acceptodds
Withdrawn from ICLR 2027

Timing Is Not Enough: Mitigating Concurrent Action Omission in Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A prompt for video generation describes actions over time, and a subject is often asked to perform several of them at once. Temporal binding methods decide when each event appears and have made sequential multi-event generation more reliable. Once events contain concurrent actions, however, many requested actions still go missing, and those that appear are often not sustained. We trace the missing actions to two places along the conditioning pathway. In the text encoder, the subject that every frame must read mixes all its events, and isolating events from each other removes the actions that the subject carries. In video-text attention, the actions requested of one subject at the same time compete for the attention at its location, and this attention shifts from one action to another across frames. We propose CoRAct, a training-free approach that keeps the original prompt intact. It anchors each subject to the actions of its current event in the text representation and supports its co-required actions together at every frame in video-text attention. We further construct a benchmark of 84 prompts that specifies an interval for every action. Our protocol judges who performs each action and for how long, and it ranks the compared methods in the same order as human annotators. Across three backbones in image-to-video and text-to-video generation, CoRAct keeps the overall omission rate on par with existing training-free methods while significantly reducing the actions missing from their interval.

Primary area:
generative models
Keywords:
Video Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8kvNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies concurrent action omission in multi-action, multi-subject video generation, where multiple actions are requested for the same subject at overlapping time intervals. The authors analyze two sources of failure, inter-event interference in the text representation and intra-event action competition in video-text attention, and propose CoRAct, a training-free method that addresses them through event-anchored subject representations and subject-anchored concurrent-action guidance. They also introduce an 84-prompt benchmark with action-level temporal annotations and an evaluation protocol that measures whether each designated subject performs each action for a sufficient portion of its requested interval. Experiments are conducted on three video-generation backbones in both image-to-video and text-to-video settings.

Critical strengths
  1. The paper tackles an important problem of multi-action, multi-subject video generation.

  2. The paper raises an important definition of "action omission" and tries to tackle it from two dimensions, while providing an end-to-end evaluation protocol specifically targeting this problem.

Critical weaknesses
  1. In the definition of Omission Rate in Appendix C, the authors state that "A performer reported as unclear is not counted as another subject, so o_a = 0 in that case", which is quite confusing and possibly erroneous. By standard, such cases should at least be excluded from the final results, and if such cases are frequent, the reliability of the rating model (Qwen) should be questioned. This also contradicts the definition of omission given by the authors in Line 134, i.e., "An action is omitted when its designated subject does not perform it over most of its interval".

  2. Typically, concurrent actions should be performed in different parts of the frame. For instance, in Figure 1, "removing glasses" happens in the face region of the subject, while "returning a high five" happens in the hand region of the subject, and I see no reason why the attention for these concurrent actions should compete with each other if an ideal model is presented. In addition, some actions are inherently more complicated than others and thus require more attention support; manually raising the support of all concurrent actions equally may therefore harm the generation of such events.

  3. The evaluation is unconvincing. First, the provided qualitative samples are far from abundant. Nevertheless, from Figure 4 first row (CoRAct), neither sample is satisfying. In the left sample, the subject does not turn back to the audience, while in the right sample, there are severe artifacts where the head of the subject turns back while its flower-holding arm does not, both of which violate the provided prompt. Therefore, I do not think these actions should be assessed as successful in the evaluation. Considering that these samples have been manually selected, this makes the overall evaluation even less convincing. This also raises questions about the qualification of the MLLM automatic evaluator used in the paper. I believe human-based evaluation is necessary for the problem setup in this paper, which is also lacking.

  4. The writing is poor, and some notations and sections need to be clarified. Overall, I find the paper difficult to follow, and several parts are quite confusing. For example, the distinction among actions, events, and co-required actions is not always clear, and the mathematical formulation in Sections 3.1 and 3.2 introduces many notations in rapid succession, making the proposed method difficult to follow. In particular, the definition of the "action direction" through a Jacobian-vector product and the subsequent construction of the subject representation are not sufficiently intuitive. Some claims are also stated too strongly, such as "The support also decides the outcome" in Section 2.4. The definitions and relationships among OM, OM_P, and OM_S also require more clarification.

  5. In Table 4, why do the statistics for backbone Wan2.2 A14B not match the statistics in Table 2 under image-to-video and mixed split?

Questions for authors

Please refer to Weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8kvN

We thank the reviewer for the detailed reading. The review questions the soundness of the evaluation. The human evaluation it asks for is already part of the paper, and a negative control now tests the judge directly. The remaining points leave every reported number unchanged, so the findings stand, including the significant reduction of absent actions over existing training-free methods.

W1. Unclear performers. When the judge cannot tell who performs an action, it reports the performer as unclear. Appendix C does not count such an action as performed by another subject, since that requires evidence of a different performer, and the action must still appear and cover its interval. It occurs only on the sequential split, for 0.6% of the judged actions, mostly when only the performing hands are visible. Counting these actions as omitted instead would raise the sequential omission rate by at most 3.2 points, with similar increases for every method, so the ranking is unchanged.

W2. Why concurrent actions compete. Actions of different body parts often share a region, as when a subject smiles while taking off glasses, and attention is not confined to the performing body part, so a broad, dominant action can take most of the subject's attention. As Section 2.4 measures, when a second action is requested together with an action, its support falls to 0.78 of its value alone, below one in 35 of 47 cases. If attention were split by body part, the support of the first action would not change. The guidance does not raise all actions equally either. Its objective is a soft minimum over the co-required actions in each frame (Eq. 4), so it lifts the action with the least support and gives little weight to actions that are already well supported.

W3. Qualitative results and human evaluation. In the left example of Figure 4, TS-Attn, SwitchCraft and TIE all leave out covering the mouth, and TS-Attn claps instead of waving, while CoRAct covers the mouth and waves. Turning back is missed by every method, CoRAct included, and in the right example CoRAct turns the head but leaves the arm with the flower still. The revised caption describes these cases, and we add randomly sampled examples.

The human evaluation (Section 4.1, Appendix C) judges every action with the frames it covers, under the same 70% rule as the judge. Annotators find more actions performed as requested with CoRAct than with any other method (Table A), significantly more than with TS-Attn and TIE. Its videos come from Wan2.2 A14B I2V, as the revised Section 4.1 specifies, except TIE, which exists only for Wan2.2 TI2V-5B. Against the annotated frames, our judge also matches the performed intervals far better than StoryEval (IoU .53 against .30), a significant difference over prompts.

Table A. Human evaluation, share of actions performed as requested (%).

MethodPerformed as requested ↑
CoRAct31.1
SwitchCraft27.7
TS-Attn19.6
TIE (Wan2.2 TI2V-5B)5.4

A negative control tests the judge itself (Table B). It reports unrequested actions far less often than requested ones, stating them in the prompt does not raise this rate, and it is no more lenient to CoRAct than to the backbone. Re-judging the same videos reproduces 99.2% of labels.

Table B. Negative control on Wan2.2 TI2V-5B I2V, 912 queries (% reported present).

VideoRequested actionsUnrequested actions
Backbone63.914.4
CoRAct71.015.5

W4. Writing and notation. Table C restates the terms of the revised Section 3.

Table C. Terms and metrics.

TermDefinition
ActionOne verb phrase performed by one subject over one time interval
EventFor one subject, a maximal interval over which its set of requested actions does not change
Co-required actionsThe actions of one event, required together throughout its interval
OM_PShare of actions that never appear in their interval
OM_SOM_P plus the actions performed by another subject
OMOM_S plus the actions that cover less than 70% of their interval, so OM_P ≤ OM_S ≤ OM

The action direction is the frozen encoder's response on the subject row to adding the verb and object embeddings, one Jacobian-vector product per action (Eq. 1), and the text-side edit rotates the subject row toward their bounded average, keeping its norm (Eqs. 2 and 3).

W5. Table 4 and Table 2. Table 4 reports all 84 prompts, as the revised caption states. Weighting the two splits of Table 2 by size reproduces Table 4 (Table D), so the tables agree.

Table D. Wan2.2 A14B I2V, Table 2 against Table 4.

Table 2, mixed (54)Table 2, sequential (30)Weighted averageTable 4
Backbone, OM64.644.557.557.5
CoRAct, OM58.238.351.151.1

We hope the clarified evaluation and the added results address the concerns of the review, and we are glad to clarify anything further.

I thank the authors for their detailed response. Although multi-action generation is an interesting and important problem, I do not think the paper is ready for publication yet.

  1. Poor qualitative results. Neither CoRAct nor the baselines appear to successfully execute the given text prompts. While the authors argue that the baselines perform worse, this does not establish that CoRAct effectively solves the problem. Perhaps the authors could consider a simpler setting. Additionally, I could not find the additional video samples mentioned in the response, despite checking both the main paper and the supplementary materials.

  2. The human evaluation involves only nine annotators and ten unique prompts, which is insufficient to establish robust conclusions. Moreover, the 83 evaluated actions are not independent, as multiple actions belong to the same prompt. This raises concerns about the validity of the reported statistical significance.

  3. I remain unconvinced on Weakness 2 by the authors' response. In many cases, such as the example in Figure 1, concurrent actions should occur in different spatial regions of the frame. Furthermore, prioritizing the least-supported action may inevitably reduce the model's attention to other, potentially more complex actions.

Overall, I will maintain my original rating.

We thank the reviewer for the follow-up and for the time spent on our response.

1. Qualitative results. We agree that the prompts are not executed in full by any method. The qualitative results show where the methods differ. The baselines can place an event at the right time and still drop one of its concurrent actions or let another subject perform it, as with the sheet music and the glasses in Fig. 1. In the same example, CoRAct performs both on the designated subjects, which is the failure the paper targets, and the larger human study below agrees with this. The method is new in two respects. It changes the representation of a subject from one event to the next, using the response of the text encoder to the actions of its current event. It also ties the attention support of several actions to one subject and one interval, whereas temporal binding methods decide when each part of the prompt is read and attention guidance for images raises a neglected concept anywhere in a frame. We apologize that the additional video samples could not be found.

2. Human evaluation. We additionally ran a user study after the response. Twenty-five annotators made 1,348 judgments of 260 actions in the videos of four methods on ten prompts, and for every action they marked whether it appears, who performs it, and over which frames. The videos of CoRAct come from the current version of our objective. Under human judgment, 13.3% of the requested actions never appear with CoRAct, against 26.6% with SwitchCraft, 34.8% with TS-Attn, and 56.6% with TIE. The differences are 13.3, 21.6, and 43.3 points. These intervals come from a bootstrap that resamples prompts, so the actions of one prompt stay together and their dependence is accounted for. CoRAct performs 49.9% of the actions as requested, against 38.6%, 32.7%, and 22.1%, and its video is preferred in 28 of 50 choices.

3. Spatial regions of concurrent actions. We agree that concurrent actions often occupy different parts of one subject, such as the mouth and each hand in Fig. 1. Our support takes, for each action separately, the maximum over the locations of its subject, so two actions can be supported at different parts of the same subject. What it does not do is assign an action to a specific part. Doing so requires knowing the body parts of the subject, which restricts the method to human subjects and to image-to-video generation, where the first frame provides them. We therefore leave part-level support to future work, for example by tracking parts across frames through the self-attention of the video model. We also agree that emphasizing the least supported action can take attention from the others. The emphasis is soft, so every co-required action keeps a positive weight, and a part-level formulation would reduce this trade-off further.

We thank the reviewer again for the careful reading. These comments will shape the next version of the paper.

Reviewer_b7rMOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies concurrent action omission in video generation, where requested actions are missing, performed by the wrong subject, or insufficiently sustained. It proposes CoRAct, a training-free method combining event-specific rotations of subject text representations with subject-anchored attention guidance during denoising. The authors introduce an 84-prompt benchmark and evaluate three backbones in both text-to-video and image-to-video settings, with human validation on a subset. Results show reductions in completely absent actions, while overall omission remains comparable to existing training-free methods and sustaining actions remains challenging.

Critical strengths
  1. The paper provides a useful formulation and empirical analysis of concurrent action omission beyond temporal event ordering. Its investigation of interference in text representations and competition in video-text attention motivates the two interventions and offers actionable insight into why temporal binding alone remains insufficient.

  2. The benchmark and evaluation explicitly examine who performs each action and how much of its requested interval is covered. Experiments across three backbones and both generation modes, supplemented by human validation on a subset, provide informative evidence of reduced absent actions and applicability without backbone retraining.

Critical weaknesses
  1. The incremental methodological contribution is insufficiently established. Temporal binding and attention enhancement have precedents, but comparisons with simpler text updates are missing. Table 4 introduces the text component together with interval masking, and individual guidance losses lack separate ablations, limiting attribution of the gains.

  2. Gains are limited relative to computational cost. In the reported image-to-video experiments, CoRAct roughly doubles peak GPU memory on both Wan models and nearly triples generation time on CogVideoX. However, under the default threshold, pooled overall-omission improvements over TS-Attn and SwitchCraft are not statistically significant.

  3. The evaluation provides limited evidence of concurrent execution. The judge agrees with humans on 64.4% of events, with human validation covering only ten prompts. Moreover, per-action coverage does not guarantee simultaneous execution when requested action intervals differ and overlap.

Questions for authors

see above

1 comment
AuthorsOpenReview ↗

Response to Reviewer b7rM

We thank the reviewer for recognizing that our analysis explains why temporal binding alone is insufficient, which is where the contribution lies. Binding only decides which part of the prompt each frame reads, which does not by itself make actions appear, and separating events in the encoder strips the subject of its actions. CoRAct therefore keeps the full prompt, anchors each subject to its current actions and guides attention so that co-required actions appear at the subject. The comparisons the reviewer requests test exactly these points.

W1. Incremental contribution. Binding and attention guidance have precedents, and three comparisons show what they leave open.

Table A compares our edit with two simpler ways to put actions into the subject row, adding the action tokens' encoder outputs or rotating the row toward their mean. Our edit instead rotates the row toward the encoder's response to each action on that subject, from a Jacobian-vector product. Alone, none of the three changes absent actions significantly. With the guidance, only our edit lowers them, to 29.1%, the fewest of all updates.

Table A. Masking and guidance with each text-side update, Wan2.2 TI2V-5B I2V, mixed split (%).

Text-side updateOM_P (absent) ↓OM ↓
None32.365.2
Add the action outputs32.067.1
Rotate toward their mean32.666.1
Rotate toward the Jacobian response (ours)29.163.5

When the losses are added one at a time on Wan2.2 A14B over all 84 prompts, the binding loss lowers overall omission from 51.6% to 49.8%, significantly so on the sequential split, and the persistence and exclusion losses change it by +1.1 and +0.2 points, within the uncertainty.

Exact timing is not enough either. LongLive (Yang et al., ICLR 2026), an autoregressive model, can give each chunk only the prompt of its own interval, yet it shows the same omission. With the whole prompt it leaves 43.0% of actions absent, and its native prompt switching does not lower this. With our components adapted to its chunk-wise generation, absent actions fall to 36.3% and overall omission from 63.8% to 58.6%, both significantly over three seeds.

Across these comparisons, deciding when an action may appear does not make it appear. Keeping the full prompt does, and guidance on the edited subject row adds to it, which is what CoRAct contributes beyond binding. Unlike Attend-and-Excite, the guidance acts only at the subject, in every frame of the event, and keeps the action off other subjects.

W2. Gain relative to cost. The cost is modest, and on CogVideoX it stems mostly from the architecture (Table B).

Table B. Time per video (s) in image-to-video generation on one H100, excluding the first video of each run.

ModelBackbone+ Masking+ Text-sideCoRAct
Wan2.2 TI2V-5B65.365.068.781.3 (1.24×)
Wan2.2 A14B484.7491.0490.3684.5 (1.41×)
CogVideoX1.5-5B91.1125.4127.0204.9 (2.25×)

Wan2.2 reads the text through cross-attention, so masking and the text-side edit are nearly free, and the guidance, run in 5 of 40 steps, brings the total to 1.24× and 1.41× the backbone. CogVideoX is an MM-DiT that processes text and video in one sequence, so per-interval text needs a copy of the text stream in every block and step. Masking alone therefore costs 1.38×, and the total is 2.25×, not 3× as in Table 18. CoRAct is also faster than SwitchCraft (409 s on TI2V-5B and 343 s on CogVideoX in Table 18). Peak memory (Table 18) can be traded for time, as host offloading does on CogVideoX.

Pooled over six settings, CoRAct leaves significantly fewer actions absent than TS-Attn and SwitchCraft, and at θ = 0.5 its overall omission is significantly lower than every baseline (Table 8).

W3. Evidence of concurrent execution. For intervals that differ and overlap, the human study tests simultaneity directly. Annotators marked the frames of every action, and among 46 pairs of actions of one subject whose requested intervals differ and overlap, every pair in which both actions pass individually also co-occurs over at least 70% of the overlap (30 of 30). The concern thus does not arise in practice. Both actions pass and co-occur in 15.4% of these pairs for CoRAct, against 8.7% for SwitchCraft and 4.8% for TS-Attn. Our window measure, which scores only the overlap of requested intervals, also ranks CoRAct first (9.2% against 5.4 to 6.7%).

On the judge, the 64.4% event-level agreement reflects its leniency, which is not specific to CoRAct, since it reports unrequested actions in 14 to 16% of 912 queries, at the same rate for CoRAct and the backbone. At the frame level, it overlaps with the annotated frames of the 83 actions at an IoU of .529, against .299 for StoryEval, a significant difference over prompts.

Reviewer_5qRcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies concurrent action omission in video generation: actions requested simultaneously may be absent, assigned to the wrong subject, or performed for an insufficient duration. The authors analyze two potential sources of this problem: inter-event interference in text representations and competition among action tokens in video-text attention. They propose CoRAct, a training-free method that combines event-specific subject representation editing with subject-anchored attention guidance. The paper also introduces an 84-prompt benchmark and an evaluation protocol that measures action presence, subject assignment, and temporal coverage. Experiments cover three backbones in both image-to-video and text-to-video generation.

Critical strengths
  1. The problem is relevant and clearly motivated. Specifying when an event should occur does not ensure that all its constituent actions are performed concurrently. Distinguishing complete absence from insufficient temporal coverage offers a useful perspective on failures that aggregate event-completion scores may obscure.

  2. The method coherently combines representation editing and attention guidance. Deriving contextual action directions through Jacobian-vector products and rotating subject representations separately for each interval is an interesting design. The attention objective also explicitly accounts for the designated subject and the least-supported actions within an event.

  3. The experimental reporting provides useful detail. The evaluation spans three backbones and two generation modes, while the appendices report uncertainty estimates, threshold sensitivity, human validation, and computational costs. The authors also acknowledge that sustaining actions remains unresolved and appropriately qualify the benefits of reducing complete absence.

Critical weaknesses
  1. The evaluation does not sufficiently establish the significance of the reported gains. The automated judge agrees with humans on only 64.4% of events, compared with 84.3% agreement among annotators. Human validation covers just 10 prompts on one backbone and does not directly validate action-level temporal coverage. Repeated judge agreement demonstrates consistency, not correctness. The pooled reductions in absent actions are modest (2.2–3.2 percentage points), while overall omission improvements over TS-Attn and SwitchCraft are not statistically significant at the default threshold. Human evaluation also shows no significant improvement over SwitchCraft.

  2. The metrics only partially capture concurrent action realization. At 70% individual coverage, three actions are guaranteed only 10% common overlap; four or more have no guaranteed common overlap, despite the benchmark including up to six actions. Thus, all actions can pass individually without occurring simultaneously. The coverage requirement may also penalize correctly completed brief transitions, such as standing up or removing glasses. This limitation is acknowledged but not quantified.

  3. The ablations do not adequately isolate component contributions. Table 4 combines interval masking with text editing and lacks an attention-only condition. Adding attention guidance reduces overall omission by only 0.5 percentage points on A14B, without uncertainty estimates. On TI2V-5B, it reduces absence but increases partial actions, leaving overall omission unchanged. A controlled comparison of masking alone, each component added separately, and the full method would clarify their contributions.

  4. The causal interpretation exceeds the available evidence. Attention correlations and interventions suggest that attention affects action realization. However, Appendix G finds no significant increase in omission for concurrent versus individual requests (44.7% versus 48.9%). The analysis therefore does not establish how much failure is caused specifically by concurrency rather than difficulty generating individual actions. The causal claims should be qualified or better supported.

Questions for authors

The paper identifies a useful problem and proposes a technically interesting approach. My main reservations concern whether the evaluation reliably establishes improved concurrent action realization and whether the ablations substantiate the contributions of the proposed components. Stronger validation of the existing results and clearer attribution of the improvements to individual components could change my recommendation.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 5qRc

We thank the reviewer for the precise review. Rich prompts ask several subjects to perform several actions at once, and temporal binding is the current answer. Our paper makes three findings about it. Binding decides only when an action may appear. Removing the shared context strips the subject of the actions it carries. Appearing and being sustained are separate failures. CoRAct keeps the full prompt and adds guidance that makes missing co-required actions appear, measured on a benchmark with an interval for every action. Below, W1 and W2 address the first condition of the review, stronger validation, and W3 and W4 the second, clearer attribution.

W1 and W2. Validity of the metric and significance. Per-action coverage is deliberate. Existing video metrics score only event order or overall text alignment, and only per-action scoring tells which action is missing, misattributed or cut short. The human study tests this directly. For its 83 actions, annotators marked the frames in which each action is performed, as the revised Section 4.1 specifies. Against these frames our judge agrees far better than StoryEval, a significant difference over prompts (Table A), and 61% of the events that StoryEval judges completed contain an action the annotators did not see within the event.

Table A. IoU of judged and annotated frames for the 288 actions that annotators saw, by the method behind the videos. Higher metric in bold.

Videos ofOmission IoU (ours) ↑Order IoU (StoryEval) ↑
CoRAct.515.404
TS-Attn.521.259
SwitchCraft.564.237
TIE.475.134
All.529.299

The case the reviewer describes, in which every action passes alone but not together, is rare. In 801 of the 812 multi-action events in which every action passes (98.6%), all actions share at least 70% of the event's frames. On a joint score that requires this, CoRAct omits fewer events than every baseline, significantly fewer than the backbone, and 4.6 points fewer than TS-Attn when three or more actions are requested, also significantly.

TS-Attn and SwitchCraft leave as many actions absent as the backbone, since binding decides only when an action may appear. CoRAct is the only method that reduces them, significantly against every baseline by 2.2 to 3.2 points (Table 8), and the annotators confirm this (Table B).

Table B. Actions that annotators judged absent on Wan2.2 A14B I2V (%). CoRAct is significantly below TS-Attn, and on no prompt does it leave more actions absent than TS-Attn.

MethodAbsent ↓
CoRAct20.9
SwitchCraft29.1
TS-Attn37.2

Overall omission differs less at θ = 0.7 because it also counts partial actions. Coverage is measured within each action's own interval, which is short for brief actions such as standing up, and at θ = 0.5 the overall omission of CoRAct is significantly lower than that of every baseline, by 2.5 to 3.3 points.

W3. Attribution. Table 4 reports all 84 prompts, as the revised caption states. On Wan2.2 A14B the backbone leaves only 13.6% of actions absent, so little remains for the guidance to recover, hence the 0.5-point step. Table C separates the components on Wan2.2 TI2V-5B I2V, where absent actions are common.

Table C. Wan2.2 TI2V-5B I2V, mixed split (%). Best in bold, second in italics.

ConfigurationOM ↓OM_P (absent) ↓Order ↑
Backbone68.637.0.178
Masking63.232.8.170
Masking + text-side edit64.333.4.206
Masking + guidance65.232.3.139
Masking + text-side edit + guidance (CoRAct)63.529.1.223

Masking gives the drop in overall omission but leaves a third of actions absent and the order of events at the level of the backbone. The two components work together. Alone, each leaves absent actions unchanged (+0.6 and −0.5 points), and added to the other, each lowers them, by 3.2 points for the edit and 4.3 points for the guidance, the latter significantly. CoRAct thus leaves the fewest actions absent with the best order among the judged rows.

W4. Causal claims. Isolating events in the encoder raises absent actions on both splits, while re-encoding each interval alone, which removes all mixing, does not reduce omission, so the revised Section 2.3 attributes the effect to the shared context. A repeat of Table 15 with an equal-area control recovers as many actions outside the subject as everywhere, as the revised Section 2.4 reports. For the 11 pairs of actions that the backbone performs when each is requested alone, it performs both in 8 within a single event but in only 5 within the full prompt, while CoRAct performs both in 8 there, so CoRAct recovers the losses that multi-event prompts cause, although the counts are small.

We hope these results meet the two conditions named in the review, and we are glad to clarify anything further.

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

Related papers

Discussion (0)

Sign in to comment.