- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper explores whether VLA models can follow negated instructions, which previous VLAs often fail to do. To address this challenge, the paper proposes NegaAlign. NegaAlign is a parameter-efficient post-training approach for VLA backbones and requires no negated robot demonstrations. NegaAlign leverages existing affirmative demonstration data, combined with manually constructed relations that pair each negated instruction with satisfying and excluded affirmative instructions. NegaAlign incorporates gated Negation Transformation Layers into the VLM backbone, which are designed to adjust the representations of negative instructions to align with their intended affirmative counterparts. Teacher-Guided Visual Alignment extends this adjustment to the visual tokens associated with the instructions. The paper also introduces NegaBench, a simulation benchmark covering five types of negation across 10 distinct scenarios. NegaAlign shows consistent gains on negated instructions compared to pi0.5 on both simulated and real-world datasets.
Critical strengths
- This paper identifies an important problem that existing VLAs often fail to follow negated instructions. To address this problem, it proposes a parameter-efficient post-training strategy NegaAlign which can be easily adapted to various VLA backbones.
- Simulation and real-world experiments demonstrate that NegaAlign significantly improves negated-instruction following while maintaining affirmative performance. Ablation studies further support design choices.
Critical weaknesses
- The generalization of the proposed method is unclear, as the negated instructions used for training appear to be the same as those used for evaluation. Therefore, it is difficult to tell whether NegaAlign learns general negation understanding or memorizes known prompts.
- NegaAlign's performance nearly matches full finefuning on pi0.5. However, it lags clearly behind on GR00T and pi0, which suggests that its effectiveness may depend on the backbone. Evaluation on more backbones is necessary to verify its general applicability.
Questions for authors
Please refer to the weaknesses.
4 comments
Response to Reviewer 2Wiz (2/2)
W2: Applicability across VLA backbones
A: We extend the evaluation to DM0.5[1]. On DM0.5, negated success increases from 4.90% to 53.40%, a gain of 48.50 percentage points. Affirmative success remains close to the affirmative-only baseline, at 52.10% versus 53.81%, a difference of 1.71 percentage points.
| Model | Object Pos. | Object Neg. | Attribute Pos. | Attribute Neg. | Comparison Pos. | Comparison Neg. | Goal Pos. | Goal Neg. | Complex Pos. | Complex Neg. | Avg. Pos. | Avg. Neg. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DM0.5 Pos. FT | 69.92 | 10.50 | 28.75 | 7.00 | 62.50 | 6.50 | 26.50 | 0.00 | 81.38 | 0.50 | 53.81 | 4.90 |
| DM0.5 Pos. & Neg. FT | 73.25 | 64.25 | 33.00 | 39.00 | 58.25 | 54.00 | 42.00 | 46.00 | 81.25 | 83.50 | 57.55 | 57.35 |
| DM0.5 w/Ours | 69.00 | 68.00 | 30.75 | 40.00 | 60.25 | 45.00 | 26.75 | 32.00 | 73.75 | 82.00 | 52.10 | 53.40 |
The averages cover all five task categories. Fine-tuning with both affirmative and negated action trajectories achieves 57.35% negated success. NegaAlign reaches 53.40% without negated action targets, leaving a gap of 3.95 percentage points. This result complements the gains on GR00T, π₀, and π₀.₅, where NegaAlign increases negated success from 1.05% to 57.75%, from 15.40% to 63.30%, and from 2.60% to 88.45%, respectively. Across the four evaluated backbones, NegaAlign consistently improves negated instruction following. Its gap to trajectory-supervised fine-tuning varies across backbones: 11.10 percentage points on GR00T, 7.95 on π₀, 0.15 on π₀.₅, and 3.95 on DM0.5. These comparisons support applicability across multiple backbones and quantify how closely the adaptation approaches the trajectory-supervised reference on each model.
[1] DM0.5: Designed for the Open World, Where Generalization Emerges. Dexmal, Technical Report, 2026.
Thank you to the author for replying to my comments. It has resolved some of my questions. I maintain my original positive score.
Dear Reviewer 2Wiz,
Thank you for your valuable suggestions, which have helped us improve our work. We sincerely appreciate your time and thoughtful feedback.
Best regards,
The Authors
Response to Reviewer 2Wiz (1/2)
Thank you for highlighting generalization and performance variation across VLA backbones.
W1 Generalization of purposed method
A: We further evaluate whether NegaAlign generalizes to OOD scenarios with changes in object instances, appearances, and scene configurations. Specifically, we directly use the checkpoint obtained after adaptation stage, without additional training on the OOD scenarios. The instruction text remains identical to that used during adaptation, while object positions, colors, sizes, and scene arrangements change, as Figure 6 illustrates in revised manuscript.
Table 3 Success rates (%) on OOD tasks with negation instructions. Bold marks the best SR in each column.
As Table 3 shows, NegaAlign achieves an average success rate of 77.8% in negation instruction tasks, compared with 67.5% for Qwen–π₀.₅-ft, an improvement of 10.3 percentage points. Qwen–π₀.₅-ft is composed with a finetuned Qwen3-VL-2B-Instruct using our negation supervision constructed in Sec 3.2 and the frozen pretrained π₀.₅. NegaAlign achieves higher success rates on eight of the ten tasks, demonstrating stronger visual generalization across scenes, object categories, and appearances. Compared with the VLM-based instruction-rewriting baseline, which depends heavily on the VLM’s visual perception and reasoning capabilities, NegaAlign follows negation instructions more reliably across visual variations.
The question of unseen negation relations concerns an additional dimension of generalization. When the negation construction and the underlying manipulation skills are individually familiar, applying them in an unseen combination tests compositional generalization. However, the compositional generalization of physical intelligence is still an unsolved problem[1]. Therefore, our setting studies how an affirmative-trained VLA acquires negated instruction following through lightweight adaptation. NegaAlign connects the supplied negation relations to existing manipulation capabilities through image–language supervision. The low negated success rate of the unadapted backbones motivates this adaptation, and the additional OOD results establish transfer of the resulting behavior across visual conditions. Please check more details in our revision.
[1] Ai B, Amin A, et al. Intelligence, Physical, et al. "${\pi} _ {0.7} $: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities." arXiv preprint arXiv:2604.15483 (2026).