- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies when a separate performance predictor can beat a model’s self-confidence at predicting whether the model answers correctly. It shows that any gain in correctness log loss over self-confidence transfers exactly to the model’s answer log loss through recalibration (Proposition 1). Assuming a lowest achievable loss at each parameter budget, this gain is bounded by the model’s suboptimality plus the loss reduction available from the predictor’s parameters (Theorem 1), and a related bound covers the extra weak-model errors a learned router can capture in a two-model cascade (Theorem 2). Experiments with 16 models on four benchmarks, before and after task-specific finetuning, find that models with higher answer loss show larger predictor and routing gains, and that these gains shrink after finetuning.
Critical strengths
A clean and correct identity with a useful reading. Proposition 1 shows that a predictor’s advantage over self-confidence is exactly loss the model itself could recover. I checked the proofs of Propositions 1–3, Lemma 1, Theorems 1–2 and both corollaries and found no errors. This reframes a common practice in routing, where separate predictors are trained without self-confidence as a baseline.
Careful and transparent experiments. Predictors are retrained for every finetuned model, and the appendix reports the controls a reader would ask for: Platt calibration, three predictor sizes, within-stage correlations, and the split into weak- and strong-model errors, including results that go against the main claim. This is why I do not recommend clear rejection.
Critical weaknesses
The theory is close to tautological as stated, numerically uninformative, and not clearly separated from prior work. Assumption 1 defines the floor as the lowest loss over all models with at most P parameters on the fixed task distribution. That is a statement about which weights exist, not about trained models, and it is unobservable. Taken literally it can sit at the Bayes loss: over a finite benchmark, a lookup table attains it with few parameters. The capacity term is then zero and Theorem 1 reduces to “the gain is at most the answer loss above the Bayes loss”, which already follows from Proposition 1. If the floor is instead the empirical scaling-law frontier, the assumption need not hold for the combined model, because the predictor is trained on labelled task data. The bounds are never compared with the data. Plugging the paper’s empirical suboptimality into Theorem 2 and dropping the non-negative capacity term gives, averaged over escalation rates, roughly 10–15 percentage points at 0.05–0.1 nats and over 30 at 0.5 nats. The measured weak-error terms are 0.17–1.80 points (Table 9). Theorems 1 and 2 follow the template of Błasiok et al. (2024): a violation found by a size-k auditor that reads the model’s output yields a network of size n+k+2 with lower loss (their Lemmas 3.1–3.2), so violations are bounded by the loss drop between sizes, with the same square-root form (their Theorem 1.2), and their Appendix B covers cross-entropy. That paper is cited only in general terms.
The experiments do not test the bound, and the headline associations are largely explained by calibration and by the finetuning setup. Empirical suboptimality is answer loss minus a constant, and answer loss contains the self-confidence loss, which is also the first term of excess predictability. Miscalibration therefore raises both axes of Figures 1 and 2. A few released models on WinoGrande and CommonsenseQA have excess predictability above log 2, meaning raw self-confidence is worse than a constant base-rate predictor. On WinoGrande these points carry the correlation: Pearson is 0.98 but Spearman 0.73 (Table 6). The appendix control confirms this. Beyond two-parameter Platt calibration, the mean predictor gain on released models is only 0.005–0.072 nats (Table 8). The direction of the finetuning effect survives, but the paired-change correlations fall to 0.24, 0.40, 0.55 and 0.25; by my calculation only GSM8K is significant at 0.05. The only intervention is supervised finetuning on the task. The predictor is trained on 1,904–6,868 labelled in-distribution questions, the released model sees five demonstrations, and the finetuned model sees 5,088–32,313. The predictor’s edge over the released model may reflect who has seen task labels, which limits the routing takeaway to cases where the weak model can be finetuned on the query distribution. On GSM8K and MMLU-Pro the weak models are finetuned on reasoning traces from the model that also serves as the strong model. The fall in routing advantage may then reflect shared errors: on MMLU-Pro after finetuning, the 0.94 points of extra weak errors captured are offset by 0.87 points of extra strong errors (Table 9).
Questions for authors
Clarifications: Is the floor a minimum over all functions with at most P parameters, or over models reachable by training? For these tasks, what do you expect the capacity term to be at P around 8B and Q = 4B? How do Theorems 1 and 2 differ from the argument of Błasiok et al. (2024)? What are the excess predictability and routing advantage for the five excluded model pairs? Their accuracy fell after finetuning, so they test the relation in the other direction.
Additional evidence that would change my assessment (within the scope of the current experiments): Figures 1 and 2 with Platt-calibrated confidence as the main-text baseline, with confidence intervals. An equal-data control, for example a predictor trained on the finetuning set, or a weak model finetuned only on the predictor’s fitting split. The numerical value of the Theorem 2 bound next to the measured weak-error terms. Routing advantage on the generation tasks with a strong model that is not the finetuning teacher, and with confidence and predictor inputs from one or a few samples instead of 128.