acceptodds
Under review as a conference paper at ICLR 2027

CALIBER: Calibrated System-One Routing for Adaptive System-Two Reasoning

21 views

Abstract

Reasoning models improve performance on difficult tasks but incur substantial latency and computational cost when invoked indiscriminately. We study whether a fast calibrated decision model can learn when deliberative reasoning is actually necessary. CALIBER introduces a System-One/System-Two architecture in which a lightweight decision model observes the question and one cheap System One pass and selects among a direct System One answer from a 4B model, a direct answer from a 27B model and extended reasoning at either scale, and can defer when no action is likely to succeed. The router returns a calibrated probability of success for every action rather than an unconstrained textual plan, enabling explicit risk-sensitive policies that trade expected success against cost. We train the router on the realized outcomes of executing every route on every question, and compare it with confidence cascades, a prompted language-model router, conventional classifiers, a binary win classifier, random routing and an outcome oracle. Experiments on MMLU-Pro and SuperGPQA measure task success, calibration, risk–coverage trade-offs, routing regret and GPU cost, both in distribution and under held-out-subject and cross-benchmark shift. The central hypothesis is that calibrated decision routing can preserve most of the capability of expensive reasoning systems while substantially reducing their invocation rate. The hypothesis holds only in part. At half the cost of always reasoning with the 27B model, retains 83% of that route's accuracy gain over the System One answer, short of the pre-specified 90% target, while invoking System Two on 41% of questions. In the pre-specified comparison under the same budget cap it is 2.9 points more accurate than the best confidence cascade in distribution (81.5% versus 78.6%; 95% interval [1.3, 4.5] points), and about 2 points at matched realized cost. The gain comes from per-action success heads over a text encoder for the large direct and large reasoning actions, not from temperature scaling, System One confidence features or the small reasoning action, which lowers accuracy under this policy. Under held-out-subject and cross-benchmark shift, reasoning becomes longer and operating points transferred from the source calibration split miss their budgets; refitting temperatures and operating points on a small target calibration set restores budget adherence, with no significant difference from the cascade.

Primary area:
foundation or frontier models, including LLMs
Keywords:
adaptive reasoning, System One, System Two, model routing, calibrated uncertainty, selective prediction, agent intelligence, test-time compute, tool use, reasoning efficiency

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_mWedOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CALIBER studies budgeted routing between direct answering and extended reasoning at two model scales. A lightweight text encoder predicts the success probability of each route from the question and a cheap initial model pass. These predictions are calibrated and combined with costs to select a route, with optional deferral. Experiments on MMLU-Pro and SuperGPQA evaluate accuracy, calibration, cost, and distribution shift. CALIBER improves over the selected confidence cascade in distribution, but misses its capability-retention target and shows no significant advantage over the cascade after target adaptation.

Critical strengths

S1. Executing all four routes on every question provides a useful basis for comparing routing policies, examining differences between routes, and measuring the gap to an outcome oracle.

S2. The evaluation separates training, calibration, and testing, includes tuned confidence cascades, and reports multiple router seeds and a paired confidence interval for the main comparison.

S3. The paper clearly reports its negative results, including the unmet retention target, the limited effect of temperature scaling, and the improvement obtained by removing the small reasoning route. These findings provide useful evidence about when additional routing options help.

Critical weaknesses

W1. The binary baseline omits the large direct route, a key contributor to CALIBER’s performance. The comparison therefore mixes routing quality with differences in available options. A text-based baseline using the same routes and budget is needed, including a direct comparison with Route to Reason.

W2. In the shift experiment, recalibration is bundled with changes to cost estimates, routing settings, and the budget. Better budget control could come from policy retuning alone. At a fixed target budget, keeping success scores unchanged while updating costs and routing settings would help isolate the contribution of recalibration.

W3. Table 4 uses the much weaker small direct model as the only deferral baseline, so the accuracy gap also reflects differences in answering ability. Compare raw and calibrated confidence on the same routed answers, and include a cascade at the same budget. Calibration should also be measured for the answers actually selected by the router.

Questions for authors

Q1. With the same encoder and training data, how does a simple router choosing between the large model's direct and reasoning modes perform at matched realized cost?

Q2. At a fixed target-domain budget, what improvement comes from probability recalibration beyond updating the action-cost estimates and operating point?

Q3. Keeping the routed answers fixed, how do raw and calibrated confidence scores compare for deferral, and how well calibrated are the selected-action scores?

2 comments

Thank you for a constructive review; the revision now includes each of the three analyses you requested.

W1 / Q1 (same routes, same encoder). We agree that the binary baseline omitted 27B direct, which mixed routing quality with the choice of available options. Every baseline now chooses among 4B direct, 27B direct and 27B reasoning under the same budget rule. The binary win classifier reaches 83.6% (73.3% in the submission), and our router minus the classifier is +1.0 [0.4, 1.7].

For Q1, we restrict the comparison to the large model's two modes at the 50% budget. A binary classifier with the same encoder and training data reaches 84.4%, our router restricted to these two actions 84.7%, and a confidence cascade from 27B direct to 27B reasoning 79.7%. Our router minus that cascade is +5.01 [3.55, 6.49]. At a matched realized cost of 1.54 GPU-s, the descriptive frontier estimates are 85.4% and 85.7% for the two learned routers and 81.2% for the cascade. The submitted router's heads without 4B reasoning (the feature variant in Table 2) minus the binary classifier is +0.10 [-0.28, 0.51]. The gain over cascades thus comes from choosing between 27B direct and 27B reasoning from the question text, and a simple binary router does almost as well. Route to Reason (reimplemented) on the same three actions reaches 83.75% (+0.9 [-0.1, 1.9]), and we claim no advantage over it (Sec. 5.2, Table 2, Appendix E).

W2 / Q2 (what restores budget adherence). We agree that the submitted recalibration combined several updates. Table 3 now fixes the target budget at half the mean 27B-reasoning cost on the target calibration split, keeps the success scores unchanged and adds one update at a time. Updating only the planning costs and the operating point, using the router's own predicted success and no labels (A), brings cost to 0.99x and 1.05x of budget. Accuracy is then 75.5% on held-out subjects and 54.9% on SuperGPQA. Choosing the operating point with labels (A+L) gives 75.1% and 54.9%. For Q2, probability recalibration on top of these updates (A+L+T) adds no accuracy relative to A: -0.01 [-0.08, 0.06] and -1.94 [-2.49, -1.40]. It lowers mean head ECE from 0.052 to 0.043 and from 0.218 to 0.101. As you suspected, better budget control comes from retuning the policy. Setting A still requires executing every action on the target calibration questions, which costs 1.87 and 3.08 GPU-h (Sec. 5.3).

W3 / Q3 (deferral on fixed answers). We agree, and we replaced the submitted comparison with 4B answers (Sec. 5.4, Table 4). On the same routed answers at the 50% budget, calibrated and raw scores rank almost identically. Both give 89.8% selective accuracy at 80% coverage, with AURC of 0.061 and 0.062. Temperature scaling is monotone within each head, so it cannot change the ranking within a head. The answering route's own confidence ranks better (93.5%; AURC 0.055), but it is known only after the route runs. The cascade at the same budget reaches 87.1% (AURC 0.097). For the selected action, calibrated and raw ECE are 0.028 and 0.033 in distribution. After target refitting, they are 0.028 and 0.063 on held-out subjects and 0.083 and 0.240 on SuperGPQA. Calibration therefore matters for the probability values, not for the ranking. Appendix G gives the same analysis for the submitted router.

Thank you for the additional experiments. The response address my main experimental concerns. As the contribution remains limited and practical efficiency gains are unclear, I will maintain my rating of 2.

Reviewer_KKRxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

CALIBER estimates per-action success probabilities and selects among direct and reasoning routes using a calibrated accuracy-cost utility. Experiments on MMLU-Pro and SuperGPQA show an in-distribution gain over a tuned cascade, but miss the primary retention target and establish no significant advantage after target recalibration under distribution shift.

Critical strengths
  1. The paper is clearly organized and easy to follow, with explicit descriptions of the routing policy and evaluation protocol.

  2. Executing every route on each question enables paired policy comparisons and an outcome oracle, providing a useful basis for analyzing routing decisions.

Critical weaknesses

W1. The contribution beyond existing joint model-and-reasoning routing is limited, and a direct comparison with Route to Reason is missing. Ablations provide little support for the added components: calibration and System One features do not improve in-distribution routing accuracy, while removing the small reasoning action improves results.

W2. The primary retention target is missed, and the accuracy advantage is not established under distribution shift. Efficiency is estimated from amortized GPU-time accounting rather than measured on an executed routed workload.

Questions for authors

Q1. What advantage remains over Route to Reason or an uncalibrated per-action router under matched actions, training data, and realized costs?

Q2. Can temperature refitting be separated from cost and operating-point updates to identify what restores budget adherence under shift?

2 comments

Thank you for the clear summary and the two precise questions, which now structure Sec. 5.2 and Table 3.

W1 / Q1 (contribution, Route to Reason, uncalibrated router). We partly agree. The components added in the submission did not help: System One features (-0.14 [-0.47, 0.18]) and temperature scaling (+0.12 [-0.04, 0.28]) add no accuracy. Removing the small reasoning action raises accuracy by 3.0 [2.2, 3.9]. With three actions, the feature variant (with System One features) reaches 84.4% with raw heads and 84.5% calibrated. The recommended method drops these components. We compared Route to Reason (reimplemented) under the same actions, training data and budget rule. It reaches 83.75% against 84.6% for our router (+0.9 [-0.1, 1.9]). At a matched realized cost of 1.54 GPU-s, the descriptive frontier estimates are 85.7% and 84.5%. We therefore claim no advantage over Route to Reason, which also has a higher point estimate at the 25% budget (75.98% against 71.70%, at 27% higher realized cost; Table 1). Our main finding is that learned routing on the question text beats confidence cascades at matched budget, with our router as one instance of it. ThinkSwitcher (reimplemented) reaches a similar 84.0% at higher cost (+0.65 [0.02, 1.30]). The advantage that remains is over confidence cascades (+5.0 [3.5, 6.4] over the strongest cascade we evaluated) and routers that read scalar confidence (logistic regression on 4B confidence features: 78.1%), not over other learned text routers. The paper now states its contribution as this empirical comparison (Sec. 2, Position; Sec. 5.2, Table 2; Appendix E).

W2 (retention, shift, executed cost). We agree with all three parts. On retention, the submitted router keeps 83% of the gain at the 50% budget, below the 90% target set before evaluation (Appendix A). The recommended three-action router reaches 0.90, but it was chosen after the submitted results by calibration-split accuracy, so we do not count it as meeting the target (Sec. 3, Sec. 6).

Under shift, transfer gives no significant difference from the cascade (+0.09 [-0.78, 1.00]; +0.80 [-0.28, 1.84]), and we no longer suggest that the advantage transfers without refitting. The label-free setting A reaches 75.5% and 54.9%, against 75.1% and 53.0% for the cascade recalibrated with labels (+0.4 [-0.5, 1.3]; +1.9 [0.8, 3.1]). Setting A requires executing every action on 702 and 1,000 target questions, which costs 1.87 and 3.08 GPU-h (Sec. 5.3).

We now measure the cost of executed workloads (Sec. 5.5, Appendix H). Executing the submitted router's 50% workload cost 2.61 GPU-s per question, 1.85x its accounting estimate of 1.41. The gap comes from the 121 questions routed to 27B reasoning, which cost 4.3x their estimate. Executed accuracy is 80.8%, against 81.5% from the stored outcomes. We also executed the router (seed 17) and the three-stage cascade (the comparison of the original submission) at their 50% operating points, one after the other on one GPU. They cost 1.62 and 1.36 GPU-s per question, a measured cost ratio of 1.19 against an accounting ratio of 1.13, and their executed accuracies are 85.0% and 78.3% (+6.7 points). The main finding is stated in accounting units, and Sec. 6 explains that the executed ratio depends on subset size and batching.

Q2. Yes; Table 3 makes this separation. We fix the target budget at half the mean 27B-reasoning cost on the target calibration split and never refit the router weights. Transfer (T) misses the budget, spending 0.70x and 1.12x of it. Refitting costs and the operating point without labels (A) brings cost to 0.99x and 1.05x of budget and accuracy to 75.5% and 54.9%. Choosing the operating point with labels (A+L) changes little: 75.1% at 0.93x, and 54.9% at 1.05x. Refitting temperatures as well (A+L+T) adds no accuracy relative to A (-0.01 [-0.08, 0.06]; -1.94 [-2.49, -1.40]). It does lower mean head ECE, from 0.052 to 0.043 and from 0.218 to 0.101. Updating costs and the operating point therefore restores budget adherence, and temperature refitting improves the probabilities but not the routing.

Thank you for the clarifications and additional experiments. While these address several experimental concerns, I still find the contribution beyond existing learned routers limited and the practical efficiency gains insufficiently established. I will maintain my rating of 1.

Reviewer_qTfFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper trains a router that picks, for each multiple-choice question, one of four options: Qwen3-4B or Qwen3.8-27B, each with or without thinking. The 4B model first answers once without thinking. A ModernBERT classifier then reads the question, predicts how likely each option is to be correct, and picks the best accuracy/cost trade-off under a compute budget. Training labels come from running all four options on about 5k MMLU-Pro questions. At half the cost of always using 27B with thinking, the router is 2.9% better than a simple confidence cascade on the training distribution.

Critical strengths
  1. The problem matters. Thinking costs about 600 times more than answering directly, but many questions don't need it.

  2. The evaluation is honest. All four options are run on every question, the hypotheses are fixed in advance, code is released.

Critical weaknesses

W1 (main reason). It only works on the kind of questions it was trained on. The router is trained on about 5k MMLU-Pro questions. When the authors test it on subjects it has not seen, or on another dataset (SuperGPQA), it does no better than a simple rule that needs no training: start with the small model and move to a bigger one only when it is unsure (Table 5). On SuperGPQA its predicted chances of success are also far off (calibration error goes from 0.045 to 0.216). To fix this, the authors take 700-1,000 questions from the new dataset, run all four options on them, and use the correct answers to re-tune. This extra cost is not counted. Worse, in real use you don't have correct answers for incoming questions, and you don't know ahead of time what they will be. So I don't see how this would work in practice.

W2. The extra router is not clearly needed, and the paper's own results go against its story. The simple rule from W1 is only 2.9 points worse (78.6 vs. 81.5) and uses about 10% less compute. The two ideas in the title don't help: without the calibration step the router gets 81.8, and without the small model's answer as input 81.4, vs. 81.5 for the full method (Table 2). The intro says letting the 4B model think is the main reason a learned router is needed, yet removing that option raises accuracy to 84.5 and lowers cost. What is left is a text classifier that guesses which model will be right, which has been done before (e.g., Shnitzer et al., 2024).

W3. No comparison with published methods. The paper says Route to Reason is the closest work. It also cites ThinkSwitcher, which is almost the same idea, and models that learn by themselves when to think (AdaptThink, Thinkless). None of them is tested. The router is only compared with simple rules the authors wrote, so we can't tell whether it beats existing work.

W4. Some comparisons are unfair. The router can choose from four options. Four of the methods it is compared with (self-assessment, binary classifier, two-step rule, random) can only choose between two: 4B without thinking or 27B with thinking (Sec. 3). They cannot use 27B without thinking, which is cheap (about 1% of the cost of 27B thinking) and much more accurate than 4B (59.1 vs. 44.0; Table 1). The router uses this option for almost all questions it does not send to thinking (Sec. 5.2). This option matters a lot: the two-step rule gets 67.0, but the same rule with 27B-without-thinking added as a middle step gets 78.6 (Table 2). So the low scores of these four methods (65.5-73.3 vs. 81.5) say more about their limited options than about the router.

Questions for authors
  1. In real use, incoming questions come without correct answers. How would you re-tune the router for a new dataset, and how much compute does running all four options on those questions take?
  2. How does the router do against Route to Reason, ThinkSwitcher and AdaptThink/Thinkless under the same compute budget?
  3. Removing the 4B-with-thinking option makes results better, and the calibration step doesn't change accuracy. So which version is the method you are proposing?
1 comment

Thank you for a precise review; W1, W4 and Q1 in particular shaped the revision.

W1 / Q1 (shift). We agree that the in-distribution advantage does not transfer without adaptation. Under transfer, the router and the cascade do not differ significantly: +0.09 [-0.78, 1.00] on held-out subjects and +0.80 [-0.28, 1.84] on SuperGPQA. The router can, however, be re-tuned without answer labels (setting A, Table 3). We execute the actions on unlabeled target questions, refit the planning costs and choose the operating point with the router's own predicted success. Cost then comes to 0.99x and 1.05x of the target budget, and accuracy to 75.5% and 54.9%. The cascade recalibrated with labels reaches 75.1% and 53.0% (+0.4 [-0.5, 1.3]; +1.9 [0.8, 3.1]). Without labels, the cascade can keep its source thresholds (72.8% and 55.1%, at 0.73x and 1.19x of the target budget); setting them by the mean confidence of its answers instead favors the 4B direct answer and gives 62.0% and 45.8%. We therefore claim no label-free advantage over the cascade, only that setting A restores the budget and matches the labeled cascade.

You are right that the submission did not count the adaptation cost. Executing all actions on the target calibration questions takes 1.87 GPU-h for 702 questions, about the cost of answering 1,941 test questions at the 50% budget. For 1,000 questions it takes 3.08 GPU-h, about 2,935 test questions. Executing only 27B direct and 27B reasoning would take 1.36 and 2.11 GPU-h. Setting A also requires a sample of the target distribution, and Sec. 6 states this requirement. On SuperGPQA, target temperatures lower the ECE of the selected action from 0.240 with raw scores to 0.083 (mean head ECE from 0.218 to 0.101). They add no routing accuracy (-1.94 [-2.49, -1.40]; Sec. 5.3).

W2. We agree. The calibration step (+0.12 [-0.04, 0.28]) and the 4B-answer features (-0.14 [-0.47, 0.18]) add no accuracy, and the 4B reasoning action lowers it: removing that action gives +3.0 [2.2, 3.9]. The recommended method and the title no longer include these components. Calibration remains only for the probability values (Sec. 5.4). On novelty, you are right that per-model correctness prediction exists (Shnitzer et al., 2024; Route to Reason). We now claim the measured comparison, not a new architecture (Sec. 2, Position). Against the strongest cascade we evaluated (27B direct to 27B reasoning), the recommended router gains +5.0 [3.5, 6.4]; against the three-stage cascade of the submission, +6.1 [4.5, 7.6]. On compute, its accounting cost is 1.39 against 1.34 GPU-s (+0.05 [-0.13, 0.23]). At a matched realized cost of 1.54 GPU-s, the descriptive frontier estimate is 85.7% against 81.2% (Sec. 5.1).

W3 / Q2. We reimplemented Route to Reason on our actions, data and budget, and it reaches 83.75%. The difference from our router, +0.9 [-0.1, 1.9], includes zero, so we claim no advantage. ThinkSwitcher reaches a similar 84.0% at 1.53 GPU-s against 1.39 for our router (+0.65 [0.02, 1.30]). Both are reimplementations. The released Route to Reason code has no routing step, so we implemented that step from the paper. ThinkSwitcher has no code release, and we train it on one outcome per question where the original averages eight (Appendix E, Sec. 6). At the 25% budget, both have higher point estimates than our router (75.98% and 73.81% against 71.70%; Table 1), at higher realized cost (0.69 and 0.80 against 0.54 GPU-s). We therefore state the main finding as a property of learned routing on the question text, which beats confidence cascades at matched budget, with our router as one instance; we do not claim that it beats the closest published method. AdaptThink and Thinkless fine-tune the 27B model with reinforcement learning and cannot run on our frozen actions (Sec. 2).

W4. We agree, and all four baselines were rerun with 27B direct as a middle tier (Table 2). The binary win classifier reaches 83.6% (from 73.3%), self-assessment 76.6% and random routing 73.8%. The two-step rule becomes the three-tier cascade (78.6%). The gap to the binary classifier shrinks to +1.0 [0.4, 1.7]. The original tables are in Appendix A.

Q3. We propose the text-only per-action router over 4B direct, 27B direct and 27B reasoning. It was chosen after the submitted results, by calibration-split accuracy: 83.8%, against 83.2% for the variant with 4B features. On test, the variant with 4B features minus the text-only router is -0.14 [-0.47, 0.18] (Sec. 3, Sec. 5.2).

Comments on the paper

2 comments

We thank the three reviewers for careful, concrete reviews that changed the paper substantially.

The reviews agree on four points, and we accept all of them: the System One features and the calibration step of the submitted router added no accuracy; several baselines lacked the 27B direct answer; published routers were not compared; and the shift adaptation combined several updates without counting its own cost. Except for the executed-cost run (Sec. 5.5), every new analysis reuses the action outcomes executed for the submission, and all new analyses are descriptive.

  • Main finding (Sec. 5.1, Fig. 2, Table 1). Learned routing on the question text over the same three actions beats confidence cascades at a matched budget. Our instance is a per-action success router: a text encoder reads the question and options, and success heads cover 4B direct, 27B direct and 27B reasoning, with no System One features. At the 50% budget it answers 84.6% of in-distribution test questions correctly, against 79.7% for the strongest cascade we evaluated (27B direct to 27B reasoning): +5.0 points [3.5, 6.4] (+2.8 [1.1, 4.4] at 25%, +1.2 [0.2, 2.2] at 75%), at 1.39 and 1.34 GPU-s per question. Against the submission's three-stage cascade (78.6%) it is +6.1 [4.5, 7.6]. We chose this configuration after seeing the submitted results, by calibration-split accuracy (Sec. 3; Sec. 6).
  • Matched-action baselines (Sec. 5.2, Table 2). Every baseline now chooses among the same three actions under the same budget rule: binary win classifier 83.6% (73.3% in the submission), logistic regression 78.1%, self-assessment 76.6%, random 73.8%. Restricted to the 27B model's two modes, a binary classifier reaches 84.4% and the text router 84.7%.
  • Published methods (Sec. 2, Sec. 5.2, Appendix E). Route to Reason, reimplemented on our actions, data and budget, reaches 83.75%; our router minus Route to Reason is +0.9 [-0.1, 1.9], so we claim no advantage over it. ThinkSwitcher reaches 84.0% at higher cost. At the 25% budget both have higher point estimates than our router, at higher realized cost. AdaptThink and Thinkless retrain the reasoning model and cannot run on a frozen action set.
  • Shift (Sec. 5.3, Table 3). Table 3 separates transfer (T), a label-free refit of planning costs and operating point (A), a labelled operating point (A+L) and a further temperature refit (A+L+T). A restores budget adherence without labels but requires executing every action on the target calibration questions (1.87 GPU-h for 702 questions, 3.08 for 1,000). Refitting temperatures adds no accuracy.
  • Deferral (Sec. 5.4, Table 4, Appendix G). Calibrated, raw and answer-confidence scores now rank the same routed answers, with the cascade at the same budget as reference.
  • Executed cost (Sec. 5.5, Appendix H). Executed on one GPU at the 50% budget, the recommended router and the three-stage cascade cost 1.62 and 1.36 GPU-s per question (measured ratio 1.19, accounting ratio 1.13), with executed accuracies of 85.0% and 78.3%. The main finding is stated in accounting units.

Summary of changes (revision round 1)

  • Title: the revised PDF is titled "When Should a Language Model Reason? Routing on Executed Outcomes versus Confidence Cascades", naming the comparison the evidence supports.
  • Main finding and recommended configuration: Sec. 1, Sec. 3, Sec. 5.1, Fig. 2, Table 1.
  • Matched-action baselines and routers over the 27B model's two modes, also at matched realized cost: Sec. 3, Sec. 5.2, Table 2.
  • Published routers: Sec. 2, Sec. 5.2, Appendix E.
  • Shift decomposition and adaptation cost: Sec. 5.3, Table 3; the shift comparator is the three-stage cascade.
  • Deferral re-analysis, with ECE of the selected action's score: Sec. 5.4, Table 4, Appendix G.
  • Executed cost: Sec. 5.5, Appendix H, including the submitted router's executed workload (1.85x its accounting estimate).
  • Figures: a concept figure contrasting the router with a cascade (Fig. 1) and accuracy against cost (Fig. 2).
  • Presentation: terms defined at first use; the original pre-specified hypotheses and tables are in Appendix A (Tables 5, 6), unchanged.
  • Limitations (Sec. 6): rewritten around models and tasks, accounting versus executed cost, adaptation under shift, the reimplemented baselines and the post-hoc choice of configuration.
Area_Chair_t8cuOpenReview ↗

Reminder: Author–Reviewer Discussion Deadline

Dear Reviewers,

This is a friendly reminder that the author–reviewer discussion period for ICLR 2027 will end on November 18, 2026 (AOE).

As the deadline approaches, I encourage you to actively engage in discussions with the authors and carefully consider their rebuttal responses. Please address any remaining concerns, seek further clarification where necessary, and update your assessments if appropriate.

Your timely and constructive participation is essential to ensuring a fair and well-informed evaluation.

Thank you very much for your time and effort!

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.