Claim-Level Reliability Assessment Beats pass@1 by 27.15 Points on GPT-OSS-20B While Spending 37% Fewer Tokens — and Needs No Training
CLR (arXiv 2608.11994, submitted Aug 12) is a training-free test-time framework that stops generating whole competing solutions and instead spends compute verifying the specific 'decision-critical claims' inside a reasoning trace, skipping routine tokens that hide errors. It exploits an asymmetry builders can reason about directly: a correct solution requires every step to hold, but refuting a wrong claim needs only one decisive flaw — which lets a nonlinear reliability score suppress confidently-wrong consensus. Across four models and four reasoning benchmarks it improved on baselines; on GPT-OSS-20B/CMIMC25 it exceeded pass@1 by 27.15 percentage points and lifted self-consistency accuracy from 77.50% to 82.19% using 37.0% fewer tokens. Training-free plus cheaper is a rare combination in the test-time-compute literature.
↳ Follow the thread