific external verifiers.
- Supervising LLMs with concise, pivot-focused reasoning naturally leads to markedly more compact reasoning generations without explicit incentives for brevity, reducing non-pivot tokens and enhancing efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.