SVR: Self-Verifying Refinement via Joint Verdict-Confidence Reinforcement Learning for Adaptive Test-Time Compute
The paper claims a language model can decide when to stop refining its own answer without an external verifier.
SVR has the model produce an answer, a correctness verdict, and a confidence score at each turn. It keeps the answer only when the verdict is Correct and confidence clears a threshold; otherwise it continues refinement. Training uses ground-truth correctness for rewards, but that signal is not shown in refinement prompts or needed at inference. On seven math reasoning benchmarks with Qwen3.5-2B, it reports 0.563 macro-average accuracy using 2.99 inference turns on average. ArXiv · AI/CL/LG's note
SVR has the model produce an answer, a correctness verdict, and a confidence score at each turn. It keeps the answer only when the verdict is Correct and confidence clears a threshold; otherwise it continues refinement. Training uses ground-truth correctness for rewards, but that signal is not shown in refinement prompts or needed at inference. On seven math reasoning benchmarks with Qwen3.5-2B, it reports 0.563 macro-average accuracy using 2.99 inference turns on average. ArXiv · AI/CL/LG's note
score 6