Think Again or Think Longer? Selective Verification for Budget-Aware Reasoning

2026-06-18

AI Summary

This work addresses the inefficiency and potential errors in inference services caused by redundant verification or blind correction. It proposes SevRA, a service-layer controller that, for the first time, formulates selective verification as a resource allocation problem at deployment time. Operating under the constraint of freezing the original large model’s outputs (e.g., Qwen3-4B), SevRA employs a recoverability-aware gating mechanism to trigger verification only when necessary. The approach requires no modification to the solver and instead trains a lightweight policy solely from log data, leveraging features of the attempted state to decide intervention. Experiments show that on MathFive, SevRA achieves 76.3% accuracy while reducing post-generation tokens by 26.8% and harmful flips by 54.5%; on GSM, it boosts accuracy to 94.5% by verifying only 3% of samples, saving 91.2% of verification overhead.

Abstract

Test-time reasoning is increasingly used as a serving-time control knob, but extra reasoning is not uniformly valuable: it can repair failed attempts, waste compute on already-correct answers, or introduce harmful answer changes. We study this as a deployment allocation problem rather than a new-verifier problem. We introduce \sevra, Selective Verification for Reasoning Allocation, a serving-layer controller that decides whether to preserve a frozen solver's initial answer or invoke active verification. Using a frozen Qwen3-4B solver, we log intervention outcomes and train recoverability-aware gates from serving-visible attempt state. On \mathfive, selective verification reaches 76.3\% accuracy, compared with 75.5\% for always verifying, while reducing post-generation tokens by 26.8\% and harmful flips from 2.2\% to 1.0\%. However, an 8,192-token initial solve reaches 76.0\% accuracy with 28\% fewer total model tokens, showing that selective recovery is useful but not the best tested cost frontier. In frozen transfer to \gsm, the selective policy verifies only 3.0\% of examples, improves accuracy from 93.4\% to 94.5\%, and reduces verification tokens by 91.2\% relative to always verifying; again, a longer initial solve matches its accuracy with fewer realized tokens. On CommonsenseQA, always-on verification hurts, while Self-Consistency@5 improves accuracy at about five times the realized token cost. The resulting deployment rule is: tune the initial budget first, then use selective recovery when explicit checks, bounded retries, auditability, or regression-risk control matter.

Problem

• test-time reasoning
• selective verification
• budget-aware reasoning
• deployment allocation
• reasoning control

Innovation

• Selective Verification
• Reasoning Allocation
• Test-time Reasoning
• Budget-aware Inference
• Recoverability-aware Gating