arXiv study finds random continuations often fix wrong reasoning answers
A new arXiv paper introduces FLCVA, a method auditing whether extra reasoning tokens actually help, and reports that even random continuations frequently repair errors.
In plain words
Large language models can either answer immediately or spend more computing effort on longer reasoning. The paper asks whether extra reasoning actually improves an answer, not just whether the current answer is wrong. Its audit method compares an immediate answer with several fixed-length continuations of the same reasoning, counting all costs.
What was announced
The authors report that on the GSM8K and MATH-500 benchmarks, a random 256-token continuation recovered a wrong answer in 26.6% of Qwen3-8B cases and 18.9% of DeepSeek-R1-Distill-Qwen-7B cases, while damaging a correct answer in only 1.4% and 1.9%. Correctness labels from independently sampled continuations predicted a separate continuation's outcome in both primary models.
However, the practical payoff looks limited. In offline replay, using those labels to allocate compute improved performance on DeepSeek, but the Qwen3 gain remained inconclusive. Under full cost accounting, the tested lightweight, myopic controllers mainly recognized current errors and added little continuation-specific value beyond that shortcut. A Llama-8B replication and ablations of actions and costs examined robustness.
What it does not show
- The source is the paper's abstract only; full methods, datasets details and complete results are not available in the supplied excerpt.
- The Qwen3 compute-allocation gain is inconclusive according to the authors, and no independent confirmation of the findings is provided.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.