How reliable is the top-2 confidence gap as a stopping signal across tasks?
This explores whether the margin between a model's top two answer/token probabilities — used to decide 'I'm confident enough, stop reasoning now' — holds up as a stopping rule, or whether its reliability swings with the task.
This reads the question as: can a confidence margin (how far the model's best guess sits above its runner-up) serve as a trustworthy 'stop here' signal across different kinds of tasks? The corpus doesn't test the top-2 gap by that exact name, but it triangulates the answer hard — and the verdict is that confidence-based stopping works, but only locally and conditionally, never as a universal threshold.
The strongest case for it is granularity. Step-level confidence catches reasoning breakdowns that whole-trace averaging smooths over, and crucially it lets you stop a trace *before it finishes* while matching the accuracy of much more expensive majority voting Does step-level confidence outperform global averaging for trace filtering?. ReBalance pushes the same idea further: confidence variance and overconfidence aren't just stop/go bits but continuous dials that steer a model away from both overthinking and underthinking, and it holds across model sizes from 0.5B to 32B Can confidence patterns reveal overthinking versus underthinking?. So as a *relative, local* signal, the confidence gap carries real information.
The catch is that the threshold itself refuses to stay put. The optimal stopping point depends on task difficulty, model training, and domain — and stays invisible until you've crossed it, with no reliable predictor that generalizes How can we predict the optimal thinking token threshold?. That's mirrored in the inverted-U shape of reasoning length: the right amount of thinking rises with difficulty but falls as models get more capable, so the same gap means different things on different tasks Why does chain of thought accuracy eventually decline with length?. A fixed top-2 threshold tuned on one benchmark is therefore quietly miscalibrated on the next.
Worse, confidence and correctness can come apart entirely. A model can be *highly* confident and still wrong — entity co-occurrence statistics from pretraining flag hallucination risk precisely in cases where the model's own confidence stays high Can pretraining data statistics detect hallucinations better than model confidence?. And consistency isn't reliability: a deterministic, repeatable output is still one draw from the distribution, so a stable confidence gap can encode a stable mistake Does setting temperature to zero actually make LLM outputs reliable?. The signal degrades exactly where you'd most want a stopping rule to be cautious.
Where does the gap become *more* trustworthy? The corpus points to a pattern: confidence tracks reliability best on objective tasks, with larger models, and with few-shot grounding — the same conditions under which confidence predicts robustness to prompt rephrasing Does model confidence predict robustness to prompt changes?. The thing you didn't know you wanted to know: the top-2 gap is less a thermometer and more a contrast knob — reliable when read as a *change* against a baseline (step-to-step, or against the data statistics) and unreliable when read as an *absolute* number you can hard-code. That's why the strongest methods treat confidence as a reward or a steering signal to be recalibrated Can model confidence work as a reward signal for reasoning?, not a fixed gate.
Sources 8 notes
Local step-level confidence catches reasoning breakdowns that global averaging masks and enables early stopping before traces complete. This approach achieves comparable accuracy gains to naive majority voting with far fewer generated traces, proving trace quality matters more than quantity.
ReBalance uses confidence variance and overconfidence as diagnostic signals to apply training-free steering vectors that reduce overthinking redundancy while promoting exploration during underthinking, improving accuracy across models from 0.5B to 32B parameters.
The overthinking threshold depends on task difficulty, model training, and domain, but remains invisible until crossed. Recent work suggests difficulty estimators and runtime confidence signals can detect thresholds dynamically.
Task accuracy peaks at intermediate CoT length, with optimal length increasing alongside task difficulty but decreasing with model capability. RL training naturally gravitates toward shorter chains as models improve, revealing that simplicity emerges from reward signals rather than explicit training.
QuCo-RAG uses entity co-occurrence patterns from training data to trigger retrieval, successfully flagging hallucination risk even when models are highly confident. This data-side approach catches the root cause (unseen combinations) rather than the symptom (low confidence).
Fixed seeds and zero temperature replicate the same output repeatedly, but that output remains one draw from the model's probability distribution. McDonald's omega testing across 100 repetitions reveals that consistency does not equal reliability.
ProSA found that when models are highly confident, they resist prompt rephrasing; low confidence causes major output swings. Larger models, few-shot examples, and objective tasks all correlate with higher confidence and greater robustness.
RLSF uses answer-span confidence to rank reasoning traces, creating synthetic preferences that strengthen step-by-step reasoning while reversing RLHF's calibration degradation—without requiring human labels or external verifiers.