- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper studies how advice-routing updates affect a language model's ability to rank its incorrect answers above correct ones. Starting from confidence-calibrated Qwen3-4B and Llama-3.1-8B checkpoints, the authors isolate selected low-rank components of the routing update and evaluate the resulting partial routers on fixed question-answer pairs. In Qwen, these components recover 51–60% of the full router's gain in error AUROC. Replacing their activity with constants retains roughly half of this gain, while restoring activity in later layers recovers some of the loss. The findings are less consistent for Llama and do not translate into a consistent improvement in advisor utility.
Critical strengths
- The fixed-answer evaluation keeps answer correctness unchanged across models, making the AUROC comparisons more interpretable. The authors also evaluate answer generation and advisor utility separately.
- The result is checked on held-out questions from two datasets and across routing seeds. The paper reports uncertainty estimates and retains the weaker Llama and utility findings.
- The activity study includes mean replacement, layer restoration, donor-question controls, and activation-magnitude-matched controls. These are useful checks on the effects of the selected updates.
Critical weaknesses
- The central motivation and task distinction are not convincing. The Introduction argues that assessing an answer and deciding to ask for advice are different abilities, but the main target is error AUROC computed from the Ask/Keep logits. A good estimate of an answer's correctness can already be used to rank errors and implement a confidence-threshold routing policy. The paper does not explain what additional routing capability is being studied, especially since the paper does not establish whether the score reflects the advisor's relative performance.
- The routing training labels are insufficiently specified. Section 2.3 and Appendix A say that 497 candidate/label pairs are used to supervise the final A/K token, but do not explain how an example is labeled Ask or Keep. Is the label determined by whether the student's answer is wrong, whether the advisor can improve it, or a cost-sensitive criterion? This is essential to understanding what the routing update has learned. The evaluation correctness labels described in the paper do not resolve this question.
- A critical confidence-based baseline is missing. The starting model has already been trained to report confidence, yet the main comparison uses its A/K logits rather than its confidence-derived error score. The paper should compare error AUROC using 1 − confidence on the same fixed answers, and compare a tuned confidence-threshold router with the learned routers under the same advisor and call costs. Without these comparisons, the incremental value of routing training over existing self-assessment is unclear.
- The Introduction and related-work positioning are too thin for the question being asked. The paper moves quickly from a short distinction between assessment and action to low-rank weight analysis, with little discussion of confidence estimation/calibration, error discrimination, selective prediction, abstention, or confidence-based deferral. Some relevant studies appear in Related Work, but the gap relative to these lines of research is not clearly established. The motivation for isolating these particular routing updates also needs to be explained more directly.
- The mechanistic and practical conclusions remain limited. The selected effect is robust mainly in Qwen, with mixed results in Llama. Replacing activity across the prompt removes positional and question-dependent variation, so the restoration effect does not identify correctness-specific information. The findings rely on two factual QA datasets and automated correctness labels. Most importantly, the improved error ranking does not yield a consistent advisor-utility advantage over the tuned full router.
Questions for authors
- How were the 497 A/K labels constructed? Did their construction use student correctness alone, advisor outcomes, or consultation costs?
- What are the error AUROC and cost-matched routing utility of the calibrated starting model when decisions are based directly on its reported confidence?
- What concrete behavior does the routing-trained model learn beyond confidence-based error ranking? Can the authors test cases with similar student confidence but different advisor outcomes?
- Can the authors further separate correctness sensitivity from question familiarity or other prompt information, for example by comparing correct and incorrect candidate answers to the same question?
2 comments
Reviewer Response
Thank you for the detailed response and additional experiments. The clarification of the A/K labels and the confidence baselines address my corresponding questions. The new results also show that the A/K score after routing training ranks errors better than the model’s reported confidence, and I acknowledge this finding.
However, my concerns about the central motivation and contribution remain. The authors confirm that the A/K labels are determined entirely by the correctness of the student’s answers, without considering the advisor’s capability or consultation costs. The additional experiments also find no advisor-relative information beyond confidence. I therefore still view this training primarily as learning to recognize the model’s own errors, rather than establishing a help-seeking capability substantively distinct from self-assessment. The higher error-recognition AUROC demonstrates that this training is effective, but does not by itself establish the proposed distinction between the tasks.
Having considered these clarifications and additional results, I maintain my original rating.
Response to Reviewer o65E on Confidence Baselines, Labels, and Advisor-Relative Information
Thank you for the detailed review. Several of your points identified real omissions, and the confidence baseline in particular changed how we present the paper. Points shared with other reviewers are in the general response, and the rest follow here.
Q1 / W2 (how the A/K labels were built). A pair is labeled K if the starting model's answer is graded correct and A otherwise. Of the 497 training pairs, 431 are A and 66 are K. Nothing about the advisor or the call cost enters the label. The routing update is trained to predict the student's own errors, and Section 2.3 now says so.
Q2 / W3 (confidence baseline and confidence-threshold router). The table compares error AUROC on identical fixed answers.
All six router-minus-confidence intervals exclude zero. For the tuned confidence-threshold router you describe (same advisor, costs and discovery-fitted thresholds), utility is on par with the tuned full router. Across the three Qwen seeds, differences range from −0.037 to +0.022, nearly all intervals include zero, and the few that do not point in both directions. Q3 explains why.
Q3 / W1 (what routing adds beyond confidence). First, the advice decision carries error information that the model's confidence does not. The router's advice score (0.909) beats both the starting model's confidence (0.782) and the confidence the router itself reports after training (0.776). We think this gap between what the model reports and what it does is the clearest statement of what routing training adds, and the revised introduction is built around it.
Second, we ran your suggested test for advisor-relative information. Among answers the student got wrong, the advice score separates advisor-correct from advisor-wrong cases with AUROC 0.32 to 0.47 across seeds and datasets, against 0.34 to 0.45 for confidence. The differences (−0.017 to +0.013) and the advice-score coefficient after controlling for confidence (−0.23 to +0.08) all have intervals including zero. So the router carries no advisor-relative information beyond confidence, which fits its training labels. Both scores sit below 0.5, which means the errors they flag most strongly tend to be the ones the advisor also gets wrong. This is why better ranking does not improve utility here, and the revision says so.
Q4 (familiarity vs. correctness). The other-question control preserves positional layout, and same-question restoration still beats it by 0.011 to 0.014 AUROC. We agree that your proposed comparison of correct and incorrect answers to the same question is the decisive test. We did not complete it within the discussion period, so the revision presents the restoration result as question-level evidence only.
W4 (motivation and related work). You are right that the introduction moved too quickly to the weight analysis. It now starts from the numbers above. The starting model's advice decision ranks errors at 0.705 although its own in-answer confidence reaches 0.782, and routing training raises the decision to 0.909. The fixed-answer design attributes that improvement to the routing update rather than to changed answers, which is why we study the update itself. Section 6 now covers selective prediction, calibration, abstention, learning to defer and LLM cascades. Those works estimate confidence or decide how to act on it. We ask how a trained update produces a better error signal inside the decision.
W5 (limits). We agree, and the revision states each one.
The contribution is explanatory, and the abstract and conclusion now say so.