acceptodds
Under review as a conference paper at ICLR 2027

LLaDA-Nav: Diffusion-Based Action Generation for Visual Navigation

Abstract

Vision-language models (VLMs) provide broad visual-semantic knowledge and promising generalization for embodied AI. To produce navigation actions, autoregressive VLM policies generate action chunks from left to right, conditioning each action on earlier predictions from the same chunk. In our evaluation, however, these policies show limited closed-loop performance in visual navigation. To test whether this feedback pathway contributes to that limitation, we conduct controlled comparisons and prefix interventions that replace model-generated action prefixes with ground-truth ones. Our results suggest that early prediction errors can propagate to later actions, especially under distribution shift. Motivated by this finding, we propose LLaDA-Nav, a visual navigation policy built on a masked diffusion VLM. LLaDA-Nav fully masks the target action block during training and predicts the entire chunk in a single forward pass, eliminating within-chunk target-action reuse. It incorporates previously executed actions as additional context to capture recent robot motion. Across four navigation datasets and unseen HM3D environments, LLaDA-Nav improves open-loop trajectory prediction and achieves 59.03% closed-loop success, exceeding the strongest evaluated baseline by 15.28 percentage points.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Diffusion Language Model, Image-Goal Navigation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qWNLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes LLaDA-Nav, a diffusion-based visual navigation framework that uses full-mask parallel action decoding to avoid error propagation from previously generated action tokens. It further introduces Action Context Conditioning (ACC) to incorporate historical action information. Experiments demonstrate improvements in both open-loop trajectory prediction and closed-loop navigation performance.

Critical strengths
  1. The experiments cover both open-loop and closed-loop evaluations, providing a relatively comprehensive assessment of the proposed framework.
  2. The figures are clear and well designed, making the proposed method easy to understand.
Critical weaknesses
  1. The paper attempts to mitigate the exposure bias caused by teacher forcing in AR models through diffusion-based full-mask parallel action decoding. However, the newly introduced ACC module may also cause a training–inference distribution mismatch, as it uses ground-truth action histories from the dataset during training but policy-executed action histories during inference. This design seems somewhat contradictory to the original motivation. Moreover, the main performance improvement appears to come from the addition of ACC (Table 3, rows 2 and 4), yet the AR baselines are not clearly evaluated with the same ACC module (or equivalent historical action information) for a fair comparison. Table A6 further raises this concern: with shorter ACC histories, the benefits of additional motion information may dominate, whereas with longer histories, the negative effects of training–inference distribution mismatch may become more pronounced, potentially explaining the substantial performance degradation. However, the paper does not further investigate this possibility.

  2. The results in Table 4 are also somewhat confusing. Iterative Decoding (T=5) reuses previously predicted action tokens during denoising, potentially allowing errors in early predictions to mislead the remaining masked tokens. However, Table 4 shows that Iterative Decoding (T=5) achieves nearly the same SR as Single-pass Decoding (T=1), with even slightly higher SPL. This leaves both the core motivation of improving closed-loop performance by eliminating generated-token feedback and the source of the proposed method's performance gains insufficiently explained.

Questions for authors
  1. Does ACC introduce a training–inference distribution mismatch similar to the exposure bias that the paper aims to address? How might this affect closed-loop performance?

  2. Given the substantial gains from ACC in Table 3, could the authors provide a comparison where the AR baselines are equipped with the same action-history information to better isolate the benefits of prefix-free decoding?

  3. Could the sharp performance degradation with longer ACC histories (k = 5) in Table A6 be related to an increased training–inference mismatch, rather than simply redundant historical information?

  4. How do the authors explain the similar SR and slightly higher SPL of iterative decoding (T = 5) compared to single-pass decoding (T = 1) in Table 4, despite the reuse of predicted action tokens?

  5. Do these results suggest that the performance gains mainly come from full-mask training rather than eliminating generated-token feedback during inference? Could the authors clarify the relative contributions of these two factors?

6 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

W1-1 and Q1. Does ACC introduce a mismatch bias?

A1. The key distinction lies in how previous actions are used. In AR decoding, unexecuted predictions are reused to generate subsequent actions within the same chunk, allowing prediction errors to propagate before execution. In contrast, ACC conditions each new action chunk on executed motions and updated observations. Even if a previous action was suboptimal, its execution history still provides valid information about how the robot actually moved, helping maintain motion consistency without reintroducing within-chunk prediction errors.

We acknowledge that policy-executed histories may differ from expert histories. To quantify this distribution shift, we compared LLaDA-Nav's executed motions with the training data (Table R1).

Table R1. Distribution differences between policy-executed motions and training action histories.

Metricdxdydyaw
JS divergence0.5820.1930.207
KL divergence2.610.720.83
Wasserstein-10.550.106.2°

Despite this substantial distribution shift, ACC improves closed-loop SR from 44.44% to 59.03% and reduces RPE from 0.49 to 0.34 (Table 3). These results show that ACC remains beneficial under the observed training–inference mismatch.

W1-2 and Q2. AR baseline with ACC.

A2. To answer this question, we evaluated Qwen3-VL-8B with and without ACC and compared standard AR fine-tuning against PAM under the same ACC setting.

Table R2. ACC ablation with AR baseline.

BackboneTrainingACCSR (%)SPL (%)
Qwen3-VL-8B (AR)Standard ARNo52.0848.95
Qwen3-VL-8B (AR)Standard ARk=344.4437.90
Qwen3-VL-8B (AR)PAMk=358.3351.28

As shown in Table R2, adding ACC alone to standard AR fine-tuning decreases SR from 52.08% to 44.44%. However, under the same ACC setting, PAM improves SR to 58.33% (+13.89 points). These results demonstrate that the performance gains cannot be attributed solely to additional motion history, supporting PAM's contribution under matched conditions.

W1-3 and Q3. Discussion about the sharp performance drop in k=5 (ACC).

A3. The performance degradation at (k=5) also occurs in open-loop evaluation, where the model receives logged action histories, as during training. Compared with (k=3), (k=5) increases ATE from 0.68 to 0.97 and RPE from 0.34 to 0.47 (Table A6). This indicates that the degradation cannot be explained solely by the training–inference mismatch arising from policy-executed histories.

To investigate whether longer ACC histories cause excessive reliance on previous motions, we analyzed repeated in-place turning. Unlike repeated forward movements, which may indicate progress toward the goal, continuously turning in the same direction can lead to a full 360° rotation, returning the robot to its original heading without making progress. This provides a clear behavioral indicator of excessive motion persistence.

Table R3. Closed-loop turning behavior across ACC history lengths (144 episodes).

Metrick=3k=5
Repeated turn chunks51.2%65.3%
Rotations continuing from 180° to 360°7.0%32.5%
Episodes with ≥360° continuous rotation4.9%23.6%

As shown in Table R3, (k=5) exhibits more repeated turn chunks than (k=3) (65.3% vs. 51.2%). While repeated turns can be necessary to reach the correct heading, excessive continuation becomes substantially more frequent: rotations continuing from 180° to 360° increase from 7.0% to 32.5%, and episodes containing a full spin increase from 4.9% to 23.6%.

These results suggest that redundant historical motion information encourages excessive persistence in previous movements, providing a plausible explanation for the substantial SR degradation at (k=5), beyond the proposed distribution mismatch.

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

W2-1 and Q4. Why the performance of one-pass and iterative decoding are similar?

A4. This similarity arises from PAM's full-mask training. Since PAM learns to predict the entire action chunk without conditioning on any target-action prefix, it does not strongly rely on previously generated actions even during iterative decoding.

To verify this, we conducted controlled prefix-corruption experiments (Table R4). We provided either a correct or an incorrect prefix (another sample's ground truth) of three actions and measured prediction errors for the remaining two actions.

Table R4. Controlled prefix corruption on the fixed PAM model (ATE/RPE).

ConditionRECONSCANDHuRoNTartanDrive
No prefix0.197 / 0.1780.408 / 0.3160.657 / 0.4950.469 / 0.366
Correct prefix0.200 / 0.1740.385 / 0.3060.585 / 0.4790.475 / 0.371
Wrong prefix0.203 / 0.1770.379 / 0.2900.599 / 0.4800.511 / 0.394

As shown in Table R4, PAM maintains similar prediction accuracy regardless of prefix correctness, indicating limited dependence on previously generated target actions.

This explains why iterative decoding provides little additional benefit: PAM already generates accurate action trajectories in the first pass referred to as Single-pass in Table A4, while subsequent iterations have limited influence on the remaining predictions. Accordingly, T=5T=5 achieves nearly identical SR 58.33% compared to Single-pass SR 59.03%. We therefore adopt single-pass decoding, achieving comparable navigation performance with only one forward pass instead of five.

W2-2 and Q5. Main contributions of PAM.

A5. As demonstrated in A4, PAM is trained without any target-action prefix, so it learns to predict action chunks without relying on previously generated actions. Consequently, even when iterative decoding reuses predicted tokens, their influence on subsequent predictions remains limited. This suggests that the main benefit comes from preventing target-prefix dependence during training, rather than simply eliminating token reuse at inference.

Table 4 further supports this interpretation. Under the same single-pass inference ((T=1)), replacing random masking with PAM improves SR from 38.89% to 59.03% (+20.14% points). In contrast, changing the decoding steps from (T=1) to (T=5) has little effect on PAM (59.03% vs. 58.33%).

Thus, as already highlighted in Section 5.3, full-mask training addresses target-prefix dependence at its source, while single-pass decoding ensures efficient, prefix-free inference.

Area_Chair_8547OpenReview ↗

Response Reminder

Dear Reviewer,

As we approach the end of the discussion period, we invite you to review the authors’ response to your comments. Please update us on whether your concerns have been sufficiently addressed and whether you wish to pose further questions to the authors.

We greatly appreciate your contribution.

Reviewer_uy1rOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates why adapting Vision-Language Models into direct action-generating policies via standard left-to-right autoregressive decoding struggles in closed-loop visual navigation. The authors hypothesize that conditioning later actions in a predicted chunk on earlier self-generated actions causes within-chunk error propagation through exposure bias. To address this, the authors adapt a pretrained masked diffusion VLM (LLaDA-V) into LLaDA-Nav using two components: Prefix-Free Action Block Modeling (PAM) and Action Context Conditioning (ACC).

PAM masks 100% of the target action block during training and predicts all discretized action tokens simultaneously in a single forward pass. ACC appends the robot's last k executed actions as unmasked context tokens in the prompt to preserve motion continuity.

Critical strengths
  1. The paper shows how the action decoding interface matters when a VLM is used as a navigation policy and demonstrates the importance of within-model interventions rather than relying only on cross-model comparisons.

  2. The paper does a great job of showing how iterative decoding and LoRA rank can improve open-loop error without improving closed-loop success.

  3. The appendix is very transparent. The k=0 results, per-dataset ablations, and refinement-step sweeps are all reported, allowing readers to see what is actually driving the results.

Critical weaknesses
  1. PAM permanently fixes the masking ratio to t=1 during training and performs a single forward pass (T=1) at inference. At both training and inference time, there is no iterative denoising process. The proposed method is therefore similar to BERT-style Non-Autoregressive (NAR) parallel token prediction built on top of a diffusion-pretrained, bidirectional VLM backbone. Predicting an entire action chunk without autoregressive compounding error is well-established in robot learning and is already the default in the navigation baselines the authors compare against (GNM and ViNT regress the full chunk in one pass; NoMaD generates the whole chunk through continuous diffusion). None of these baselines suffer from the within-chunk autoregressive action-reuse problem that motivates this paper.

  2. The primary claim of the paper is that conditioning on self-generated actions causes error propagation due to train/test exposure bias. While ACC avoids directly conditioning on predicted target tokens, it introduces a potential distribution shift across planning steps. During training, ACC conditions on logged action histories from the dataset, while closed-loop evaluation conditions on measured executed-action histories induced by the learned policy. These distributions may differ, particularly when the policy deviates from expert trajectories. The paper does not sufficiently investigate whether ACC remains robust to such deviations.

  3. The results show that PAM alone without action history achieves a closed-loop success rate of 44.44%, which is only slightly higher than existing specialist and world-model baselines (NWM at 43.75% and ViNT at 42.36%). Furthermore, its average open-loop ATE (1.13) is slightly worse than GNM (1.08). Adding ACC to PAM raises success substantially to 59.03%, whereas ACC alone achieves only 38.89%. This suggests a strong interaction between PAM and ACC, rather than a substantial benefit from prefix-free parallel decoding alone. It remains unclear which aspect of their interaction accounts for the large closed-loop improvement.

Questions for authors
  1. During training, ACC conditions on logged action histories, while during closed-loop evaluation it uses odometry-derived executed-action histories from the learned policy. Could this introduce a distribution shift similar to the one the paper aims to mitigate? Have the authors evaluated ACC's robustness when the robot deviates from the demonstrated trajectories or when the action history contains execution errors?

  2. Since PAM uses t=1 masking and T=1 single-pass decoding, how does its action-generation mechanism fundamentally differ from fine-tuning a standard AR VLM with bidirectional attention over masked action-query tokens? While the AR-style LLaDA-V intervention demonstrates that generated-prefix sensitivity also arises within a diffusion-pretrained backbone, the converse comparison is missing. Could the authors evaluate Llama3-V or Qwen3-VL using prefix-free parallel action-block prediction, by replacing target actions with learned/masked action-query slots, allowing non-causal attention within the target block, and fine-tuning for joint prediction of all action tokens? Ideally, this comparison would include ACC under matched training and evaluation settings. This would better isolate whether the closed-loop gains arise from PAM/ACC and the removal of within-chunk action reuse, or from properties specific to LLaDA-V's diffusion pretraining.

6 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We appreciate Reviewer uy1r for your constructive feedback. We hope our response effectively addresses your concerns and questions.

W1-1. Difference to BERT-style NAR.

A1. We thank the reviewer for highlighting this relevant prior work and will incorporate it into the final manuscript. The main difference lies in the training protocol. BERT-style NAR randomly masks target tokens during training and predicts the masked tokens, which corresponds to our Random Masking baseline. In contrast, PAM fully masks the target action block throughout training. As shown in Table 4 in the manuscript, replacing Random Masking with PAM substantially improves navigation performance. While random masking is effective in language modeling, our results demonstrate that full-mask training is more effective for action prediction in our navigation setting.

Table 4 in the manuscript.

MaskTATE ↓RPE ↓SR (%) ↑SPL (%) ↑
Random11.230.5838.8930.48
Random50.670.3038.8933.03
PAM10.680.3459.0350.15
PAM50.670.3158.3351.70

W1-2. Compare to existing baselines.

A2. We agree that GNM, ViNT, and NoMaD already generate action chunks without autoregressive error propagation. Our work instead investigates within-chunk error propagation in autoregressive VLM-based navigation policies, where previously generated actions can affect subsequent predictions. We identify and mitigate this limitation through prefix-free action prediction, aiming to leverage the broad visual-semantic knowledge of pretrained VLMs for navigation.

W2 and Q1-1. ACC's training–inference distribution shift.

A3. The key distinction lies in how previous actions are used. In AR decoding, unexecuted predictions are reused to generate subsequent actions within the same chunk, allowing prediction errors to propagate before execution. In contrast, ACC conditions each new action chunk on executed motions and updated observations. Even if a previous action was suboptimal, its execution history still provides valid information about how the robot actually moved, helping maintain motion consistency without reintroducing within-chunk prediction errors.

We acknowledge that policy-executed histories may differ from expert histories. To quantify this distribution shift, we compared LLaDA-Nav's executed motions with the training data (Table R1).

Table R1. Distribution differences between policy-executed motions and training action histories.

Metricdxdydyaw
JS divergence0.5820.1930.207
KL divergence2.610.720.83
Wasserstein-10.550.106.2°

Despite this substantial distribution shift, ACC improves closed-loop SR from 44.44% to 59.03% and reduces RPE from 0.49 to 0.34 (Table 3). These results show that ACC remains beneficial even when policy-executed histories differ from training histories. We further evaluate robustness to odometry and actuation noise in A4.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

W3. Complementary roles of PAM and ACC.

A4. As shown in Table 3, ACC improves SR by 14.58 points with PAM (44.44% → 59.03%), but by only 2.08 points with random masking (36.81% → 38.89%). We attribute this difference to the complementary roles of PAM in action prediction and ACC in temporal consistency.

-PAM improves action prediction. Under random-mask training, revealed target actions (e.g., a₁ [MASK] a₃ [MASK] a₅) allow the model to interpolate missing actions from neighboring tokens rather than relying on visual observations and goals. In contrast, PAM predicts all actions from a fully masked block, encouraging visually grounded action prediction. Our controlled prefix-corruption experiments (Table R2 and detailed in A.2 of reviewer 8XQy) further show that PAM maintains similar prediction accuracy regardless the prefix, indicating limited dependence on previously generated target actions.

Table R2. Controlled prefix corruption on the fixed PAM model (ATE/RPE).

ConditionRECONSCANDHuRoNTartanDrive
No prefix0.197 / 0.1780.408 / 0.3160.657 / 0.4950.469 / 0.366
Correct prefix0.200 / 0.1740.385 / 0.3060.585 / 0.4790.475 / 0.371
Wrong prefix0.203 / 0.1770.379 / 0.2900.599 / 0.4800.511 / 0.394

-ACC improves temporal consistency. Without ACC, action chunks are predicted without explicit motion history, potentially causing inconsistent movements across planning steps. Our analysis shows that turning-direction reversals decrease from 56.2 to 16.3 per 100 re-plans when ACC is added to PAM. Notably, random masking with ACC achieves a similar reversal frequency (16.2), yet its SR remains at 38.89%. This suggests that improving temporal consistency alone is insufficient without reliable action prediction.

Together, these results explain the complementary effects of PAM and ACC: PAM enables reliable action prediction from visual conditions, while ACC maintains consistency between successive action chunks. Consequently, ACC's temporal context translates into greater navigation success when combined with PAM's reliable action prediction (44.44% → 59.03%). We will incorporate this analysis into Section 5.3.

Q1-2. ACC's robustness against deviation or noisy.

A5. To further evaluate the robustness of ACC during deployment, we conducted two additional experiments involving (i) unexpected trajectory deviations and (ii) odometry and actuation noise.

(i) Forced trajectory deviations. At the 4th and 9th planning steps of each active episode, we replaced the policy's actions with a 90° turn (random direction) followed by a 0.4 m forward movement. The executed motions were recorded through odometry and provided to ACC for subsequent planning.

Table R2. Closed-loop evaluation under forced deviations (144 episodes).

MethodSR clean → deviated (%)ΔSRLLaDA-Nav advantage
ViNT42.36 → 29.17−13.19+12.50
PAM-only44.44 → 30.56−13.89+11.11
LLaDA-Nav59.03 → 41.67−17.36—

Although forced deviations reduce SR across all methods, LLaDA-Nav maintains an advantage over PAM-only (+11.11 points) and ViNT (+12.50 points). This indicates that the benefit of ACC persists even under severe unexpected motion disturbances.

(ii) Odometry and actuation noise. We further evaluated two types of Gaussian noise: odometry noise applied to measured motions provided to ACC and actuation noise applied to executed actions.

Table R3. Closed-loop evaluation under noise (mean over 3 seeds).

ConditionSR (%)SPL (%)ΔSR
Clean (paper)59.0350.15—
Odometry σ = 0.02 m / 2°55.5647.45−3.5
Odometry σ = 0.05 m / 5°54.6342.67−4.4
Actuation σ = 10% / 3°58.8049.28−0.2
Actuation σ = 20% / 6°55.0946.73−3.9

As shown in Table R3, LLaDA-Nav maintains SR above 54% across all tested noise conditions, with a maximum reduction of 4.4 percentage points. Together, these results show that ACC's benefit persists under forced trajectory deviations, while LLaDA-Nav maintains relatively stable performance under the evaluated motion errors.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Q2. Qwen3-VL with PAM and ACC.

A6. To directly answer this question, we conducted the suggested experiment on Qwen3-VL-8B, an AR-pretrained VLM. Specifically, we replaced the 15 target action tokens with a shared learned action-query embedding, enabled bidirectional attention within the action block, and trained the model to predict all action tokens jointly in a single forward pass.

Table R1. Standard fine-tuning vs. PAM on two pretraining families. All experiments include ACC.

BackboneTrainingSR (%)SPL (%)
Qwen3-VL-8BAR44.4437.90
Qwen3-VL-8BPAM58.3351.28
LLaDA-VRandom masking38.8930.48
LLaDA-VPAM59.0350.15
MMaDAPAM70.1464.51

As shown in Table R1, on the AR backbone, PAM improves SR over standard AR fine-tuning (+13.89 points). These results suggest that our performance gains mainly arise from PAM/ACC.

Reviewer_8XQyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LLaDA-Nav adapts a diffusion-pretrained VLM to image-goal navigation through fully masked, single-pass action prediction. It also supplies recent executed actions as motion context and examines sensitivity to generated prefixes through within-model interventions. The method substantially improves reported closed-loop success, with an additional MMaDA experiment suggesting that the recipe extends beyond one backbone.

Critical strengths
  • The closed-loop improvement over the evaluated baselines is substantial.
  • Within-model prefix interventions strengthen the analysis beyond comparisons between unrelated backbones.
  • Training-mask and decoding-step controls help distinguish adaptation from inference effects.
  • The MMaDA result provides useful evidence that the approach is not specific to LLaDA-V.
Critical weaknesses
  • The closest parallel-decoding precedent is missing. OpenVLA-OFT [1] already predicts action chunks from empty embeddings using bidirectional attention in one pass. Compare against an AR-pretrained VLM adapted this way, with the same action vocabulary, training data, and motion history.
  • The mechanism behind the gain remains unresolved. With PAM training, one-pass and iterative decoding reach similar success rates: 59.03% and 58.33%. This result supports full-mask training more directly than eliminating target-token reuse as the decisive explanation; controlled prefix corruption would help separate these effects.
  • The reported closed-loop benchmark also determines the defaults. Appendix C.4 selects hyperparameters using performance on the same 144 episodes. Separate tuning from final evaluation and report variation across training seeds, with confidence intervals computed at the scene level.

[1] Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

Questions for authors
  • Does the advantage persist when competing decoders receive identical executed-action context?
  • How sensitive is ACC to noisy odometry or imperfect action execution?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

We appreciate Reviewer 8XQy for the positive feedback. We try to address your comments and questions below.

W1. Adpatation of OpenVLA-OFT style to AR-pretrained VLM.

A1. We thank the reviewer for highlighting OpenVLA-OFT and will incorporate it into the related work in the final manuscript.

As requested, we adapted OFT-style parallel decoding to Qwen3-VL-8B, an AR-pretrained VLM, using empty action embeddings and bidirectional attention. We used the same action vocabulary, training data, and motion-history inputs for a controlled comparison.

Table R1. Closed-loop evaluation results (following WIW, HM3D on Habitat-Sim).

MethodsSR (%, ↑\uparrow)SPL (%, ↑\uparrow)
Qwen3-VL + OFT59.7254.10
LLaDA-Nav59.0350.15
MMaDA-Nav70.1464.51

As shown in Table R1, OFT-style parallel decoding achieves performance comparable to LLaDA-Nav, confirming that eliminating within-chunk autoregressive feedback is also effective with AR-pretrained VLMs. MMaDA-Nav (see Section C.5 in Appendix) further improves SR and SPL by 10.42 and 10.41 percentage points, respectively, over the OFT-style baseline. These results support the effectiveness of prefix-free action prediction while demonstrating the strong closed-loop performance of our masked-diffusion navigation framework.

W2. Why the performance of one-pass and iterative decoding are similar?

A2. As discussed in Section 5.3, the primary benefit of PAM stems from its full-mask training, which prevents reliance on previously generated target tokens. Consequently, even when iterative decoding is applied at inference, the model does not strongly rely on previously generated actions.

To further verify the reviewer's concern, we conducted controlled prefix-corruption experiments (Table R2). We provided the correct or incorrect prefix (another sample's ground truth) of three actions and measured the prediction error for the remaining two actions. LLaDA-Nav maintains similar prediction errors regardless of prefix correctness

Table R2. Controlled prefix corruption on the fixed PAM model (ATE/RPE).

ConditionRECONSCANDHuRoNTartanDrive
No prefix0.197 / 0.1780.408 / 0.3160.657 / 0.4950.469 / 0.366
Correct prefix0.200 / 0.1740.385 / 0.3060.585 / 0.4790.475 / 0.371
Wrong prefix0.203 / 0.1770.379 / 0.2900.599 / 0.4800.511 / 0.394

These results indicate that PAM relies little on target-action prefixes, consistent with the similar closed-loop SR of single-pass and iterative decoding (59.03% vs. 58.33%). We therefore adopt the more efficient single-pass decoding (one vs. five forward passes).

W3-1. Using separate tuning set.

A3. We built a tuning split of 129 episodes from 22 HM3D validation scenes unused by the test split, and re-ran the configuration selection of Appendix C.4. The selected configuration (k = 1) still significantly outperforms all seven baselines with per-episode test results (Table R3).

Table R3. Closed-loop evaluation results of the configuration selected on the tuning split (95% scene-level CIs).

MethodSR (%, ↑\uparrow)SPL (%, ↑\uparrow)ΔSR (Ours − baseline)
NWM43.7532.62+12.50 [+1.47, +23.24]
ViNT42.3636.61+13.89 [+4.03, +23.78]
LLaDA-Nav56.2551.04—

W3-2. Additional training seed and scene level CI.

A4. We trained LLaDA-Nav again with a different training seed. Averaged over the two seeds, LLaDA-Nav remains significantly ahead of the strongest baselines.

Table R4. Training-seed variation of LLaDA-Nav. Brackets indicate 95% confidence intervals for SR.

Training seedSR (%, ↑\uparrow)SPL (%, ↑\uparrow)
42 (paper)59.03 [51.59, 66.39]50.15
153.47 [45.71, 61.31]47.37
Mean56.2548.76

About scene-level CI, LLaDA-Nav reaches 59.03% [51.59, 66.39], +15.28 [+4.38, +25.77] over NWM and +16.67 [+6.98, +26.39] over ViNT, respectively.

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

Q1. Using executed-action context to the competing decoder.

A5. To address the reviewer's question, we additionally evaluated executed-action conditioning (ACC) on Qwen3-VL-8B. As shown in Table R5, adding ACC alone reduces SR from 52.08% to 44.44%. Notably, combining ACC with PAM on Qwen3-VL improves SR to 58.33%, approaching LLaDA-Nav. These results show that executed-action context alone does not explain LLaDA-Nav's performance gains.

Table R5. Using executed-action context in AR.

Backbone (pretraining)TrainingSR (%)SPL (%)
Qwen3-VL-8B (AR)AR fine-tuning52.0848.95
Qwen3-VL-8B (AR)+ ACC44.4437.90
Qwen3-VL-8B (AR)+ ACC and PAM58.3351.28

Q2. Robustness of ACC.

A6. To answer this question, we injected two types of gaussian noise in closed loop, with three noise seeds per level: odometry noise added noise to the measured motions fed to ACC, and actuation noise perturbs the executed distance and turning angle.

Table R6. Closed-loop evaluation results under noise.

ConditionSR (%)SPL (%)
Clean (paper)59.0350.15
Odometry σ = 0.02 m / 2°55.5647.45
Odometry σ = 0.05 m / 5°54.6342.67
Actuation σ = 10% / 3°58.8049.28
Actuation σ = 20% / 6°55.0946.73

As shown in Table R6, LLaDA-Nav maintains relatively stable closed-loop performance under both odometry and actuation noise. Even at the highest tested noise levels, the SR decreases by at most 4.4 points. Moreover, the SR remains above 54% across all noise conditions, exceeding the clean-condition performance of the strongest baselines (NWM: 43.75%, ViNT: 42.36%). These results demonstrate that ACC maintains stable navigation performance under the tested levels of motion estimation and execution errors.

Area_Chair_8547OpenReview ↗

Response Reminder

Dear Reviewer,

As we approach the end of the discussion period, we invite you to review the authors’ response to your comments. Please update us on whether your concerns have been sufficiently addressed and whether you wish to pose further questions to the authors.

We greatly appreciate your contribution.

Thank you again for your time and thoughtful feedback. As the discussion deadline approaches, we would appreciate hearing about any remaining questions or concerns while there is still time for us to address them. Please feel free to let us know if any part of our response needs further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.