Large language models are increasingly deployed with test-time reasoning as a control knob to improve accuracy. But the extra compute is not uniformly valuable: it can fix failed attempts, waste cycles on already-correct answers, or even introduce harmful answer changes, according to a new preprint on arXiv.
Researchers 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. Rather than building a new verifier from scratch, the team frames the problem as a deployment allocation decision.
The SEVRA Approach
Using a frozen Qwen3-4B solver as the baseline, the researchers logged intervention outcomes and trained recoverability-aware gates from serving-visible attempt state. The controller decides based on the state of the initial reasoning attempt whether verification is likely to help or hurt.
Performance Gains Across Benchmarks
On MATH-500, selective verification reached 76.3% accuracy, compared with 75.5% for always verifying — a modest but consistent improvement. At the same time, it reduced post-generation tokens by 26.8% and cut harmful flips from 2.2% to 1.0%.
| Method | Accuracy | Post-generation tokens (relative) | Harmful flips |
|---|---|---|---|
| Selective verification (SEVRA) | 76.3% | 26.8% fewer than always verify | 1.0% |
| Always verify | 75.5% | Baseline | 2.2% |
| Initial solve (8,192 tokens) | 76.0% | 28% fewer total model tokens than always verify | N/A |
However, the researchers note that an 8,192-token initial solve alone achieved 76.0% accuracy with 28% fewer total model tokens than always verifying, showing that selective recovery is useful but not necessarily the best cost frontier.
On GSM8K, the selective policy verified only 3.0% of examples, improving accuracy from 93.4% to 94.5%, and reduced verification tokens by 91.2% relative to always verifying. Again, a longer initial solve matched that accuracy with fewer realized tokens.
On CommonsenseQA, always-on verification actually hurt performance, while Self-Consistency@5 improved accuracy at about five times the realized token cost.
The Deployment Rule
The study concludes with a clear deployment recommendation: first tune the initial reasoning budget, then apply selective recovery when explicit checks, bounded retries, auditability, or regression-risk control are important. The selective verification approach provides a middle ground between no verification and full verification, offering controlled compute costs with accuracy guarantees.
Implications for Enterprise AI
For enterprise technology buyers deploying large language models in production, this research offers a practical technique to manage inference costs without sacrificing output quality. By selectively verifying only the most uncertain or error-prone outputs, organizations can reduce compute expenditure while maintaining high accuracy standards — a key consideration for applications such as automated document processing, customer support, or any workflow where output correctness is critical.