acceptodds
Withdrawn from ICLR 2027

Think-to-Act: Grounded Reasoning Supervision for Generalizable Vision-and-Language Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The dominant recipe for continuous vision-and-language navigation (VLN) fine-tunes a pretrained vision-language model (VLM) with an action cross-entropy loss. This loss rewards the correct action but places no semantic constraint on the model. The policy therefore overfits to action and pixel-goal outputs and loses the pretrained semantics that generalisation relies on. We call this failure action overfitting. It goes unnoticed in training environments but hurts navigation in unseen ones. We present Think-to-Act (T2A), which uses structured reasoning as a training-time semantic anchor. At every decision point, T2A supervises four types of navigation semantics (scene understanding, instruction progress, decision rationale, and stop condition). The supervision comes from T2A-GR, a grounded reasoning dataset that covers nearly the entire R2R-CE training split. An action-first target layout prevents teacher-forced leakage. Reasoning dropout and a path-consistency objective train the reasoning-free path used at deployment, so no reasoning token is decoded at test time. Under matched training, T2A improves val-unseen success rate by 5.7 points on R2R-CE and by 4.8 points on RxR-CE, whose instructions are never annotated. It is also more effective than data replay and single-field auxiliary objectives under the same protocol. Trained only on three simulated datasets, the reasoning-free semantic system transfers zero-shot to a quadruped robot in unseen physical scenes. It succeeds in 86.7% of trials, against 38.3% for the InternVLA-N1. Code and data will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-and-language navigation, embodied AI, vision-language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zZpYOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Think-to-Act trains a navigation model to predict an action followed by a structured explanation covering observations, instruction progress, the decision, and stopping. Reasoning dropout and a latent-consistency objective connect this training format to deployment without reasoning text. The paper contributes a large annotation resource and a training recipe that improves unseen navigation while retaining a short action output at inference.

Critical strengths
  1. The action-first layout addresses a specific training problem (Section 3.3, Eq. (3); Figure 7). Causal attention prevents the action from depending on a gold explanation that will be absent at deployment. Combining this ordering with reasoning dropout in Section 3.4 makes the training and deployment paths fit together naturally.

  2. The matched results support the practical value of the method (Table 2). R2R-CE unseen SR improves from 53.3 to 59.0 with the same initialization, mixture, and training steps. Observation and progress supervision each recover part of the gain, while shuffled rationales hurt performance. These comparisons give a useful account of why the content of the auxiliary supervision matters.

  3. The RxR result extends the value of the annotations (Section 4.3; Table 3). SR improves from 50.8 to 55.6 even though the reasoning annotations come from R2R. RxR action supervision remains in the training mixture, but the result still shows that the benefit of R2R reasoning targets carries beyond the annotated dataset.

  4. T2A-GR is a useful contribution in its own right (Section 3.2; Appendix D). The 166,997 annotated decision points turn trajectories into supervision about intermediate navigation judgments. The prompt, examples, and human audit explain what the fields are intended to capture and should help others build on the resource.

Critical weaknesses
  1. The semantic-retention explanation is stronger than the result establishes (Section 1; Section 4.3). The smaller seen/unseen gap is consistent with retaining useful pretrained knowledge, but T2A could also teach new navigation-specific relationships between observations, progress, and actions. Table 2 demonstrates the value of richer supervision without distinguishing these explanations. I would present semantic retention as an interpretation and center the claim on the demonstrated improvement in generalization.

  2. The decision targets explain demonstrated actions (Section 3.2; Appendix D.1). The annotator receives the ground-truth next action and explains why it is appropriate. This is sensible supervision, but it has a different role from the observation and progress fields, which describe the scene and task state. The paper would benefit from making this distinction explicit: what is the decision field intended to teach beyond a richer description of the action label?

  3. The results favor a stronger emphasis on supervision design (Section 4.6; Table 2). Four-field supervision without path consistency already reaches 58.2 unseen SR, compared with 53.3 for action-only training and 59.0 for full T2A. Most of the gain therefore comes before adding Eq. (5). Path consistency is a reasonable refinement, but the statement that every component is necessary gives it more weight than these results suggest. The annotation design and action-first/dropout training format appear to be the main contribution.

Questions for authors
  1. Does “semantic retention” in Sections 1 and 4.3 mean preserving abilities from pretraining, or learning navigation semantics that action-only supervision does not encourage? The distinction would help clarify the proposed mechanism.
  2. In the four-field design of Section 3.2, how do the authors distinguish the intended contribution of the action-conditioned decision explanation from that of observation and progress supervision?
Reviewer_YVxHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper hypothesizes that action-only fine-tuning can erode a VLM's pretrained semantics and hurt generalisation to unseen environments, because the action loss does not explicitly constrain semantic representations. It calls this hypothesised effect action overfitting.

The authors propose using grounded reasoning as a training-time semantic anchor in Think-to-Act (T2A). T2A-GR provides rationales covering scene understanding, instruction progress, action decisions and stopping. An action-first layout prevents action tokens from reading the gold rationale, while reasoning dropout and latent consistency train the model to act without generating reasoning at test time. Under matched training, T2A improves val-unseen success rates by 5.7 points on R2R-CE and 4.8 points on RxR-CE. The paper also reports zero-shot transfer to a real robot.

Critical strengths
  1. Table 2 includes useful controls for understanding the training objective. The comparisons share initialisation, a 15k-step budget, seed 42 and reasoning-free evaluation, and cover replay, single-field targets, shuffled rationales and training-layout variants. At this checkpoint, four-field supervision without consistency reaches 58.2 unseen SR, compared with 52.1 for shuffled rationales and 53.3 for action-only training. This is suggestive evidence that supervision content matters. The replay rows replace 25% of the navigation mixture, so their navigation exposure differs. The design is informative, while the stability of the detailed rankings remains open (see W1). The action-first causal mask structurally blocks action tokens from reading the gold rationale.

  2. The reported models show promising results on two benchmarks without generating reasoning at deployment. At the reported checkpoints, R2R-CE unseen SR is 59.0 vs 53.3 for action-only training, and RxR-CE SR is 55.6 vs 50.8 (Tables 2-3). RxR receives no rationale labels, although its action targets are in training. The second evaluation distribution broadens the observed benefit, but does not provide an independent training replication or establish robustness across seeds or checkpoints (see W1). No rationale is decoded at deployment, avoiding reasoning-generation overhead. The final system's reported 65.8 SR after on-policy fine-tuning is also competitive on R2R-CE.

  3. The real-robot study is reported transparently. Inference runs on the robot's Jetson Thor. Each of six shared tasks is run ten times per system, with difficulty tiers fixed before testing, props on taped footprints, an independent safety observer, and stop distances measured by tape (Appendix G.3). Per-task results, the stop distance of every trial and a failure taxonomy are all given (Tables 9-10, Figure 13), so a reader can check the robot numbers in detail.

  4. T2A-GR is a carefully controlled annotation resource. The annotation inputs separate what the student will see from teacher-only information, and the full prompt is reproduced (Appendix D.1). This level of quality control makes the four-field labels reusable by other VLN work.

Critical weaknesses
  1. The reported gain is positive, but its stability across training runs and checkpoints is not established. The experiment in Table 2 and Figure 3 does not cover training randomness or checkpoint selection. Figure 3c shows T2A unseen SR at about 52 at 9k, 50 at 12k, 51 at 13k, 59.0 at 15k and 55 at 17k; action-only scores about 47 at 7k, 53.3 at 15k and 52 at 16k. The only common evaluated step is 15k, the highest plotted unseen point for both runs. The paper does not fully explain how the evaluation steps and reported checkpoint were chosen. These observations warrant clarification. The val-seen curves do not show that the unseen gain is checkpoint-independent. Only the main runs have plotted unseen trajectories, so Table 2 does not show whether the small ablation differences persist across checkpoints. Please explain the checkpoint rules and report both main models' unseen SR/SPL at common saved steps; existing additional-seed results would also help.

  2. The relationship to closely related work needs clearer citation and discussion of the actual contribution. Aux-Think (Wang et al., NeurIPS 2025) is cited but appears only as cross-paper rows in Tables 1 and 3. It already built R2R-CoT-320k using a VLM that sees the history, current view and ground-truth next action (its Appendix A.2); this dataset and its relation to T2A-GR should be discussed. Aux-Think also uses CoT and instruction reconstruction as training-only auxiliary tasks with reasoning-free testing (its Section 3.4), and evaluates an action-then-reasoning Post-Think variant (its Section 3.3, Appendix A.4). The uncited Training Strategies for Efficient Embodied Reasoning (Chen et al., CoRL 2025) introduces ECoT-Lite recipes. Its reasoning-dropout variant places reasoning before actions, randomly removes zero to all reasoning steps, applies losses to both reasoning and actions, and can start directly at ACTION at test time (its Section 5, Appendix E.3). This differs from T2A's fixed-probability whole-rationale dropout, alongside differences in models, tasks and objectives. Its Section 4 proposes representation, curriculum and expressivity hypotheses, and Section 6.3 analyses why standard action-only training may learn a poor representation-to-action mapping. This makes the blanket claim at L99-101 about prior work too broad, without establishing T2A's particular semantic-forgetting hypothesis. Please cite this work and explain the connections and differences. The contribution of T2A's particular combination of action-first layout and dropout, consistency objective, grounded labels and matched comparisons should be positioned precisely.

  3. Semantic preservation is a plausible explanation, but the causal wording is stronger than the evidence. L42-50 attribute poor unseen-environment performance to pretrained semantics lost during action-only fine-tuning, and contribution (1) claims to identify this as a cause. The cited Aux-Think, ERVLA, AwareVLN, and Progress-Think provide useful results and analyses supporting reasoning-augmented policies. These results reasonably motivate the hypothesis, but do not directly establish the proposed chain from semantic erosion to unseen failures. An action loss lacking an explicit semantic constraint does not itself show that pretrained semantics are lost. T2A's narrower seen/unseen gap at the reported checkpoint (Section 4.3) is consistent with this hypothesis, but improved task-specific representations, teacher-derived navigation knowledge, or regularisation could also explain the gain. The paper does not compare semantic abilities before and after fine-tuning, and leaves hidden-state probing to future work. I encourage the authors to present semantic preservation as a hypothesis and qualify the causal wording in the introduction and contribution (1), unless more direct evidence is available. This clarification would preserve the value of the reported navigation gains.

  4. The real-robot gap cannot be attributed to grounded reasoning supervision, and the baseline integration needs a sanity check. The robot comparison is between the T2A checkpoint and the released InternNav S2, which was trained on different data with a different recipe. No matched action-only checkpoint was run on the robot. Even so, Section 4.5 attributes the drop in route errors from 19 to 1 to the semantic types that T2A-GR supervises.

  5. The claim that the semantic term removes the need to freeze the slow system has little support. Section 3.5 and Section 4.3 conclude that the control gradient "helps rather than harms, so freezing is not required". The evidence is joint NextDiT at 59.5 against frozen transfer at 59.1. Appendix B shows these are read at different checkpoints (14k for joint, 20k for frozen), and both come from a single seed. NavDP*'s 60.1 also uses simulator depth (Table 5). There is no jointly trained action-only backbone, so the paper never shows that control gradients cause harm when the semantic term is absent. Unless such a run exists, the claim should be restated as "joint training did not hurt in our setting".

  6. Minor. Appendix E has a heading but no text.

Questions for authors
  1. How were the evaluated steps and reported checkpoints in Tables 2-3 chosen? Was the 15k checkpoint for Table 2 fixed before val-unseen evaluation, and how were unseen results used in selection? Please report both main models' val-unseen SR/SPL at a common set of saved steps. If additional-seed runs already exist, please include them.

  2. Was the rationale-first row of Table 2 trained with p=0.2p = 0.2 dropout, and with or without the consistency term? Do reasoning-bearing samples use a different prompt from the deployment prompt, and does this explain the drop of the no-dropout row? Please clarify T2A-GR's differences from R2R-CoT-320k and discuss T2A's connections and differences to ECoT-Lite.

  3. Could you distinguish the hypothesis that action-only fine-tuning erodes pretrained semantics from what the experiments establish, and qualify the causal wording in L42-50 and contribution (1)? If supporting evaluations already exist, please compare the pretrained VLM, action-only, and T2A checkpoints on a suitable semantic probe, ideally relevant to the four supervised fields. Reframing semantic preservation as a hypothesis would also address this concern without additional experiments.

  4. For the robot study, was the action-only checkpoint tested on any robot task?

Reviewer_u3S3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Think-to-Act (T2A), a training framework for reducing action overfitting and improving generalization in continuous vision-and-language navigation. It introduces T2A-GR, a dataset with grounded reasoning supervision for scene understanding, instruction progress, action decisions, and stopping conditions. The framework uses an action-first target format to avoid leakage from teacher-forced reasoning, together with reasoning dropout and latent path consistency to support reasoning-free inference. The learned semantic representations can also condition a diffusion-based controller for continuous navigation.

Experiments on R2R-CE and RxR-CE show improved navigation performance in unseen environments. The paper also includes component ablations and real-world deployment on a quadruped robot.

Critical strengths
  1. The paper addresses a relevant problem in VLN with a coherent training framework. It introduces navigation-specific reasoning supervision during training without requiring reasoning generation at inference. The action-first format, reasoning dropout, and path consistency work together toward this objective.

  2. The gains under matched training are meaningful. T2A improves val-unseen success rate by 5.7 percentage points on R2R-CE and 4.8 percentage points on RxR-CE. The ablations support the effectiveness of the training strategy, while the real-robot experiments demonstrate deployment feasibility.

Critical weaknesses
  1. The methodological novelty is limited. Training-time reasoning supervision with reasoning-free inference has already been explored in Aux-Think. The action-first format and reasoning dropout also build on existing techniques. T2A introduces structured navigation reasoning and path consistency, but the paper does not establish their additional value over prior reasoning-supervision methods through matched comparisons. This makes it difficult to assess what the proposed design contributes beyond adapting existing techniques to navigation.

  2. The experiments do not directly substantiate the action-overfitting hypothesis. The paper argues that action-only fine-tuning degrades the pretrained VLM’s semantic capabilities and that T2A prevents this degradation. However, the experiments primarily evaluate navigation performance rather than semantic retention. The observed gains could also result from learning representations that are better suited to navigation. They do not, by themselves, establish that T2A preserves the pretrained model’s semantic capabilities.

Questions for authors
  1. Comparison with prior reasoning supervision: How does T2A compare with an Aux-Think-style approach under the same backbone, training data, and computational budget? Which gains can be attributed specifically to the proposed design?

  2. Evidence for semantic retention: Can you directly compare the semantic capabilities of the pretrained VLM, the action-only model, and the T2A model before and after fine-tuning? This would help substantiate the claim that T2A prevents semantic degradation.

  3. Evidence for the real-world contribution: How does the matched action-only checkpoint perform on the same real-robot tasks using the same execution pipeline? This would help isolate the real-world gains attributable to T2A’s reasoning supervision.

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

Related papers

Discussion (0)

Sign in to comment.