acceptodds
Under review as a conference paper at ICLR 2027

Evidence-Guided Counterfactual Learning for Long-Tailed Robotic Policies

Abstract

In the practical deployment of vision–language–action policies, handling long-tailed task distributions is a common challenge due to the highly imbalanced composition of robot demonstration datasets. To investigate this issue, we study the head-task action bias in shared robot policies, particularly focusing on its relationship with task-specific visual evidence. Our experimental analysis shows that, when the visual evidence required by a tail task is insufficient, the policy increasingly falls back on action patterns reinforced by data-rich head tasks, resulting in degraded performance on under-represented tasks. Motivated by this observation, we propose Evidence-Guided Counterfactual Learning, a training framework for long-tailed robot policies. Our method identifies decision-relevant visual evidence using instruction-related and action-related signals, and constructs counterfactual observations by removing the selected evidence. The resulting counterfactual predictions are used to strengthen the association among task instructions, visual evidence, and actions, with the supervision adaptively adjusted according to the number of demonstrations available for each task. The proposed framework requires neither additional training data nor region-level annotations and incurs no additional computation during deployment. Extensive experiments across simulated benchmarks, different VLA architectures, and real-world robot platforms demonstrate that our method robustly and consistently improves long-tailed robot policy learning, particularly for tail tasks, while maintaining reliable performance on head tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Long-Tailed Learning, Counterfactual Learning, Vision-Language-Action Policies

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dqptOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies head-task action bias in long-tailed robot policies and proposes Evidence-Guided Counterfactual Learning (ECL). A two-task diagnostic shows that tail-task instructions increasingly elicit head-task actions as imbalance grows, with visual masking further strengthening this tendency. ECL selects visual tokens through joint instruction–action attention, masks the selected evidence, and supervises factual predictions, masked predictions, and their contrast z+−λz−z^+-\lambda z^-. Counterfactual supervision is weighted by task rarity. Experiments cover LIBERO-Core-LT with MiniVLA and π0.5\pi_{0.5}, two real-world robot platforms, and context-level distribution shifts. The additional branches are used only during training.

Critical strengths

1. Coherent problem–method connection: the two-task diagnostic isolates the action dimension distinguishing head and tail tasks while holding the scene and manipulation procedure fixed. Reversing the head–tail assignment reverses the directional preference, arguing against a fixed spatial bias. Although imbalance and tail-data coverage change together, the experiment provides a useful motivation for strengthening task-specific visual grounding rather than treating long-tailed learning solely as a sampling problem.

2. Simple intervention with broad empirical support: ECL requires neither additional demonstrations nor region annotations and leaves deployment unchanged. The reported MiniVLA success improves from 40.8% to 53.0%, with additional gains across both real-world platforms and context-shift settings. Figure 4 supports the joint token selector over single-signal and random alternatives. The λ=0\lambda=0 comparison is also informative: this setting retains rarity-weighted factual supervision and masked supervision, providing evidence that prediction subtraction contributes beyond those components.

Critical weaknesses

1. The debiasing interpretation is stronger than the mechanism evidence. Removing selected tokens provides a useful evidence-limited reference, but z−z^- still contains task-relevant information and is explicitly trained toward the expert action. Subtracting it therefore does not isolate head-task bias. The method can be justified as an effective auxiliary training objective without a strict causal interpretation, but post-training HPS and Head-Action Rate measurements on factual observations would more directly support the claim that ECL reduces the diagnosed bias.

2. Cross-architecture implementation is underspecified. The generic notation zz does not establish whether subtraction operates on action logits, continuous actions, or flow predictions, which have different interpretations. The masking location, attention aggregation, and gradient routing through the two branches also need to be stated. Appendix C uses semantic-mask and saved-attention proxies rather than actual selector outputs, so its illustration cannot substitute for documenting the implemented selection procedure.

Questions for authors
  1. What are z+z^+ and z−z^- for each backbone, and how is LCF\mathcal L_{\mathrm{CF}} implemented? Please specify masking location, attention aggregation, gradient routing, and whether noise and timesteps are shared when contrasting flow predictions.

  2. Does ECL reduce HPS and Head-Action Rate on intact observations after training? This would strengthen the connection between the motivating diagnostic and the learned policy.

  3. Please clarify the result reporting. What does “synthetic” mean in the π0.5\pi_{0.5} header of Table 1? Appendix B.1 states 675 episodes per task, whereas Table 6 specifies 225; which count underlies the main results, and how are the reported uncertainties calculated? Please also reconcile the tail-task definitions in Section 4.1 (Tasks 7–9) and Appendix A (Tasks 3–9), ideally with per-seed success counts.

4 comments
AuthorsOpenReview ↗

Response to Reviewer — dqpt [1/2]

q1: What are z+z^+ and z−z^- for each backbone, and how is LCF\mathcal{L}_{\mathrm{CF}} implemented? specify masking location, attention aggregation, gradient routing, and whether noise and timesteps are shared when contrasting flow predictions.

a1:

We thank the reviewer for this question. We clarify the implementation details below.

  1. Prediction definitions. For MiniVLA, z+z^+ and z−z^- denote the action-token logits predicted from intact and masked visual observations, respectively. For π0.5\pi_{0.5}, they denote the corresponding velocity predictions in Flow Matching. Both branches share the same policy parameters.

  2. Masking and attention aggregation. We perform a separate forward pass without gradient tracking to extract attention from instruction-related and action-related queries to visual tokens. The attention signals are aggregated into token-level scores αjL\alpha_j^L and αjA\alpha_j^A, respectively. Their joint relevance is computed as rj=αjLαjAr_j=\alpha_j^L\alpha_j^A, and the top 10% of visual tokens are selected for masking. Masking is applied only to the selected visual-token positions in the observation, while the language instruction and action targets remain unchanged. The selected masked observation is then processed by the same policy to obtain z−z^-. The attention-based selection and masking decisions do not receive gradients.

  3. Counterfactual loss and gradient routing. We construct the contrastive prediction as zeffect=z+−λz−z^{\mathrm{effect}}=z^+-\lambda z^-, with λ=0.3\lambda=0.3, and define LCF=ℓ(zeffect,a∗)\mathcal{L}{\mathrm{CF}}=\ell(z^{\mathrm{effect}},a^*), where ℓ\ell is the backbone-specific action supervision loss (cross-entropy for MiniVLA and Flow Matching regression for π0.5\pi{0.5}). The overall objective is L=LBC+μLmask+riLCF\mathcal{L}=\mathcal{L}{\mathrm{BC}}+\mu\mathcal{L}{\mathrm{mask}}+r_i\mathcal{L}{\mathrm{CF}}, where rir_i adjusts counterfactual supervision according to task frequency. LBC\mathcal{L}{\mathrm{BC}} and Lmask\mathcal{L}{\mathrm{mask}} supervise the intact and masked branches, respectively, while LCF\mathcal{L}{\mathrm{CF}} backpropagates through both predictions without stop-gradient, jointly optimizing the shared policy.

  4. Shared Flow Matching conditions. For π0.5\pi_{0.5}, both branches use the same sampled noise and Flow Matching timestep. Thus, their velocity predictions are contrasted under identical stochastic conditions, ensuring that the prediction difference primarily reflects the removal of task-relevant visual evidence rather than variations in noise or timestep.


q2: Does ECL reduce HPS and Head-Action Rate on intact observations after training?

a2:

  1. To verify that ECL can alter the scale or bias of the original predictions, we applied the ECL method to the toy experiment described in Section 3.1. Upon re-evaluating the HPS and Head action rate, we observed a significant decrease, which indicates that our method effectively reduces the head-action bias affecting tail tasks.

    γ = Nₕ/NₜHPS (LT)HPS ↓ (ECL)Head-Action Rate (LT)Head-Action Rate ↓ (ECL)
    150.0% ± 0.050.0% ± 0.050.0% ± 0.050.0% ± 0.0
    253.3% ± 2.051.7% ± 2.951.7% ± 2.951.7% ± 2.9
    561.7% ± 3.058.3% ± 5.855.0% ± 0.053.3% ± 2.9
    1070.0% ± 3.563.3% ± 2.963.3% ± 2.961.7% ± 2.9
    2081.7% ± 2.470.0% ± 5.073.3% ± 2.965.0% ± 5.0
  2. We have added ablation experiments in response to your comments, and the relevant content has been supplemented in the main text (see 4.3.1). We focus on analyzing the success rates across tasks for the full objective with factual-plus-masked supervision and ablation was performed for this condition.

    Supervision SettingLBC\mathcal{L}_{BC}Lmask\mathcal{L}_{mask}LCF\mathcal{L}_{CF}average success rates (%)
    Factual-only✓——40.8
    Factual + Masked✓✓—43.2
    Factual + Counterfactual✓—✓41.6
    r<sub>i</sub>=1✓✓✓45.5
    ECL (Ours)✓✓✓53.0
  3. Furthermore, the attention heatmap analysis in Appendix D demonstrates that ECL improves the accuracy of visual evidence grounding for underrepresented tail tasks, enabling the policy to focus more consistently on task-relevant objects, interaction regions, and target locations.

AuthorsOpenReview ↗

Response to Reviewer — dqpt [2/2]

q3: What does “synthetic” mean? Which count underlies the main results, and how are the reported uncertainties calculated? Please also reconcile the tail-task definitions in Section 4.1 (Tasks 7–9) and Appendix A (Tasks 3–9), ideally with per-seed success counts.

a3:

We thank you for this question. We clarify the details below.

  1. "synthetic" refers to the intermediate action states synthesized during the Flow Matching training process of π0.5\pi_{0.5}. We originally used this term to distinguish the action modeling and training mechanisms of π0.5\pi_{0.5} and MiniVLA. Specifically, π0.5\pi_{0.5} adopts a hybrid architecture combining a vision-language model and an action expert. During action generation, intermediate action states are constructed by interpolating between random Gaussian noise and ground-truth demonstration actions:

    at=(1−t)ϵ+tagt.a_t = (1-t)\epsilon + ta_{gt}.

In contrast, MiniVLA employs autoregressive action-token prediction. We acknowledge that the term "synthetic" in the original Table 1 may cause confusion, particularly regarding the use of additional synthetic data. To avoid this misunderstanding, we have removed the "synthetic" label from Table 1 and clarified the different action modeling mechanisms in the experimental setup.

  1. The tasks in Section 4.1 refer to the extreme tail tasks.We want to distinguish between different degrees of tail tasks.And each task is evaluated with three training seeds and 225 closed-loop episodes per seed, 675 episodes per task per method.

  2. Let ksk_s denote the number of successful episodes obtained by seed s∈{7,14,21}s \in \{7,14,21\} on a given task, out of the 225225 closed-loop episodes evaluated for that seed. Each task is thus evaluated with 675675 episodes per method.

    The per-seed success rate is ps=ks/225p_s = k_s / 225, and the reported per-task success rate is

    \qquad K = \sum_{s \in \{7,14,21\}} k_s.$$ The equality on the right holds because all three seeds are evaluated with the same number of episodes. The value after $\pm$ is the sample standard deviation of the three seed-level rates (denominator $n-1=2$): $$\sigma = \sqrt{\frac{1}{2} \sum_{s \in \{7,14,21\}} (p_s - \bar{p})^2}.$$
  3. Table R1 — Per-seed success counts.Each cell reports k1/k2/k3k_1/k_2/k_3, the number of successful episodes out of 225 for the three seeds.

    MiniVLA

    TaskFULLLT (Baseline)APAECL (Ours)
    Task 0118/128/136148/157/157159/162/169175/180/185
    Task 1167/176/185103/113/119149/150/155152/155/158
    Task 2154/167/170132/142/146144/151/151164/168/172
    Task 3114/122/131118/122/133152/159/161184/188/192
    Task 4142/146/16096/107/112123/128/130142/147/148
    Task 5121/135/13983/92/9896/101/103105/108/111
    Task 6114/123/13276/81/9179/86/8687/89/92
    Task 798/112/11615/23/3323/27/3029/30/33
    Task 8116/132/13379/84/9444/47/50120/122/125
    Task 950/59/680/0/01/3/54/5/7

    π0.5\pi_{0.5}

    TaskFULLLT (Baseline)APAECL (Ours)
    Task 0122/128/133155/164/170153/156/160164/171/175
    Task 1132/136/147141/146/148143/150/151157/162/165
    Task 2157/166/167126/134/136126/132/132148/149/151
    Task 3109/114/116114/116/125120/126/128133/134/135
    Task 4149/154/15899/105/10589/99/104109/119/124
    Task 5128/131/13877/81/8988/89/9594/102/104
    Task 6116/123/12558/70/7359/63/6679/82/92
    Task 7101/103/11050/53/5357/62/6766/70/71
    Task 8125/134/13426/38/3942/43/4645/54/55
    Task 954/59/648/10/1121/23/2416/28/32
  4. About Appendix C: We would like to clarify that the Object-only, Action-only, and Object–Action Relation visualizations are all directly obtained by exporting the corresponding selection results from the full ECL selector. Only the Random baseline is presented as an illustrative example due to its stochastic nature. This was our original intention, but our previous wording may have caused confusion. We have revised the corresponding description in the manuscript to explicitly clarify this distinction. We apologize for the ambiguity and thank you for pointing it out.

Robot demonstration datasets often contain vastly different amounts of data for different tasks, causing VLA policies trained on them to perform poorly on tail tasks. The paper argues that when the visual evidence needed for a tail task is lacking, the policy tends to favor actions associated with data-rich head tasks. To mitigate this, it scores visual tokens using instruction- and action-related attention, masks the highest-scoring tokens, and adds a counterfactual loss on the contrast between intact and masked predictions to the standard behavior cloning loss, weighted according to task frequency. The idea is interesting, the experiments are well structured, and the authors’ response addressed my questions about implementation details and how the results were aggregated.

However, a few concerns remain. The additional measurements suggest that head-task bias is clearly reduced only in the most imbalanced setting, and remains substantial even there. It would also help to clarify how the Head-Action Rate should be interpreted. Under a tail-task instruction, a policy that follows the instruction would presumably move toward the tail object, giving a rate near zero; yet the rate is 50% in the balanced setting (γ = 1). What value is desirable, and does 50% indicate the absence of head-task bias or the absence of instruction grounding? Finally, the new ablation calls for an explanation of why frequency-aware weighting lowers success on some tail tasks relative to fixed weighting (e.g., Task 7: 13.7% vs. 44.7% in Table 3).

AuthorsOpenReview ↗

Response to Reviewer — dqpt

Thank you once again for your response. You clearly have a strong grasp of our work, and we are very grateful for the excellent suggestions and questions you have raised regarding our paper.

q1: How the Head-Action Rate should be interpreted? What does 50% indicate?

a1: We want to clarify that the "Head-Action Rate" represents the extent to which the focus of a given task shifts toward the head task when the 10% of tokens deemed most important are masked during the execution phase, given a set of random token-based instructions. When the proportions of head and tail tasks are equal, our experimental results indicate that the model perceives the important segments for these random instructions to be distributed evenly; consequently, the shift toward the head and the tail is balanced. And 50% indicates the absence of head-task bias.As the proportion of head tasks increases, the primary basis for model predictions shifts toward these head classes. This is evidenced by an increase in the HPS and the fact that the Head-Action Rate is calculated by masking out the most critical subset of these tasks. Notably, the deviation in the model's performance on head tasks,relative to the HPS,is lower, which indicates that when tasks are distributed unevenly, the model tends to develop more evidence associated with the head classes. This, in turn, explains the origin of head action bias.(Note: The term "tail instructions" in the original text was a clerical error; we have corrected this in the main body of the paper.)


q2: Why frequency-aware weighting lowers success on some tail tasks relative to fixed weighting?

a2:

We would like to point out that when rir_i is held constant, certain tail tasks can attain a larger proportion of counterfactual branches, which to some extent boosts their success rate. However, this comes at a significant cost to head tasks; excessive counterfactual learning causes the performance of head tasks to deteriorate substantially. Even so, our results are still better than those of LT .To strike a better balance and achieve greater overall gains, we opted for frequency-aware weighting,a key reason why we adopted this approach.

Reviewer_1pkSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Evidence-Guided Counterfactual Learning (ECL) for vision–language–action policies trained on imbalanced task demonstrations. It combines instruction-related and action-related attention to select visual tokens for masking, trains a shared policy on intact and masked observations, and supervises the prediction contrast. The paper reports improvements over long-tailed training and Approaching-Phase Augmentation (APA) on LIBERO-Core, together with real-world comparisons on two robot platforms.

Critical strengths
  1. The paper addresses a relevant problem in robot learning: shared policies may default to frequent behaviors when evidence for an infrequent task is insufficient. The two-task diagnostic along with the toy experiment is useful, and reversing the head–tail assignment helps rule out a fixed directional preference in that setting.
  2. ECL is conceptually simple and uses existing demonstrations without requiring additional region annotations or deployment-time computation.
  3. The experiments are broad in empirical scope and encouraging, with both simulation results and real-world results which cover two policy backbones and two platforms.
Critical weaknesses
  1. The empirical evidence contains material reporting ambiguities. The π0.5 block in Table 1 is explicitly labeled “Success Rate (%, synthetic),” but the paper does not explain what “synthetic” means or how these numbers were obtained.
  2. The contribution appears to be the combination of joint relevance masking, prediction-difference supervision, and frequency-dependent weighting for VLA training. Counterfactual prediction subtraction for debiasing has precedents, including “Unbiased Scene Graph Generation From Biased Training” (CVPR 2020). And ECL's training-only use is a meaningful distinction, but the paper needs evidence that its particular combination offers benefits beyond simpler controls.
  3. The contribution of the proposed objective is not adequately isolated. The paper does not report a clean comparison between $L_{BC} + μ\muL_{mask}$ and the full objective, or between uniform and frequency-dependent counterfactual weighting. Importantly, setting λ=0\lambda = 0 is not equivalent to removing LCFL_{CF}: Equation (9) becomes (1+ri)LBC+μLmask(1 + r_i)L_{BC} + \mu L_{mask}. Thus, that ablation retains task-dependent reweighting. The current results do not establish how much of the benefit comes from prediction subtraction, ordinary masked-observation training, or sample reweighting.
  4. The head-task-bias explanation remains only partially validated. The diagnostic changes both relative task frequency and tail-data coverage, as the paper acknowledges. More importantly, it does not show that ECL reduces HPS or Head-Action Rate relative to LT and masking-only training under the same diagnostic protocol.
Questions for authors
  1. What does “synthetic” in Table 1 denote?
  2. How does ECL compare with LBC+μLmaskL_{BC} + \mu L_{mask} alone and with uniform counterfactual weighting under the same training protocol?
3 comments
AuthorsOpenReview ↗

Response to Reviewer — 1pkS [1/2]

q1: What does "synthetic" denote?

a1:

We sincerely appreciate you for carefully reviewing our manuscript and providing valuable comments."Synthetic" refers to the synthetic intermediate action states arising in the Flow Matching action-generation process of π0.5\pi_{0.5}. Because these states differ from auto-aggressive states, we use this term to distinguish the two policies in their action modeling and training mechanisms. Specifically, π0.5\pi_{0.5} adopts a hybrid architecture that combines a vision-language model with an action expert. In its action generation, intermediate action states are obtained by interpolating between random Gaussian noise and the ground-truth demonstration action, and these synthetic intermediate states are used for training:

at=(1−t)ε+t agta_t = (1-t)\varepsilon + t\,a_{gt}

On this basis, our method ECL applies a unified counterfactual learning framework to both models while adapting to each model's original action mechanism: for MiniVLA, counterfactual supervision acts on its auto-regressive action prediction; for π0.5\pi_{0.5}, it acts in the action-prediction space. To avoid such misunderstanding (and to distinguish it from MIniVLA), we will remove this label from the header of Table 1 in the revised manuscript and explicitly explain the mechanism in the experimental-setup section.


q2: Counterfactual prediction subtraction has prior precedents(CVPR 2020). and further evidence is needed to demonstrate its benefits over simpler alternatives.

a2:

  1. Although the CVPR 2020 method performs counterfactual debiasing by extracting object-level features from detected bounding boxes, intervening on these features through zeroing or mean replacement, and computing the Total Direct Effect (TDE), this approach is not directly transferable to VLA tasks. Specifically, it relies on prior object localization to identify the regions for intervention. While object detection and region feature extraction are inherent components of the scene graph generation (SGG) pipeline, introducing an additional detection stage into VLA would raise two major concerns. First, it would introduce extra inference latency, which is particularly undesirable for real-time robotic manipulation, such as grasping. Second, it would introduce additional uncertainty, as the counterfactual intervention would depend on the quality of the detector. Detection errors and variations across different detection models could propagate to VLA predictions, making the experimental results less consistent and reliable. In contrast, our ECL identifies task-relevant counterfactual evidence directly from the VLA model's internal representations, leveraging language and action related attention signals without relying on external object detectors. Moreover, the counterfactual branch is introduced only during training and removed at deployment. This design avoids additional inference-time computation and dependence on external detection modules, providing a more efficient, consistent, and architecture-integrated solution for counterfactual learning in VLA.

  2. To further validate the effectiveness of our method, we implemented the CVPR 2020 approach using additional object detectors (YOLOv8 and DETR) for comparison. The experimental results are presented below:

    MethodAverage success Rate (%)Head tasks (%)Torso tasks (%)Extreme tail tasks (%)
    ECL (Ours)53.074.559.023.5
    CVPR 2020 (YOLOv8)45.365.350.318.5
    CVPR 2020 (DETR)45.469.253.710.6

q3: How does ECL compare with LBO+μLmask\mathcal{L}_{BO} + \mu \mathcal{L}_{mask} alone and with uniform counterfactual weighting under the same training protocol?

a3:

We thank the reviewer's suggestion. We have added ablation experiments in response to your comments, and the relevant content has been supplemented in the main text (see 4.3.1). We focus on analyzing the success rates across tasks for the full objective with factual-plus-masked supervision and ablation was performed for this condition.

Supervision SettingLBC\mathcal{L}_{BC}Lmask\mathcal{L}_{mask}LCF\mathcal{L}_{CF}average success rates (%)
Factual-only✓——40.8
Factual + Masked✓✓—43.2
Factual + Counterfactual✓—✓41.6
ri=1r_i=1✓✓✓45.5
ECL (Ours)✓✓✓53.0
AuthorsOpenReview ↗

Response to Reviewer — 1pkS [2/2]

q4: It does not show that ECL reduces HPS or Head-Action Rate relative to LT and masking-only training under the same diagnostic protocol.

a4:

To verify that ECL can alter the scale or bias of the original predictions, we applied the ECL method to the toy experiment described in Section 3.1. Upon re-evaluating the HPS and Head action rate, we observed a significant decrease, which indicates that our method effectively reduces the head-action bias affecting tail tasks.

γ = Nh/NtN_h/N_tHPS (LT)HPS ↓ (ECL)Head-Action Rate (LT)Head-Action Rate ↓ (ECL)
150.0% ± 0.050.0% ± 0.050.0% ± 0.050.0% ± 0.0
253.3% ± 2.051.7% ± 2.951.7% ± 2.951.7% ± 2.9
561.7% ± 3.058.3% ± 5.855.0% ± 0.053.3% ± 2.9
1070.0% ± 3.563.3% ± 2.963.3% ± 2.961.7% ± 2.9
2081.7% ± 2.470.0% ± 5.073.3% ± 2.965.0% ± 5.0

We sincerely thank the reviewer for the suggestions.

Reviewer_1pkSOpenReview ↗

the response addresses my concerns

I thank the authors for clarifying the terminology and providing additional ablation studies and bias diagnostics. These explanations and new results have alleviated my concerns, particularly the gains of full ECL over the masked-supervision baseline and the reduction in bias metrics relative to LT. I am therefore willing to raise my score to 3.

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

ECL addresses imbalanced VLA training by masking task-relevant visual tokens and supervising the difference between intact and masked predictions. Rare tasks receive stronger counterfactual supervision, while inference uses only the intact observation.

Critical strengths

s1. Task imbalance is a practical problem, and the method requires no additional demonstrations or region annotations.

s2. The paper reports improvements across two policy architectures and two real-robot platforms.

s3. The additional computation is confined to training, with no extra inference cost.

Critical weaknesses

w1. The three supervision targets can conflict. For continuous regression, if both predictions match the expert target, their weighted difference does not. The paper does not explain why this compromise improves the intact prediction used at deployment.

w2. The gains could partly come from masking augmentation or frequency-based weighting. Comparisons with simple balancing methods are missing, so the claimed reduction in head-task action bias is not fully established.

w3. Some experimental details need clarification. Table 1 labels the π0.5 results “synthetic,” and the appendix gives conflicting evaluation counts of 675 and 225 episodes per task.

Questions for authors

q1. What exactly does zz represent in each architecture? For continuous outputs, how do you reconcile supervising z+z^{+}, z−z^{-}, and z+−0.3z−z^{+}-0.3z^{-} with the same target?

q2. Could you compare the full objective with factual-plus-masked supervision alone, and report whether the subtraction changes the scale or bias of the intact predictions?

q3. What does “synthetic” mean in the π0.5 table, and which evaluation count is correct?

2 comments
AuthorsOpenReview ↗

Response to Reviewer — ujbJ [1/2]

Q1: What does zz represent in each architecture? How do you reconcile supervising z+z^+, z−z^-, and z+−0.3z−z^+ - 0.3z^- with the same target?

A1:

  1. zz denotes the prediction that each base policy uses to compute its action supervision loss. For MiniVLA (auto-regressive), zz denotes the predicted logits over discrete action tokens, which are supervised with a cross-entropy loss. For π0.5\pi_{0.5}, zz denotes the continuous velocity vector predicted during flow matching, which is supervised with a regression loss.We have added further details in the main text Section 3.2.1

  2. We thank you for pointing out the potential conflicts among these three objectives. Our goal is not to require all three outputs to independently reach the same optimal solution. Rather, our primary objective is to improve action prediction for the factual branch (z+z^+) during deployment— by helping the model establish more reliable links between task-relevant visual evidence and actions, particularly for tail tasks.

  3. More precisely, LBCL_{\mathrm{BC}} provides direct supervision for z+z^+, ensuring that the predictions used during deployment are preserved and optimized.LCFL_{\mathrm{CF}} leverages the discrepancy in predictions between the full-observation and occluded-observation scenarios to provide an additional training signal, encouraging the model to make more effective use of task-relevant visual evidence.LmaskL_{\mathrm{mask}} aligns the predictions from the occluded branch with expert actions. We do not require the masked input to recover action predictions with the same accuracy as the full input; however, without this supervision, z−z^- could degenerate into noise or approach a near-zero value. This would cause zeffectz_{\mathrm{effect}} to collapse toward z+z^+, thereby undermining the effectiveness of the counterfactual supervision.Through the interplay of these three loss functions, the model is able to maintain its action-prediction capability while increasingly basing its decisions on visual evidence relevant to the current task, ultimately enhancing the performance of z+z^+.

  4. We have included the loss curves from the training process in appendix E. As shown in the plot , the LmaskL_{\mathrm{mask}} curve does not drop to a near-zero value; instead, it stabilizes and fluctuates around a specific level.


Q2: What does “synthetic” mean and which evaluation count is correct?

A2:

  1. "Synthetic" refers to the synthetic intermediate action states arising in the Flow Matching action-generation process of π0.5\pi_{0.5}. Because these states differ from auto-aggressive states, we use this term to distinguish the two policies in their action modeling and training mechanisms. Specifically, π0.5\pi_{0.5} adopts a hybrid architecture that combines a vision-language model with an action expert. In its action generation, intermediate action states are obtained by interpolating between random Gaussian noise and the ground-truth demonstration action, and these synthetic intermediate states are used for training:

    at=(1−t)ε+t agta_t = (1-t)\varepsilon + t\,a_{gt} On this basis, our method ECL applies a unified counterfactual learning framework to both models while adapting to each model's original action mechanism: for MiniVLA, counterfactual supervision acts on its auto-regressive action prediction; for π0.5\pi_{0.5}, it acts in the action-prediction space.

    To avoid such misunderstanding (and to distinguish it from MIniVLA), we will remove this label from the header of Table 1 in the revised manuscript and explicitly explain the mechanism in the experimental-setup section.

  2. In addition, regarding the number of rollouts in Table 6, the correct configuration is Episodes/task = 675 and Episodes/seed = 225.The relevant expressions in the article have been modified.

AuthorsOpenReview ↗

Response to Reviewer — ujbJ [2/2]

q3: Compare the full objective with factual-plus-masked supervision , and report whether the subtraction changes the scale or bias ?

a3:

  1. We really thank your suggestion. We have added ablation experiments in response to your comments, and the relevant content has been supplemented in the main text (see 4.3.1). We focus on analyzing the success rates across tasks for the full objective with factual-plus-masked supervision and ablation was performed for this condition.

    Supervision SettingLBC\mathcal{L}_{BC}Lmask\mathcal{L}_{mask}LCF\mathcal{L}_{CF}average success rates (%)
    Factual-only✓——40.8
    Factual + Masked✓✓—43.2
    Factual + Counterfactual✓—✓41.6
    r<sub>i</sub>=1✓✓✓45.5
    ECL (Ours)✓✓✓53.0
  2. To verify that ECL can alter the scale or bias of the original predictions, we applied the ECL method to the toy experiment described in Section 3.1. Upon re-evaluating the HPS and Head action rate, we observed a significant decrease, which indicates that our method effectively reduces the head-action bias affecting tail tasks.

    γ = Nₕ/NₜHPS (LT)HPS ↓ (ECL)Head-Action Rate (LT)Head-Action Rate ↓ (ECL)
    150.0% ± 0.050.0% ± 0.050.0% ± 0.050.0% ± 0.0
    253.3% ± 2.051.7% ± 2.951.7% ± 2.951.7% ± 2.9
    561.7% ± 3.058.3% ± 5.855.0% ± 0.053.3% ± 2.9
    1070.0% ± 3.563.3% ± 2.963.3% ± 2.961.7% ± 2.9
    2081.7% ± 2.470.0% ± 5.073.3% ± 2.965.0% ± 5.0
  3. We additionally evaluate conventional resampling- and reweighting-based strategies under the same experimental protocol to examine whether the performance gains can be attributed to simple task balancing. Furthermore, previous experiments in APA have shown that conventional long-tail learning methods remain insufficient for VLA tasks, which further motivates our investigation of this challenging problem.

    Per-task success rates (%)

    TaskLT (Baseline)ResamplingReweightingECL (Ours)
    Task 068.4 ± 2.369.1 ± 2.364.3 ± 2.680.0 ± 2.2
    Task 149.6 ± 3.647.3 ± 3.250.4 ± 3.068.9 ± 1.3
    Task 262.2 ± 3.259.6 ± 2.859.1 ± 2.374.7 ± 1.8
    Task 355.2 ± 3.560.2 ± 2.662.2 ± 2.783.5 ± 1.8
    Task 446.7 ± 3.648.4 ± 2.842.9 ± 2.364.7 ± 1.4
    Task 540.5 ± 3.434.5 ± 3.240.1 ± 3.648.0 ± 1.3
    Task 636.7 ± 3.435.5 ± 2.430.6 ± 2.739.7 ± 1.1
    Task 710.5 ± 4.08.2 ± 1.67.7 ± 2.413.7 ± 0.9
    Task 838.1 ± 3.437.9 ± 2.230.4 ± 2.854.3 ± 1.1
    Task 90.0 ± 0.00.0 ± 0.00.0 ± 0.02.4 ± 0.7
    Average40.8 ± 3.040.1 ± 2.338.8 ± 2.453.0 ± 1.4

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear reviewers and AC,

We sincerely appreciate your valuable time and effort spent reviewing our manuscript. We propose Evidence-Guided Counterfactual Learning (ECL) for long-tailed VLA policies, which leverages task-relevant visual evidence and counterfactual supervision to mitigate head-task action bias and improve the performance of underrepresented tasks. As reviewers highlighted, our method addresses an important practical problem in robotic policy learning (1pkS, dqpt), has a coherent problem-method motivation and a simple yet effective training framework (ujbJ, 1pkS, dqpt), and achieves notable improvements across simulation benchmarks, different policy architectures, and real-world robotic platforms (ujbJ, dqpt). We sincerely appreciate your constructive feedback on our manuscript. In response to the comments, we have carefully revised and enhanced the manuscript as follows:

  • We add ablation experiments on different supervision settings, including factual-only, factual-plus-masked, factual-plus-counterfactual, uniform counterfactual weighting, and the full ECL objective, to demonstrate the contribution of each component in our proposed framework.

  • We add experiments comparing ECL with conventional long-tailed learning methods, including resampling and reweighting, to demonstrate the advantages of our counterfactual learning framework over standard task-balancing strategies.

  • We add experiments comparing ECL with a counterfactual debiasing baseline inspired by previous work on unbiased scene graph generation, implemented with YOLOv8 and DETR, to further validate the effectiveness of our proposed method.

  • We add experiments on Head Prediction Similarity (HPS) and Head-Action Rate under different task imbalance ratios, which demonstrate that ECL effectively mitigates head-task action bias, particularly for tail tasks.

Meanwhile, we give more clarifications of the algorithmic details and experiments as follows:

  • We add an explanation of the roles and interactions of factual, masked, and counterfactual supervision objectives, clarifying how they jointly improve the factual action predictions during deployment.

  • We add an explanation of the prediction representations and counterfactual learning implementation across different VLA architectures, including MiniVLA and π0.5.

  • We add an explanation of the joint instruction-related and action-related attention aggregation, visual token selection and masking procedures, and gradient propagation through the two branches.

  • We add an explanation of the Flow Matching implementation in π0.5, including the shared noise samples and timesteps between factual and masked branches.

  • We add an explanation of the meaning of "synthetic" in the original π0.5 results and clarify the definitions of different degrees of tail tasks.

  • We add an explanation of the evaluation protocol, including three random seeds and 225 episodes per seed, together with per-seed success counts and the calculation of standard deviations.

Meanwhile, we wish to highlight that the two reviewers (ujbJ, 1pkS) who gave negative evaluations (Weak Rejection) have not provided further feedback following our detailed responses during the rebuttal period. We believe that these additional experiments and clarifications have successfully addressed the major concerns raised by the reviewers and substantially strengthened the technical soundness and empirical validation of our submission. We also note that Reviewer dqpt gave a positive evaluation (Weak Acceptance) and recognized the coherent problem-method connection and broad empirical support of our approach. We kindly request that the AC give careful consideration to our rebuttal responses and the substantial improvements made to the revised manuscript.

We firmly believe that our extensive revisions and additional experiments have addressed all the concerns raised by the reviewers and further demonstrated the strong performance and practical significance of our proposed ECL framework. This work has involved substantial research efforts, computational resources, engineering costs, and real-world experimental expenses. Furthermore, we have publicly released our source code to facilitate reproducibility and sincerely welcome the research community to reproduce, evaluate, and build upon our work.

If you have any further questions or require additional clarifications, please feel free to raise them. We would be more than happy to provide further explanations and address any remaining concerns.

Thanks for your efforts very much again.

Authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.