acceptodds
Under review as a conference paper at ICLR 2027

LANCE: Learning What to See in Vision-Language-Action Policies via Language-Anchored Action Queries

Abstract

Vision-language-action (VLA) policies have advanced robotic manipulation, yet they still struggle with reliable instruction following. These policies tend to exploit vision-action correlations instead of consistently using language to identify task-relevant visual evidence for action prediction. To address this issue, we propose **LANCE** (**Language-Anchored Action Queries**), an action-query framework that explicitly couples scene-conditioned action representations to task semantics derived from language instructions. LANCE employs two sets of learnable action queries: one reads only the language tokens to construct a *language anchor*, while the other reads both language and visual tokens to form a scene-conditioned action representation. To explicitly couple the scene-conditioned representation with the language anchor, *Reciprocal Alignment* aligns their latent distributions. Crucially, alignment alone does not ensure that the language anchor captures the action distinctions between instructions. We therefore devise *PriorAction*, an auxiliary action prediction loss conditioned on the language anchor. The two objectives jointly steer visual processing to prioritize task-relevant evidence for instruction-consistent action prediction. Experiments across diverse unseen settings show that LANCE improves instruction-consistent target selection. Moreover, under zero-shot composition of learned skills, LANCE achieves higher semantic coverage and subtask completion rates.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision–Language–Action, Vision Shortcut, Vision-Language grounding

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_DYiROpenReview ↗
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
  1. Clear entry point: insufficient language conditioning is an important and under-explored gap in VLA research.
  2. Sound method design: task semantics are decoupled from scene evidence, with KL alignment and prior action supervision playing complementary roles.
  3. Sound evaluation: the real-robot evaluation separates semantic accuracy from task success rate, providing convincing evidence of enhanced language capabilities.
Critical weaknesses
  1. 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.
  2. 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.
  3. 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
  1. 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.
  2. Do the ablation results imply that some components are ineffective on their own? Please provide further analysis.
  3. How exactly are the attention visualizations computed? And beyond such relatively subjective evidence, is there any further interpretability analysis?
8 comments

We thank you for your positive assessment and helpful suggestions for improving the reproducibility and presentation of our work.

W1|Q1: Robustness under non-language perturbations.

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\pi_{0.5} 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 %):

MethodCameraRobotLanguageLightBackgroundNoiseLayoutTotal
DC (Baseline)44.962.588.991.592.654.975.271.0
LANCE (Ours)58.063.190.892.990.860.373.174.0

W2|Q2: Complementarity of the objectives and statistical reliability.

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:

  • Alignment alone (87.7%): Without PriorAction, aligning q(u∣ot,ℓ)q(u\mid o_t,\ell) to a language anchor p(u∣ℓ)p(u\mid\ell) that lacks action-discriminative information can weaken action prediction and lead to a performance drop.
  • PriorAction alone (88.7%): Without Alignment, the task semantics learned by the language anchor lack an explicit constraint to guide the scene-conditioned representation.

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:

VariantPriorActionReciprocal AlignmentSeed 42Seed 43Seed 44Mean ± Std (%) ↑
DC (Baseline)✗✗88.4--88.4
Gaussian-only✗✗89.0--89.0
w/ PriorAction only✓✗88.789.1488.9788.94 ± 0.22
w/ Alignment only (BiKL)✗✓87.787.7687.7087.72 ± 0.03
LANCE (Ours Full)✓✓90.8--90.8

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.

  1. 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: mi=∑k=1K(1D⋅HD⋅A∑d,h,aAd,h,a,kDiT)⏟Action-to-Query(1L⋅HV∑ℓ,hAℓ,h,k,iVLM)⏟Query-to-Image.\displaystyle m_i=\sum_{k=1}^{K}\underbrace{\left(\frac{1}{D\cdot H_D\cdot A}\sum_{d,h,a}A^{\mathrm{DiT}}_{d,h,a,k}\right)}_{\text{Action-to-Query}}\underbrace{\left(\frac{1}{L\cdot H_V}\sum_{\ell,h}A^{\mathrm{VLM}}_{\ell,h,k,i}\right)}_{\text{Query-to-Image}}.
    • Queries & Layers: We compose attention from the A=8A=8 decoded action tokens to the K=32K=32 vision-language queries QvlQ_{vl} using the last L=12L=12 layers (L24–35) and all D=8D=8 DiT cross-attention blocks at the final update of the 4-step sampler.
    • Aggregation & Normalization: We average all HV=32H_V=32 heads and selected layers, and all HD=12H_D=12 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.
  2. Quantitative Interpretability: Evaluating across all LIBERO suites on the Instruction object mask 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.

Reviewer_6mW5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LANCE explicitly couples scene-conditioned action representations to task semantics derived from language instructions. It improves paraphrase robustness, unseen-object performance, and real-robot target selection

Critical strengths
  1. The problem is important and supported by several evaluations: 90.8% on Plus Language and 63% on Arena unseen objects.
  2. The ablations are comprehensive.
Critical weaknesses

1.The paper should cover more results on different benchmarks, such as robotwin and robodojo. 2. The paper lack quantitative task-region grounding or localized interventions, leaving visual selection indirectly supported

Questions for authors
  1. Could you report more results on different benchmarks?
  2. Could you provide quantitative task-region grounding or localized interventions?
4 comments

We sincerely thank you for your constructive comments and valuable suggestions. Here are our responses to your comments:

W1|Q1: Additional benchmarks.

Thank you for this suggestion. To broaden our evaluation, we conducted additional experiments on RoboTwin and VLA-Arena, and will add the results to our revised paper:

  • RoboTwin: Evaluated under the Clean2Clean setting (trained on Clean dataset, evaluated on Clean test set):
MethodClean2Clean SR (%) ↑
π0\pi_{0}46.4
Qwen3VL-GR00T51.2
π0.5\pi_{0.5}60.2
Qwen3VL-PI60.5
Qwen3VL-OFT61.7
LANCE (Ours)64.6
  • VLA-Arena [1]: Evaluated under W0–W4 language-command perturbations (average SR % on StatePreservation and UnseenObjects):
Level (StatePreservation)L0L1L2Overall
W084.766.756.069.1
W187.368.752.069.3
W284.069.344.766.0
W384.063.350.065.8
W485.359.348.064.2
Level (UnseenObjects)L0L1L2Overall
W060.057.341.352.9
W162.759.340.754.2
W258.051.332.747.3
W354.045.326.742.0
W452.748.025.342.0

While existing baselines in Fig.2 of VLA-Arena suffer steep performance drops under progressive language perturbations (W0→W4), LANCE maintains strong stability across all perturbation levels.

  • RoboDojo: We recognize RoboDojo as an important benchmark for evaluating language-conditioned generalization. However, the substantial cost of full-benchmark training (~12 days on 8×A800 GPUs) exceeds our computational resources. We will complete the evaluation and release the results in the revised paper.

W2|Q2: Quantitative task-region grounding and localized interventions.

Thank you for your valuable suggestion. Following the diagnostic protocol of VLA-Trace [2], we quantitatively evaluate task region grounding across all 40 tasks from four LIBERO suites with 10 fixed initial states per task under clean, M1 (manipulated-object masking), and M2 (goal-receptacle masking) conditions using Attention Mass, IoU@90, and Hit.

Here, the target region is defined as the Instruction Object to measure attention. The results are summarized below (%, higher is better; each entry reports Clean / M1 / M2):

ModelAttention Mass ↑IoU@90 ↑Hit ↑
GR00T N1.77.26 / 7.28 / 7.262.62 / 2.68 / 2.946.75 / 6.00 / 7.00
π0.5\pi_{0.5}7.23 / 6.81 / 7.225.09 / 4.95 / 5.053.25 / 2.75 / 2.00
LangForce1.04 / 1.01 / 2.707.01 / 6.36 / 5.850.00 / 0.00 / 1.25
APT7.63 / 7.65 / 7.275.25 / 5.98 / 4.685.75 / 4.25 / 3.50
DC (Baseline)2.59 / 2.79 / 2.4915.20 / 14.40 / 11.060.75 / 1.25 / 1.25
DC + Align.16.87 / 15.87 / 14.1419.32 / 17.46 / 14.7836.75 / 30.50 / 27.25
DC + Align. + direct zlz_l5.27 / 5.32 / 3.8414.42 / 14.26 / 10.214.00 / 5.25 / 0.75
LANCE (Ours)19.58 / 18.47 / 15.3719.83 / 18.75 / 15.5841.00 / 37.00 / 25.25

LANCE achieves the highest Attention Mass and IoU@90 across all three conditions, providing direct quantitative evidence of task-relevant attention under localized interventions. Full protocols are added to the Appendix C.5.

[1] VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models.

[2] VLA-Trace: Diagnosing Vision-Language-Action Models Through Representation and Behavior Tracing.

Dear Reviewer 6mW5,

Have our rebuttal and additional experiments resolved your concerns? If you have any further questions, please let us know and we would be happy to discuss them.

Otherwise, we respectfully request you to consider raising your score to reflect the contributions of our work.

Best,

Authors

Thanks for the author's response.The performance improvement on robotwin clean2clean is too minor to demonstrate the effectness of the method, therefore I will keep my score.

Thank you for your reply. We would like to clarify the role of RoboTwin Clean2Clean and the performance gains validating our method. RoboTwin Clean2Clean evaluates general bimanual manipulation under fixed instructions, rather than testing instruction following generalization under language variations. Nevertheless, LANCE still improves overall performance to 64.6% (surpassing Qwen3VL-GR00T by +13.4%, pretrained π0.5\pi_{0.5} by +4.4%, and Qwen3VL-OFT by +2.9%).

On language variation benchmarks, LANCE delivers larger gains: it outperforms the strongest baseline by +16.9% on VLA-Arena UnseenObjects (Figure R1), and ranks 1st on LIBERO-Plus Language (90.8%)

We fully respect your assessment and sincerely hope that these clarifications and experimental evidence address your concerns.

Reviewer_TtJKOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes LANCE, which constructs scene-independent language anchors and scene- conditioned action representations respectively through two sets of action queries, language only and visual language. PriorAction shapes language anchors with auxiliary action predictions, and reciprocal alignment aligns two latent distributions through bidirectional KL divergence, guiding the model to utilize visual information related to instructions. The authors evaluate the method on multiple simulation benchmarks and real robots, and analyze its effect in combination with module ablation, action separability, and attention maps. The results show that LANCE achieves better performance in zero-shot combination tasks of language generalization, target selection, and learned skills.

Critical strengths
  1. This article focuses on the problem that VLA models over-rely on visual shortcuts and fail to fully utilize language instructions. This directly affects the robot's generalization ability under visual interference, unseen objects and combined instructions, and has clear research value.

  2. The experiments cover multiple simulation benchmarks and real robots, and are implemented through the ablation analysis module. LANCE performs well in language rewriting and unseen object settings

Critical weaknesses
  1. The conceptual novelty of LANCE remains limited. LangForce already explores language-dependent action representations through differently conditioned distributions, while LA4VLA investigates vision free text action supervision. LANCE combines a scene independent language anchor, auxiliary action supervision, and bidirectional KL alignment. The paper does not clearly establish what fundamentally new mechanism this combination introduces, beyond changing the conditioning structure and alignment objective. The reported improvements demonstrate empirical effectiveness but do not sufficiently substantiate the claimed methodological novelty.

  2. Real robot comparisons are not yet able to fully isolate the contribution of the proposed mechanism. LANCE uses Qwen3-VL-4B with DiT-B, and the main experiment only compares with π0.5 and GR00T-N1.7. Different backbones and pretraining sources affect language understanding and action prediction. The paper lacks quantitative comparison of real robot with the same backbone variant. It is recommended to report real-world ablation studies on the same vlm backbone to rule out the impact of model configuration differences.

  3. The semantic justification of PriorAction remains insufficient. The same language instruction and proprioceptive state can require different actions under different scene configurations, yet PriorAction supervises a language only branch using ground-truth actions. The language anchor may therefore capture dataset-specific action patterns rather than task-level action semantics. The reported action space separability analysis does not rule out this alternative explanation or establish that the learned distinctions reflect task level semantics. Moreover, the specific advantages of PriorAction over LA4VLA's text action pretraining remain unclear.

  4. Table 8 reports 63% average success, inconsistent with the 64% arithmetic mean of L0 to L2. The comparison also omits MolmoAct2 [1] (92.67%) and ACoT-VLA [2] (64.45%) from the official leaderboard, although protocol differences need to be considered. Furthermore, only UnseenObjects is evaluated, without the language command and visual observation diagnostics that directly relate to the paper's claims. The inconsistent performance under visual perturbations in Appendix Table 7 also remains unexplained.

[1] MolmoAct2: Action Reasoning Models for Real-world Deployment(corl 2026)

[2] ACoT-VLA: Action Chain-of-Thought for Vision-Language-Action Models (cvpr 2026)

Questions for authors
  1. How does PriorAction resolve conflicting action supervision without visual input? For example, the same instruction, 'Pick up the black pen', may require moving left or right depending on the pen's position. How can the language anchor capture task-level action semantics rather than dataset-specific action patterns under such ambiguity?

  2. Figure 5(c) shows improved action space separability, but does separating predicted actions across instructions necessarily imply better language grounding, rather than stronger instruction-conditioned action biases?

  3. In Figure 7(c), blue marker appears in both training and unseen testing groups. Are they the same instance, different instances of the same model, or have the illustrations been misplaced?

5 comments

W1: Contribution and methodological novelty.

Our core contribution is introducing an active language-guided perception and action mechanism: it actively directs visual attention toward instruction-relevant regions and simultaneously shapes action representations to preserve instruction-consistent task semantics. While prior instruction following methods rarely address both, their visual attention often fails to attend to the correct objects relevant to the instruction (Fig. 1(a)). As shown in Table 5 and Fig. 1(a), our method achieves the highest IoU@90 on target regions (19.83%), thereby enabling the highest success rate on LIBERO-Plus Language in Table 3.

Essential Design Differences from Prior Work:

  • vs. LangForce: LangForce introduces action queries that attend solely to vision as a vision-only prior and forces this prior to predict the language. In contrast, we use a language-only anchor to directly constrain the vision-language query by distribution alignment. Moreover, LangForce requires two separate forward passes to optimize its objective, whereas our method only needs a single forward pass to optimize the model. (Note: LangForce's vision-only prior is only dependent on the vision inputs instead of language inputs).
  • vs. LA4VLA: LA4VLA requires costly video segmentation and dense atomic sub-action instruction annotations. In contrast, our method only requires a single global task instruction without expensive annotations.

Beyond Naive Additive Combination: Rather than an additive combination of independent tricks, our Distribution Modeling, Reciprocal Alignment, and PriorAction must work together to simultaneously preserve linguistic semantics (cos⁡(zl,L)=0.583\cos(z_l, L)=0.583) and focus visual attention on instruction-relevant regions (41.00% Hit, 19.58% Attention Mass). Simply combining only auxiliary action supervision and KL alignment (Align. + direct zlz_l) fails to achieve this: as shown in Fig. 5(b) and Table 5, it causes language forgetting (cos⁡(zl,L)\cos(z_l, L) drops to 0.064) and results in poor visual grounding (only 4.00% Hit, 5.27% Attention Mass).

W2: Fairness of real-robot comparisons.

To ensure strictly controlled evaluations, our real-robot experiments are directly conducted using the π0.5\pi_{0.5} vlm backbone. We will add the implementation details to the revised paper. We additionally evaluate on the GR00T backbone below:

  • Single-task Real-robot Evaluation
DimensionSettingTrialsGR00TLANCE
Cross-conditionAttribute-Near2085/40100/75
Attribute-Far2075/2595/55
State-Lying2085/6095/65
State-Standing20100/55100/65
Visual distractorSame Category4090/5597.5/65
Stacking/Clutter2075/3580/60
Partial Occlusion1090/4070/60
Unseen conditionsUnseen Object1060/3080/50
Unseen Container1080/5080/50
Unseen Background1080/8080/70
Unseen Instruction1070/7080/80
  • Long-horizon Real-robot Evaluation
TaskHModelSem. Cov.Task Cov.Ord. Prog.Full SR
Multi-target composition
Multiple Black Pens3GR00T13/152/15—0/5
LANCE14/157/15—1/5
Multiple State-conditioned Objects3GR00T14/159/15—2/5
LANCE13/157/15—1/5
Composite-instruction composition
Black Pen → Standing Sharpener → Tape3GR00T5/152/150/150/5
LANCE11/159/155/151/5
Standing Sharpener → Lying Sharpener → Lying Eraser → Standing Eraser4GR00T5/204/204/200/5
LANCE14/2010/206/200/5

Across these pretrained backbones, our method consistently improves instruction-consistent target selection and zero-shot skill composition.Regarding Qwen3-VL-4B with DiT-B, it lacks real-robot pretraining and fails on basic manipulation tasks, a foundational capability deficit that our method cannot compensate for.

W3|Q1,Q2: PriorAction under ambiguous supervision, task-level semantics, and visual grounding.

Thank you for your deeply insightful comments. Here we clarify the motivation and advantages of PriorAction:

1. How PriorAction Handles Conflicts (W3 & Q1): In our framework, we do not expect the language-only branch to accurately predict the action trajectory alone. Instead, we use a small loss weight (λPA=0.1\lambda_{PA}=0.1) to build a language anchor that encodes full instruction information while learning discriminative action semantics across instructions. By aligning the main branch q(u∣ot,ℓ)q(u\mid o_t,\ell) with this language anchor, PriorAction helps the main branch better capture instruction-relevant visual evidence (Table 5, +2.71% Attention Mass and +4.25% Hit) and enhance action-space discriminability under the same instruction across different scenes (as shown below).

We verify that adding PriorAction enhances scene-adaptive action-space separability (SepR, defined in Appendix C.2) at two levels:

  • Intra-Task Action Discriminability: Under the same instruction across different scenes, adding PriorAction increases Intra-Task SepR from 3.078 to 3.348 (+8.8%) compared to Alignment only, showing stronger scene-adaptive action discriminability.
  • Inter-Task Action Discriminability: Across all 40 tasks, adding PriorAction increases Inter-Task SepR from 39.569 to 43.094 (+8.9%). It also strengthens the separability of predicted actions across varied visual scene configurations in Fig. 5(c) (+7.2 on Paired, +16.1 on Removed).

2. Action space separability (Q2): We agree that action-space separability (SepR, Fig. 5(c)) does not necessarily imply better language grounding. We use it to show that our method enhances representation discriminability and action-space diversity across instructions. Meanwhile, our results show that SepR and language grounding improve together in our method: in Table 5, adding PriorAction increases target region Attention Mass from 16.87% to 19.58% and Hit from 36.75% to 41.00%.

3. Advantages over LA4VLA: LA4VLA relies on text-action pretraining with costly video segmentation and dense atomic sub-action instruction annotations, where fine-grained sub-actions across time are also ambiguous to define. In contrast, PriorAction only requires a single global task instruction without any extra segmentation or annotation cost. This supervision preserves linguistic semantics (cos⁡(zl,L)\cos(z_l, L) improves from 0.064 to 0.583).

W4: VLA-Arena comparisons, numerical consistency, and visual robustness.

Thank you for your careful reading and constructive suggestions. We address each point below:

  1. Table 8: We have corrected the arithmetic mean typo to 64% in the revised Table 8.

  2. MolmoAct2 and ACoT-VLA comparisons: Thank you for pointing this out. These two methods were recently added to the leaderboard (in the August update), and we will add both to Table 8 and discuss them in our revised paper. Compared to ACoT-VLA, our method achieves comparable overall performance (64.0%) and +3% higher success on the more challenging L2 generalization split. Meanwhile, MolmoAct2 uses a stronger foundation model with extensive pretraining on embodied reasoning and multi-robot data (as well as different fine-tuning/evaluation protocols), whereas our model is trained strictly on VLA-Arena's L0 data without action pretraining.

  3. Language and visual diagnostics: We additionally evaluated our method on StatePreservation and UnseenObjects under Language Command (W0–W4) and Visual Observation (V0–V4) perturbations:

    • StatePreservation & UnseenObjects (Success Rate %)
    Level (StatePreservation)L0L1L2OverallLevel (StatePreservation)L0L1L2Overall
    W084.766.756.069.1V084.764.048.065.6
    W187.368.752.069.3V182.045.316.047.8
    W284.069.344.766.0V278.720.020.039.6
    W384.063.350.065.8V367.324.012.734.7
    W485.359.348.064.2V42.00.00.70.9
    Level (UnseenObjects)L0L1L2OverallLevel (UnseenObjects)L0L1L2Overall
    W060.057.341.352.9V063.357.349.356.7
    W162.759.340.754.2V166.770.045.360.7
    W258.051.332.747.3V237.349.330.038.9
    W354.045.326.742.0V342.746.018.035.6
    W452.748.025.342.0V42.02.00.01.3

    Compared to the baseline curves in VLA-Arena (Fig. 2), where existing models suffer drops on UnseenObjects across W0→W4 and V0→V3, our method maintains a more resilient curve under both language rewritings (retaining 42.0% overall at W4) and visual shifts (35.6% overall at V3, whereas V4 stacks severe sensor noise, causing both existing baselines and our model to drop near zero). Since VLA-Arena only plots curves without releasing exact tables, we will re-evaluate the baselines and include the full comparison in our revised paper.

  4. Performance explanation: Compared to baseline DC, our method improves overall success rate by +3.0%. Beyond achieving rank 1 on Language, our method consistently ranks in the top-4 across Camera, Robot subsets among all compared methods.

    Results on LIBERO-Plus (same-backbone comparison, SR in %):

    MethodCameraRobotLanguageLightBackgroundNoiseLayoutTotal
    DC (Baseline)44.962.588.991.592.654.975.271.0
    LANCE (Ours)58.063.190.892.990.860.373.174.0

Q3: Figure 7(c).

We apologize for the confusion. The illustration was misplaced; the blue marker belongs to the unseen test set. We have revised the figure.

Comments on the paper

2 comments

We deeply appreciate the time and effort you have spent reviewing our paper. Please note that all revised parts are highlighted in blue in the updated paper.

Dear Area Chair and Reviewers,

Thank you for the time and effort spent reviewing our paper. We have thoroughly considered all the comments to improve our work, with all revisions highlighted in blue in the updated paper.

We are thankful to Reviewer DYiR for the positive feedback and for recognizing the potential of our research. We have also diligently addressed the issues raised by Reviewer TtJK and Reviewer 6mW5 and we hope our clarifications and additional analyses have resolved their concerns.

The constructive suggestions from all reviewers have undoubtedly enriched our paper, and we are truly appreciative of this collaborative review process. Once again, thank you all for your dedication!

Warm regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.