acceptodds
Under review as a conference paper at ICLR 2027

Balancing Timeliness and Firing Precision in Proactive Streaming Video Agents

12 views

Abstract

Streaming video understanding is becoming increasingly important for real-world interactive systems. Such systems must be proactive for natural and useful human interaction: they should respond as soon as sufficient evidence becomes available while avoiding premature, redundant, or incorrect responses. We present PACE, a framework unifying data, metrics, harness, and training around a shared formulation of proactivity as effective causal response decisions balancing the two competing objectives of responding vs silence. First, we construct PACE-780K dataset containing human annotations of proactive response content and timing, which defines this desired behavior. We then propose PAUC-F, a metric to effectively measure our desired proactive behavior. PAUC-F combines existing response timeliness metrics with firing precision to reward early useful responses without rewarding excessive firing. With our desired behavior defined and measurable, we finally investigate model-agnostic inference harnesses that draw out proactive streaming capabilities of offline vision language models (VLMs), together with training pipelines that learn to further improve their proactive responsiveness. Across both settings, we find that combinations of sliding visual windows, online interaction memory, and structured response outputs consistently yield strongest proactive performance. Our experiments result in two models, with our PACE-Qwen model achieving state-of-the-art performance across four proactive streaming benchmarks, with gains up to 44% over baselines and outperforming multiple closed API models. Our code and data will be released publicly.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Proactive Streaming Video

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_cM8YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies proactive streaming video understanding. In this setting, a video agent observes frames causally and must decide at each timestep whether to respond to an earlier user query or stay silent. The authors frame proactivity as a sequence of causal response decisions that trade off response timeliness against firing precision, and build a framework (PACE) around this formulation with four components:

  1. PACE-780K dataset. About 785K responses over 178K videos. Gemini generates the query and response/silence sequences, which human annotators then correct and expert reviewers audit. Response content, response intervals, and silence states are all verified.
  2. PAUC-F metric. It uses PAUC as recall and pairs it with firing precision via a harmonic mean, penalizing both redundant and out-of-window responses.
  3. Inference harness. It adapts frozen offline VLMs to streaming inference with a sliding frame window, interaction memory, and structured outputs. The design was found by agent-assisted search on validation PAUC-F.
  4. Streaming SFT pipeline. It trains on the same interface with Silence/Response/Hold state tokens and loss masking, yielding PACE-Gemma and PACE-Qwen.
Critical strengths
  1. Clear motivation. PAUC can be inflated by excessive firing. This is documented in prior work (MMDuet2) and demonstrated by the oracle/spam simulation in Fig. 2, so adding a precision term is well justified.
  2. Valuable dataset. Human verification covers response content, response intervals, and when the model should stay silent. Existing proactive datasets generally lack this (Table 1).
  3. Comprehensive experiments. The evaluation spans four proactive streaming benchmarks and different baselines.
Critical weaknesses
  1. Typos and writing issues. For example, “Casual Decision States” in Fig. 1 should be “Causal”, and “its roles is” in Sec. 2 is ungrammatical. The paper would benefit from careful proofreading.
  2. Hold ablation. Only 18% of Response states are followed by Hold, yet removing Hold drops WEB from 56.5 to 32.9 (Table 10). The mechanism behind this large effect needs explaining. It is also unclear how Hold is used at inference, since the output format (Table A.9) only distinguishes silence from a response.
  3. Train/inference mismatch. Training memory holds the last k annotated responses (Sec. 5.2.2), but inference keeps only the last accepted one (Table A.9). Likewise, training chunks use up to 32 frames versus 3 at inference.
  4. Scale mismatch in PAUC-F. PAUC has a silence floor of 0.25 (Appendix D), while precision has a floor of 0. The harmonic mean therefore combines R ∈ [0.25, 1] with P ∈ [0, 1], two terms on different scales. The authors should consider normalizing PAUC, e.g., (PAUC − 0.25)/0.75, or show that this choice does not affect model rankings.
  5. Possible column swap in Table 8. PACE-Gemma scores WEB 56.46 and EGO 43.80 in Table 3, consistent with Table 10. Table 8, however, reports the same model as EGO 56.5 and WEB 43.8, suggesting the EGO/WEB labels are swapped.
Questions for authors
  1. Is PAUC-F aggregated at the macro level or the micro level?
  2. How is the Hold state used at inference? Why does removing it cause such a large performance drop?
8 comments
AuthorsOpenReview ↗

Rebuttal [1/1]

We thank the reviewer for their time and feedback. All concerns are addressed below.

1. Typos

  • Thanks for the pointers, we have fixed these.

2. “Hold” State

  • The “Hold” state is only used with our finetuned models and their inference

  • As described in Table A.11, our prompt contains “</Hold> when the event continues but no new answer is due” and this prompt is used to finetune our model (and run inference with finetuned models).

  • We provide additional results on “Hold” ablation experiments below

    SplitModelPAUC-FFiresFire precisionPAUC (recall)Repeated fires (%)
    WEBWith Hold56.4697875.1%45.26%15%
    WEBWithout Hold32.902,84025.8%45.39%65%
    EGOWith Hold43.801,95542.2%45.52%40%
    EGOWithout Hold24.404,95016.7%45.28%75%
  • Note how this "Hold" state mainly reduces repeated redundant responses.

  • We will explain these clearly during revision.

3. Training/inference configuration

  • Table A.9 describes the selected inference harness (from search) used by the base Gemma4-A4B and Qwen3.5-A3B models, not our fine-tuned models that uses k=1 for memory and window size 3.
  • Our finetuned models use the 32 size window and the same k=1 across both training and inference.
  • We will clearly highlight this during revision.

4. PAUC metric definition and behavior

  • “which has a non-zero floor of 0.25 for a fully silent response” - note that for proactive streaming video, a response with multiple wrong responses can be worse than a fully silent response (see analysis and human studies in Wang et al. (2025b)); in fact this is how PAUC score is below 0.25 for some methods in Table A.1.
  • The reviewer’s proposal that “R should be rescaled, for example to (PAUC − 0.25)/0.75” would provide negative values that would break assumptions about the harmonic mean calculation in PAUC-F.
  • In summary, both the recall (PAUC) and precision terms in PAUC-F are bounded in [0,1] with values able to reach both ends and there is no explicit scale mismatch.

5. Typo (Column Swap)

  • Thanks for the pointer, fixed this.

Q1. Score aggregation

  • We aggregate at micro level. For each benchmark split or task, precision is computed from firing counts pooled across all videos, while the recall component follows the official PAUC protocol by averaging over all gold response turns. PAUC-F is then the harmonic mean of these two aggregate quantities.

Q2. Hold State

  • The “Hold” state is only used in our trained models and their inference. Our inference harness does not use this.
  • As described in Table A.11, a <Hold> state occurs when the event continues but no new answer is due
  • The large performance drop is due to increased repeated fires when removing the Hold state, as observed in the earlier table.

Thank you again for your feedback. We have uploaded a revised PDF, with changes highlighted in green, clarifying the Hold state, training/inference configurations, and metric definition, alongside the additional Hold ablation analysis in our rebuttal. We would appreciate your thoughts on whether these updates address your concerns.

We thank the reviewer again for the feedback.

We are wondering if the reviewer had a chance to see our rebuttal, and if any concerns still remain.

Reviewer_NbzpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses proactive streaming video agents, which must decide at each timestep whether to respond to an earlier query or stay silent, and must avoid both late and redundant responses. PACE aligns data, metric, inference harness and training under one formulation of proactivity as causal response decisions. First, PACE-780K (177,789 LLaVA-Video-178K videos, 784,944 responses) is built in three steps: Gemini generates one query and a 1-FPS sequence of response or silence states while prompted to behave causally; human annotators correct the timing and content; and PhD-level reviewers audit batches with a per-second F1 > 0.95 acceptance rule. Second, PAUC-F combines the quality-weighted timeliness of PAUC (as recall) with a firing-precision term (the earliest valid in-window fire per gold turn divided by all fires) through a harmonic mean. Simulated spam and a 10-rater preference study show that PAUC and ESTP-F1 reward excessive firing while PAUC-F penalizes it. Third, a model-agnostic harness (a 3-frame sliding window at 1 Hz, memory of the last accepted response, JSON-structured output) is selected through a 10-round agent-assisted search over 100 candidates on Gemma4-26B-A4B and Qwen3.5-35B-A3B. Other offline models receive only a small calibration. Fourth, chunked teacher-forced SFT with Silence/Response/Hold state tokens, response-preserving loss masking and memory produces PACE-Gemma and PACE-Qwen. PACE-Qwen reports the best PAUC-F on ProactiveVideoQA (49.08 vs 34.01 for MMDuet2) and on ESTP-Bench Explicit, and the best default-metric scores on ESTP-Bench, StreamingBench-PO and OVO-Bench CRR. Ablations cover data curation and scale, harness components and training components.

Critical strengths

1. The paper identifies a real flaw in current evaluation. PAUC rewards in-window repetition and ignores out-of-window firing, and the oracle-plus-spam simulation (Fig. 2) shows this convincingly. Reporting fire counts relative to gold and repeat statistics (Tables 2, A.4) is a useful practice.

2. The dataset contributes human-verified timing for responses and for silence, with a defined audit protocol (per-second TP/FP/FN/TN against an F1 threshold). This is a meaningful improvement over purely LLM-derived timing labels.

3. The evaluation is broad and largely held out: four external benchmarks, many proprietary, open-weight and streaming baselines, and the official metrics reported alongside PAUC-F (Tables A.1, A.2), including cases where the proposed models do not win.

4. The training ablations are informative. Without loss masking the model collapses to silence (0.0), and removing the Hold state or memory markedly increases repetition, which gives concrete guidance for streaming SFT.

Critical weaknesses

1. PAUC-F combines terms on different scales. Recall is PAUC, which has a non-zero floor of 0.25 for a fully silent response, while precision lies in [0, 1] with a true zero. A model that fires once, correctly, on a small fraction of gold turns and is otherwise silent obtains P = 1 and R ≈ 0.25, so PAUC-F ≈ 0.40 even though it misses almost every event. The harmonic-mean argument in Appendix D assumes both inputs are rates with a common zero. R should be rescaled, for example to (PAUC − 0.25)/0.75, and the main tables recomputed to check whether the rankings change.

2. The human validation of PAUC-F does not support the claim of "over 90% better alignment". The study compares only two harness variants (v1 with up to 48× gold fires and 186 repeats, versus v2), on 20 videos, with 10 raters. Almost any repetition-aware metric would prefer v2. A proper validation would collect pairwise human preferences over many model pairs with less extreme differences, and compare rank agreement (e.g., Kendall τ) for PAUC-F, PAUC, ESTP-F1 and simple alternatives such as PAUC with a repetition penalty.

3. Metric, model selection and evaluation share the same objective, and tuning budgets are unequal. The harness search, the calibration and checkpoint selection all maximize validation PAUC-F, and the headline results are reported in PAUC-F. Gemma and Qwen receive a 100-candidate search, whereas the proprietary and other open models receive five prompt and schema edits on 20 videos. The near-zero VAD scores of several Gemini models (0.33–0.66) suggest harness or parsing failures rather than model weakness. Under the benchmark's official PAUC (Table A.1), GPT-4.1 (58.89) and Gemini-2.5-Flash (64.51) exceed PACE-Qwen (45.07). Please report how many outputs failed to parse for each model, and give at least one strong proprietary model the full search budget.

4. The data contribution is not separated from the base-model contribution. Most of the reported 44% gain over MMDuet2 comes from the backbone and harness: the untrained Qwen3.5-A3B with the PACE harness already reaches 42.94 versus 34.01 for MMDuet2, and SFT adds about 6 points. MMDuet2 and VideoChat3 use older, smaller backbones. To support the claim that PACE-780K and the training recipe matter, fine-tune at least one baseline backbone (or the MMDuet2 recipe applied to Qwen3.5-A3B) on its original data versus PACE-780K under the same harness.

5. Evidence for the value of the human-verified dataset, presented as the first contribution, is modest. Table 8 shows that raw Gemini-generated labels reach 54.8 / 42.4, versus 56.5 / 43.8 after full human verification, a gain of about 1.5 points. Training on 10% of the data loses only about 2.5 points. The paper should report the human edit rates (the fraction of queries rejected and of states retimed, added or deleted) and test whether verification matters more for timing-sensitive metrics, such as the firing-delay distribution, than for PAUC-F. Note also that the WEB/EGO columns in Table 8 appear to be swapped relative to Table 10 and Table 3.

6. The source of hindsight in the labels is a concern. Stage-1 labels are generated by Gemini with access to the full video, under only a prompt-level instruction to behave causally. Annotators are told to enforce causality but correct only what they notice. The paper reports no measurement of residual anticipation, for example how often a response timestamp precedes the first frame where the event is visually identifiable, estimated on the audited subset. Systematic early labels would teach the model to fire before sufficient evidence appears.

7. The Hold state is under-analyzed relative to its effect. Only 18% of responses are followed by Hold states, yet removing Hold drops WEB/EGO from 56.5/43.8 to 32.9/24.4, a larger drop than removing memory. This suggests Hold does more than suppress repetition, perhaps by changing the Silence/Response balance under the subsampled loss. Please report fire counts and precision for the no-Hold model, and the distribution of predicted states at inference.

8. Potential overlap between training and test sources is not addressed. PACE-780K uses LLaVA-Video-178K source videos, while the test sets draw on web, egocentric (ESTP-Bench is built on Ego4D) and TV sources. The paper should state whether any test video, or near-duplicate clip, appears in the training sources, and describe any deduplication.

Questions for authors

1. In PAUC-F, a "value-adding" response needs a positive judge score. What is the judge's score scale (S = 2 in PAUC), and is a partially correct response with score 1 counted as fully valid in the precision term?

2. Memory holds only the last accepted response. For queries with several distinct events, how does the model avoid re-reporting an event that occurred two responses earlier and has left the 3-frame window?

3. At a 1 Hz decision rate with 3 frames of context, timeliness is quantized to 1 s, while ESTP-Bench Explicit gold windows average 4.8 s. How sensitive are the results to the decision rate and window size, given that the harness search converged to 3 frames?

4. The harness search was carried out by an agent with access to validation PAUC-F. Were any changes accepted that are specific to the PACE-780K query style, and how stable is the final harness when the search is repeated with a different random validation subset?

6 comments
AuthorsOpenReview ↗

Rebuttal [1/2]

We thank the reviewer for their time and feedback. All concerns are addressed below.

1. PAUC metric definition and behavior

  • “which has a non-zero floor of 0.25 for a fully silent response” - note that for proactive streaming video, a response with multiple wrong responses can be worse than a fully silent response (see analysis and human studies in Wang et al. (2025b)); in fact this is how PAUC score is below 0.25 for some methods in Table A.1.

  • The reviewer’s proposal that “R should be rescaled, for example to (PAUC − 0.25)/0.75” would provide negative values that would break assumptions about the harmonic mean calculation in PAUC-F.

  • For the given example, all variants that fire once correctly (for some golden windows) and a lot more incorrectly out of windows or redundant correct responses in gold windows score the same 0.25 PAUC while ours PAUC-F maps those worse variants to the [0, 0.40] range. See example below:

    Policy over 100 gold turnsPAUCPrecisionPAUC-F
    One perfect response at window start; otherwise silent0.25751.00000.4095
    Same response + 1 out-of-window fire0.25750.50000.3399
    Same response + 5 out-of-window fires0.25750.16670.2024
    Same response + 10 out-of-window fires0.25750.09090.1344
    Same response + 100 out-of-window fires0.25750.00990.0191
  • In summary, both the recall (PAUC) and precision terms in PAUC-F are bounded in [0,1] with values able to reach both ends. Our PAUC-F jointly measures both timeliness and firing precision through a suitable harmonic mean based combination.

2. Human validation

  • We reiterate that our human validation in Table 4 aims to highlight how “models game these metrics with undesirable traits”. The human study is simply to verify these traits are undesirable to humans. We will clarify this better during revision.
  • As observed in Table 4, PAUC-F and PAUC pick different models as better. For the two tasks (WEB, EGO) evaluated, humans pick that same model as PAUC-F 94% and 97%.
  • The reviewer mentions simpler alternatives without any reference to such existing metrics. Our proposed PAUC-F is a simple combination of PAUC with a repetition penalty. In Figure 2, we clearly illustrate benefits of PAUC-F against existing metrics PAUC and ESTP-F1.

3. Evaluation, tuning budgets, and baseline diagnostics

  • Prior work like MMDuet2 train their model with the same PAUC metric they used for evaluation. We do not train any of our models to directly optimize the PAUC-F metric. We only select models on held-out validation sets using the same downstream evaluation metric, which is standard across research.
  • Our improvements across other metrics such as ESTP-F1, Time Accuracy, Accuracy, and PAUC in Tables 5, 6, 7, A1, and A2 further establish how there is no selection bias or metric overfitting/gaming.
  • We explore per-model harness tuning when adapting baseline models to proactive streaming tasks. Prior works do not perform any per baseline systemic harness search, e.g. Wang et al. (2025b). This is another strength of our experimental setup compared to prior work. We did run full searches on Gemini-3.8-Flash and GPT-Sol-5.6 models, but observed no improvements in results.
  • There are zero parsing failures across all proprietary models tested. Our evaluation re-runs parse failures with the models up to 3 times, and zero cases failed 3 times in row.

4. Data Contribution

  • Thank you for the suggestion. We include two tables to separate the dataset and backbone effects more clearly below.

    Training data ablation: All methods share the same Qwen3.5-A3B backbone, and our training strategy. Only difference is training data. We report PAUC-F at ω=0.5. We also report the Base Qwen3.5-A3B (no finetuning) with our PACE harness for comparison.

    Training datasetWEBEGOTVVADMean
    Base model (no finetuning)54.2439.2934.4743.7542.94
    MMDuet2 (finetuned on MMDuet2 data)27.3424.3916.497.4618.92
    VideoChat3 (finetuned on VideoChat3 data)55.1539.7140.5345.7245.28
    PACE-780K (finetuned on our data)57.4846.0543.1149.6749.08
AuthorsOpenReview ↗

Rebuttal [2/2]

4. Data Contribution (cont.)

  • Second Table

    Training method comparison: All methods use the same Qwen3.5-A3B backbone but use their original data and training recipe. The two baselines use their default datasets and training strategies. We report PAUC-F at ω=0.5.

    MethodWEBEGOTVVADMean
    MMDuet2 (Qwen3.5-A3B)32.7629.5722.2013.0324.39
    VideoChat3 (Qwen3.5-A3B)35.8630.3640.7524.6732.91
    PACE-Qwen (Qwen3.5-A3B)57.4846.0543.1149.6749.08

5. Annotation quality and data scale

  • We include human edit rates of our dataset in the following tables. Note that deleted counts here include data deduplication as well.

    Record typeExported countFinal dataset countNet reductionReduction (%)
    Queries186,751177,7898,9624.80
    Response records893,562784,944108,61812.16
    Annotation outcomeCountPercentage of exported response records (%)
    Unchanged427,15047.80
    Modified104,46811.69
    Added253,32628.35
    Deleted361,94440.51
    Total893,562100.00
  • We also reiterate that PAUC-F is a timing sensitive metric, with varying omega changing the timeliness weight.

  • Thanks for the pointer, we have fixed that typo.

6. Causal labeling

  • While hindsight may appear in Gemini-annotated labels, labels in our dataset are cleaned by human annotators in the end.
  • Clear results improvements over prior work establishes how our proposed dataset does not lead to any such excessive early firing before sufficient evidence appears.
  • In fact, results indicate the exact opposite: models trained with our PACE-780K dataset exhibit much lower out of window firing compared to existing methods, which our proposed PAUC-F metric explicitly measures.

7. Hold State Analysis

  • Removing Hold substantially increases firing and repetition while leaving recall essentially unchanged

  • Hold primarily improves firing selectivity by suppressing repeated responses

    SplitModelPAUC-FFiresFire precisionPAUC (recall)Repeated fires (%)
    WEBWith Hold56.4697875.1%45.26%15%
    WEBWithout Hold32.902,84025.8%45.39%65%
    EGOWith Hold43.801,95542.2%45.52%40%
    EGOWithout Hold24.404,95016.7%45.28%75%

8. Dataset Deduplication

  • For all experiments, our PACE train and validation splits were filtered after checking for overlap with videos from ProactiveVideoQA, ESTP-Bench, StreamingBench (PO), and OVOBench (FAR tasks).
  • Filtering was performed by matching normalized source-video IDs (including parent-video IDs) and file checksums.

Q1. Value Adding for P term

  • Yes, we use the same S=2 used for the recall (PAUC) calculation.
  • Yes, both 1 & 2 are considered value adding.
  • We will better clarify both these in our updated paper.

Q2. Response memory

  • The 3 second window and single response memory was learned by the harness auto-search
  • Models do not demonstrate issues in repeating earlier responses so we believe this was learned for that reason

Q3. Decision rate and window size

  • We align the decision rate to prior work (e.g. MMDuet2) and fix to 1 FPS, which allows fair comparison. A higher decision rate could allow better timing and possibly improve performance.
  • The inference harness did not see improvements for longer window sizes (our search strategy included many variants with longer window sizes). This could be due to how the model uses longer visual context. Our trained model however benefited from a visual context up to 32 frames long, which we use as the default for both training and inference in our two finetuned models.

Q4. Search robustness

  • The harness design remains stable across searches, as indicated by two different searches using different model backbones with different pretraining led to same high level harness design
  • Also, as observed in Figure A.1, the last iteration of harness search barely changes performance
  • Our harness selected on PACE-780K achieves clear performance improvements against methods that are explicitly trained (e.g. MMDuet2, VideoChat3) across 10 different tasks from 4 benchmarks. We take this as clear evidence that all changes accepted in harness search generalize across downstream tasks.
  • Repeating the harness search with different validation subsets would require substantial resources that are not currently available during the rebuttal period. We will make every effort to conduct this analysis and include the results in the camera-ready version.

Thank you again for your detailed feedback. We have uploaded a revised PDF, with changes highlighted in green, incorporating clarifications and additional analyses discussed in our rebuttal, including controlled training comparisons, annotation statistics, and Hold ablations. We would appreciate your thoughts on these updates and any remaining concerns.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Overall Rebuttal

We thank reviewers for the feedback and positive comments:

  1. Clear motivation (cM8Y), identifies a real flaw in current evaluation (Nbzp)
  2. Valuable dataset (cM8Y), a meaningful improvement over purely LLM-derived timing labels. (Nbzp)
  3. Comprehensive experiments (cM8Y), evaluation is broad and largely held out (Nbzp), training ablations are informative (Nbzp)

We first point out three clarifications regarding our work.

1. Clarification on PAUC & Proactive Streaming Tasks

  • PAUC metric from Wang et al. (2025b)) has a floor of 0.0
  • For proactive streaming tasks, all silence is better than all wrong answers (see analysis in Wang et al. (2025b))
  • An entirely silent response has a 0.25 PAUC floor, but a response with only in-window wrong answers can score lower
  • PAUC scores below 0.25 naturally occur too - see Table A.1
  • Proposal to rescale “(PAUC − 0.25)/0.75” would result in negative values, breaking assumptions of the harmonic mean calculation, while our existing PAUC-F accounts for full range of PAUC.

2. “Hold” State Clarification

  • The training-free inference harness uses a binary silence/response interface (Table A.9). Hold is introduced in the fine-tuning pipeline only.
  • Hold distinguishes an ongoing, already-reported event from irrelevant frames labeled Silence, providing explicit supervision against repeating the same response.
  • Fine-tuned models can predict Hold, which produces no user-facing response and leaves response memory unchanged. Thus, both Hold and Silence correspond to remaining silent externally.

3. Dataset Deduplication

  • For all experiments, our PACE train and validation splits were filtered after checking for overlap with videos from ProactiveVideoQA, ESTP-Bench, StreamingBench (PO), and OVOBench (FAR tasks).
  • Filtering was performed by matching normalized source-video IDs (including parent-video IDs) and file checksums.

All other concerns are addressed in individual responses.

We have also uploaded a revised PDF (containing revision in green) to address clarifications and concerns.

We thank the reviewer for the Weak Accept rating, but must unfortunately mention that the review appears to concern a different paper. Our submission presents PACE, whereas the “Paper Summary” describes CommitStream. The strengths, weaknesses, and questions also appear to address that paper. Could the reviewer please check whether the intended review was posted to the correct submission?

Friendly reminder if the reviewer could verify that comments are for this paper (PACE) and not for another paper (CommitStream).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.