CROP wins EMNLP 2026 Outstanding Paper: conformal certification for reasoning trace prefixes
hanjie_chen · x · 2026-10-28
A Rice University team (Matt Y. Cheung, Ashok Veeraraghavan, Hanjie Chen, Guha Balakrishnan) won an EMNLP 2026 Outstanding Paper Award for CROP (Conformal Reasoning Output Prefixes).
Key points:
- Reasoning traces usually contain valid intermediate steps before a critical error, but existing UQ methods certify only final answers or whole responses. CROP instead certifies the longest contiguous "clean prefix" whose step-level risk proxy stays below a calibrated threshold, routing the uncertified suffix to review or repair.
- Under exchangeability, CROP rigorously controls the marginal probability that the returned prefix contains an annotated error.
- Across six process-labeled reasoning datasets, standard step metrics like AUROC fail to capture prefix utility — verifiers should be judged by certified prefix length. CROP balances over/under-withholding and improves downstream repair accuracy.
The work positions prefix certification as a practical bridge between process supervision, abstention, and repair.
More from Research
- Shaun Maguire hails three Harmonic discoveries in math, CS, and physics as jaw-dropping — inductionheads · 2026-10-28
- VitronEdit open-sources a unified model for image, video, and 3D scene editing — Desperate_Day_5416 · 2026-10-28
- Eric Topol highlights immunome paper combining multi-assay data with AI to model immune aging — EricTopol · 2026-10-28
- Google Research's 'sleep framework' lets LLMs keep learning, retaining up to 10M tokens — behrouz_ali · 2026-10-28
- CATCH: ~200 examples of adversarial training make LLMs robustly honest across benchmarks — irinarish · 2026-10-28
- Andrew Critch's dubious credences: P≠NP at 80%, Riemann hypothesis at just 60% — AndrewCritchPhD · 2026-10-28