Claim-Level Reliability Assessment for Efficient Test-Time Reasoning
CLR spends test-time compute on falsifying key claims in a reasoning trace, not on sampling more full answers.
The paper argues that whole-trace scoring can miss decisive mistakes because routine tokens dilute the signal. CLR compresses a model’s reasoning into decision-critical claims, then looks for semantic evidence that any claim is false. In the reported tests across four models and four reasoning benchmarks, it generally beats pass@1 and self-consistency under matched budgets. One cited result: on GPT-OSS-20B/CMIMC25, it raises self-consistency accuracy from 77.50% to 82.19% while using 37.0% fewer tokens. HF Daily Papers' note
The paper argues that whole-trace scoring can miss decisive mistakes because routine tokens dilute the signal. CLR compresses a model’s reasoning into decision-critical claims, then looks for semantic evidence that any claim is false. In the reported tests across four models and four reasoning benchmarks, it generally beats pass@1 and self-consistency under matched budgets. One cited result: on GPT-OSS-20B/CMIMC25, it raises self-consistency accuracy from 77.50% to 82.19% while using 37.0% fewer tokens. HF Daily Papers' note
score 5