acceptodds
Under review as a conference paper at ICLR 2027

Video Evidence Indexing: Learning Where to Look from Video Previews for Token-Budgeted Long-Video Question Answering

13 views

Abstract

Long-video question answering is limited by the high cost of visual tokens and by the fixed context width of current VLMs. A long-video question may require broad temporal coverage, but the answer is often supported by only a compact set of moments. To locate these moments efficiently, we propose token-budgeted Video Evidence Indexing (VEI): given a dense low-resolution Video Preview, the model constructs a compact high-resolution Evidence Set for final reasoning. We treat VEI as a policy that must jointly solve evidence localization, which finds question-relevant moments, and budget planning, which decides where to spend the limited high-resolution frame budget. We implement this idea with an inference pipeline: the Video Preview provides cheap global coverage, Video Evidence Indexing constructs the Evidence Set, and Answer Generation combines both inputs for final VQA. To address missing frame-level supervision, we adopt privileged self-distillation, where an answer-aware teacher guides the normal test-time policy on student-generated indexing traces. We explore previews at 1, 6, 12, and 24 visual tokens per frame, training a single policy that supports all four resolutions. Experiments show that Video Evidence Indexing improves accuracy under limited visual budgets, and self-distillation further improves both QA accuracy and temporal evidence localization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
VLM, VideoQA, Long Video, Knowledge Distillation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_xot9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VEI, which first goes through a spare preview, lists the evidence needed for the question, and selects timestamps for each target. The lists are merged and filled or reduced to K frames, which are reloaded at higher resolution. During training, a frozen teacher model sees the correct answer and a 24-token-per-frame preview. Reverse KL trains the student's indexing trace. The trained model is used for both selection and answering. Experiments on VideoMME, LVBench, and MLVU demonstrate the effectiveness.

Critical strengths
  1. The method improves on uniform sampling with the same visual input limit. At 24 tokens per preview frame, VEI achieves 66.9 average accuracy, versus 59.4 for uniform sampling. This supports using a cheap preview to decide which frames need a closer look.

  2. It addresses missing frame-level labels by deriving supervision for VEI decisions from an answer-aware teacher. Privileged self-distillation transfers this guidance across preview resolutions.

Critical weaknesses
  1. Some questions might have small text or a small object change, which can disappear in resizing. If the preview gives no clue about that moment, the model may never select it. The paper acknowledges this limit, but does not show how much of this gap comes from missing preview information, wrong selection, or wrong answering.
  2. The proposed method is quite similar to zoom-in methods which first have a global view of a video and the predict a temporal span. Recent recent adaptive selectors also have similar idea. The novelty is quite incremental compared with those methods.
  3. There is no comparison with zoom-in methods (VideoZoomer, VideoChat-R1.5) about whether selecting segment or select frames is better.
  4. Missing comparisons with recent adaptive selectors. FOCUS uses uncertainty to guide sampling under a frame budget, which is cited but has no quantitative comparison. FFS also learns variable frame counts without frame annotations, though its model and training setup differ.
  5. The experiments only show the final accuracy, but there is no quantitative analysis on the frame selection correctness.
Questions for authors
  1. For the one-token VideoMME Long result, what causes the drop below uniform sampling? Is the needed information missing from the preview, visible but indexed incorrectly, or selected correctly but answered incorrectly?
  2. Can you provide results from other frame selections methods or zoom-in methods [1-5].
  3. After merging and subsampling, how often does a target lose all its requested frames? Can you report target coverage on the final evidence set?
  4. CG-Bench is a challenging long video understanding benchmark which contains the temporal spans for the question. The paper should also check the selection result on this benchmark instead of just reporting accuracy.

[1] Thinking With Videos: Multimodal Tool-Augmented Reinforcement Learning for Long Video Reasoning, CVPR 2026 [2] Flexible Frame Selection for Efficient Video Reasoning, CVPR 2025 [3] FOCUS: Efficient Keyframe Selection for Long Video Understanding, ICLR 2026 [4] VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning, ICLR 2026 [5] VideoChat-R1.5: Visual Test-Time Scaling to Reinforce Multimodal Reasoning by Iterative Perception, NeurIPS 2025

5 comments
AuthorsOpenReview ↗

Response to Reviewer xot9 Part I

Thank you for recognizing the matched-budget gains and privileged guidance for learning evidence acquisition. We clarify the existing analyses and provide additional comparisons, localization results, and target-retention measurements below.


W1 / Q1. Resizing may remove small text or object changes before indexing. For the one-token VideoMME Long regression against uniform sampling, is the evidence missing from the preview, selected incorrectly, or selected but answered incorrectly?

Our supplementary one-token LVBench analysis points to evidence acquisition as a bottleneck: Training-Free VEI hits fewer annotated windows but answers comparably well on hit cases. With very limited visual detail, it may formulate an incorrect target and repeatedly sample around it, yielding less useful evidence coverage than uniform sampling. In one inspected trace, the question asks about a phone's previous generation, but the targets focus on the phone itself. This illustrates how an early interpretation error can misdirect subsequent selection.

Because VideoMME does not provide ground-truth evidence windows, we use LVBench's temporal annotations to support this analysis, examining the two one-token settings already reported in Table 1:

MethodAnswer accuracy (%)Window hitEvidence recallEvidence precisionCorrect given hit (%)Correct given miss (%)
Training-Free VEI36.726.914.214.457.829.0
Self-Distilled VEI44.939.219.520.955.638.2

Training-Free VEI answers comparably well when an evidence window is hit (57.8% vs. 55.6%), but reaches these windows much less often. Its lower overall accuracy is therefore associated with fewer evidence hits and poorer performance on non-hit cases (29.0% vs. 38.2%), both of which improve after distillation.


W2. Existing zoom-in methods already use a global video view to locate relevant spans, and adaptive selectors share this idea. What is substantively new in VEI beyond this common framework?

Our contribution is the VEI framework, which learns to turn a dense, compressed Video Preview into explicit observation targets and target-wise evidence acquisition through privileged self-distillation. Coarse global preview cues guide where to inspect, while evidence localization and budget planning jointly determine how to allocate limited high-resolution observations across complementary evidence needs. An answer-aware teacher supplies guidance for these intermediate decisions without frame-level annotations, teaching the policy what evidence to seek and how to distribute its observations.

FOCUS uses uncertainty-guided search, while FFS learns flexible frame selection. VEI makes the evidence needs explicit and adopts privileged self-distillation to train target formulation and timestamp allocation without frame-level labels. Section 4.3 tests this coupling: target-wise budget planning adds 2.4/2.6 accuracy points for Training-Free/Self-Distilled VEI beyond target formulation with one global selection list. One policy trained on MLVU's training split supports all four tested preview resolutions.


W3 / W4 / Q2. Comparisons with recent zoom-in methods and adaptive selectors are missing, leaving frame selection versus segment acquisition unclear. Please compare with FOCUS, FFS, VITAL, VideoZoomer, and VideoChat-R1.5.

We would like to clarify that Figure 2 and Table 2(c) already compare representative evidence-acquisition approaches: predictive compression (AdaCodec), multi-agent aggregation (MACF), encoder-based selection (AKS and AdaQ), and RL-trained video-tool acquisition (ParaVT).

We now include four of the specifically requested methods; Self-Distilled VEI achieves the highest accuracy on all three datasets in this expanded comparison. Additional evaluations and reported results are summarized below; VEI uses the 50K / 24-token setting.

MethodVideoMMELVBenchMLVUtest\mathrm{MLVU}_{\mathrm{test}}Average
FOCUS69.253.367.563.3
VITAL63.445.156.054.8
VideoZoomer65.241.555.854.2
VideoChat-R1.567.148.457.857.7
Training-Free VEI70.553.566.763.6
Self-Distilled VEI71.157.372.366.9

For FFS, we could not locate a publicly released implementation or checkpoint; its methodological distinction is discussed above. Further details are provided in Appendix I.2, Table 21.


AuthorsOpenReview ↗

Response to Reviewer xot9 Part II

Q3. After merging and subsampling to the final evidence budget, how often does an observation target lose all its requested frames? Please report target coverage.

Only 0.27% of analyzable traces lose all requested frames for at least one target; overall target retention exceeds 99.9% in our offline replay. We replayed the 32-frame cap once per analyzable trace across all three datasets at 50K / 24 preview tokens per frame, using seed 42:

MethodQuestions losing at least one targetTarget retention
Training-Free VEI14 / 4,712 (0.30%)99.91%
Self-Distilled VEI11 / 4,681 (0.23%)99.92%

Among the 1,343 traces requiring truncation, 25 (1.86%) lost at least one target. Losing every requested frame for a target is uncommon, although not explicitly prevented. The per-dataset breakdown is provided in Appendix I.3, Table 22.


W5. There is no quantitative analysis of frame-selection correctness. Q4. CG-Bench provides temporal evidence spans for long-video questions. Please evaluate selection correctness on this benchmark rather than reporting only answer accuracy.

We would like to kindly clarify that the submitted Section 4.4, Figure 4, and Table 5 already quantitatively evaluate frame-selection correctness using LVBench's temporal annotations. They report evidence-window hit rate, recall, precision, and the relationship between localization and answer correctness.

The added evaluation on all 3,000 CG-Bench-mini questions also shows improved evidence recall, precision, and answer accuracy after distillation. Both methods use final 32-frame Evidence Sets at 50K / 24 preview tokens per frame:

MethodAnswer accuracy (%)Window hitEvidence recallEvidence precision
Training-Free VEI50.957.134.927.4
Self-Distilled VEI53.962.641.933.2

Values are percentages. Scoring uses five-second tolerance and the union of annotated clue intervals, preserving gaps. With the same evidence-frame budget, distillation improves evidence coverage and precision alongside QA. The full protocol and conditional QA results are provided in Appendix I.4, Table 23.


Please let us know if you have any further questions.

AuthorsOpenReview ↗

Friendly reminder from authors

Dear Reviewer xot9,

As the rebuttal deadline approaches, we would like to take this opportunity to ask whether our response has adequately addressed your concerns and whether there is any additional information or clarification we could provide.

Thank you for taking the time to review our paper and for your valuable feedback.

Best regards, The Authors

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

The paper studies long-video multiple-choice QA under a visual-token budget. A low-resolution preview covers the video; the model uses it to request timestamps, reloads 32 frames at higher resolution, and answers using both views. The indexing step is trained by distillation from an answer-aware teacher. At a 50K context limit, average accuracy across VideoMME, LVBench, and MLVU is 66.9%, versus 63.6% for the training-free indexing pipeline and 59.4% for uniform frame sampling.

Critical strengths
  1. The matched-budget comparisons are useful. At 12 and 24 preview tokens per frame, training-free indexing beats uniform sampling on all three datasets (Table 1), so the selection pipeline has value before distillation. The result also appears on two other 8B backbones in Table 2.

  2. The paper measures selected timestamps against annotated LVBench evidence windows. At 24 preview tokens per frame, distillation raises window hit rate from 48.8% to 53.3% and evidence recall from 31.2% to 35.8% (Table 5). That gives direct, if modest, evidence that the trained model looks in better places.

Critical weaknesses
  1. The distilled checkpoint is used both to select frames and to produce the final answer. The main QA gain therefore mixes a better indexer with a changed answerer. The localization results partly address this, but they do not tell us how much of the 63.6% to 66.9% QA gain comes from the indexing policy itself. A selector/answerer swap is needed for that attribution.

  2. The advantage over other frame selectors is less settled. In the matched Qwen3-VL comparison, training-free VEI is only 0.8 points above CLIP retrieval on average and is slightly worse on LVBench (Table 2). Most comparisons with other adaptive methods use published results under different observation protocols. The method also depends on preview quality: at the tightest 15K budget, training-free VEI falls below uniform sampling, 47.2% versus 52.1% (Table 1).

Questions for authors
  1. Can you report all four combinations of original/distilled indexer and original/distilled answerer under the same preview and 32-frame budget?
  2. For the cross-backbone self-distilled results in Table 2(d), was a separate policy trained for each backbone? Please give the corresponding training setup.
  3. Does the Table 7 cost include both model calls, repeated preview processing, and generation of the indexing trace? End-to-end latency would make the efficiency claim easier to assess.
5 comments
AuthorsOpenReview ↗

Response to Reviewer svpX Part I

Thank you for recognizing the training-free gains and the direct evidence-localization analysis. We provide the requested indexer–answerer control, separate-backbone training settings, and end-to-end timing below.


W1 / Q1. Because the distilled checkpoint both selects frames and answers, the 63.6-to-66.9 gain mixes indexing and answering improvements. Please report all four original/distilled combinations under the same preview and 32-frame budget to separate their contributions.

Both stages contribute: at 50K, changing only the indexer raises average accuracy from 63.6 to 66.2, while changing only the answerer raises it to 64.9. Using both reaches 66.9. Both components also improve average accuracy at 15K:

Original denotes the pretrained model; Distilled denotes the student trained on indexing traces from the separate MLVU training split.

ContextIDIndexerAnswererVideoMMELVBenchMLVUtest\mathrm{MLVU}_{\mathrm{test}}Average
15KAOriginalOriginal58.636.746.247.2
15KBDistilledOriginal62.942.553.653.0
15KCOriginalDistilled59.337.954.450.5
15KDDistilledDistilled63.944.958.855.9
50KAOriginalOriginal70.553.566.763.6
50KBDistilledOriginal70.756.271.766.2
50KCOriginalDistilled70.254.669.964.9
50KDDistilledDistilled71.157.372.366.9

A/C reuse the same final 32 frames from the original indexer; B/D reuse the same final 32 frames from the distilled indexer, including supplementation or subsampling. Thus, the answerer comparison holds the visual input fixed.

The indexer improves the usefulness of the selected evidence, while the answerer makes better use of the same visual input. This suggests that learning to index from compressed previews also helps the model extract global cues from the preview during answering. The gains therefore reflect both better evidence acquisition and better use of the available visual information, with the combined system performing best at both budgets. The complete comparison is in Appendix H.4, Table 18.


W2. Training-Free VEI is only 0.8 points above CLIP on average and slightly worse on LVBench; most other comparisons use published results under different observation protocols. At the tightest 15K budget (one-token previews), it falls below uniform sampling, 47.2% versus 52.1%.

The one-token regression does not erase the training-free gains: VEI improves average accuracy at 6/12/24 tokens and exceeds matched-budget CLIP by 0.8 points. The gains over Uniform FS are 3.0, 3.4, and 4.2 points, respectively; the 12- and 24-token settings improve all three datasets.

The CLIP comparison uses the same preview and evidence budget; CLIP remains slightly stronger on LVBench. CLIP relies on an additional pretrained encoder for frame ranking; VEI uses the existing VLM with a structured indexing prompt, without an auxiliary retrieval model or parameter updates. The gain therefore demonstrates useful evidence selection from the framework itself, before distillation.

One token per frame is the most extreme compression setting tested. Limited visual cues can lead to uncertain targets and concentrated selection of unhelpful moments. Distillation improves this capability: with the original answerer fixed, the one-token distilled indexer reaches 53.0 average accuracy, compared with 47.2 for Training-Free VEI and 52.1 for Uniform FS. The controlled CLIP comparison and indexer swap directly address selection quality; the broader method comparisons provide additional performance context.


Q2. For the cross-backbone results in Table 2(d), was a separate policy trained for each backbone? Please provide the corresponding training settings.

Yes. Each backbone has its own separately trained policy and performs its own indexing and answering. All three policies are trained on MLVU's training split. Expanded Table 10 lists the settings for Qwen3-VL-8B, LLaVA-OneVision-2-8B, and Penguin-VL-8B.

We train language-model LoRA adapters with rank 64 and alpha 128, keeping the base parameters and same-backbone teacher frozen. The teacher uses 24 preview tokens/frame; the student samples from {1, 6, 12, 24}. Table 10 provides the visual processing, optimizer, learning rate, batch size, rollout settings, training steps, and hardware.


AuthorsOpenReview ↗

Response to Reviewer svpX Part II

Q3. Does the cost include both model calls, repeated preview processing, and indexing-trace generation? What is the end-to-end latency?

End-to-end latency decreases from 35.03 to 20.90 seconds per question, including preprocessing, both VEI calls, and indexing-trace generation. These measurements cover all 4,751 evaluation questions:

MethodVideo preprocessing (s/question)Total model time (s/question)End-to-end (s/question)Generated tokens/question
Full-context Qwen3-VL-8B2.6732.3635.032
Self-Distilled VEI4.9215.9920.90approximately 302

VEI's model time includes both indexing and answering. The revised illustrative token-cost estimate includes approximately 300 generated tokens per question and shared-preview cache reuse. Hardware and serving settings, the cost calculation, and timing details are provided in Appendix C.


Please let us know if you have any further questions.

AuthorsOpenReview ↗

Friendly reminder from authors

Dear Reviewer svpX,

As the rebuttal deadline approaches, we would like to take this opportunity to ask whether our response has adequately addressed your concerns and whether there is any additional information or clarification we could provide.

Thank you for taking the time to review our paper and for your valuable feedback.

Best regards, The Authors

Comments on the paper

7 comments
AuthorsOpenReview ↗

Response to Reviewer uqUP Part I

Thank you for recognizing the coupling of evidence localization and budget planning, and the value of directly evaluating evidence quality. We address the concerns below with additional training seeds, privilege controls, and evidence-acquisition comparisons.


W1 / Q1. Greedy evaluation does not remove training randomness, and external comparisons use different budgets and protocols. With only a 0.8-point margin over matched-budget CLIP, how robust are the gains? Please report 2–3 training seeds at 15K and 50K.

Training-run variability.

All three training seeds outperform Uniform FS and Training-Free VEI in average accuracy at both budgets; the standard deviation of average accuracy is 0.3/0.4 points at 15K/50K. We trained two additional policies with seeds 43 and 44, alongside the original seed 42. Each policy supports both settings; evaluation remains greedy.

ContextTraining seedVideoMMELVBenchMLVUtest\mathrm{MLVU}_{\mathrm{test}}Average
15K4263.944.958.855.9
15K4363.744.557.855.3
15K4464.044.958.655.8
15KMean ± SD63.9 ± 0.244.8 ± 0.258.4 ± 0.555.7 ± 0.3
50K4271.157.372.366.9
50K4370.956.172.166.4
50K4470.957.672.767.1
50KMean ± SD71.0 ± 0.157.0 ± 0.872.4 ± 0.366.8 ± 0.4

SD is the sample standard deviation across training seeds. For the one-token LVBench result specifically, all three runs reach 44.5–44.9%, versus Training-Free VEI's 36.7%. Individual runs and summaries are in Appendix H.1, Table 14.

Comparing VEI with CLIP retrieval.

Training-Free VEI exceeds CLIP retrieval, 63.6 versus 62.8, under the same preview and 32-frame evidence budget. CLIP uses an additional pretrained encoder to rank frames. VEI obtains its gain by prompting the existing VLM to formulate observation targets and select corresponding timestamps, without an auxiliary retrieval model or parameter updates. This supports the value of the VEI design itself before task-specific training.


W2 / Q2. The largest gain is on MLVU, which also supplies the training data, while Training-Free VEI regresses at one-token previews. How can policy learning be distinguished from in-domain adaptation? What happens without answer privilege, or when training on VideoMME and testing on MLVU?

Training-Free VEI has value before distillation.

Before distillation, VEI already improves average accuracy over uniform sampling at 6, 12, and 24 preview tokens/frame. The gains are 3.0, 3.4, and 4.2 points, respectively; the 12- and 24-token settings improve all three datasets (Table 1).

The one-token regression identifies a limitation at the most extreme compression level tested. Sparse visual detail can make target formulation uncertain, directing high-resolution observations toward the wrong moments. It does not negate the training-free gains at the other resolutions. Distillation extends this capability: at one token, the distilled indexer with the original answerer reaches 53.0 average accuracy, above Uniform FS's 52.1, and using the distilled model at both stages reaches 55.9.

Evidence localization and indexer–answerer controls.

The paper already shows improved localization outside the training dataset: LVBench evidence recall rises from 31.2% to 35.8%. The MLVU-trained policy also raises window hit rate from 48.8% to 53.3% (Table 5), evaluated against temporal annotations independently of answer correctness.

With the answerer fixed, the distilled indexer raises average accuracy from 63.6 to 66.2 at 50K. To separate the effects of selection and answering, we independently replace the original indexer, the original answerer, or both with the distilled model. The table reports average answer accuracy across the three datasets:

IDIndexerAnswerer15K50K
AOriginalOriginal47.263.6
BDistilledOriginal53.066.2
COriginalDistilled50.564.9
DDistilledDistilled55.966.9

For each indexer, both answerers receive exactly the same final 32 frames (A/C and B/D), so the answerer comparison does not change the evidence. At 50K, changing only the indexer (A → B) raises accuracy from 63.6 to 66.2; changing only the answerer (A → C) raises it to 64.9. Both changes help, and combining them performs best at both budgets. The fixed-answerer gain, together with improved localization on LVBench, shows that distillation improves evidence acquisition beyond the MLVU training domain.

AuthorsOpenReview ↗

Response to Reviewer uqUP Part II

Contribution of answer privilege.

Training with an answer-aware teacher improves student average accuracy by 1.2/1.8 points at 15K/50K. We removed only the answer from the teacher input, retaining its 24-token preview and available question type. Both conditions use MLVU's training split and seed 42; neither student receives privilege at inference.

(1) Distilled-student evaluation: students trained with answer-removed or answer-aware teachers perform both indexing and answering.

Teacher answer during distillationStudent contextVideoMMELVBenchMLVUtest\mathrm{MLVU}_{\mathrm{test}}Average
Removed15K62.843.457.854.7
Available15K63.944.958.855.9
Removed50K70.654.170.565.1
Available50K71.157.372.366.9

(2) Frozen-teacher selection evaluation: no student distillation is involved in this comparison. At 50K, each frozen indexer selects the evidence, and the same original, nonprivileged model answers from it. This tests the usefulness of the teacher's selections directly.

Frozen indexer conditionVideoMMELVBenchMLVUtest\mathrm{MLVU}_{\mathrm{test}}Average
Ordinary Training-Free VEI70.553.566.763.6
Teacher: question type only70.353.767.563.9
Teacher: answer + question type71.357.570.566.4

Together, these controls show that the answer-aware teacher selects more useful evidence and transfers this benefit to the student. Adaptation on the separate MLVU training split may help the distilled system surpass the frozen-teacher reference, which is not a strict upper bound on QA performance. Removing the answer does not remove all guidance: the teacher still has a clearer preview when the student's input is compressed. Full results are in Appendix H.2, Tables 15–16.

Changing the training dataset.

MLVU test accuracy improves at both budgets even when the policy is trained on VideoMME instead of MLVU. We trained on short/medium VideoMME videos up to 25 minutes and evaluated on LVBench and the MLVU test split:

Training sourceContextLVBenchMLVUtest\mathrm{MLVU}_{\mathrm{test}}
None: Training-Free VEI15K36.746.2
VideoMME15K40.353.4
MLVUtrain\mathrm{MLVU}_{\mathrm{train}}15K44.958.8
None: Training-Free VEI50K53.566.7
VideoMME50K53.467.3
MLVUtrain\mathrm{MLVU}_{\mathrm{train}}50K57.372.3

At 15K, both external datasets improve over Training-Free VEI. At 50K, MLVU improves modestly while LVBench remains similar. These results support transfer of the learned behavior, with gains depending on the training source and preview resolution. Details are in Appendix H.3, Table 17.


Q3. Does the budget include indexing generation, and what are the per-dataset parsing and timestamp-validity failure rates?

Context budget and computation.

The indexing trace does not carry over into the answering context; its generation is included in the updated cost and latency accounting. The reported 15K/50K settings describe context width, not cumulative tokens processed across calls. Answering receives the preview, Evidence Set, question, and choices. Indexing generation occupies context during the first call and incurs computation, but its discarded text does not enlarge the second call's input. Appendix C reports the calculation and timing for both calls.

Parsing and timestamp validity.

Both failure types are uncommon, and their rates decrease after distillation. The table summarizes the unweighted mean of the three dataset rates; Appendix H.5, Table 19 provides the complete per-dataset counts.

MethodContextParse failure (%)Any out-of-range timestamp (%)
Training-Free VEI15K1.320.62
Self-Distilled VEI15K0.400.12
Training-Free VEI50K1.500.10
Self-Distilled VEI50K0.860.09

Please let us know if you have any further questions.

Area_Chair_3q4TOpenReview ↗

Friendly reminder from AC

Dear Reviewers,

Thank you for your reviews. The initial ratings are 3 / 2 / 2. The authors have posted responses to each review. Please read the response and let us know in the discussion whether your concerns have been addressed and whether your assessment has changed. Reviewers svpX and xot9, your views on whether the new attribution and comparison results are convincing would be especially helpful.

Thank you for your time and contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.