acceptodds
Under review as a conference paper at ICLR 2027

ECR: Evidence-Certified Revision for Self-Correcting Long-Video Agents

Abstract

Long-video agents increasingly answer questions by actively searching for task-relevant evidence. Once an answer has been formed, a separate decision arises: when does new evidence justify revising it? Evidence that supports an alternative answer does not necessarily refute the current one. We introduce ECR (Evidence-Certified Revision), a training-free framework for post-answer revision. ECR keeps the current answer as an anchor and treats a new answer as a challenger. It separately evaluates challenger support, anchor refutation, and evidence admissibility, records these judgments in a structured certificate, and applies a fixed reason-conditioned policy to retain the anchor, switch to the challenger, or escalate selected residual conflicts to blind pairwise arbitration that does not reveal anchor identity. On Video-MME Long, the full evaluation pipeline improves a frozen AVP agent from 52.11% to 62.33% accuracy while preserving 94.9% of initially correct answers. On a 655-question exact-replay cohort with frozen base executions and proposal outputs, ECR reaches 65.50% accuracy with 99 fixes and 13 breaks, compared with 65.19%, 141 fixes, and 57 breaks under unguarded proposal adoption. Under matched evidence access and nearly matched inference budgets, ECR-K2 reaches 55.29% accuracy versus 52.16% for a compute-matched symmetric selector, with more fixes and fewer breaks. Across five heterogeneous base agents on a shared 128-question cohort, ECR yields positive point estimates in all five settings. These results support treating the decision to revise as distinct from the process that produces complementary evidence.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Long Video Understanding, Multimodal Agents, Evidence-Certified Revision, Belief Revision, Evidence Reasoning, Video Question Answering

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_P8LROpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ECR, a training-free framework that uses structured evidence judgments and a fixed policy to retain, revise, or arbitrate a long-video agent’s answer. Experiments show improved accuracy and fewer harmful revisions.

Critical strengths
  1. Relevant problem. The paper addresses harmful revisions in self-correction and makes a sensible distinction between supporting a new answer and refuting the original one.

  2. Useful empirical results. ECR substantially reduces harmful revisions while maintaining accuracy comparable to unconditional challenger adoption.

  3. Informative evaluation. The paper reports fixes and breaks separately and includes replay, ablation, and approximately compute-matched comparisons.

Critical weaknesses

My expertise is outside long-video understanding, so I may be missing some field-specific context. My main concerns are about the methodological contribution and the experimental evidence supporting it.

  1. Limited methodological novelty. The method appears implementable as a reusable agent skill combining prompting and fixed decision rules. It is unclear what methodological contribution it offers beyond such a workflow.

  2. Insufficient compute-matched baselines. The main gains include additional search and verification. Comparisons with equally resourced AVP search/reflection or generic self-correction are needed to establish ECR’s specific benefit.

  3. Limited backbone validation. Most experiments rely on Qwen3-VL-Plus. The 48-question GPT-5.5 experiment is insufficient to establish effectiveness across stronger contemporary multimodal models.

Questions for authors

As weaknesses

4 comments

Thank you for your review and comments! We respond to each point below, with new results on extra search, reflection and a stronger backbone. The changes in the revised paper are marked in blue with a bar in the margin, and each change carries a margin tag naming the corresponding comment. The tags for your comments are purple (for example, P8LR·W1), and those for the other reviewers use other colors.

W1: Methodological contribution beyond a workflow of prompts and fixed rules

Thank you for raising this question. ECR is implemented with prompts and a fixed policy, as you observe, and its contribution lies in how the revision decision is formulated. A symmetric selector asks which candidate is more credible. ECR asks whether the new evidence gives sufficient grounds to overturn the current anchor, which makes revision a directional decision on evidence permission. Point (1) describes what this formulation adds, and point (2) gives the evidence that it matters.

(1) What ECR offers beyond prompting and fixed decision rules. It offers a criterion that determines what the prompts judge and what the rules permit, in three respects. First, support for the challenger (SPS_P) and refutation of the anchor (RAR_A) are judged separately, so that local support for the challenger is not mistaken for grounds to overturn the existing answer. Second, DPD_P checks whether the evidence is sufficient in time, logic and coverage. For example, an event not being observed in a local segment does not mean that it never occurs in the whole video. Third, the certificate records the evidential state and its reason, and the fixed policy maps them to RETAIN, SWITCH or ESCALATE. Every action can thus be traced to a specific evidence relation, which a single preference score would not allow.

(2) Empirical evidence that this criterion makes a difference. The criterion changes the outcome when the form of the workflow is the same. A simple conservative gate and the symmetric selector are also built from prompts and fixed rules. At the same revision rate on 491 questions, ECR obtains 57 fixes and 13 breaks against 52 and 17 for the gate, and is 1.83 points more accurate (exact McNemar p = 0.0352). On the extended matched control of 655 cases, ECR reaches 54.81% against 51.91% for the symmetric selector (exact McNemar p = 0.0271). The core contribution of ECR is therefore an explicit, traceable and comparable criterion for revision decisions, which goes beyond adding one more general-purpose agent skill.

Revision: Contribution (3) in Section 1 now includes the rate-matched control, and Section 4.4 reports both comparisons.

W2: Comparison with extra search and reflection for AVP

As you note, the full pipeline adds search and verification on top of AVP. To establish what ECR itself contributes, we gave AVP extra search or generic reflection and compared them with ECR on the same initial AVP answers. Neither extra search nor generic reflection improves on AVP, while ECR raises accuracy by 8.59 points at nearly the same added cost (point (1)). The specific benefit of ECR thus lies in how the extra computation is used (point (2)).

(1) AVP with extra search or generic reflection compared with AVP with ECR. Extra search and generic reflection do not reproduce the gain of ECR. We used a newly sampled set of 256 questions, independent of ECR-SCOPE-256, and applied each method to the same fixed initial AVP answers, under the same cap of 15,000 added tokens per question.

Method (256 questions)AccuracyRelative to AVPNet correctionAdded calls / tokens per question
AVP134/256 (52.34%)
AVP + extra search133/256 (51.95%)−0.39 points−12.64 / 12,390
AVP + generic reflection121/256 (47.27%)−5.08 points−132.63 / 12,420
AVP + ECR156/256 (60.94%)+8.59 points+222.62 / 12,450

Extra search changes accuracy by −0.39 points and generic reflection by −5.08 points, with net corrections (fixes minus breaks) of −1 and −13. ECR gains 8.59 points with a net correction of +22, which is 8.98 points above extra search. Paired tests give p = 0.00047 for ECR against AVP and p = 0.00082 for ECR against extra search. The cost in the last column is what each method adds on top of the frozen AVP run. It is nearly the same for the three methods, about 2.6 model calls and 12.4k tokens per question, so the gain of ECR does not come from a larger budget.

(2) What this shows about the specific benefit of ECR. The gain depends on how the extra computation is used. Searching further and answering again without distinction does not improve the initial answers in this comparison. Placing explicit evidence constraints on the revision of an existing answer keeps a better balance between fixes and breaks.

Revision: Section 4.4 summarizes this comparison, and Appendix I.3 gives the full results with the added cost (Table 24).

W3: Validation with stronger backbones

Thank you for this comment on backbone coverage. The GPT-5.5 experiment in Appendix E.3 was a small check on 48 questions. To test ECR on stronger models more thoroughly, we extended it to 209 questions and then added 100 more. On the 309 questions ECR improves GPT-5.5 by 3.88 points with p = 0.0227 (point (1)), and we also replaced the arbitration model with GPT-5.6 Luna (point (2)).

(1) GPT-5.5 as the backbone on 309 questions. On 309 questions, ECR raises the accuracy of GPT-5.5 from 81.23% to 85.11% (+3.88 points), with 18 fixes and 6 breaks (exact McNemar p = 0.0227). On the first 209 questions the gain is +3.83 points with 12 fixes and 4 breaks (p = 0.0768). To strengthen the test, we added 100 questions that do not overlap these 209 and kept the GPT-5.5 backbone, the ECR rules and the evaluation protocol unchanged. The added questions give +4.00 points with 6 fixes and 2 breaks (p = 0.289), in line with the first 209. Because they were added after the first 209 had been evaluated, the table reports the two parts next to the pooled result.

GPT-5.5First 209 questionsAdded 100 questionsAll 309 questions
Base correct17081251
ECR correct17885263
Base accuracy81.34%81.00%81.23%
ECR accuracy85.17%85.00%85.11%
Gain+3.83 points+4.00 points+3.88 points
Fixes12618
Breaks426
Net correction+8+4+12
Exact McNemar p0.07680.2890.0227

On the 309 questions, the correction precision is 75.0% and 97.61% of the initially correct answers are kept. Fixes outnumber breaks three to one in both parts, so the balance of fixes and breaks remains favorable on a stronger backbone.

(2) A different model in the arbitration stage. On 151 questions, replacing Qwen3-VL-Plus with GPT-5.6 Luna as the blind verifier raises accuracy from 61.59% to 64.24%. The two results support the applicability of ECR across models, at the level of the backbone and at the level of the arbitration model.

Revision: Appendix E.3 now reports the extension to 309 questions in Table 20, Appendix I.6 reports the verifier replacement, and Section 4.5 refers to both.

Your comments led us to compare ECR with extra search and generic reflection on the same initial answers, and to extend the test on a stronger backbone. The questions you referred to are answered in our replies to W1 to W3. We hope these results clarify what ECR contributes, and we would be glad to discuss any point further.

Thank you for your response, I have updated the rating.

Thank you very much for reading our response and for updating your rating. Your support is very encouraging to us. Your review also helped us improve the paper. The comparisons with extra search and generic reflection, and the larger GPT-5.5 experiment, were both added in response to your comments, and they give clearer evidence of what ECR itself contributes. We will keep these results in the final version.

Thank you again for your time and your careful review!

Reviewer_X3RCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Evidence-Certified Revision (ECR), which formalizes the decision of whether to revise an already submitted answer in long-form video QA as an asymmetric revision decision: the current answer A serves as an anchor, and a challenger P must provide sufficient evidence before A can be replaced. The method has three components: evidence qualification (S_P, R_A, D_P), which assesses challenger support, anchor refutation, and evidence admissibility; structured revision certificates (z, ρ), which record VALID/REFUTED/UNRESOLVED status and reasons; and a fixed reason-conditioned policy that maps certificates to RETAIN/SWITCH/ESCALATE actions, invoking anonymous blind pairwise arbitration only for residual conflicts. Experiments cover Video-MME Long Full900, 655 exact-replay questions, the ECR-SCOPE-256 ablation set, 255 matched controls, an additional benchmark, and five base agents. The paper emphasizes that ECR improves accuracy while greatly reducing harmful flips, and it reports that several key comparisons are not statistically significant.

Critical strengths
  1. The problem reframing is intuitive and is carried through to a concrete pipeline. Recasting post-answer revision from "which answer is better" to "whether new evidence is sufficient to overturn an already submitted answer," with challenger support separated from anchor refutation, is the paper's core contribution. The principle explains directly why a revision that only supports the challenger, without refuting the anchor, should be rejected, and it makes the large drop in harmful flips in exact replay mechanistically interpretable. This is the most valuable part of the paper.

  2. The exact-replay control is careful, the fail-safe design is sensible, and the reporting is honest. Fixing base execution and proposal outputs makes guarded and unguarded revision comparable at the item level; falling back to RETAIN on parsing failures and restricting blind arbitration to residual conflicts limits systemic risk. The paper also states plainly that the matched controls are not significant, that transfer is not significant, and that some ablation comparisons are not significant. These points support the claim that the method is practically usable, but they do not justify acceptance on their own.

Critical weaknesses
  1. The core causal claim lacks statistically robust support, and the Full900 gain is not cleanly attributed. The decisive comparison is ECR-K2 55.29% vs. Compute-Matched Symmetric-K2+ 52.16% under matched evidence and compute, but McNemar p = 0.1849 and the 95% CI [−0.78, +7.06] includes 0; the authors themselves write that this "does not establish a causal effect." The +10.22-point gain on Full900 (52.11% → 62.33%) comes from the full pipeline, including proposal generation and evidence acquisition, while in exact replay ECR improves over unguarded adoption by only +0.31 points (65.50% vs. 65.19%). So the central claim, that a revision-specific criterion has a robust advantage over symmetric selection or simple gating, currently rests on point estimates and mechanistic reasoning, with no statistically solid causal evidence behind it. For ICLR this is a rejection-level issue.

  2. Key components lack independent quality evaluation and robustness analysis, and strong baselines are missing. The paper does not report consistency between the certificate's (S_P, R_A, D_P) and human annotations or gold evidence; it does not analyze the parser rejection rate or the causes of U_keep (in Table 14, U_keep accounts for 90/268 ≈ 33.6%, which is high); it does not quantify the consistency of blind arbitration under candidate 1/2 swapping or the winner change rate between anonymous and non-anonymous arbitration; K = 2 was selected on a 64-question development set with no sensitivity analysis for K = 1, 3, 4; the experiments include no matched-budget comparison against strong self-correction/verification methods such as PAG, ReVISE, or VideoHV-Agent; and end-to-end latency and model call counts are not reported. Without these, there is no way to tell whether ECR's gains come from the asymmetric certificate design, the extra computation, the conservative fallback, or proposal quality. Together with W1, this is a substantive obstacle to acceptance.

Questions for authors

Please see weakness

3 comments

Thank you for your careful review and valuable comments! Your comments led us to add several controlled comparisons and diagnostics, which we report point by point below. The changes in the revised paper are marked in blue with a bar in the margin, and each change carries a margin tag naming the corresponding comment. The tags for your comments are orange (for example, X3RC·W1), and those for the other reviewers use other colors.

W1: Statistical support for the central claim and attribution of the Full900 gain

Thank you for this detailed comment. We added controlled comparisons that test the central claim directly, and it now has statistical support. On an extended matched control of 655 cases, ECR-K2 is ahead of Symmetric-K2+ with p = 0.0271, and at the same revision rate ECR is ahead of a simple conservative gate with p = 0.0352 (point (1)). Points (2) and (3) explain what the Full900 gain and the exact-replay result each measure, and point (4) reports a comparison with extra search for AVP.

(1) Evidence for the advantage over symmetric selection and simple gating. Both advantages are now supported by controlled comparisons, and each is statistically significant in its own paired test.

ComparisonSet and controlResult
ECR-K2 vs. Symmetric-K2+Extended matched control, 655 cases (a different cohort from the exact replay)54.81% vs. 51.91%, paired 43 vs. 24, exact McNemar p = 0.0271
ECR vs. conservative switching491 new questions, same evidence, both methods switch 96 timesECR 57 fixes / 13 breaks, conservative rule 52 / 17, paired 12 vs. 3, exact McNemar p = 0.0352

The first comparison extends the 255-case control of Section 4.4 to 655 cases, and the advantage observed on 255 cases is now statistically supported (+2.90 points, p = 0.0271). In the second, both methods change the same number of answers, so the fixes and breaks of ECR cannot be explained only by its making fewer changes, and its accuracy is 1.83 points higher (p = 0.0352). The two tests use different cohorts and are reported separately.

(2) What the +10.22-point gain on Full900 measures. The Full900 gain measures the full pipeline of challenger generation, evidence acquisition and ECR revision, in which AVP rises from 469/900 (52.11%) to 561/900 (62.33%). We do not attribute this gain to the revision rule alone, and Section 5 states so.

(3) What the 0.31-point difference from unguarded adoption in exact replay shows. The small difference in accuracy reflects how ECR controls harmful revisions: the two policies reach similar accuracy, but ECR breaks 13 correct answers against 57. With the candidates frozen, unguarded adoption makes 141 fixes and 57 breaks, and ECR makes 99 fixes and 13 breaks. ECR gives up 42 fixes and avoids 44 breaks, so final accuracy is close (65.50% against 65.19%) while far fewer correct answers are broken. What ECR improves here is the control of risk when answers are changed, which final accuracy alone does not show.

(4) Whether extra search by AVP alone gives the same gain. In our test, extra search by AVP does not give the gain of ECR. On a newly sampled set of 256 questions with the initial AVP answers fixed, AVP reaches 52.34%, AVP with extra search 51.95% and AVP with ECR 60.94%. Paired tests give p = 0.00047 for ECR against AVP and p = 0.00082 for ECR against extra search.

Revision: Section 4.4 now summarizes the rate-matched control, the extended matched control and the AVP comparison, and Appendices I.1 to I.3 give the full results. The statements on the matched control in the introduction and Section 5 are updated.

W2 (part 1): Independent evaluation and robustness of the key components

We added the four analyses you list, one per point below. In brief, the three judgments agree with human labels in 71.0% to 77.0% of cases, none of the 90 U_keep cases comes from a parse failure, blind arbitration is stable when the candidates are swapped, and accuracy changes little beyond K = 2.

(1) Consistency of SPS_P, RAR_A and DPD_P with human annotations. We drew 100 cases from the 268 live disagreements, and two annotators labelled them independently, with disagreements adjudicated. The qualifier agrees with the human labels on 76.0% of the SPS_P judgments, 77.0% of the RAR_A judgments and 71.0% of the DPD_P judgments. The two annotators agree with each other at a macro-averaged Cohen's κ of 0.69 (0.71, 0.69 and 0.67 for the three fields). For RAR_A, the annotators confirm 85.7% of the refutations reported by the qualifier (precision), and the qualifier finds 72.0% of the refutations identified by the annotators (recall). In the remaining 28.0% the qualifier outputs no (20%) or uncertain (8%). This share quantifies the risk that conservative revision keeps an answer which the evidence could overturn.

(2) Parser rejection and the causes of the 90 U_keep cases. All 90 U_keep cases are regular UNRESOLVED to RETAIN decisions, and none comes from a parser rejection. A rejected parse falls back to U_keep (Appendix A.7), so we audited the routing records of the 268 live disagreements. The audit shows that all 90 U_keep cases in Table 14 (33.58%) are such regular decisions, with no hidden fallback after a parse failure. Since a rejected output can only end in U_keep, no parser rejection affected the routes of these 268 disagreements. The share therefore reflects how often the fixed policy keeps the anchor when the evidence is insufficient. Such retention may miss some answers that could have been corrected, which is the trade-off between risk and benefit that the paper studies.

(3) Consistency of blind arbitration under candidate 1/2 swapping and anonymization. On the 73 escalated cases, the verifier selects the same actual answer in 69 cases (94.52%) after the two candidates are swapped. For anonymity, what matters is whether the anonymized input still reveals which candidate is the anchor, so we asked a model to guess the original anchor from the two anonymous candidates and their evidence. It is correct in 35 of the 73 cases (47.95%), close to chance (two-sided binomial test, p ≈ 0.815). The verifier is thus stable to candidate order, and the source of a candidate is hard to detect from the anonymized input.

(4) Sensitivity to K = 1, 3 and 4. The method is not fragile to K. On the 268 live disagreements, accuracy rises clearly from K = 1 to K = 2 (45.90% to 53.36%), and K = 3 and K = 4 bring smaller, diminishing gains (55.22% and 55.60%).

KCorrect / 268AccuracyFixesBreaksCorrection precision
112345.90%841882.35%
2 (frozen in the paper)14353.36%991388.39%
314855.22%1031289.57%
414955.60%1041289.66%

K = 2 was selected on DEV64 and frozen before evaluation as a trade-off between resources and effect, and it is not required to give the highest accuracy under every budget. We do not reselect K from these results, since that would use evaluation labels. A larger packet adds a small gain at the cost of acquiring and judging more evidence.

Revision: Appendix I.4 reports the human validation (Table 25), Appendix B.3 the U_keep audit, Appendix A.8 the arbitration diagnostics, and Appendix I.5 the sensitivity to K (Table 26). Section 5 is updated accordingly.

W2 (part 2): Strong baselines, inference cost and the source of the gains

PAG, ReVISE and VideoHV-Agent control different stages of the task from ECR, so we test the revision decision with controls on the same interface (point (1)). We now report model calls, tokens and latency (point (2)), and the added controls address the four possible sources of the gain one by one (point (3)).

(1) Matched-budget comparison with PAG, ReVISE and VideoHV-Agent. Their published results cannot be compared directly with ours, so we carry out the matched comparison with three controls on the interface of ECR. These methods do not share this interface, the revision of a frozen anchor and challenger. Self-correction frameworks such as PAG involve training or multi-round revision, and long-video verification systems such as VideoHV-Agent also include hypothesis generation and evidence search. Their published accuracies are obtained under different training conditions and evidence budgets. Placed next to ours, they would not show whether a difference comes from candidate generation or from the revision decision. The three controls isolate the revision decision: conservative switching at the same revision rate (491 questions), symmetric evidence selection with Symmetric-K2+ (655 cases), and extra search or generic reflection for AVP (256 new questions). They are reported separately from the performance of the full pipeline.

(2) End-to-end latency and model calls. Over 655 questions run through the full pipeline, a question takes on average 8.11 model calls, 45,960 tokens and 90.8 seconds from end to end. These figures cover the video agent, the proposal stage and revision, while the matched token budgets in Section 4.4 cover only the revision module. Of the 268 live disagreements in these runs, 136 are retained, 59 are switched and 73 are escalated. The 189 logical adjudications in Appendix C count adjudication workload and are a different measure from the 73 escalations. The AVP comparison on 256 questions is reported with another measure, the cost added on top of the frozen AVP run, which is about 2.6 model calls and 12.4k tokens per question for each of the three methods (see our reply to Reviewer P8LR, W2).

(3) Whether the gains come from the certificate design, extra computation, the conservative fallback or proposal quality. The results favor the certificate design and its routing over the other three explanations, which the added controls examine from complementary angles. For the conservative fallback, a conservative rule with as many switches as ECR (96) obtains 52 fixes and 17 breaks against 57 and 13 for ECR, and is 1.83 points less accurate (p = 0.0352). For extra computation, extra search and generic reflection for AVP add about the same calls and tokens per question as ECR and neither gives its gain, and Symmetric-K2+ is behind ECR-K2 on the 655 paired cases (p = 0.0271). For proposal quality, exact replay and the matched controls give all policies the same anchor and challenger pairs. For the certificate itself, the human annotation measures its three fields directly, and the U_keep audit, the order swap and the sensitivity to K describe how the implementation behaves. Together, these results support treating the revision decision as a separate, asymmetric question of whether the evidence permits a revision.

Revision: Section 2 now discusses PAG and ReVISE. Section 4.4 and Appendix I.3 report the AVP comparison, and Appendix C now reports the inference cost of the full pipeline.

Your comments led us to test the central claim on a larger matched set and at a matched revision rate, and to examine each component of ECR on its own. Our replies to W1 and W2 also cover the points referred to in your questions. We hope the revised paper now answers your concerns, and we would welcome further discussion of any point that remains open.

Reviewer_cy3KOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ECR (Evidence-Certified Revision), a training-free post-answer revision framework for long-video agents. Its core observation is that evidence supporting an alternative answer does not necessarily refute the current one — Support(P, EP) does not imply Refute(A, EP). ECR formalizes this asymmetry via three separate judgments: challenger support (SP), anchor refutation (RA), and evidence admissibility (DP). These are recorded in a structured certificate and mapped by a fixed deterministic policy to RETAIN, SWITCH, or ESCALATE (the last sending anonymized candidates to a blind pairwise verifier). On Video-MME Long (900 questions), the full pipeline raises a frozen AVP agent from 52.11% to 62.33% while preserving 94.9% of initially correct answers. On a 655-question exact-replay cohort with frozen inputs, ECR achieves 65.50% with 99 fixes and 13 breaks, versus 65.19%, 141 fixes, and 57 breaks under unguarded proposal adoption. The same frozen rules yield positive point estimates across five heterogeneous base agents.

Critical strengths

S1. A well-motivated and novel problem decomposition

The "support ≠ refutation" asymmetry is sharp and, to my knowledge, has not been formalized in the long-video or self-correction literature. Existing self-correction work (Self-Refine, Reflexion, CRITIC, Reflect-R1, VideoSeal) treats revision as either iterative refinement or symmetric comparison. ECR isolates a narrower decision — whether new evidence warrants overturning an already-committed answer — and argues convincingly that this requires separate evaluation of challenger support and anchor counterevidence. The running example in Figure 1 makes the distinction intuitive: seeing a woman walk with a dog does not prove the man didn't take it home. Formalizing this into three orthogonal judgments kept separate through a structured certificate is a clean contribution that stands independently of the specific experimental results.

S2. Unusually disciplined experimental methodology

The paper earns trust through its control structure. The exact-replay cohort freezes base executions and proposal outputs so all revision policies see identical inputs. The matched-compute control (Symmetric-K2+ at ~1.97M tokens vs. ECR-K2 at 1.96M) isolates the revision-specific criterion from raw inference budget. The component ablation adds pieces one at a time on a fixed cohort. The 64-question dev set is excluded from all evaluation cohorts. This level of experimental rigor is well above average for agent papers and makes the results more trustworthy despite the modest sample sizes.

S3. The break-reduction result is the paper's strongest empirical claim

The drop from 57 breaks (unguarded) to 13 breaks (ECR) on the exact-replay cohort, while maintaining comparable accuracy (65.19% vs. 65.50%), directly demonstrates the value of the revision criterion. Correction precision rises from .712 to .884 and harmful-flip rate drops from 8.70% to 1.98%. This is not a side effect — it is exactly what the framework is designed to do, and it does it convincingly. For deployment scenarios where breaking a correct answer is costlier than missing a potential fix, this operating point is clearly preferable.

Critical weaknesses

W1. The qualification judgments (SP, RA, DP) are LLM calls with no per-judgment validation (Major)

The entire framework rests on the quality of three LLM judgments per live disagreement. If the qualifier consistently misjudges support, refutation, or admissibility, the certificate is wrong and the downstream policy acts on faulty input. The paper explicitly acknowledges this ("model-based qualification, whose individual judgments are not separately validated against human annotations") but never measures it.

How often does the qualifier say "no anchor refutation" when refutation is actually warranted? Without per-judgment accuracy on even a small annotated subset, we cannot distinguish two explanations for the empirical success: (a) the SP/RA/DP decomposition is the right abstraction and the qualifier executes it well, or (b) Qwen3-VL-Plus happens to be conservative enough that defaulting to RETAIN usually works regardless of the specific decomposition. These have very different implications for generality. Even 50–100 human-annotated cases with per-field agreement rates would substantially strengthen the paper's central claim.

W2. The matched-compute comparison does not reach statistical significance (Major)

The most important causal claim — that revision-specific qualification outperforms a generic symmetric criterion when evidence access and compute are matched — rests on a 3.14-point difference that is statistically unresolved (p = 0.1849, CI95 [−0.78, +7.06]). The confidence interval includes zero. On 255 cases, the study is underpowered for the observed effect size.

The authors are commendably honest about this, but it leaves the door open to an alternative explanation: ECR's advantage comes from spending tokens on revision-specific prompts rather than from the conceptual decomposition itself. Running the matched comparison on the full 655-question replay cohort would likely resolve this — the effect size suggests ~400+ cases would be needed for 80% power at α = 0.05.

W3. Single backbone for all stages (Moderate)

Qwen3-VL-Plus serves as qualifier, evidence ranker, packet selector, and blind verifier (Table 7). Correlated errors across these stages are likely. A qualifier that misjudges admissibility and a verifier that shares the same biases will produce systematically aligned mistakes. Testing with a different model for at least the arbitration stage (GPT-5.5 is already available in the paper's ecosystem) would clarify model-specificity.

W4. Blind arbitration is not fully blind (Moderate)

The paper acknowledges that hiding the anchor label "does not guarantee invariance to candidate order or other indirect source cues," but this is undersold. Anchor evidence comes from a full agentic search (rich, multi-step), while complementary evidence comes from a targeted follow-up (focused, shorter). These structural differences are detectable. No order-swap consistency check is reported. A simple diagnostic — running the verifier with both orderings on the 73 escalated cases and measuring agreement — would quantify the residual position bias.

Questions for authors

Q1. Can you report per-judgment accuracy for SP, RA, and DP on even a small annotated subset? Specifically: how often does the qualifier say RA = "no refutation" when a human would disagree? This is the most important missing validation.

Q2. On the 73 escalated cases, what is the verifier's order-swap agreement rate? If a model is shown both evidence packets and asked to guess which is the anchor, how often does it succeed?

Q3. What prevents running the matched-compute comparison on the full 655-question replay cohort? The current 255-case sample is underpowered.

Q4. In Table 6, ECR reduces accuracy relative to unguarded adoption for TimeChat (48.44% vs. 50.78%) and VideoMind (53.91% vs. 56.25%). Does the framework's value diminish or reverse when proposals are reliably strong? Is there a principled way to adjust the conservatism of the policy based on estimated proposal quality?

Q5. What happens with a different model for arbitration (e.g., GPT-5.5 as verifier while keeping Qwen3-VL-Plus as qualifier)? Even a small-scale test would speak to model-specificity of the results.

3 comments

Thank you for your detailed review and concrete suggestions! We have run the checks you proposed, and we respond to each point below. The changes in the revised paper are marked in blue with a bar in the margin, and each change carries a margin tag naming the corresponding comment. The tags for your comments are green (for example, cy3K·W1), and those for the other reviewers use other colors.

W1 & Q1: Per-judgment validation of SPS_P, RAR_A and DPD_P against human annotations

We appreciate this careful comment on the reliability of the three judgments. The certificate and the policy both act on these judgments, so their quality matters for the whole framework. To measure it directly, we drew 100 cases from the 268 live disagreements. Two annotators labelled SPS_P, RAR_A and DPD_P independently, their disagreements were adjudicated, and we compared the qualifier with the adjudicated labels field by field. Point (1) reports the agreement for each field, point (2) the rate at which the qualifier misses a refutation, and point (3) addresses explanations (a) and (b).

(1) Per-judgment accuracy of SPS_P, RAR_A and DPD_P on a human-annotated subset. On the 100 annotated cases, the qualifier agrees with the human labels on 76.0% of the SPS_P judgments, 77.0% of the RAR_A judgments and 71.0% of the DPD_P judgments. The two annotators agree with each other at Cohen's κ = 0.71 for SPS_P, 0.69 for RAR_A and 0.67 for DPD_P, with a macro-average of 0.69.

MeasureResult
Annotated cases (two annotators)100
Agreement with human labels, SPS_P76/100 (76.0%)
Agreement with human labels, RAR_A77/100 (77.0%)
Agreement with human labels, DPD_P71/100 (71.0%)
Precision of RAR_A = yes85.7%
Recall of RAR_A = yes72.0%
False negative rate of RAR_A28.0%
Cohen's κ between the two annotators, SPS_P / RAR_A / DPD_P0.71 / 0.69 / 0.67
Cohen's κ, macro-average0.69

(2) How often the qualifier outputs no refutation when a human finds one. When the annotators find a valid refutation of the anchor, the qualifier outputs RAR_A = no in 20% of the cases. The table compares the two sets of RAR_A labels on the 100 cases, with the adjudicated human label in the rows and the qualifier output in the columns.

Human \ QualifierYesNoUncertainTotal
Yes185225
No240850
Uncertain151925
Total215029100

The annotators find a valid refutation in 25 of the 100 cases. The qualifier also reports it in 18 of them (recall 72.0%), outputs no in 5 (20%) and outputs uncertain in 2 (8%). The false negative rate of 28.0% counts both kinds of miss, and an explicit no accounts for 20%. In the other direction, the qualifier reports a refutation in 21 cases and the annotators confirm 18 of them (precision 85.7%). The missed refutations are the main opportunity cost of conservative revision, because some answers with real counterevidence are then not switched.

(3) Whether the results come from the decomposition (a) or from a conservative qualifier (b). Conservatism alone (b) does not account for the results, and they are consistent with the decomposition (a). If a conservative qualifier were enough, a rule that switches as rarely as ECR but uses no certificate labels should do as well. In the rate-matched control (see our reply to Reviewer 8spP, W1 & Q1), such a rule makes 96 switches on 491 questions, as many as ECR, yet it obtains 52 fixes and 17 breaks against 57 and 13 for ECR, and the paired difference in accuracy is significant (p = 0.0352). The three fields also do different jobs. SPS_P judges whether the evidence supports the challenger, RAR_A whether it refutes the anchor, and DPD_P whether its temporal and logical coverage is sufficient for the judgment. A single score of which answer is more likely to be correct cannot take over these roles.

Revision: Appendix I.4 reports the human validation with the sample, the per-field κ and the confusion matrix of RAR_A (Table 25), and the limitation in Section 5 is revised accordingly.

W2 & Q3: Matched comparison on a larger set

Thank you for the power estimate. With 255 cases the comparison has limited power for the observed effect, and your estimate indicates that about 400 or more cases are needed. Following it, we added 400 cases to the original 255, and on the resulting 655 cases the advantage of ECR-K2 over Symmetric-K2+ is statistically supported (exact McNemar p = 0.0271). These 655 cases are the extended matched-control set, which is a different cohort from the 655-question exact replay. The replay cohort itself contains 268 live disagreements, the cases on which a matched comparison can be run, which is below the 400 or more cases of your estimate.

(1) Result of the matched comparison on 655 cases. ECR-K2 is correct on 359 cases and Symmetric-K2+ on 340, which is 54.81% against 51.91%. ECR-K2 also has more fixes and fewer breaks.

655 matched casesSymmetric-K2+ECR-K2
Correct340359
Accuracy51.91%54.81%
Fixes6882
Breaks2419
Correction precision73.91%81.19%

In the paired comparison, ECR-K2 alone is correct on 43 cases and Symmetric-K2+ alone on 24, with both correct on 316 and both wrong on 272. The net difference is 19 of the 655 cases (+2.90 points), and the exact two-sided McNemar test gives p = 0.0271.

(2) Whether a larger sample resolves the lack of statistical support. On 655 cases, the difference is statistically significant (p = 0.0271). The observed advantage is close to that on the original 255 cases (2.90 against 3.14 points), and p falls from 0.1849 to 0.0271. Symmetric-K2+ receives extra generic review until its token use is within 0.49% of ECR-K2 (Section 4.4), so the difference does not come from the number of tokens spent. This strengthens the controlled evidence for ECR-K2 over a generic symmetric criterion.

Revision: Section 4.4 now reports the comparison on 655 cases, Appendix I.2 gives the full results (Table 23), and the statements on this comparison in the introduction and Section 5 are updated.

W3 & Q5: A different model for the arbitration stage

In the primary implementation, Qwen3-VL-Plus serves every ECR stage (Table 7). As you point out, errors in qualification and arbitration could then be correlated, and replacing the model in one stage shows whether the results depend on this. To test it, we replaced the blind verifier with GPT-5.6 Luna and kept the other ECR modules unchanged (point (1)). The revision effect of ECR holds with the new verifier, so the results do not require every stage to use Qwen3-VL-Plus (point (2)).

(1) Result with a different model for arbitration. On 151 questions, replacing Qwen3-VL-Plus with GPT-5.6 Luna as the verifier raises accuracy from 61.59% to 64.24% (+2.65 points). Fixes increase from 23 to 26, and breaks fall from 13 to 12.

151 questions, blind verifierQwen3-VL-PlusGPT-5.6 Luna
Correct93/15197/151
Accuracy61.59%64.24%
Fixes2326
Breaks1312
Correction precision63.89%68.42%
Harmful-flip rate8.61%7.95%

(2) Whether the results are specific to using Qwen3-VL-Plus in all stages. The results do not depend on using Qwen3-VL-Plus in every stage. With GPT-5.6 Luna as the arbitration model, ECR keeps its revision effect and reaches higher accuracy.

Revision: Appendix I.6 reports the verifier replacement (Table 27), and Section 4.5 refers to it.

W4 & Q2: Order-swap agreement and source cues in blind arbitration

Blind arbitration hides the anchor label so that the verifier does not know which candidate is the original answer, and Section 3.5 notes that this alone does not guarantee invariance to candidate order or to indirect source cues. The two diagnostics you propose measure how large these remaining effects are, so we ran both on the 73 escalated cases. The verifier keeps its choice in 94.52% of the cases when the candidate order is swapped (point (1)), and a model asked to identify the anchor from the two evidence packets is at chance level (point (2)).

(1) Order-swap agreement rate of the verifier on the 73 escalated cases. The order-swap agreement rate of the verifier is 94.52%. After Candidate 1 and Candidate 2 are swapped, the verifier selects the same actual answer in 69 of the 73 cases, and its choice changes in 4 cases. In this test the verifier is therefore largely stable to candidate order.

(2) How often a model shown both evidence packets can guess which one is the anchor. The model guesses correctly in 35 of the 73 cases (47.95%), close to chance. The model sees the two anonymous candidates with their evidence packets and no explicit anchor label, and a two-sided binomial test against 50% gives p ≈ 0.815. For the anonymized inputs we tested, the differences between the two evidence packets therefore give no clear cue to which candidate is the anchor.

Revision: Section 3.5 now reports these diagnostics, and Appendix A.8 gives the details.

Q4: Strong proposals and adjusting the conservatism of the policy

Thanks for the question, and for pointing to the two cases in Table 6. When proposals are strong, conservative revision gives up some fixes, which is why ECR is below unguarded adoption for TimeChat and VideoMind (point (1)). The conservatism of the policy can be calibrated to the reliability of the proposals (point (2)).

(1) Whether the value of the framework diminishes or reverses when proposals are reliably strong. Relative to unguarded adoption, the accuracy advantage diminishes and can reverse, as for TimeChat (48.44% against 50.78%) and VideoMind (53.91% against 56.25%) in Table 6. Relative to the base agents, ECR still gains 25.00 and 13.28 points on these two agents. ECR aims to make each answer change rest on sufficient evidence and does not aim to adopt as many challengers as possible, so it gives up some fixes when challengers have strong correction potential. When the quality of the challengers is unstable or below that of the anchors, the same retention mechanism removes breaks effectively. Both effects follow from the trade-off between keeping correct answers and taking every chance to correct a wrong one, which the asymmetric formulation makes explicit.

(2) A principled way to adjust the conservatism to estimated proposal quality. The conservatism can be adjusted in a principled way. From a decision-theoretic view, it can be calibrated from the empirical reliability of the challengers and the costs of fixes and breaks. For example, the revision gain under each evidence state can be estimated on separate calibration data, and the evidence strength required for a switch is then chosen for the application. The paper keeps one frozen policy so that comparisons across agents and datasets are not affected by tuning on each test set.

Revision: Appendix E.2 now discusses how the conservatism of the policy can be calibrated, and Section 4.5 refers to it.

We appreciate your power estimate and your suggestions on human validation, the arbitration model and the diagnostics for blind arbitration. All of these checks are now in the paper. If anything is still unclear, we would be glad to explain it further.

Reviewer_8spPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Long video question answering agents search for relevant evidence, but new evidence can lead them to change an initially correct answer. The proposed method ECR address this exact step and decides when such a revision is justified. It separates evidence supporting a new answer from evidence against the original one. A model-generated certificate and fixed policy decide whether to keep the answer, replace it, or send both candidates to a blind verifier. ECR improves the accuracy by 10.22.

Critical strengths
  • The replay result shows a useful preservation benefit. With frozen anchors and proposals, ECR reduces breaks while keeping accuracy almost unchanged compared with unconditional proposal adoption. That matters when harmful revisions are costly.

  • As a reader, I appreciate that the evaluation goes beyond final accuracy, reporting fixes, breaks & preservation metrics. Providing comparison of ECR with a baseline using the same evidence access and nearly the same compute budget, making the comparison more informative.

Critical weaknesses
  • It is unclear whether ECR avoids harmful revisions because its certificates identify which changes are justified, or mainly because it changes fewer answers. In replay, ECR produces 44 fewer breaks but also 42 fewer fixes than unconditional proposal adoption. A useful comparison would be a rule that switches only when a verifier strongly prefers the challenger, without using ECR's certificate labels. Its threshold should be chosen on held-out data to give a similar revision rate. With the same evidence access, does ECR then fix more wrong answers or break fewer correct ones? The symmetric comparison in section 4.4 matches evidence access and nearly matches compute, but does not answer this question about comparable revision rates.
Questions for authors
  • Could you address the comparison with a simpler conservative switching rule described in Weakness 1?

Minor clarification:

  • L259-262 describe an "Unresolved" reason that permits a direct switch. Could you give the exact rule or prompt instruction used to assign this reason, along with one worked example showing the anchor, challenger, evidence, and support/refutation/admissibility judgments? Please explain what makes that example eligible for a direct switch rather than retention or arbitration.
3 comments

Thank you for your careful review and positive assessment! Your questions have been very helpful to us. We respond to each point below. The changes in the revised paper are marked in blue with a bar in the margin, and each change carries a margin tag naming the corresponding comment. The tags for your comments are red (for example, 8spP·W1), and those for the other reviewers use other colors.

W1 & Q1: Comparison with a conservative switching rule at a similar revision rate

Thank you for this comment on what drives the reduction in breaks. A method that changes fewer answers will also break fewer, so the question is whether ECR still does better when the number of changes is the same. We ran the comparison you describe to test this. At the same revision rate, ECR fixes more wrong answers and breaks fewer correct ones than the conservative rule, and the paired difference is significant (p = 0.0352, point (1)). Changing fewer answers therefore does not explain its fewer breaks, and the advantage is accounted for by the certificate constraints and the routing built on them (point (2)).

(1) Whether ECR fixes more wrong answers or breaks fewer correct ones at a similar revision rate. At the same revision rate, ECR does both: it fixes more wrong answers and also breaks fewer correct ones. The conservative rule follows your description: it uses neither the certificate labels nor the judgments SPS_P, RAR_A and DPD_P, and it reads the same candidates and evidence as ECR. For each question, the verifier gives token log-probabilities lPl_P for the challenger and lAl_A for the anchor, and the preference score is s = exp(lPl_P) / (exp(lAl_A) + exp(lPl_P)). The rule switches when s > τ, and τ is its only parameter. As you suggested, we set τ on held-out data, namely DEV64, the 64-question development set, which shares no question with the 491 test questions. ECR switches on 13 of these 64 questions (20.31%). Thresholds of 0.68, 0.72 and 0.76 give 15, 13 and 11 switches, so we took τ = 0.72, at which the rule revises exactly as often as ECR. Only the revision rate on DEV64 entered this choice. τ was then frozen, and no accuracy, fix or break count from the 491 questions was used to adjust it. With τ fixed, the rule switches on 96 of the 491 test questions, the same number as ECR (19.55%). The table shows what these 96 changes do.

491 questionsECRConservative switching
Switches (revision rate)96 (19.55%)96 (19.55%)
Fixes (wrong → correct)5752
Breaks (correct → wrong)1317
Wrong → wrong2627
Correction precision81.4%75.4%
Fixes − breaks+44+35

Of its 96 changes, ECR turns 57 wrong answers into correct ones and 13 correct answers into wrong ones, against 52 and 17 for the conservative rule. ECR thus makes 5 more fixes and 4 fewer breaks. Its net correction, fixes minus breaks, is 9 questions higher (1.83 points on the 491 questions), and its correction precision is 81.4% against 75.4%. Compared question by question, ECR is correct and the conservative rule wrong on 12 questions, the reverse holds on 3, and the two agree in correctness on the other 476. The exact two-sided McNemar test gives p = 0.0352.

(2) Whether ECR avoids harmful revisions through its certificates or mainly by changing fewer answers. Changing fewer answers does not explain the result, and the advantage we observe is accounted for by the certificate constraints and the routing built on them. If the fewer breaks came only from changing fewer answers, the difference would disappear once the number of changes is matched. It does not: with 96 changes each, ECR still has 4 fewer breaks and 5 more fixes. The difference therefore lies in which answers are changed. ECR decides from what the evidence establishes about the anchor, whereas the conservative rule decides from how strongly the verifier prefers the challenger, and the evidence constraints of ECR select the answers worth revising more precisely. This comparison holds both conditions you name, the same evidence access and the same revision rate, and under them ECR is 1.83 points more accurate than the simple conservative rule we tested (p = 0.0352). A separate control in Section 4.4 compares ECR with a symmetric selector under matched evidence and nearly matched compute, and on 655 cases ECR-K2 is ahead with p = 0.0271 (see our reply to Reviewer cy3K, W2 & Q3). The two controls use different cohorts and are reported separately, and they point the same way.

Revision: Section 4.4 now summarizes this comparison. Appendix I.1 gives the preference score, the calibration of the threshold on DEV64 and the full results (Table 22). Appendix H records this use of DEV64, and Table 21 there now lists the cohorts added in the revision.

Minor clarification: The Unresolved reason that permits a direct switch (L259 to 262)

In ECR the certificate state and the final action are separate. A certificate consists of a state and a reason. UNRESOLVED means that the evidence relation falls into neither the direct VALID case nor the direct REFUTED case, and it does not by itself require keeping the anchor. The action is decided by the reason. Point (1) gives the rule, point (2) a worked example, and point (3) explains why that example is switched.

(1) The rule that assigns this reason and permits a direct switch. The qualifier assigns this reason, and the fixed map in Table 13 (Appendix B) turns it into a switch. Under the instruction in Appendix A.5, the qualifier returns the state and the reason together with SPS_P, RAR_A and DPD_P in one structured output, and the parser accepts only the state and reason pairs listed in Table 9. The map then sends UNRESOLVED with U_reason_switch to SWITCH. This route differs from the usual switch, which requires admissible evidence that refutes a legal anchor. U_reason_switch is a normalized label that groups three raw reasons. The table gives the condition and the outcomes of each, which add up to the 26 cases, 17 fixes, 2 breaks and 7 wrong-to-wrong changes reported in Table 14.

Raw reasonConditionCasesFixes / breaks / wrong → wrong
anchor_is_not_a_legal_optionThe anchor is not a legal option of the question, and the challenger is.87 / 0 / 1
question_constraint_conflictThe anchor is a legal option, but under the evidence it violates an explicit constraint stated in the question, while the challenger satisfies that constraint.97 / 0 / 2
temporal_consistency_overrideThe question asks for a specific temporal relation, admissible evidence directly establishes the relation asserted by the challenger, and the usual RAR_A judgment remains non-decisive.93 / 2 / 4

The three conditions are judged from the question, the candidates and the evidence, without reference to the answer label. In the first two, the anchor is either not an option of the question or in conflict with a constraint that the question states. The third is a narrowly scoped exception for questions that ask for a specific temporal relation. It permits a switch when admissible evidence directly establishes the challenger's temporal claim while the standard anchor-refutation judgment remains non-decisive. This is a permission that the frozen policy grants for this case only, and it is not a claim that support for the challenger amounts to a refutation of the anchor. Our distinction between the two is unchanged: generic support alone does not authorize a revision. The branch is not error-free. Its nine cases give three fixes, two breaks and four wrong-to-wrong changes, and both breaks of the category fall here. The map was frozen before evaluation, and we report the branch as it is.

(2) A worked example with the anchor, challenger, evidence and the three judgments (Video-MME 644-3). We take an example from the legality-fallback branch. The question asks within how long a white-haired speaker expects extraterrestrial intelligence to be discovered. The legal options are A (within 10 years), B (within 20 years), C (within 30 years) and D (within 40 years), and the gold answer is B.

ItemVideo-MME 644-3
AnchorE (not stated). E is not a legal option, so the parsed anchor is None.
ChallengerB (within 20 years)
EvidenceItem T003 (2600 to 2670 seconds) records the speaker's explicit prediction that extraterrestrial intelligence will be discovered within the next 20 years.
Challenger support (SPS_P)The evidence supports B.
Anchor refutation (RAR_A)Anchor E is not a legal option, so this is not a refutation of a valid anchor proposition in the usual sense. The route uses a separate condition on answer legality.
Evidence admissibility (DPD_P)The timestamped evidence directly covers the time frame asked about, with matching temporal and semantic scope.
Certificate and actionUNRESOLVED with raw reason anchor_is_not_a_legal_option (normalized as U_reason_switch). Action SWITCH, final answer B, a fix.

Minor clarification: The Unresolved reason that permits a direct switch (L259 to 262) (continued)

(3) Why this example is switched directly instead of being retained or sent to arbitration. This example is switched directly because only the challenger is a legal option. RETAIN would keep an answer that the question does not allow, and ESCALATE handles an unresolved evidence conflict between two legal candidates, so neither applies here. The frozen legality fallback therefore already determines the action. Section 3.1 now states this case explicitly: the raw response E lies outside the legal answer set, and the parsed anchor takes the invalid value ⊥ (None in the implementation). The fallback thus acts on an invalid anchor and does not attempt to refute a valid answer. The example also shows why the state and the routing are kept apart: the state describes the evidence relation, and the reason-conditioned policy decides which revision is permitted.

Revision: Appendix B.3 now explains how this reason is assigned, names the three raw reasons with their conditions and outcomes, and gives this worked example. Appendix A.6 and the caption of Table 14 state that U_reason_switch is a normalized category, and Section 3.1 now allows the parsed anchor to take an invalid value.

The comparison you proposed at a matched revision rate, together with your question on the Unresolved route, helped us show more clearly where the reduction in breaks comes from and how the policy acts on each certificate. We are glad to discuss any of these points further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.