- Rating
- 4(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
VLA policies under-utilize language and rely on vision-action shortcuts. LANCE addresses this with dual-context action queries, where a language-anchored distribution is aligned with a scene-conditioned one via reciprocal KL, together with an auxiliary prior action loss. This yields consistent gains in instruction following and language generalization in simulation and on real robots.
Critical strengths
- Clear entry point: insufficient language conditioning is an important and under-explored gap in VLA research.
- Sound method design: task semantics are decoupled from scene evidence, with KL alignment and prior action supervision playing complementary roles.
- Sound evaluation: the real-robot evaluation separates semantic accuracy from task success rate, providing convincing evidence of enhanced language capabilities.
Critical weaknesses
- Robustness trade-off: Appendix Table 7 shows LANCE peaks only on Language (90.8%) while lagging behind π0.5 on Camera, Robot, Noise, and Layout (74.0% vs 85.7% on average). The source of this trade-off and possible mitigations should be discussed.
- Inconclusive ablations: in Table 4, PriorAction alone (88.7%) and Reciprocal Alignment alone (87.7%) underperform the Gaussian-only variant (89.0%); only their combination helps. With gaps of 1–2 points and no multi-seed variance reported, the differences may be within noise.
- Opaque attention visualizations: Figures 1(a) and 4(b) are central to the motivation and qualitative evidence, but the computation (layers, queries, aggregation, normalization) is unexplained, and attention maps across different architectures are not directly comparable.
Questions for authors
- Is the performance drop under non-language perturbations inherent to language anchoring or an artifact of the training setup? If the former, please discuss and characterize the limit; if the latter, please provide further experimental analysis.
- Do the ablation results imply that some components are ineffective on their own? Please provide further analysis.
- How exactly are the attention visualizations computed? And beyond such relatively subjective evidence, is there any further interpretability analysis?
8 comments
W3|Q3: Attention-map computation and interpretability.
Thank you for pointing this out. We have clarified our attention protocol in Appendix C.5 and added the quantitative grounding analysis in Table 5.
- Attention Computation Protocol: We visualize action-pathway attention to image tokens by composing action-to-query and query-to-image attention via a two-hop formulation:
- Queries & Layers: We compose attention from the decoded action tokens to the vision-language queries using the last layers (L24–35) and all DiT cross-attention blocks at the final update of the 4-step sampler.
- Aggregation & Normalization: We average all heads and selected layers, and all DiT heads, blocks, and action tokens equally. Maps are computed via native softmax, max-normalized for visualization, and mass-normalized per view for evaluation.
- Cross-Architecture Comparability: For baselines with direct visual access, we extract action-to-image attention from the last third of their image-reading layers at the final denoising forward, evaluating all models on matched observations with scale-invariant metrics.
- Quantitative Interpretability: Evaluating across all LIBERO suites on the
Instruction objectmask under Clean, M1, and M2 interventions (see table in Response to Reviewer 6mW5, W2), our method achieves the highest Attention Mass (19.58%), IoU@90 (19.83%), and Hit (41.00%), far surpassing GR00T N1.7 (6.75% Hit) and APT (5.75% Hit).
I thank the authors for their detailed response.
W3 is withdrawn. The attention protocol in C.5 is now sufficiently clear (two-hop composition, L24–35, final denoising step, softmax over the full key sequence without intermediate renormalization), and Table 5 supplies the quantitative evidence. The table could still be improved: under the strong M1/M2 interventions, GR00T's Attention Mass is 7.26 / 7.28 / 7.26 and π0.5's is 7.23 / 6.81 / 7.22, i.e. essentially unchanged, which suggests the diagnostic is not sensitive to the intervention; and DC's IoU@90 (15.20, second best overall) is inconsistent with its Peak-Hit (0.75, second worst), indicating that argmax over an 8×8 grid with first-index tie-breaking is unstable.
W1 is partially withdrawn. I accept that within the same backbone and the same training data LANCE outperforms DC on the perturbed axes (Camera 58.0 vs 44.9), and I therefore withdraw the causal claim that language anchoring degrades visual robustness. Two points are still worth correcting: (a) the "top-4" ranking covers only Camera and Robot, whereas Noise (60.3) and Layout (73.1) rank 11th and 10th, which should be stated alongside it; (b) if the gap to π0.5 is explained by action pretraining, then Qwen3-VL-OFT, which has no action pretraining and still reaches a Total of 75.0 against LANCE's 74.0, is inconsistent with that explanation and deserves a sentence.
One new reservation. The paper attributes the gain to language anchoring, but the anchor query bank is never allowed to read visual tokens, so the "language" factor itself is not ablated. In Table 6, replacing KL with MSE still yields 89.7 (only 1.1 below the full model), which suggests the probabilistic modeling contributes little and that the remaining gain may come from the regularization effect of a second, restricted query bank plus an auxiliary action loss rather than from the language constraint. A single control would settle this: let the anchor query bank also receive visual tokens, with everything else unchanged. If performance is unchanged, the mechanistic account needs rewriting; if it drops substantially, the anchoring hypothesis is supported.
We thank you for your positive assessment and helpful suggestions for improving the reproducibility and presentation of our work.
Thank you for pointing this out. Our method improves overall success rate by +3.0% over the DC baseline, outperforming DC on 5 out of 7 perturbation axes. Beyond achieving rank 1 on Language, our method consistently ranks in the top-4 across Camera, Robot subsets among all compared methods. The few models outperforming ours on specific visual subsets (π0.5 and ABot-M0) are backed by large-scale multi-robot action pretraining, whereas our method is built on standard Qwen3-VL-4B without action pretraining.
Results on LIBERO-Plus (same-backbone comparison, SR in %):
Thank you for this comment. All ablations in Table 4 were conducted using the same fixed random seed (42) for controlled comparison. Because PriorAction and Reciprocal Alignment are complementary, using either alone is suboptimal:
Combining both (90.8%) shapes an action-discriminative anchor to guide scene-conditioned learning. Following your suggestion, we repeated the ablations across two additional seeds (Seeds 43 and 44) to confirm statistical stability: