- 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 . Counterfactual supervision is weighted by task rarity. Experiments cover LIBERO-Core-LT with MiniVLA and , 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 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 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 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
-
What are and for each backbone, and how is implemented? Please specify masking location, attention aggregation, gradient routing, and whether noise and timesteps are shared when contrasting flow predictions.
-
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.
-
Please clarify the result reporting. What does “synthetic” mean in the 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
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.
-
"synthetic" refers to the intermediate action states synthesized during the Flow Matching training process of . We originally used this term to distinguish the action modeling and training mechanisms of and MiniVLA. Specifically, 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:
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.
-
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.
-
Let denote the number of successful episodes obtained by seed on a given task, out of the closed-loop episodes evaluated for that seed. Each task is thus evaluated with episodes per method.
The per-seed success rate is , 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}.$$ -
Table R1 — Per-seed success counts.Each cell reports , the number of successful episodes out of 225 for the three seeds.
MiniVLA
Task FULL LT (Baseline) APA ECL (Ours) Task 0 118/128/136 148/157/157 159/162/169 175/180/185 Task 1 167/176/185 103/113/119 149/150/155 152/155/158 Task 2 154/167/170 132/142/146 144/151/151 164/168/172 Task 3 114/122/131 118/122/133 152/159/161 184/188/192 Task 4 142/146/160 96/107/112 123/128/130 142/147/148 Task 5 121/135/139 83/92/98 96/101/103 105/108/111 Task 6 114/123/132 76/81/91 79/86/86 87/89/92 Task 7 98/112/116 15/23/33 23/27/30 29/30/33 Task 8 116/132/133 79/84/94 44/47/50 120/122/125 Task 9 50/59/68 0/0/0 1/3/5 4/5/7 Task FULL LT (Baseline) APA ECL (Ours) Task 0 122/128/133 155/164/170 153/156/160 164/171/175 Task 1 132/136/147 141/146/148 143/150/151 157/162/165 Task 2 157/166/167 126/134/136 126/132/132 148/149/151 Task 3 109/114/116 114/116/125 120/126/128 133/134/135 Task 4 149/154/158 99/105/105 89/99/104 109/119/124 Task 5 128/131/138 77/81/89 88/89/95 94/102/104 Task 6 116/123/125 58/70/73 59/63/66 79/82/92 Task 7 101/103/110 50/53/53 57/62/67 66/70/71 Task 8 125/134/134 26/38/39 42/43/46 45/54/55 Task 9 54/59/64 8/10/11 21/23/24 16/28/32 -
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).
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 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.
Response to Reviewer — dqpt [1/2]
q1: What are z+ and z− for each backbone, and how is LCF 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.
Prediction definitions. For MiniVLA, z+ and z− denote the action-token logits predicted from intact and masked visual observations, respectively. For π0.5, they denote the corresponding velocity predictions in Flow Matching. Both branches share the same policy parameters.
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 and αjA, respectively. Their joint relevance is computed as rj=αjLαjA, 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−. The attention-based selection and masking decisions do not receive gradients.
Counterfactual loss and gradient routing. We construct the contrastive prediction as zeffect=z+−λz−, with λ=0.3, and define LCF=ℓ(zeffect,a∗), where ℓ is the backbone-specific action supervision loss (cross-entropy for MiniVLA and Flow Matching regression for π0.5). The overall objective is L=LBC+μLmask+riLCF, where ri adjusts counterfactual supervision according to task frequency. LBC and Lmask supervise the intact and masked branches, respectively, while LCF backpropagates through both predictions without stop-gradient, jointly optimizing the shared policy.
Shared Flow Matching conditions. For π0.5, 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:
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.
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.
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.