acceptodds
Under review as a conference paper at ICLR 2027

How Advice-Routing Updates Shape Error Recognition in Language Models

Abstract

Language models that can consult an advisor must decide which of their answers need help. We study how training a model to request advice changes its ability to rank incorrect answers above correct ones, which we call error recognition. Starting from models trained for confidence calibration, we isolate part of the learned routing update and intervene on the activity it produces. In Qwen, a distributed subset of the update improves error ranking when all model variants judge the same proposed answers. Replacing this activity throughout the prompt with fitted constants preserves roughly half of the partial router's improvement over its starting model. Restoring activity in a later group of selected layers recovers much of the lost performance, and activity from the same question gives greater recovery than activity from another question. This pattern repeats across two factual question-answering datasets and a second routing seed. The results reveal a substantial gain under constant internal additions and a smaller, reproducible benefit from question-specific restoration. They do not establish a dedicated correctness representation or a consistent advantage in advisor decisions.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, language models, error detection, confidence calibration, learning to defer, low-rank adaptation, causal interventions

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_o65EOpenReview ↗
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
AuthorsOpenReview ↗

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.

PopQATriviaQA
1 − in-answer confidence, starting model0.7820.757
1 − confidence readout, starting model0.6810.737
Advice score, full router (3 seeds)0.902 to 0.9090.853 to 0.859
Full router minus in-answer confidence+0.120 to +0.127+0.097 to +0.102

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 mechanism is shown only for Qwen.
  • Mean replacement removes positional and question-level variation, so the paper no longer claims answer-specific information.
  • Labels are automated, although the reference-only rule reproduces every primary result.
  • There is no utility advantage, for the reason given in Q3.

The contribution is explanatory, and the abstract and conclusion now say so.

Reviewer_o65EOpenReview ↗

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.

Reviewer_RPEuOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates how weight updates learned during advice-routing training affect error recognition in language models. Whether selected components of the routing update improve the ability to rank incorrect answers above correct ones is the core idea of the work. Using SVD-based weight selection and controlled activation replacement and restoration experiments, the work identifies a distributed subset of routing updates in Qwen that contributes to improved error recognition.

Critical strengths
  • The fixed-answer evaluation provides a controlled setting for studying error recognition independently of changes in generated answers.

  • The study includes relevant controls involving mean replacement, layer-specific restoration, other-question activity etc

  • The statistical analysis uses paired bootstrap comparisons with multiple-testing adjustments.

Critical weaknesses
  • they key weakness of the paper is mathematical clarity and notations. Several important mathematical quantities are introduced without immediately specifying their shapes, scaling factors, and connections to previously defined objects.

  • The focused mechanistic results are established only for Qwen, while Llama exhibits substantially weaker and less consistent effects.

  • The activation-restoration experiments establish that certain input-dependent activities contribute to error-ranking performance, but do not show that these activities specifically encode correctness.

Questions for authors
  • Why are the mechanistic findings substantially less consistent for Llama? Do the results generalize to independently calibrated checkpoints and other types of reasoning tasks?

  • What practical benefit do the identified components provide when downstream advisor-selection policies are already optimized using the full router?

2 comments
AuthorsOpenReview ↗

Response to Reviewer RPEu on Notation, Llama, and Practical Value

Thank you for the review. We agree that the notation was hard to follow. The revised Section 3.1 now opens with a short notation box, and the later sections refer back to it.

W1 (notation). The new box in Section 3.1 reads as follows.

  • Routing update. Each adapted projection jj (query, value, output, up or down) has input and output dimensions din,jd_{\text{in},j} and dout,jd_{\text{out},j}. Its LoRA factors are Aj∈Rr×din,jA_j\in\mathbb{R}^{r\times d_{\text{in},j}} and Bj∈Rdout,j×rB_j\in\mathbb{R}^{d_{\text{out},j}\times r}, with rank r=8r=8, scaling sj=α/r=16/8=2s_j=\alpha/r=16/8=2, and update ΔWj=sjBjAj\Delta W_j=s_jB_jA_j.
  • Dimensions (din→doutd_{\text{in}}\to d_{\text{out}}). For Qwen3-4B, query 2560→4096, value 2560→1024, output 4096→2560, up 2560→9728 and down 9728→2560. For Llama-3.1-8B, query 4096→4096, value 4096→1024, output 4096→4096, up 4096→14336 and down 14336→4096.
  • Selected components. With the SVD ΔWj=UjΣjVj⊤\Delta W_j=U_j\Sigma_jV_j^\top, we keep the top two components, ΔWj(2)=Uj,2Σj,2Vj,2⊤\Delta W_j^{(2)}=U_{j,2}\Sigma_{j,2}V_{j,2}^\top.
  • Selected branch. Aj(S)=Σj,21/2Vj,2⊤∈R2×din,jA_j^{(S)}=\Sigma_{j,2}^{1/2}V_{j,2}^\top\in\mathbb{R}^{2\times d_{\text{in},j}} and Bj(S)=Uj,2Σj,21/2/sj∈Rdout,j×2B_j^{(S)}=U_{j,2}\Sigma_{j,2}^{1/2}/s_j\in\mathbb{R}^{d_{\text{out},j}\times 2}, so that sjBj(S)Aj(S)=ΔWj(2)s_jB_j^{(S)}A_j^{(S)}=\Delta W_j^{(2)} exactly. The partial router adds this branch to the starting model in 45 projections (layers 9 to 17).
  • Selected activity. At token tt, with incoming hidden state hj,th_{j,t}, the activity is zj,t=Aj(S)hj,t∈R2z_{j,t}=A_j^{(S)}h_{j,t}\in\mathbb{R}^2, and the branch adds sjBj(S)zj,ts_jB_j^{(S)}z_{j,t}.
  • Mean replacement. zj,tz_{j,t} is replaced by a fitted mean μj,d,w∈R2\mu_{j,d,w}\in\mathbb{R}^2 (Eq. 4), so the branch adds the fixed vector sjBj(S)μj,d,ws_jB_j^{(S)}\mu_{j,d,w} at every prompt position.
  • Advice score. z(q,a)=ℓA−ℓKz(q,a)=\ell_A-\ell_K, the difference between the next-token logits for A and K at the end of the prompt.

W2 / Q1 (Llama and generalization). The clearest difference is headroom. Llama's untrained advice score already ranks errors better than either of its confidence measures, with 0.869 against 0.812 and 0.750 on PopQA, and 0.778 against 0.612 and 0.693 on TriviaQA. Routing training then adds only 0.009 to 0.059 AUROC, against 0.085 to 0.204 for Qwen, so there is much less to localize. All routing seeds also share one calibrated checkpoint per family. We have not tested independently calibrated checkpoints or other task types such as reasoning, and the revision makes no mechanistic claims beyond Qwen and factual QA.

W3 (do the restored activities encode correctness?). We agree that restoration alone does not show this. The other-question control keeps the prompt's positional layout, and same-question restoration still beats it by 0.011 to 0.014 AUROC, so position alone does not explain the effect. Separating correctness from question familiarity requires scoring correct and incorrect answers to the same question, which we did not complete within the discussion period. The revision therefore describes the restored activity as carrying question-level information and makes no claim that it encodes correctness specifically.

Q2 (practical benefit). The selected components do not beat a tuned full router, and we do not claim they do; the paper is explanatory. The routing score itself ranks errors far better than the model's own confidence, at 0.902 to 0.909 against 0.782 for the in-answer confidence on PopQA, with intervals excluding zero (general response, point 1). At the decision level, confidence-threshold routers perform on par with the tuned full router, because this advisor tends to fail on the same questions the router flags. The paper explains part of how the ranking improvement is produced inside the model.

If any definition is still unclear, we would be glad to clarify it.

Thank you for the response - I appreciate the improvements to the mathematical notation this helps understand the paper a bit better than before - yet some areas could be written in a more easier to readable manner.

Nevertheless, some important concerns remain unresolved,

  • the limited generalization to Llama (without this the work is simply evaluated on Qwen which seems to narrow to my end - ideally the method should be generalizable across multiple LLMs right ?). The authors' explanation involving Llama's limited headroom is plausible, not a demonstrated causal explanation
  • The inability to distinguish correctness-specific information from general question-dependent activity,
  • the weaker evidence that the selected components are necessary within the full router.

It seems to me that the primary contribution is explanatory rather than focused on downstream performance improvements, which helps clarify the paper's intended significance.

Additionally the most important experiment remains missing:

  • Same question, correct answer.
  • Same question, incorrect answer.
  • Compare the corresponding restored activities and their effects on error-ranking performance. Without this, the mechanistic interpretation remains limited.

Given these clarifications and the authors' acknowledgment of the limitations, I am inclined to retain my overall rating and increase the soundness score by 1, noting that the methodological section requires significant notation overhaul.

Reviewer_yKwcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how parameter updates affect a language model's ability to identify incorrect answers after routing training involving either retaining the proposed answer or seeking help from another model. Starting with confidence-calibrated Qwen and Llama models, the authors fix candidate answers and analyze changes in error-ranking ability by selecting partial LoRA updates, replacing internal activities, and restoring local activities. Results show that in Qwen, selected partial updates can recover a significant portion of the gains from full routing training; replacing its activity with fitted constants retains about half of the gains, while restoring the normal activity of the current question in some layers recovers most of the loss, and is more effective than restoring the activity of other questions. These findings are validated on two factual question-answering datasets and another routing training seed, but parameter experiments on Llama are inconsistent, and actual help evaluation does not show a stable gain advantage over the threshold-optimized full route model.

Critical strengths
  1. Replacing the selected branches’ activity with fixed signals retained approximately half of the partial router’s gain, while restoring same-question activity further improved error ranking. These interventions provides evidence beyond overall performance comparisons.
  2. Fixed candidate answers, separate discovery and confirmation sets, replication across routing training seeds, and controls for activity source, layer location, and magnitude strengthen the credibility of the results.
Critical weaknesses
  1. The cross-model applicability of the core findings remains uncertain. The main results of full instant mean replacement and activity recovery are only validated on the Qwen dataset, and both route training versions share the same starting checkpoint. Parametric experiments on the Llama dataset show poor consistency, so it is unclear which models or training conditions these findings depend on, thus limiting the generalizability of the conclusions.
  2. The differences between the results of adding and removing the update lack further explanation. Adding the selected update consistently improves the false detection rate, but when the update is removed from the full route, only one of the six Qwen comparative experiments shows a significant decrease after multiple comparison correction. The authors do not further differentiate between compensation for other updates, parameter interactions, or statistical uncertainties, so the role of this update in the full route remains unclear.
  3. The reason why activity recovery is effective is not elucidated. The effectiveness of the current hint is more helpful than other hints, but the information obtained could be the correctness of the answer, familiarity with the question, the content of the answer, or the positional structure. The lack of targeted comparisons (e.g., fixing the questions while changing the correctness of candidate answers) limits the interpretation of the internal mechanisms revealed by the current results.
Questions for authors
  1. Do the weak and inconsistent parameter experimental results in Llama stem primarily from the initial model's capabilities, small routing training gains, or an unsuitability of the selected layer region and component structure for the model? Are there existing analyses that can differentiate between these explanations and explain the conditions under which Qwen's findings might hold?
  2. Regarding the deletion effect of selected updates in the full router, are there further analyses that can distinguish between compensation by other updates, parameter interactions, and statistical uncertainties? Will similar effects still be observed when performing corresponding activity replacements or restorations within the full router?
  3. If the question remains unchanged and only the correctness of the candidate answers is altered, do the selected activities and their restoration benefits change accordingly? Can this type of validation distinguish the contribution of answer correctness information from question familiarity or general content information?
2 comments
AuthorsOpenReview ↗

Response to Reviewer yKwc on Redundancy in the Routing Update, Restored Activity, and Llama

Thank you for the careful summary and for questions that pointed to concrete analyses.

Q2 / W2 (why removal is weaker than addition). The two halves of the routing update overlap heavily. In all six Qwen panels, the selected components alone recover 51 to 60% of the full router's gain, and the remaining components alone recover 74 to 91%, so the two gains add up to well over the full gain. The interaction [A(S)−A(C)]−[A(R)−A(R∖S)][A(S)-A(C)]-[A(R)-A(R\setminus S)] is positive in all six panels, from +0.035 [+0.012, +0.058] to +0.078 [+0.039, +0.117] AUROC, and every interval excludes zero. When the selected components are removed, the rest of the update covers most of what they did. Of the explanations you list, this points to compensation through overlap rather than statistical noise. Section 3.2 now includes this analysis.

Running the same activity intervention inside the complete router, as you suggest, is the natural next test. We did not complete it within the discussion period, so the revision states that the activity results are established in the partial router only and lists this test as a next step.

Q3 / W3 (what the restored activity carries). We address each alternative you raise in turn.

  • Position. The other-question control restores activity from another question aligned by relative position. It keeps the positional layout, and same-question restoration still beats it in all four panels (0.011 to 0.014 AUROC).
  • Correctness vs. familiarity. Your proposed comparison, which fixes the question and varies only the answer's correctness, is the decisive test. We did not complete it within the discussion period, so the revision limits the claim to question-level evidence.
  • Answer content. In natural answer pairs, content and correctness change together, so the two cannot be fully separated. The revision says so explicitly.

Q1 / W1 (Llama). The data point mostly to headroom. Llama's untrained advice score already ranks errors better than either of its confidence measures, reaching 0.869 on PopQA against 0.812 for in-answer confidence and 0.750 for the readout, and routing training adds only 0.009 to 0.059 AUROC, against 0.085 to 0.204 for Qwen. Consistent with this, Llama shows no reliable redundancy pattern, and its interaction intervals include zero in five of six panels. We cannot rule out that the layer region and component structure suit Llama less well, since they were chosen on Qwen and Llama's own discovery screen did not pass; with this little routing gain, the data cannot separate the two explanations. You are also right that the routing seeds share one calibrated starting checkpoint. The revision states both limits and keeps mechanistic claims to Qwen.

Thanks again for the constructive questions.

Thank you for your reply, I will keep my rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Confidence Baselines, Label Construction, Llama, and What Restoration Shows

We thank the reviewers for their time and comments. A few concerns came up in more than one review, so we answer them once here. All new numbers come from analyses of the paper's existing records, with paired question-bootstrap 95% intervals. The revised PDF marks all changes in blue.

1. Does routing training add anything beyond the model's own confidence? (o65E, RPEu)

Yes. We should have reported this comparison in the submission. On the same fixed answers (original Qwen router, canonical labels):

Score used to rank errorsPopQATriviaQA
Starting model, in-answer confidence (what calibration training optimized)0.7820.757
Starting model, separate confidence readout0.6810.737
Starting model, advice score0.7050.768
Partial router, advice score0.8180.818
Full router, advice score0.9090.853
Full router, its own confidence readout0.7760.750

Across the three routing seeds, the full router's advice score beats the in-answer confidence by 0.120–0.127 AUROC on PopQA and 0.097–0.102 on TriviaQA, and every interval excludes zero (e.g., +0.127 [+0.090, +0.166] for the original router on PopQA). The gap holds under reference-only labels (0.747 vs. 0.870–0.875 on PopQA). Two further points stand out. On PopQA, the starting model's own in-answer confidence (0.782) already ranks errors better than its advice decision (0.705), and routing training closes that gap and goes beyond it. After routing training, the router's own confidence report still ranks errors much worse than its advice score (0.776 vs. 0.909). The error signal that routing training builds into the decision is not one the model expresses as confidence, and that signal is what the paper sets out to explain.

Better rankings do not become better decisions with this advisor, and we now say so plainly. With the same advisor, call costs and discovery-fitted thresholds, routers that threshold the starting model's confidence perform on par with the tuned full router. Across the three Qwen seeds, both confidence measures and both datasets, utility differences range from −0.037 to +0.022; nearly all intervals include zero, and the few that do not point in both directions. The advisor explains this. On PopQA it fixes 47 student errors but breaks 17 correct answers (of 376), and among student errors both the advice score and confidence flag most strongly the questions the advisor also gets wrong (reply to o65E, Q3).

2. How the routing labels were built (o65E)

A training pair is labeled K when the automated grader marks the starting model's answer correct, and A otherwise: 431 A and 66 K among the 497 pairs (316 PopQA and 181 SimpleQA questions, none used in evaluation). Advisor outcomes and call costs never enter the label, so the router is trained to predict the student's own errors. Section 2.3 now states this.

3. Llama (all reviewers)

Llama's untrained advice score already ranks errors better than either of its confidence measures (PopQA: 0.869 vs. 0.812 in-answer and 0.750 readout; TriviaQA: 0.778 vs. 0.612 and 0.693). Routing training then adds only 0.009–0.059 AUROC, against 0.085–0.204 for Qwen, so the Llama runs leave little to localize. We now limit mechanistic claims to Qwen and state that the routing seeds in each family share one calibrated checkpoint.

4. Answer correctness or question familiarity? (all reviewers)

On position, the other-question control restores activity from a different question aligned by relative position, so it keeps the prompt's positional layout. Same-question restoration still beats it in all four panels (0.011–0.014 AUROC). On familiarity, the decisive test is to score correct and incorrect answers to the same question, as several of you proposed. We did not complete it within the discussion period, so the revision presents the restoration result as evidence about question-level information, not answer-specific error detection, and lists this test first among next steps.

Other changes. A notation box now defines every shape and scaling factor (RPEu). A redundancy analysis with intervals explains the addition/removal asymmetry (yKwc). Related work now covers selective prediction, calibration, abstention and confidence-based deferral (o65E).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.