模型推論
SVR Lets a 2B Model Decide When to Stop Revising, Completing Mathematical Reasoning in 2.99 Rounds on Average
SVR jointly trains answer generation, correctness judgments, and confidence estimation, allowing the model to decide at inference time whether to retain an answer or continue refining it. Qwen3.5-2B achieved 56.3% average accuracy across seven mathematics benchmarks, although its stopping signal may still be affected by the model’s own calibration biases.
Increasing test-time compute typically relies on a fixed number of samples or refinement rounds, which can waste tokens on easy problems. Using an external verifier, meanwhile, adds model, latency, and operational costs. Self-Verifying Refinement (SVR) attempts to make a single model serve as both the problem solver and the stopping controller: in each round, it outputs not only an answer but also a discrete Correct/Incorrect judgment and a confidence score. The system retains the current answer and stops only when the model judges it correct and its confidence exceeds a threshold; otherwise, the self-evaluation results are fed into the next refinement round.
Training uses fixed-length trajectories and GRPO. The reward jointly measures answer correctness, the reliability of the correctness judgment, confidence calibration, and whether the model can reach a correct state in which stopping is appropriate. Reference answers are used only to construct training rewards and never appear in refinement prompts; deployment also requires neither an oracle nor a separate reward model. In other words, SVR learns not only to “think again,” but also when further computation is expected to be beneficial.
The study evaluated SVR on Qwen3.5-2B across seven mathematical reasoning benchmarks, including AIME, AMC, and MATH-500. SVR achieved a macro-average accuracy of 0.563 while running an average of 2.99 rounds per problem. The paper claims that it outperformed standard GRPO, multi-round refinement baselines, and a fixed ten-round approach while using significantly fewer rounds. This is practical for batch inference services because individual requests can consume different compute budgets based on difficulty instead of all paying the worst-case cost.
However, internal verification is not a free source of ground truth. The model may retain an incorrect answer with high confidence or repeatedly reject a correct one. Mathematical problems provide verifiable training signals, so the results may not generalize directly to code modification, open-ended question answering, or agentic operations. The paper also does not currently provide reproducible code or model weights. Engineering teams should next examine the stability of confidence thresholds on out-of-distribution data, actual token usage and latency rather than round counts alone, and whether low-cost external checkers can mitigate incorrect stopping decisions.