A new research paper from Hugging Face introduces Claim-Level Reliability Assessment (CLR), a training-free framework that reallocates test-time compute from additional solution sampling to targeted verification. The method is designed to improve reasoning accuracy in large language models (LLMs) without requiring a stronger model.
The core principle is claim-level falsification. Instead of evaluating entire reasoning traces—which can dilute decisive errors with routine tokens—CLR condenses each trace into a compact set of decision-critical claims. This isolates the logical anchors of the reasoning process, making it easier to spot flaws.
CLR exploits an asymmetry between solution construction and claim refutation: constructing a valid solution requires a flawless reasoning path, whereas refuting an incorrect claim requires identifying only a single decisive flaw. This targeted search for negative evidence compresses the survival space of high-confidence incorrect traces, suppressing erroneous consensus via nonlinear reliability scoring.
Empirically, across four LLMs and four reasoning benchmarks under matched budgets, CLR generally improves upon pass@1 and self-consistency. For instance, on GPT-OSS-20B/CMIMC25, CLR exceeds pass@1 by 27.15 percentage points and raises self-consistency accuracy from 77.50% to 82.19% while using 37.0% fewer tokens. The method also rescues ~37% of failed consensus cases on average when at least one correct trace is present but standard self-consistency fails.
The paper is available on arXiv, and code is released on GitHub.