AI models that "think longer" at inference time don't always think better.
Researchers introduced SEVRA, a serving-layer controller that decides, per query, whether a frozen language model should stick with its first answer or trigger a verification pass. Using a frozen Qwen3-4B solver on the MATH5 benchmark, SEVRA reached 76.3% accuracy versus 75.5% for always-verify runs, while cutting post-generation tokens by 26.8% and halving harmful answer flips — cases where verification changes a correct answer to a wrong one — from 2.2% to 1.0%. On GSM, the selective policy verified only 3.0% of examples yet still improved accuracy from 93.4% to 94.5%, slashing verification token use by 91.2% compared to blanket re-checking.
The finding matters because "test-time compute" — letting models reason longer at serving time rather than baking more into training — has become a dominant scaling lever for labs like OpenAI and DeepSeek. The implicit assumption has been that more reasoning tokens are always worth the cost. SEVRA's results challenge that: on CommonsenseQA, always-on verification actively hurt performance, while a simple five-sample self-consistency approach improved it at about five times the realized token cost.
The honest caveat from the paper is that simply tuning the initial token budget often matches or beats selective recovery with fewer total tokens — making SEVRA most useful in constrained scenarios like auditable pipelines, regression-risk deployments, or systems where bounded retries are a hard requirement, not a general-purpose cure for over-spending on inference.