acceptodds
Under review as a conference paper at ICLR 2027

Rewiring Semantics, Dynamics, and Control: A Simple yet Effective Action-Centric Tri-Stream Transformer

Abstract

Vision-Language-Action (VLA) models have emerged as a prominent framework for complex robotic manipulation, building on the strong semantic understanding of pretrained Vision–Language Models (VLMs). However, such VLM backbones offer insufficient physical dynamics priors, which limits the generalization capabilities of robot policies. Recent efforts therefore integrate video-generation World Models (WMs) into robot policies through various strategies, using predictive dynamics to facilitate action generation. Despite these advances, harnessing semantic understanding and dynamics prediction as complementary guidance for action generation remains challenging. In this paper, we introduce ACT, a simple yet effective Action-Centric Tri-Stream Transformer that fuses semantic and dynamics information into control action while preserving the distinct roles of context streams. Specifically, ACT enables the dedicated action expert to access VLM and WM representations through layerwise attention, with each backbone attending only within its own streams. This straightforward interaction design maintains independent forward propagation in the context streams while allowing both backbones to be updated through control supervision. Experiments on both simulated and real-world robotic manipulation benchmarks show that the proposed ACT yields results superior to its counterparts.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language-Action Models, Video-Generation World Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_QLUAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an architecture, ACT3, leveraging a VLM and a world model as the foundational models. ACT3 keeps the semantic and dynamic stream independent, while the action expert cross-attends to both streams through layerwise cross-attention.

It uses pi05 and cosmos-predict video model as the backbone for the semantic and the dynamic stream and trained joint end-to-end for action generation and future visual prediction.

ACT3 reports results on LIBERO, LIBERO-PLUS, RoboCasa and fine-tuning on real-world robot data.

The central claims are

  1. How should the semantic and dynamic streams be updated ?
  2. Layerwise cross-attention is better suited for action-generation than final layer features
  3. Combining VLM with WM improves general manipulation performance.
Critical strengths
  1. ACT3 proposes a well motivated design question about how the semantic and dynamic stream interact and influence the action expert.
  2. Success rate improvement over baseline pi05, especially on RoboCasa and robust across shifts in LIBERO-Plus.
  3. Ablations are well motivated. Tab.3 Ablations are convincing about how the stream should interact. Table 7. Ablation aims to investigate which streams of the model should be kept frozen (authors might consider moving it to main paper). Sec 4.3 aims to analyze the how the action query attend to the three streams.
Critical weaknesses
  1. This paper is missing analysis on how the dynamics and semantic streams affect the action generations. What are the effect of degraded dynamics generation ? What happens when two of the streams decide on conflicting paths ? e.g. ‘for put the vegetables in the bowl, VLM might have a bias towards carrots, and WM towards potato, how does ACT3 action stream handle such cases ? Controlled experiments are missing to distinguish the benefits of the dynamics stream vs just additional features.

  2. This work fails to clearly isolate the source of performance gains. ACT3 relies on both a pretrained world-model and pi05. The work further claims generated future obs, trainable backbone, prediction supervision all contribute to improved performance. The real-world finetuning experiments compared with pi05 might not be fairest comparison. A baseline with both a VLA + dynamics model/vgm would provide more evidence supporting ACT3 architecture.

  3. Although an expensive experiment, missing ablation on where this approach would generalize to various sizes of VLM and WM and perform equally well with other foundational models. (least priority)

Questions for authors
  1. The novelty of ACT3 needs clearer claim. The architectural classification in Fig1 is helpful but still needs more stronger claims. Novelty is mainly architectural about layerwise attention and independent contexts. It would be useful to have a fair comparison on the various architecutural design, e.g. TriVLA vs ACT3 with the same backbones. An experiment comparing ACT3 with the prior works architecure with the same training and evaluation protocol would be very informative to the readers.
  2. How much does ACT rely on realistic dynamics ? How does action-stream cope with non-plausible generated future frames ?
6 comments
AuthorsOpenReview ↗

Response to Reviewer [Part 1/3]

We sincerely thank the reviewer for your insightful comments and for highlighting the strengths of ACT3\mathrm{ACT}^{3}, including its well motivated design question, success rate improvement, and convincing ablations. Below we address the main concerns raised in the review.


[W1&Q2] This paper is missing analysis on how the dynamics and semantic streams affect the action generations. How much does ACT rely on realistic dynamics? How does action-stream cope with non-plausible generated future frames?

Thank you for this insightful observation.

(1) We would like to clarify that the two streams are designed to provide complementary guidance for action generation.

  • The semantic stream provides global task guidance from current observations and the task instruction, while the dynamics stream predicts fine-grained local state changes. Both streams are conditioned on the same task instruction and are intended to support the same task objective through complementary global and local information. The evaluated benchmarks further provide clearly specified goals and target objects, with instructions fixed throughout execution. The carrot-versus-potato ambiguity is therefore absent from the task specification, and we expect goal-level conflicts to be limited in this setting.

  • Even when the two streams provide inconsistent information, the action expert learns to balance their contributions through attention. To better understand how action generation utilizes these contexts, we have analyzed event-dependent attention allocation in Figure 5 of the original paper. We observe that dynamics attention decreases during approach and initial contact, but increases around grasping, object motion, and articulation onset, where local state transitions become more important. In contrast, semantic attention changes more modestly, consistent with providing stable global task guidance as the scene evolves. These observations suggest that the action expert adaptively utilizes semantic and dynamics information across manipulation stages, with greater reliance on dynamics during object interactions.

(2) In Section 4.3 of our original paper, we have conducted ablation experiments to examine whether the benefits of the dynamics stream go beyond simply providing additional features. Specifically, we compare our layerwise access to WM representations with a variant that uses only the final-layer WM representations as additional context for the action expert. For convenience, we present the relevant results from Figure 4(a) of the original paper in Table R1. The results show that layerwise access performs better than using only final-layer WM features. This suggests that the improvements are not simply due to providing additional WM features, but also benefit from integrating dynamics information across layers.

Table R1. Success rates (%) with different WM layer access on RoboCasa.

WM Layer AccessSuccess Rate
Final layer63.5
Layerwise (ACT3\mathrm{ACT}^{3})71.2

(3) As suggested, we have additionally conducted experiments to examine the robustness of the action stream to non-plausible generated future frames by introducing different interventions on future representations. Specifically, we introduce three interventions on future representations throughout each rollout, while keeping all other settings fixed: First, we reduce the number of sampling steps used for future generation from the default of two to one. Second, we perturb the default future latent (zz) using zα=z+α(σc⊙ϵ),α∈0.25,0.5z_{\alpha}=z+\alpha(\sigma_c\odot\epsilon),\alpha\in\\{0.25,0.5\\} where σc\sigma_c is the per-channel standard deviation estimated from independent samples and ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I). Both perturbation levels use the same Gaussian noise realization for each corresponding query. Third, we introduce mismatched futures by replacing zz with the corresponding latent slice predicted from an image in a normal rollout of the same task, preferably from a different task stage.

The results in Table R2 show that the policy is relatively robust to reduced generation steps and the tested latent perturbations, but is sensitive to mismatched future representations. This contrast suggests that the effectiveness of predictive guidance depends on its alignment with current task progress. The additional results have been updated in Appendix C.6 (Page 21) of our revised manuscript.

Table R2. Success rates (%) under different interventions on future representations.

ConditionSuccess Rate
Default two-step generation (ACT3\mathrm{ACT}^{3})71.2
Single-step generation70.9
Latent perturbation, (α=0.25\alpha=0.25)70.6
Latent perturbation, (α=0.5\alpha=0.5)70.3
Mismatched futures54.8
AuthorsOpenReview ↗

Response to Reviewer [Part 2/3]

[W2&Q1] This work fails to clearly isolate the source of performance gains. ACT3 relies on both a pretrained world-model and pi05. The work further claims generated future obs, trainable backbone, prediction supervision all contribute to improved performance. The real-world finetuning experiments compared with pi05 might not be fairest comparison. A baseline with both a VLA + dynamics model/vgm would provide more evidence supporting ACT3 architecture. The novelty of ACT3 needs clearer claim. The architectural classification in Fig1 is helpful but still needs more stronger claims. Novelty is mainly architectural about layerwise attention and independent contexts. It would be useful to have a fair comparison on the various architecutural design, e.g. TriVLA vs ACT3 with the same backbones. An experiment comparing ACT3 with the prior works architecure with the same training and evaluation protocol would be very informative to the readers.

We thank the reviewer for the thoughtful comment.

(1) We would like to emphasize that our goal is not to develop an entirely new VLA architecture that requires training from scratch, but to explore a more practical yet challenging question: Can we push an already state-of-the-art pretrained VLA to new heights by directly leveraging an existing pretrained world model?

  • This is a very important direction. π0.5\pi_{0.5} is widely recognized in the research community as a state-of-the-art VLA, with strong generalization capabilities acquired through large-scale robot pretraining. Further improving such a strong model is far from trivial. Meanwhile, Cosmos has learned rich predictive dynamics from large-scale video pretraining. Rather than relearning these capabilities from scratch, we aim to unlock the complementary strengths of both pretrained models, leveraging the learned dynamics of Cosmos to further enhance π0.5\pi_{0.5}.

  • Simplicity is exactly what we strive for, and we believe it is not at odds with novelty. To our knowledge, our action-centric architecture has not been explored before, as highlighted by the comparison with prior methods in Figure 1. Our design preserves the independent forward computations of the pretrained context streams, while allowing the action stream to leverage both representation hierarchies through layerwise attention. Achieving meaningful improvements over an already state-of-the-art VLA through such a lightweight design is valuable in itself. More importantly, our approach provides the community with a simple, effective, and readily adoptable baseline that requires no additional large-scale pretraining and can be directly fine-tuned on downstream tasks after integration.

(2) We have provided an architectural comparison between ACT3\mathrm{ACT}^{3} and existing multi-stream approaches in Figure 1 of the original paper. The key difference from TriVLA [1] lies in how semantic and dynamics contexts are provided to the action policy. TriVLA uses VLM features and dynamics tokens aggregated from multi-level video features as conditioning inputs to a downstream action policy. In contrast, ACT3\mathrm{ACT}^{3} allows action queries to directly access the keys and values from the corresponding VLM and WM layers, enabling layer-wise integration of semantic and dynamics information.

(3) Moreover, we have additionally compared against a TriVLA-style feature-conditioning baseline suggested by the reviewer. Since the official training code and model checkpoints of TriVLA [1] are not publicly available, we implement this baseline using π0.5\pi_{0.5} and Cosmos under the same downstream training setup as ACT3\mathrm{ACT}^{3}. Specifically, semantic tokens are extracted from the final vision-language layer of π0.5\pi_{0.5}, while predictive tokens are aggregated from selected Cosmos blocks using learned queries. Each action block attends to these two token sets through separate cross-attention modules. As shown in Table R3, ACT3\mathrm{ACT}^{3} achieves a higher success rate than this TriVLA-style baseline, further supporting the effectiveness of our action-centric integration strategy. The additional results have been updated in Appendix C.2 (Page 19) of our revised manuscript.

Table R3. Comparison with the TriVLA-style baseline on RoboCasa.

MethodSuccess Rates (%)
TriVLA-style baseline64.9
ACT3\mathrm{ACT}^{3}71.2
AuthorsOpenReview ↗

Response to Reviewer [Part 3/3]

[W3] Although an expensive experiment, missing ablation on where this approach would generalize to various sizes of VLM and WM and perform equally well with other foundational models. (least priority)

Thanks for the constructive suggestion. Due to the high experimental cost, we focus on evaluating the generalizability of ACT3\mathrm{ACT}^{3} across different world models. Specifically, we have additionally designed an ACT3\mathrm{ACT}^{3} variant that pairs π0.5\pi_{0.5} with Wan2.2-TI2V-5B. We use learned linear projections to align the layerwise keys and values of Wan2.2 with the attention space of the action stream. We evaluate this variant on all 24 kitchen tasks in RoboCasa using the same training and evaluation settings. The results in Table R4 show that ACT3\mathrm{ACT}^{3} with Wan2.2 achieves comparable performance to ACT3\mathrm{ACT}^{3} with the Cosmos world model and outperforms the π0.5\pi_{0.5} baseline. These findings suggest that the effectiveness of ACT3\mathrm{ACT}^{3} is not limited to a specific world-model backbone and can generalize across different world models. The additional results have been updated in Appendix C.4 (Page 20) of our revised manuscript.

Table R4. Success rates (%) with different world-model backbones on RoboCasa.

MethodDynamics BackboneSuccess Rates
π0.5\pi_{0.5}None59.4
ACT3\mathrm{ACT}^{3}Wan2.2-TI2V-5B69.1
ACT3\mathrm{ACT}^{3}Cosmos-Predict2.5-2B71.2

Reference

[1] TriVLA: A Triple-System-Based Unified Vision-Language-Action Model with Episodic World Modeling for General Robot Control. arXiv 2025.

Reviewer_8W4mOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ACT³, an Action-Centric Tri-Stream Transformer that integrates semantic understanding and predictive dynamics for robotic manipulation. Motivated by the complementary capabilities of pretrained vision-language models (VLMs) and video-generation world models (WMs), the proposed framework aims to effectively combine semantic and dynamics representations for action generation.

The architecture consists of three specialized streams: a semantic stream initialized from the VLM backbone of π0.5, a dynamics stream initialized from Cosmos-Predict2.5-2B, and an action expert responsible for generating executable action chunks. The central design is an asymmetric attention mechanism in which action queries access both semantic and dynamics representations at each Transformer layer, while the two context streams remain independent during forward computation. The framework jointly optimizes all three streams through action supervision, together with an auxiliary future-prediction objective for the dynamics stream.

Experiments on RoboCasa, LIBERO, LIBERO-Plus, and three real-world bimanual manipulation tasks demonstrate consistent improvements over a matched π0.5 baseline. Additional ablations investigate the effects of attention topology, layerwise context access, generated future representations, and joint backbone optimization.

Critical strengths
  1. The paper addresses an important problem in vision-language-action learning: how to effectively combine semantic understanding from pretrained VLMs with predictive dynamics representations from video-generation world models. The proposed action-centric architecture is conceptually simple and well motivated. Rather than introducing extensive cross-stream interactions, ACT³ preserves independent semantic and dynamics context construction while allowing the action expert to integrate both sources of information through layerwise attention. I find this asymmetric interaction design interesting, particularly because it provides a clear separation between context representation learning and action generation while still allowing both pretrained backbones to adapt to the downstream control objective.

  2. The experimental results are encouraging, particularly on the more challenging RoboCasa and LIBERO-Plus benchmarks. On RoboCasa, ACT³ improves the average success rate from 59.4% to 71.2% compared with the matched π0.5 baseline. On LIBERO-Plus, the method achieves 82.4% success, outperforming the baseline by 14.0 percentage points without additional fine-tuning. These improvements are particularly meaningful because they extend beyond the nearly saturated standard LIBERO benchmark and demonstrate better robustness under distribution shifts. The paper also includes real-world evaluations on three bimanual manipulation tasks, where ACT³ consistently improves over π0.5. The gains on Block Collection and Flower Arrangement are especially encouraging, providing additional evidence that the proposed integration of semantic and dynamics information can benefit physical manipulation.

  3. A notable strength of the paper is its relatively comprehensive experimental analysis. The attention-topology ablations directly compare different information-flow patterns among the semantic, dynamics, and action streams. The results provide useful evidence that the proposed action-centric interaction can outperform more tightly coupled alternatives under the evaluated architecture and training protocol. The layerwise context experiments further demonstrate the value of accessing intermediate representations from both pretrained backbones rather than relying exclusively on their final-layer features. Moreover, the factorial ablation separating generated future context from auxiliary dynamics-prediction supervision is particularly informative. The results indicate that both components contribute to manipulation performance, helping distinguish the benefits of future representation conditioning from those of predictive supervision. The additional backbone-freezing and optimization controls further strengthen the empirical support for jointly adapting the semantic and dynamics streams to downstream control.

Critical weaknesses
  1. My primary concern is the extent of the methodological novelty relative to existing multi-stream VLA and WAM architectures. Several previous works, including TriVLA, F1, DUST, and related approaches, have already explored integrating semantic representations from VLMs with predictive dynamics from world models for robotic action generation. ACT³ distinguishes itself primarily through its asymmetric attention topology, layerwise access to independently constructed semantic and dynamics features, and joint optimization of the context backbones under action supervision. Although these design choices are reasonable and supported by the reported ablations, the overall methodological contribution appears somewhat incremental. In particular, the paper does not provide a direct, fully controlled comparison against a TriVLA-style feature-conditioning approach using the same pretrained π0.5 and Cosmos backbones, matched downstream training data, and comparable optimization and computational budgets. While the existing final-layer and attention-topology ablations are informative, they do not completely isolate the advantages of the proposed architecture over closely related feature-conditioning approaches. I believe a more direct comparison would help clarify whether the observed improvements primarily arise from the action-centric layerwise interaction design, the joint adaptation of the additional pretrained world model, or their combination.

  2. The proposed framework uses generated future visual representations from Cosmos as an important source of dynamics information for action generation. However, the future generation process is conditioned on the current observation and language instruction rather than on candidate robot actions. Consequently, it remains unclear to what extent the generated representations encode action-dependent physical transitions, as opposed to providing task-conditioned future-state priors. The experiments comparing current-frame and generated-future context provide encouraging evidence that future representations improve downstream control. However, they do not fully characterize the physical information captured by these predictions or explain how their accuracy and consistency affect action generation. For instance, the framework may generate a plausible future state that is inconsistent with the robot's actual subsequent actions or environmental evolution. It is unclear how robust the action expert is to such prediction discrepancies. Additional analyses relating future-prediction quality to downstream task success, or controlled experiments involving inaccurate or mismatched future representations, would help establish a clearer connection between predictive dynamics and improved control. Such evidence would strengthen the interpretation of the world model as providing physically meaningful guidance rather than simply additional pretrained visual representations.

  3. Although ACT³ demonstrates strong improvements in manipulation performance, the integration of an additional pretrained video-generation world model introduces substantial computational overhead. According to the reported timing measurements, ACT³ requires approximately 418 ms per policy query on RoboCasa using an H100 GPU, compared with 160 ms for the matched π0.5 baseline. This corresponds to approximately 2.6 times the inference latency. While the performance improvements on RoboCasa and LIBERO-Plus are meaningful, the paper provides limited analysis of whether similar benefits could be achieved with a more computationally efficient dynamics representation or a smaller world-model backbone. The real-world evaluations are appreciated, but their scope is also relatively limited. The experiments cover three manipulation tasks with 30 trials per method and task, using separately fine-tuned policies. Some improvements, such as the difference between 27/30 and 26/30 successful trials on Bowl Stacking, are relatively modest. Furthermore, the paper does not provide detailed end-to-end inference and control latency measurements on the physical robotic platform. A more comprehensive performance-efficiency analysis, together with additional real-world evaluations and uncertainty estimates, would help establish the practical scalability and deployment value of the proposed architecture.

Questions for authors
  1. Could the authors further clarify the methodological differences between ACT³ and closely related approaches such as TriVLA? In particular, could the authors provide a controlled comparison with a feature-conditioning baseline that uses the same pretrained π0.5 and Cosmos backbones, downstream demonstrations, and training protocol? Such a comparison would help isolate the benefits specifically attributable to layerwise action-centric attention and independent context construction, beyond the additional capacity and pretrained knowledge introduced by the world model.

  2. Since the future representations are generated without conditioning on candidate robot actions, how should their contribution to physical dynamics modeling be interpreted? Could the authors provide additional analyses relating future-prediction quality to downstream action prediction and manipulation success? It would also be valuable to investigate whether the policy remains robust when generated future representations are inaccurate or inconsistent with the actual state transitions, and whether these failure cases reveal limitations of the proposed conditioning mechanism.

  3. Could the authors provide a more detailed analysis of the performance-efficiency trade-off introduced by the additional world model? For example, how would the method perform with a smaller dynamics backbone, fewer future-generation steps, or a more lightweight dynamics encoder under matched inference budgets? Additionally, could the authors report end-to-end inference latency on the real-robot platform and provide uncertainty estimates or additional evaluations to better establish the consistency of the observed physical manipulation improvements?

4 comments
AuthorsOpenReview ↗

Response to Reviewer [Part 1/3]

We are glad that the reviewer appreciates the strengths of ACT3\mathrm{ACT}^{3}, including its conceptually simple and well-motivated architecture, encouraging experimental results, and comprehensive experimental analysis. We also sincerely thank you for your constructive comments and positive support. Below, we address the main points raised in the review.


[W1&Q1] Methodological novelty and controlled comparison with a TriVLA-style feature-conditioning baseline.

We thank the reviewer for the thoughtful comment.

(1) We would like to emphasize that our goal is not to develop an entirely new VLA architecture that requires training from scratch, but to explore a more practical yet challenging question: Can we push an already state-of-the-art pretrained VLA to new heights by directly leveraging an existing pretrained world model?

  • This is a very important direction. π0.5\pi_{0.5} is widely recognized in the research community as a state-of-the-art VLA, with strong generalization capabilities acquired through large-scale robot pretraining. Further improving such a strong model is far from trivial. Meanwhile, Cosmos has learned rich predictive dynamics from large-scale video pretraining. Rather than relearning these capabilities from scratch, we aim to unlock the complementary strengths of both pretrained models, leveraging the learned dynamics of Cosmos to further enhance π0.5\pi_{0.5}.

  • Simplicity is exactly what we strive for, and we believe it is not at odds with novelty. To our knowledge, our action-centric architecture has not been explored before, as highlighted by the comparison with prior methods in Figure 1. Our design preserves the independent forward computations of the pretrained context streams, while allowing the action stream to leverage both representation hierarchies through layerwise attention. Achieving meaningful improvements over an already state-of-the-art VLA through such a lightweight design is valuable in itself. More importantly, our approach provides the community with a simple, effective, and readily adoptable baseline that requires no additional large-scale pretraining and can be directly fine-tuned on downstream tasks after integration.

(2) We have provided an architectural comparison between ACT3\mathrm{ACT}^{3} and existing multi-stream approaches in Figure 1 of the original paper. The key difference from TriVLA [1] lies in how semantic and dynamics contexts are provided to the action policy. TriVLA uses VLM features and dynamics tokens aggregated from multi-level video features as conditioning inputs to a downstream action policy. In contrast, ACT3\mathrm{ACT}^{3} allows action queries to directly access the keys and values from the corresponding VLM and WM layers, enabling layer-wise integration of semantic and dynamics information.

(3) Moreover, we have additionally compared against a TriVLA-style feature-conditioning baseline suggested by the reviewer. Since the official training code and model checkpoints of TriVLA [1] are not publicly available, we implement this baseline using π0.5\pi_{0.5} and Cosmos under the same downstream training setup as ACT3\mathrm{ACT}^{3}. Specifically, semantic tokens are extracted from the final vision-language layer of π0.5\pi_{0.5}, while predictive tokens are aggregated from selected Cosmos blocks using learned queries. Each action block attends to these two token sets through separate cross-attention modules. As shown in Table R1, ACT3\mathrm{ACT}^{3} achieves a higher success rate than this TriVLA-style baseline, further supporting the effectiveness of our action-centric integration strategy. The additional results have been updated in Appendix C.2 (Page 19) of our revised manuscript.

Table R1. Comparison with the TriVLA-style baseline on RoboCasa.

MethodSuccess Rate (%)
TriVLA-style baseline64.9
ACT3\mathrm{ACT}^{3}71.2
AuthorsOpenReview ↗

Response to Reviewer [Part 2/3]

[W2&Q2] Predictive dynamics modeling and policy robustness to inaccurate or mismatched future representations.

Thank you for this insightful observation.

(1) We would like to clarify that our future generation does not explicitly predict action-dependent physical transitions conditioned on candidate robot actions, but instead provides task-conditioned future-state priors. In our imitation-learning setting, the dynamics stream is conditioned on the current observation and language instruction and supervised using future observations from successful expert trajectories. This encourages the generated future representations to capture short-horizon scene changes associated with successful task execution. These representations provide anticipatory cues about task progression, complementing the representation of the current observation and instruction from the semantic stream.

(2) In Figure 4(b) of our original paper, we have shown that both guiding action generation with predicted futures and supervising future prediction with demonstrated futures can improve task success rates. As suggested, we have conducted additional experiments to investigate how future-prediction quality affects task success. Specifically, we introduce three interventions on future representations throughout each rollout, while keeping all other settings fixed: First, we reduce the number of sampling steps used for future generation from the default of two to one. Second, we perturb the default future latent (zz) using zα=z+α(σc⊙ϵ),α∈0.25,0.5z_{\alpha}=z+\alpha(\sigma_c\odot\epsilon),\alpha\in\\{0.25,0.5\\} where σc\sigma_c is the per-channel standard deviation estimated from independent samples and ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I). Both perturbation levels use the same Gaussian noise realization for each corresponding query. Third, we introduce mismatched futures by replacing zz with the corresponding latent slice predicted from an image in a normal rollout of the same task, preferably from a different task stage.

The results in Table R2 show that the policy is relatively robust to reduced generation steps and the tested latent perturbations, but is sensitive to mismatched future representations. This contrast suggests that the effectiveness of predictive guidance depends on its alignment with current task progress. The additional results have been updated in Appendix C.6 (Page 21) of our revised manuscript.

Table R2. Success rates (%) under different interventions on future representations.

ConditionSuccess Rates
Default two-step generation (ACT3\mathrm{ACT}^{3})71.2
Single-step generation70.9
Latent perturbation, (α=0.25\alpha=0.25)70.6
Latent perturbation, (α=0.5\alpha=0.5)70.3
Mismatched futures54.8
AuthorsOpenReview ↗

Response to Reviewer [Part 3/3]

[W3&Q3] Performance-efficiency trade-off, end-to-end inference latency, and consistency of real-world manipulation improvements.

Thanks for the valuable feedback.

(1) We are pleased that the reviewer recognizes the meaningful performance improvements of ACT3\mathrm{ACT}^{3} despite its higher inference latency. We would like to clarify that our implementation includes no dedicated inference-speed optimizations beyond context caching, leaving room for further acceleration. Given the high experimental cost of evaluating alternative dynamics backbones, we instead explore a simple acceleration strategy by reducing the number of WM denoising steps from two to one. As shown in Table R3, this reduces inference latency from 418 ms to 374 ms, with only a marginal decrease in success rate from 71.2% to 70.9%. The additional results have been updated in Appendix B.4 (Page 18) of our revised manuscript.

Table R3. Success rates (%) and inference latency (ms/query) on RoboCasa.

MethodModel CapacityWM Denoising StepsInference LatencySuccess Rates
π0.5\pi_{0.5}3BN/A16059.4
ACT3\mathrm{ACT}^{3}5B137470.9
ACT3\mathrm{ACT}^{3}5B241871.2

(2) To quantify end-to-end inference latency during real-world evaluation, we measure the elapsed time from observation availability to dispatch of the first command in the newly generated action chunk on Block Collection. Table R4 reports the mean policy latency and replanning interval for both methods at a target command rate of 30 Hz. The additional results have been updated in Appendix F.2 (Page 26) of our revised manuscript.

Table R4. End-to-end inference latency (s) and replanning intervals (s) on Block Collection.

MethodMean inference latencyMean replanning interval
π0.5\pi_{0.5}1.253.49
ACT3\mathrm{ACT}^{3}1.714.10

(3) To provide uncertainty estimates for the real-world evaluations, we compute two-sided 95% Wilson confidence intervals from 30 trials per method and task. As shown in Table R5, the improvement on Bowl Stacking is modest (27/30 vs. 26/30), while larger gains are observed on Block Collection and Flower Arrangement, suggesting greater benefits on more challenging manipulation tasks. The additional results have been updated in Appendix F.2 (Page 26) of our revised manuscript.

Table R5. Real-world success rates with 95% Wilson confidence intervals. Entries are success rate [95% CI], all in percent.

Taskπ0.5\pi_{0.5}ACT3\mathrm{ACT}^{3}
Bowl Stacking86.7 [70.3, 94.7]90.0 [74.4, 96.5]
Block Collection73.3 [55.6, 85.8]83.3 [66.4, 92.7]
Flower Arrangement46.7 [30.2, 63.9]63.3 [45.5, 78.1]

Reference

[1] TriVLA: A Triple-System-Based Unified Vision-Language-Action Model with Episodic World Modeling for General Robot Control. arXiv 2025.

AuthorsOpenReview ↗

Response to Reviewer

Dear Reviewer,

We sincerely appreciate the time and effort you have put into reviewing our work. Your constructive feedback has been valuable in helping us further improve the paper. If you have any further questions or concerns, please feel free to let us know. We would be happy to address them before the discussion period ends. Thank you again, and have a wonderful day :)

Best regards,

The Authors of ACT3

Reviewer_wfc3OpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ACT3ACT^3 an Action-Centric Tri-Stream Transformer that combines a pretrained VLM and a pretrained video world model for robot action generation. Built as a Mixture-of-Transformers, ACT3ACT^3 initializes the semantic and action streams from π0.5\pi_{0.5} and the dynamics stream from Cosmos. The core design is a block attention mask under which the semantic and dynamics streams attend only within themselves, while action queries attend to both context streams at every layer. This makes the action expert the sole integration point while keeping the two context streams independent in the forward pass. All three streams are trained jointly with a flow-matching action loss plus an auxiliary future-latent prediction loss, and action gradients update both context backbones. The dynamics context comes from future latents generated from the current observation. On RoboCasa, LIBERO, and LIBERO-Plus, as well as three real-world bimanual tasks, ACT3ACT^3 outperforms an in-house π0.5\pi_{0.5} baseline.

Critical strengths

Strengths

  • Simple and practical design. The method combines existing pretrained checkpoints (π0.5\pi_{0.5} and Cosmos) through a single attention mask, without new modules or multi-stage training, making it easy to adopt on top of π0.5\pi_{0.5}.

  • Consistent empirical gains with supporting ablations. ACT3ACT^{3} improves over the π0.5\pi_{0.5} baseline across all evaluated settings, including RoboCasa (+11.8%p), LIBERO-Plus (+14.0%p), and three real-world tasks, with the largest real-world gain on the precision-demanding Flower Arrangement task.

  • Informative ablations. The design choices are each supported by controlled experiments: attention topology, layerwise vs. final-layer context access, generated vs. current-frame context with and without prediction loss, and freezing of each backbone.

Critical weaknesses

Weakness

  • Questionable baseline strength. Proprioceptive input is disabled for all policies without justification. The in-house π0.5\pi_{0.5} reaches only 68.4% on LIBERO-Plus, whereas a recent robustness study (Zhang et al., 2026) [1] reports 85.7% for π0.5\pi_{0.5}, which is higher than ACT3ACT^3(82.4%). This raises concerns that the main gains are measured against a weakened baseline.

  • Missing comparison with dynamics-only models. The method combines a VLM and a world model, but controlled comparisons are made only against the VLM-only baseline π0.5\pi_{0.5}. The Cosmos Policy is compared only via numbers reported in prior work, and not at all on LIBERO-Plus, where Zhang et al. (2026) [1] report 82.2% for Cosmos Policy, comparable to ACT3ACT^3. It is therefore unclear whether the gains come from combining the two streams or from the Cosmos prior alone.

  • Limited novelty. Each component already exists in prior work: independent context streams (TriVLA), layerwise MoT attention (π0\pi_0), and joint tuning of context backbones. The contribution reduces largely to a specific attention mask, with no analysis of why this topology should be preferable beyond empirical results.

Questions for authors

Questions

  1. Proprioception. Why is proprioceptive input disabled in the main experiments? Since π0.5\pi_{0.5} natively supports state input and the datasets already provide it, could the authors report results with proprioception enabled for both π0.5\pi_{0.5} and ACT3ACT^3, especially on LIBERO-Plus?

  2. Comparison with Cosmos Policy. Could the authors evaluate the publicly released Cosmos Policy checkpoint on LIBERO-Plus? This would help determine whether the gains come from combining both streams or from the Cosmos prior alone.

  3. Compute and latency. ACT3ACT^3 adds a 2B-parameter model and incurs about 2.6× inference latency compared to π0.5\pi_{0.5} (420 ms vs. 160 ms). How do the gains compare against a capacity- or compute-matched baseline?

6 comments
AuthorsOpenReview ↗

Response to Reviewer [Part 1/4]

We sincerely thank the reviewer for your insightful comments and for highlighting the strengths of ACT3\mathrm{ACT}^{3}, including its simple and practical design, consistent empirical gains, and informative ablations. Below we address the main concerns raised in the review.


[W1&Q1] Questionable baseline strength. Proprioceptive input is disabled for all policies without justification. The in-house π0.5\pi_{0.5} reaches only 68.4% on LIBERO-Plus, whereas a recent robustness study (Zhang et al., 2026) [1] reports 85.7% for π0.5\pi_{0.5}, which is higher than ACT3\mathrm{ACT}^{3} (82.4%). This raises concerns that the main gains are measured against a weakened baseline. Proprioception. Why is proprioceptive input disabled in the main experiments? Since π0.5\pi_{0.5} natively supports state input and the datasets already provide it, could the authors report results with proprioception enabled for both π0.5\pi_{0.5} and ACT3\mathrm{ACT}^{3}, especially on LIBERO-Plus?

Thanks for the constructive comments.

(1) Sorry for the confusion. Proprioceptive conditioning is optional in our framework. For LIBERO training, we follow the original configuration [2] in the official π0.5\pi_{0.5} repository, which does not use proprioceptive input. We apply the same setting to both π0.5\pi_{0.5} and ACT3\mathrm{ACT}^{3}. We have clarified this in Section 3.1 (Page 3) of our revised manuscript.

(2) The discrepancy in reported results stems from different implementations, including PyTorch and JAX, both of which are provided in the official π0.5\pi_{0.5} repository [3].

  • Previous studies also report performance differences between the two implementations. On LIBERO-Plus, Zhang et al. [1] report 85.7% for π0.5\pi_{0.5} using the JAX implementation, whereas ActionUNet [4] reports 65.8% for π0.5\pi_{0.5} using the PyTorch implementation. Moreover, Issue #989 in the official π0.5\pi_{0.5} repository [5] reports 82.6% with JAX and 71.9% with PyTorch on LIBERO-Plus.
  • Our work investigates how to integrate π0.5\pi_{0.5} with a world model. Since world models are commonly implemented in PyTorch, we adopt the PyTorch implementation provided in the official π0.5\pi_{0.5} repository to facilitate this integration. Note that our π0.5\pi_{0.5} and ACT3\mathrm{ACT}^{3} use the same PyTorch implementation. Therefore, the performance comparison in our experiments is fair.

(3) As suggested, we have additionally evaluated both π0.5\pi_{0.5} and ACT3\mathrm{ACT}^{3} on LIBERO and LIBERO-Plus with proprioceptive input. The results in Table R1 show that proprioception input improves the success rate on LIBERO-Plus for both methods, while ACT3\mathrm{ACT}^{3} maintains a substantial advantage over π0.5\pi_{0.5}. The additional results have been updated in Appendix E (Page 25) of our revised manuscript.

Table R1. Success rates (%) on LIBERO and LIBERO-Plus with and without proprioceptive input.

BenchmarkMethodProprioceptionSpatialObjectGoalLongOverall
LIBEROπ0.5\pi_{0.5}No98.098.497.593.596.9
LIBEROACT3\mathrm{ACT}^{3}No99.599.098.597.498.6
LIBEROπ0.5\pi_{0.5}Yes97.099.096.690.495.8
LIBEROACT3\mathrm{ACT}^{3}Yes99.098.899.096.298.3
LIBERO-Plusπ0.5\pi_{0.5}No77.171.463.662.068.4
LIBERO-PlusACT3\mathrm{ACT}^{3}No87.987.876.477.882.4
LIBERO-Plusπ0.5\pi_{0.5}Yes86.882.868.560.174.4
LIBERO-PlusACT3\mathrm{ACT}^{3}Yes87.590.877.979.583.9
AuthorsOpenReview ↗

Response to Reviewer [Part 2/4]

[W2&Q2] Missing comparison with dynamics-only models. The method combines a VLM and a world model, but controlled comparisons are made only against the VLM-only baseline π0.5\pi_{0.5}. The Cosmos Policy is compared only via numbers reported in prior work, and not at all on LIBERO-Plus, where Zhang et al. (2026) [1] report 82.2% for Cosmos Policy, comparable to ACT3\mathrm{ACT}^{3}. It is therefore unclear whether the gains come from combining the two streams or from the Cosmos prior alone. Comparison with Cosmos Policy. Could the authors evaluate the publicly released Cosmos Policy checkpoint on LIBERO-Plus? This would help determine whether the gains come from combining both streams or from the Cosmos prior alone.

Thanks for the valuable feedback.

(1) We would like to clarify that Cosmos Policy and the Cosmos world model are distinct models. Cosmos Policy uses a pretrained Cosmos world model as its foundation and fine-tunes it on robot demonstrations [6]. Similarly, we use a pretrained Cosmos world model as the foundation for our dynamics stream and also fine-tune it on robot demonstrations. Note that we do not initialize our model directly from Cosmos Policy. As suggested, we have evaluated the publicly released Cosmos Policy checkpoint on LIBERO-Plus. It achieves a success rate of 80.3%, compared with 82.4% for our ACT3\mathrm{ACT}^{3}. Thus, our ACT3\mathrm{ACT}^{3} achieves comparable performance to this strong dynamics-only baseline.

(2) To examine whether the gains come from combining the semantic and dynamics streams or from the Cosmos prior alone, we have conducted controlled ablations by freezing each pretrained context backbone in turn while keeping the other backbone and the action expert trainable. For convenience, we present the relevant results from Table 7 of the original paper in Table R2. With the VLM frozen, performance drops even when WM (Cosmos) is allowed to adapt, indicating that the dynamics stream cannot compensate for the lack of semantic adaptation. With WM (Cosmos) frozen, its pretrained prior combined with an adapted semantic stream is still not enough to match the performance of the fully fine-tuned model. Together, these results show the benefit of learning to combine semantic and dynamics information in ACT3\mathrm{ACT}^{3}, beyond simply using the pretrained Cosmos prior.

Table R2. Context-backbone adaptation ablations on RoboCasa.

VariantSuccess Rate (%)
Frozen VLM32.7
Frozen WM (Cosmos)62.3
ACT3\mathrm{ACT}^{3}71.2
AuthorsOpenReview ↗

Response to Reviewer [Part 3/4]

[W3] Limited novelty. Each component already exists in prior work: independent context streams (TriVLA), layerwise MoT attention (π0\pi_0), and joint tuning of context backbones. The contribution reduces largely to a specific attention mask, with no analysis of why this topology should be preferable beyond empirical results.

We thank the reviewer for the thoughtful comment.

(1) We would like to emphasize that our goal is not to develop an entirely new VLA architecture that requires training from scratch, but to explore a more practical yet challenging question: Can we push an already state-of-the-art pretrained VLA to new heights by directly leveraging an existing pretrained world model?

  • This is a very important direction. π0.5\pi_{0.5} is widely recognized in the research community as a state-of-the-art VLA, with strong generalization capabilities acquired through large-scale robot pretraining. Further improving such a strong model is far from trivial. Meanwhile, Cosmos has learned rich predictive dynamics from large-scale video pretraining. Rather than relearning these capabilities from scratch, we aim to unlock the complementary strengths of both pretrained models, leveraging the learned dynamics of Cosmos to further enhance π0.5\pi_{0.5}.

  • Simplicity is exactly what we strive for, and we believe it is not at odds with novelty. To our knowledge, our action-centric architecture has not been explored before, as highlighted by the comparison with prior methods in Figure 1. Our design preserves the independent forward computations of the pretrained context streams, while allowing the action stream to leverage both representation hierarchies through layerwise attention. Achieving meaningful improvements over an already state-of-the-art VLA through such a lightweight design is valuable in itself. More importantly, our approach provides the community with a simple, effective, and readily adoptable baseline that requires no additional large-scale pretraining and can be directly fine-tuned on downstream tasks after integration.

(2) To evaluate the benefits of the proposed topology, we have conducted controlled attention-topology ablations. For convenience, we present the relevant results from Table 3 of the original paper in Table R3. Specifically, we compare independent context construction with three alternatives: semantic-to-dynamics coupling, additional action-to-dynamics feedback, and fully coupled attention. The proposed topology achieves the highest success rate, while each tested form of additional coupling reduces performance. These results support combining semantic and dynamics information through the action stream without requiring the context backbones to interact during representation construction.

Table R3. Attention-topology ablations on RoboCasa. S, A, and D denote the semantic, action, and dynamics streams, respectively. "X reads Y" means that queries from X attend to keys and values from Y.

VariantS readsA readsD readsSuccess Rate (%)
Semantic-to-dynamics couplingSS, A, DS, D64.6
Additional action-to-dynamics feedbackSS, A, DS, A, D64.0
Fully coupled attentionS, A, DS, A, DS, A, D59.1
ACT3\mathrm{ACT}^{3}SS, A, DD71.2

(3) Moreover, we have additionally compared against a TriVLA-style feature-conditioning baseline suggested by the reviewer. Since the official training code and model checkpoints of TriVLA [7] are not publicly available, we implement this baseline using π0.5\pi_{0.5} and Cosmos under the same downstream training setup as ACT3\mathrm{ACT}^{3}. Specifically, semantic tokens are extracted from the final vision-language layer of π0.5\pi_{0.5}, while predictive tokens are aggregated from selected Cosmos blocks using learned queries. Each action block attends to these two token sets through separate cross-attention modules. As shown in Table R4, ACT3\mathrm{ACT}^{3} achieves a higher success rate than this TriVLA-style baseline, further supporting the effectiveness of our action-centric integration strategy. The additional results have been updated in Appendix C.2 (Page 19) of our revised manuscript.

Table R4. Comparison with the TriVLA-style baseline on RoboCasa.

MethodSuccess Rate (%)
TriVLA-style baseline64.9
ACT3\mathrm{ACT}^{3}71.2
Reviewer_dR57OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ACT³ combines independently computed semantic and dynamics contexts through layerwise action attention, jointly optimizing action generation and future prediction. Experiments show improvements over π0.5 across simulation benchmarks and real-robot tasks, supported by controlled component ablations.

Critical strengths
  1. Clear performance gains. Under matched training and evaluation protocols, ACT³ improves success from 59.4% to 71.2% on RoboCasa and from 68.4% to 82.4% on LIBERO-Plus, supporting improved task execution and generalization.

  2. Controlled evidence supports the design. With the same backbones, the action-centric topology outperforms three alternatives with additional coupling. Further ablations examine layerwise feature access, future context, and prediction supervision separately.

  3. The gains extend beyond one world-model backbone. Replacing Cosmos with Wan2.2 still achieves 69.1% success on RoboCasa versus 59.4% for π0.5, supporting applicability across different world models.

Critical weaknesses
  1. Inference overhead increases substantially. Compared with π0.5, query latency on an H100 rises from 160 ms to 418 ms, and real-robot replanning intervals increase from 3.49 s to 4.10 s (Tables 7 and 18).

  2. The policy is sensitive to mismatched future context. In the authors’ stress test, replacing future representations with those generated from other states of the same task reduces success from 71.2% to 54.8%, showing substantial disruption from incorrect context (Table 11).

Questions for authors
  1. Does future context still outperform current-frame context when the number of world-model forward passes is matched?

  2. How does future-prediction error relate to task failure during normal rollouts? Does the sensitivity observed in the mismatch test also occur with naturally generated predictions?

2 comments
AuthorsOpenReview ↗

Response to Reviewer [Part 1/2]

We sincerely thank the reviewer for the constructive comments and for highlighting the strengths of ACT3\mathrm{ACT}^{3}, including its clear performance gains, controlled ablations supporting the design, and applicability across different world-model backbones. Below, we address the main concerns and questions raised in the review.


[W1] Inference overhead increases substantially. Compared with π0.5, query latency on an H100 rises from 160 ms to 418 ms, and real-robot replanning intervals increase from 3.49 s to 4.10 s (Tables 7 and 18).

Thank you for the constructive comment.

The inference overhead of ACT3\mathrm{ACT}^{3} reflects a trade-off between task performance and inference efficiency: our method achieves higher task success rates than π0.5\pi_{0.5} at the cost of increased inference latency. Tables 7 and 18 in our revised manuscript report the inference and deployment latencies, respectively. For convenience, we summarize these results alongside the corresponding success rates in Tables R1 and R2 below.

We would also like to clarify that our current implementation includes no dedicated inference-speed optimizations beyond context caching, leaving room for further acceleration. As a simple acceleration strategy, we reduce the number of WM denoising steps from two to one. As shown in Table R1, this reduces inference latency while retaining strong task performance.

Table R1. Success rates (%) and inference latency (ms/query) on RoboCasa.

MethodNominal capacityWM denoising stepsInference latencySuccess rates
π0.5\pi_{0.5}3BN/A16059.4
ACT3\mathrm{ACT}^{3}5B137470.9
ACT3\mathrm{ACT}^{3}5B241871.2

Table R2. End-to-end inference latency (s), replanning intervals (s), and success rates (%) on Block Collection.

MethodInference latencyReplanning intervalSuccess rates
π0.5\pi_{0.5}1.253.4973.3
ACT3\mathrm{ACT}^{3}1.714.1083.3
AuthorsOpenReview ↗

Response to Reviewer [Part 2/2]

[W2&Q2] The policy is sensitive to mismatched future context. In the authors’ stress test, replacing future representations with those generated from other states of the same task reduces success from 71.2% to 54.8%, showing substantial disruption from incorrect context (Table 11). How does future-prediction error relate to task failure during normal rollouts? Does the sensitivity observed in the mismatch test also occur with naturally generated predictions?

Thanks for the valuable feedback.

(1) The mismatched-future experiment was introduced in response to the concerns raised by Reviewer 8W4m (W2/Q2) regarding how inaccurate or inconsistent future representations affect policy performance. To investigate this, we evaluate three interventions on the generated future representations used as dynamics context: reducing the number of world-model denoising steps from two to one, applying Gaussian latent perturbations at different strengths, and introducing mismatched futures. Each intervention is applied throughout each rollout while keeping all other settings fixed. Table 11 of our revised manuscript reports the results, which we present in Table R3 below for convenience.

ACT3\mathrm{ACT}^{3} demonstrates robustness to the tested sampling and latent perturbations. In contrast, the mismatched-future experiment is designed as a severe stress test: we replace generated future latents with those from different states of the same task, preferably at different task stages, throughout each rollout. This deliberately misaligns the predictive context with the current observation, potentially misleading the action expert about task progress and local state changes. The resulting performance drop is therefore consistent with how ACT3\mathrm{ACT}^{3} uses state-aligned predictive context to guide action generation.

Table R3. Success rates (%) under different interventions on future representations.

ConditionSuccess rates
Default two-step generation (ACT3\mathrm{ACT}^{3})71.2
Single-step generation70.9
Gaussian latent perturbation, α=0.25\alpha = 0.2570.6
Gaussian latent perturbation, α=0.5\alpha = 0.570.3
Mismatched futures54.8

(2) During normal rollouts, future representations are generated from the current observation and task instruction, and refreshed at each policy query. This differs from the mismatch test, which deliberately introduces persistent inconsistencies between the current state and future context. To investigate how future-prediction error supervision affects task success, we conduct an ablation in Figure 4(b). Specifically, our auxiliary future-prediction loss minimizes the reconstruction error between predicted and ground-truth future latents during training. For convenience, we present these results in Table R4 below. The results show that incorporating this prediction-error supervision consistently improves success rates with both current-frame and generated-future context, corresponding to fewer task failures during normal evaluation.

Table R4. Success rates (%) with different dynamics contexts and prediction-supervision settings.

Dynamics contextWithout prediction supervisionWith prediction supervision
Current-frame context61.366.0
Generated-future context65.871.2

[Q1] Does future context still outperform current-frame context when the number of world-model forward passes is matched?

Thanks for the insightful comment.

(1) We have conducted an ablation comparing current-frame and generated-future context to evaluate the benefit of future context for action generation in Table R4 above. Future context improves task success both with and without auxiliary prediction supervision. Both variants use the same architecture, with current-frame conditioning requiring fewer world-model (WM) forward passes since it omits future sampling at inference.

(2) To examine whether generated-future context remains effective with fewer WM forward passes, we reduce the number of WM denoising steps from two to one. For convenience, we summarize the relevant results from Table 7 of our revised manuscript in Table R5 below. The results show that generated-future context maintains strong task performance even with fewer WM denoising steps, suggesting that its benefits for action generation can be largely preserved with reduced WM computation.

Table R5. RoboCasa success rates with different numbers of WM denoising steps.

WM denoising stepsSuccess rates (%)
2 (default)71.2
170.9

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.