acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Reasoning Context for Vision-Language Navigation

Abstract

Reasoning in vision-language navigation (VLN) can serve as an action-generation trace, auxiliary supervision, or policy context. Policy context is particularly suited to sequential navigation, making task-level guidance explicit at execution time and reusable across action decisions. Yet what this context should contain remains underexplored. Scene description, progress assessment, and subgoal planning address complementary navigation needs, suggesting that combining them should provide better guidance. Surprisingly, we find that more comprehensive reasoning context can lead to worse navigation. Through component-wise comparisons on matched navigation trajectories and action annotations, we find that subgoal planning alone outperforms the full reasoning context. However, the best way to complete a subgoal can depend on what follows it. We therefore propose Context-Think, which uses final-goal replanning to regenerate a language plan for the remaining route from the agent's current state to the original destination. Each updated plan explicitly represents the remaining subgoals, allowing future route requirements to inform current action selection while adapting to execution progress. Context-Think improves navigation success and path efficiency over both subgoal planning and full reasoning context on the widely used R2R-CE and RxR-CE benchmarks. Our findings challenge completeness as a design principle for policy context: include fewer reasoning components, but plan all the way to the final goal.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language navigation, Reasoning context, Planning horizon, Final-goal replanning, Embodied AI

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_kb89OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper investigates what reasoning context is useful for vision-language navigation. It finds that subgoal planning alone outperforms combining scene descriptions, progress assessments, and planning. The authors then propose Context-Think, which extends planning to cover the remaining route to the final destination and updates it during execution. Experiments on R2R-CE and RxR-CE report improvements in navigation success and path efficiency over local subgoal planning and combined reasoning context.

Critical strengths

The paper studies an important problem: how reasoning context should guide long-horizon navigation.

The paper is easy to follow. Given the result shown in Fig. 3, it is natural to investigate whether extending planning beyond the immediate subgoal can further improve navigation.

Table 1 provides a broad comparison with existing methods and shows that the proposed method achieves strong performance.

Critical weaknesses
  1. How robust are the findings in Fig. 3 across environments, prompt designs, and base models? Could reasoning-context degradation reflect limitations of this specific VLM, and would it persist with a more capable model in the future?

  2. The narrow y-axis range in Fig. 3 visually emphasizes relatively small performance differences. Are these differences statistically significant? Please report confidence intervals or variability across training runs, particularly to support the claim that subgoal planning is the strongest component.

  3. Equation (2) is underspecified. The authors should explain how LactionL_{\mathrm{action}} and LreasonL_{\mathrm{reason}} are computed and normalized, and specify the value of λreason\lambda_{\mathrm{reason}}.

Questions for authors

Can a more capable frontier VLM (e.g., GPT6-Astra) achieve comparable performance through zero-shot agentic planning?

What motivated the choice of Llama-3 8B as the backbone, and how might this choice affect the findings in Fig. 3?

4 comments

Response to Reviewer kb89 (1/2)

We thank Reviewer kb89 for suggesting tests across prompts, backbones, and training seeds, and a comparison with zero-shot frontier-model planning. These suggestions help establish the scope of our findings and distinguish reproducible effects from implementation-specific outcomes.

1. Validate robustness across environments, prompts, and backbones (W1)

Reasoning-context degradation is consistently observed across diverse environments, instruction complexities, prompt designs, and VLM backbones.

Cross-environment and benchmark validation. R2R-CE and RxR-CE are two widely used VLN benchmarks built on Matterport3D, which contains 90 buildings and 2,056 rooms (Chang et al., 2017). Our evaluations use their val-unseen splits, covering 11 unseen buildings per benchmark and 12,845 navigation episodes. The benchmarks also differ substantially in task complexity, with average instruction lengths of 32 vs. 120 words and reference path lengths of 9.89 vs. 15.23 m, respectively. On both benchmarks, Full underperforms Progress and Subgoal on R2R-CE, and every single component on RxR-CE.

Additional prompt and backbone tests. Following the reviewer's suggestion, we further evaluate an alternative prompt design and a different VLM backbone. With the alternative prompt, Full again achieves the lowest SR on RxR-CE (Table S1). With Qwen2.5-VL-7B replacing the original Llama-3 8B backbone, Full has the lowest SR and SPL, falling 1.65–3.92 SR points below every individual component (Table S2). All conditions within each experiment share the same trajectories, teacher targets, reasoning-update positions, and training settings.

Table S1. Alternative prompt design on RxR-CE Val-Unseen.

ConditionSR (%) ↑SPL (%) ↑NE (m) ↓
Full56.1347.896.75
Scene56.4747.736.55
Progress57.9049.296.04
Subgoal58.0749.476.29

Table S2. Qwen2.5-VL-7B backbone on RxR-CE Val-Unseen.

ConditionSR (%) ↑SPL (%) ↑NE (m) ↓
Full39.9532.978.25
Scene43.8736.617.83
Progress41.8535.008.37
Subgoal41.6034.718.26

Scope of our claim. Our finding is that more comprehensive reasoning context does not necessarily improve navigation, rather than that it always causes degradation. The consistent results across benchmarks, prompts, and backbones challenge the assumption that more reasoning is always beneficial and motivate more selective, decision-oriented reasoning context.

Response to Reviewer kb89 (2/2)

2. Add training-seed variability and revise Fig. 3 (W2)

The main Full–Subgoal gap is substantially larger than the observed between-seed variation. We retrain all five conditions with two additional seeds and report three-seed statistics below.

Table S3. R2R-CE and RxR-CE val-unseen results across three training seeds (mean ± standard deviation).

BenchmarkMetricFullSceneProgressSubgoalFinal-goal
R2R-CESR (%) ↑50.44 ± 0.1950.34 ± 0.1651.47 ± 0.2252.18 ± 0.3053.15 ± 0.70
R2R-CESPL (%) ↑45.39 ± 0.2545.51 ± 0.1546.03 ± 0.1246.55 ± 0.1847.62 ± 0.45
RxR-CESR (%) ↑52.90 ± 0.4255.92 ± 0.1756.19 ± 0.2856.48 ± 0.2358.90 ± 0.31
RxR-CESPL (%) ↑45.15 ± 0.5447.40 ± 0.1647.87 ± 0.3148.01 ± 0.2249.82 ± 0.36

Subgoal exceeds Full by 1.74/3.58 SR points on R2R-CE/RxR-CE. Final-goal replanning further improves over Subgoal by 0.97/2.42 points. These results support reproducibility across the tested seeds.

We will add exact values and standard-deviation error bars to Fig. 3 and make the restricted axis range explicit, so readers can assess both the magnitude and variability of the differences.

3. Specify the loss definitions, normalization, and weight (W3)

Both losses are token-level cross-entropy losses, each normalized by its own number of valid supervised output tokens, with λreason=1\lambda_{\mathrm{reason}}=1 throughout. Let Da\mathcal D_a and Dr\mathcal D_r denote the action and reasoning queries, respectively, and let x=(I,Ht,Ct)x=(I,H_t,C_t) be the query input:

Laction=−1Na∑(x,a)∈Da∑jlog⁡pθ(aj∣x,a<j),L_{\mathrm{action}}=-\frac{1}{N_a}\sum_{(x,a)\in\mathcal D_a}\sum_j\log p_\theta(a_j\mid x,a_{<j}),

Lreason=−1Nr∑(x,r)∈Dr∑jlog⁡pθ(rj∣x,r<j).L_{\mathrm{reason}}=-\frac{1}{N_r}\sum_{(x,r)\in\mathcal D_r}\sum_j\log p_\theta(r_j\mid x,r_{<j}).

Here, NaN_a and NrN_r count the valid supervised tokens in their respective query sets. We will add these definitions below Eq. (2) and include the loss weight in the hyperparameter table.

4. Compare with zero-shot frontier-model planning (Q1)

Context-Think outperforms the tested zero-shot GPT-6-Astra configuration on the same 50 randomly sampled RxR-CE val-unseen episodes. Both use the same prompt; the baseline replaces the policy model with the GPT-6-Astra API.

Table S4. Matched-prompt comparison on the same 50 episodes.

MethodSR (%) ↑SPL (%) ↑NE (m) ↓
GPT-6-Astra (zero-shot)54.0050.535.50
Context-Think72.0063.594.07

Context-Think improves SR/SPL by 18.00/13.06 points and reduces NE by 1.43 m. This result concerns the tested model-replacement baseline.

5. Explain the backbone choice and its relation to the findings (Q2)

We adopt the same NaVILA Llama-3 8B backbone as AwareVLN (CVPR 2026), the most directly relevant prior work on reasoning context for VLN. AwareVLN also combines scene description, progress assessment, and planning, and uses the same underlying navigation datasets (R2R, RxR). This alignment allows us to investigate reasoning-context composition within an established and directly relevant framework, rather than introducing additional differences in backbone or data sources.

More importantly, our additional Qwen2.5-VL-7B experiments (Table S2) confirm that reasoning-context degradation is not specific to Llama-3 8B.

The authors’ response has largely addressed my concern. I am therefore raising my recommendation to Weak Accept.

Thank you for your positive feedback and for raising your recommendation. We are glad that our response has addressed your concerns.

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

This paper investigates the impact of different reasoning contexts on navigation decisions in vision-and-language navigation (VLN). Existing methods typically combine multiple types of reasoning information, including scene descriptions, task progress assessment, and subgoal planning, to guide navigation. Through experimental analysis of the contributions of different reasoning components, the authors find that incorporating more reasoning information does not necessarily improve navigation performance, while using subgoal planning alone can achieve better results. Based on this observation, the authors propose Context-Think, which employs final-goal replanning to regenerate a language-based plan covering the entire remaining route based on the agent's current state. This allows future subgoals and the final destination to guide current action selection.

Critical strengths
  1. The paper presents a clear research motivation and addresses a meaningful research question. Existing vision-and-language navigation methods often incorporate multiple types of reasoning information to improve navigation decisions. However, through controlled experiments, the authors find that more comprehensive reasoning contexts do not necessarily lead to better navigation performance.

  2. Building on the findings from the controlled experiments, the authors propose a relatively comprehensive framework that enables the model to consider future navigation constraints when making current action decisions. The proposed method also achieves competitive performance across multiple benchmarks.

Critical weaknesses
  1. The paper finds that combining scene descriptions, progress assessment, and subgoal planning may lead to performance degradation, and attributes this phenomenon to redundant reasoning information and the temporal relevance of historical context. However, could this degradation also be influenced by other factors, such as context length, information interference, or the quality of generated reasoning? The current experiments mainly demonstrate performance differences among different reasoning components, without directly analyzing these potential factors. Therefore, the underlying causes of reasoning-context degradation remain insufficiently understood.

  2. Existing vision-and-language navigation methods, such as NavGPT, MapGPT, and AwareVLN, have already explored long-horizon task planning, dynamic subgoal updates, and adapting future plans based on navigation progress. The authors should further clarify the key technical distinctions between Context-Think and existing hierarchical planning and dynamic replanning approaches.

  3. The independent performance gains of Final-goal Replanning are relatively limited in the current experiments. Further evidence is needed to demonstrate whether this mechanism can consistently provide stable performance improvements.

Questions for authors
  1. Could the authors further verify whether the observed performance degradation is indeed caused by redundant reasoning information or outdated historical context, rather than interference from longer contexts? As the amount of information increases, interactions among different reasoning components and the increased context length may both affect navigation performance. Therefore, the observed performance gap between the full reasoning context and single-component subgoal planning cannot be simply attributed to reasoning-context degradation without further analysis.

  2. Could the authors further analyze the accuracy of the generated remaining-route plans and how planning errors affect subsequent navigation decisions? Particularly in longer or more complex navigation tasks, long-horizon planning may suffer from error accumulation. A quantitative analysis of planning quality and navigation failure cases would help provide a clearer understanding of the effectiveness of the proposed method.

4 comments

Response to Reviewer Zkvk (1/2)

We appreciate Reviewer Zkvk's insightful questions about the causes of reasoning-context degradation, connections to prior planning methods, and the impact of planning errors. These suggestions help deepen our analysis of how reasoning context supports navigation decisions.

1. Explain reasoning-context degradation through context composition and decision usefulness (W1, Q1)

Our additional analysis reveals that the usefulness of reasoning context depends more on its content than on its length, relevance, or correctness. Scene descriptions and progress assessments often repeat information available from observations and instructions, and may become outdated between reasoning updates. In contrast, planning provides prospective guidance for subsequent actions. This difference offers a plausible explanation for why combining individually useful reasoning components can degrade navigation.

To investigate the factors raised by the reviewer, we analyze context length, information relevance, and reasoning correctness on 50 randomly sampled RxR-CE Val-Unseen episodes, with manual annotations performed blind to the condition.

Context length. All three shorter single-component contexts (19.2–36.9 tokens) outperform Combined Context (76.0 tokens). Meanwhile, Context-Think achieves 5.6 points higher SR with nearly the same length (74.9 tokens). This highlights the importance of reasoning content beyond token count.

Information relevance. All contexts achieve similarly high relevance (97.4–100.0%), yet their SR ranges from 53.0% to 58.6%. High relevance alone therefore does not guarantee effective navigation guidance.

Reasoning correctness. Scene-only achieves the highest correctness (86.4%), while Context-Think achieves the highest SR (58.6%) with 80.9% correctness. Combined Context also has higher correctness than Progress-only and Subgoal-only, yet lower SR. Thus, reasoning accuracy and downstream decision quality are distinct.

Table S5. Reasoning-context analysis on RxR-CE. Target tokens: average teacher-target length; Correct and Relevant: manual annotations on 50 sampled episodes; SR: full Val-Unseen evaluation.

ConditionTarget tokensCorrect (%)Relevant (%)SR (%) ↑
Combined Context (Scene + Progress + Subgoal)76.083.099.253.0
Context-Think (Ours)74.980.9100.058.6
Scene-only36.986.498.456.1
Progress-only19.264.097.456.0
Subgoal-only20.060.2100.056.3

Together, these findings support our central insight: effective reasoning context should prioritize information that guides future decisions, rather than simply maximizing the amount or apparent quality of reasoning.

Response to Reviewer Zkvk (2/2)

2. Clarify the distinction from existing planning methods (W2)

Context-Think retains a plan covering the remaining route to the final goal, rather than relying on local subgoal planning or comprehensive reasoning context. Its contribution combines a systematic study of reasoning-context composition with a plan-only, final-goal replanning design.

Table S6. Comparison with existing planning methods. Retained context refers to information beyond the original instruction and navigation history.

MethodPlanning horizonRetained contextPlan source
NavGPTSubgoalsReasoning tracesZero-shot
MapGPTMap nodes to exploreOnline map and previous planZero-shot
AwareVLNNext navigation phaseScene, Progress, and PlanSupervised
Context-Think (Ours)Remaining route to final goalFinal-goal plan onlySupervised

3. Validate final-goal replanning across seeds and task lengths (W3)

Final-goal replanning consistently improves navigation across three training seeds, both benchmarks, and longer navigation tasks. It outperforms Subgoal in every training seed on R2R-CE and RxR-CE, with mean SR gains of 0.97/2.42 points and SPL gains of 1.07/1.81 points, respectively.

Table S3. Final-goal replanning versus Subgoal across three training seeds (mean ± standard deviation).

BenchmarkMetricSubgoalContext-Think (Ours)
R2R-CESR (%) ↑52.18 ± 0.3053.15 ± 0.70
R2R-CESPL (%) ↑46.55 ± 0.1847.62 ± 0.45
RxR-CESR (%) ↑56.48 ± 0.2358.90 ± 0.31
RxR-CESPL (%) ↑48.01 ± 0.2249.82 ± 0.36

Long-horizon navigation. On RxR-CE, the SR gains remain positive for longer instructions and trajectories: +2.13 points for 140–199-token instructions and +3.87 points for 20–25 m reference trajectories, compared with +2.27 points overall in the original checkpoint.

These results demonstrate that the benefit of extending the planning horizon is reproducible across training runs and remains effective on longer navigation tasks.

4. Quantify plan accuracy and explain error handling (Q2)

Context-Think achieves 80.9% plan correctness and 100.0% relevance, compared with 60.2% correctness for Subgoal planning, based on manual annotations of 50 RxR-CE Val-Unseen episodes (Table S5).

Error handling through replanning. At each reasoning update, Context-Think regenerates the remaining-route plan from the original instruction and observed navigation history. This allows the policy to revise earlier planning errors as new observations become available, rather than continuing to follow a fixed plan. By retaining the original destination and updating the unfinished route, the model can adapt subsequent actions to the state actually reached.

The authors' response has largely addressed my concerns. The additional analyses of context length, information relevance, and reasoning correctness provide further support for the research motivation, while the supplementary experiments also validate the stability of the proposed method. Therefore, I maintain my positive rating.

Thank you for your positive feedback and for engaging in the discussion. We are glad that the additional analyses and experiments have addressed your concerns.

Reviewer_6HsmOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a method for vision-language navigation models, with the following main claims

  • it is better to keep a concise "context" during both reasoning step (as prediction target) and act step (as input), and plan as context works the best.
  • when using plan as context, it is better to use not just the next step plan, but also the plan until the end of the episode.
Critical strengths
  • The method is intuitive and easy to follow.
  • The controlled experiment setup in 3.1 and 3.2 is clean, although I do have concerns on the significance of the results.
Critical weaknesses

The paper uses teacher models to regenerate labels for Table 1 but never specified what it is. Because this essentially changes the dataset, it is hard to attribute whether the better performance is from the teacher labels, or the methods. The weak ablation makes this issue worse.

Table 2 (a) ablates over the main method of this paper, but I have major concerns about its significance:

  • The improvement is small (~1, with no error bars). At this small level of delta, it is important to have some measures to show that the improvements are real.
  • Both methods are worse than the un-tuned model NaVILA on R2R-CE (comparing numbers between Table 2 (a) and Table 1). Then I am not sure how informative this ablation setup is. What's the reason to not perform ablation using the same dataset as in Table 1?

Similarly, the claim that subgoal plan is the best (Fig 3) is only true for R2R-CE and the gap between subgoal and progress is negligible on RxR-CE.

Questions for authors
  • What's the teacher model?
  • In Fig 3, what if you train the model with the same outputs in reasoning steps, but only feed the Subgoal plan to the act step as context?
  • What's the reason to not use the same setup as Fig 1 for ablations? The performance drops a lot from Table 1 to Table 2. So maybe the gap will be a lot more visible if you use the dataset setup of Table 1 to do ablations.
2 comments

Response to Reviewer 6Hsm (1/2)

We thank Reviewer 6Hsm for the review. We would first like to clarify a fundamental distinction that appears to have been overlooked in the review: Subgoal Planning is an intermediate finding of our analysis, not the proposed Context-Think method.

Our paper makes two distinct contributions: (1) systematically studying reasoning-context composition and identifying Subgoal Planning as the strongest individual component, and (2) proposing Context-Think with Final-goal Replanning, which extends planning from the immediate subgoal to the entire remaining route. The latter is our central methodological contribution, evaluated in Table 2(a).

We provide additional experimental evidence and clarifications below.

1. Clarify the teacher model and controlled supervision (W1, Q1)

We use Qwen3.8 27B as a fixed teacher for all reasoning-context configurations. Scene, Progress, and Subgoal annotations are generated independently at the same trajectory-aligned reasoning-update positions. Combined Context (denoted as "Full" in the paper) concatenates these same annotations. All configurations share identical navigation trajectories, action labels, update positions, and training settings.

We will explicitly specify the teacher model and controlled supervision protocol in the revision.

2. Validate Final-goal Replanning across training seeds (W2.1)

Context-Think consistently outperforms Subgoal Planning across three training seeds on both benchmarks. Mean SR improves by 0.97/2.42 points and SPL by 1.07/1.81 points on R2R-CE/RxR-CE, respectively. The gains are positive in every evaluated seed.

Table S3. R2R-CE and RxR-CE Val-Unseen results across three training seeds (mean ± standard deviation).

BenchmarkMetricCombinedSceneProgressSubgoalContext-Think
R2R-CESR (%) ↑50.44 ± 0.1950.34 ± 0.1651.47 ± 0.2252.18 ± 0.3053.15 ± 0.70
R2R-CESPL (%) ↑45.39 ± 0.2545.51 ± 0.1546.03 ± 0.1246.55 ± 0.1847.62 ± 0.45
RxR-CESR (%) ↑52.90 ± 0.4255.92 ± 0.1756.19 ± 0.2856.48 ± 0.2358.90 ± 0.31
RxR-CESPL (%) ↑45.15 ± 0.5447.40 ± 0.1647.87 ± 0.3148.01 ± 0.2249.82 ± 0.36

Relative to Combined Context, Context-Think improves mean SR by 2.71/6.00 points on R2R-CE/RxR-CE. We will include seed variability in the corresponding figures and tables.

3. Add ablations with the full training data, including DAgger (W2.2, Q3)

Our findings remain consistent when extended from the original controlled setting to the full training recipe with DAgger data. DAgger requires an initial trained policy to collect additional on-policy demonstrations, so our original reasoning-context study was conducted without this subsequent data-collection stage to isolate the effects of context design.

To address the reviewer's suggestion, we further evaluate Combined Context, Subgoal Planning, and Context-Think using the full training data, including DAgger.

Table S7. Reasoning-context comparison with the full training data, including DAgger.

ConditionR2R-CE SR (%) ↑R2R-CE SPL (%) ↑R2R-CE NE (m) ↓RxR-CE SR (%) ↑RxR-CE SPL (%) ↑RxR-CE NE (m) ↓
Combined Context62.753.94.2167.155.74.64
Subgoal Planning64.255.24.0167.756.04.31
Context-Think (Ours)66.557.03.7270.357.83.92

The same performance ordering holds with DAgger: Context-Think outperforms Subgoal Planning by 2.3/2.6 SR points on R2R-CE/RxR-CE, while both outperform Combined Context. This confirms that our findings extend to the full training recipe rather than depending on the absence of DAgger.

Response to Reviewer 6Hsm (2/2)

4. Clarify the distinction between prior Subgoal Planning and our proposed method (W3)

The review appears to conflate Subgoal Planning, an existing reasoning component, with our proposed Context-Think method. Subgoal Planning has been adopted in prior work, and our controlled study finds that using it alone outperforms combining multiple reasoning components. Building on this finding, we further investigate planning horizon and propose Context-Think with Final-goal Replanning.

Among individual reasoning components, Subgoal Planning achieves the highest mean SR and SPL on both benchmarks. Its advantage over Progress on RxR-CE is modest (56.48 vs. 56.19 SR across three seeds), but all three individual components outperform Combined Context on RxR-CE.

More importantly, Context-Think further improves over Subgoal Planning by 0.97/2.42 SR points on R2R-CE/RxR-CE across three training seeds, demonstrating the additional benefit of Final-goal Replanning.

5. Evaluate combined reasoning with plan-only action context (Q2)

We thank the reviewer for suggesting this controlled experiment to distinguish the effects of reasoning supervision and action-time context.

Plan-only reasoning and action conditioning outperform generating Combined Context and retaining only its plan for action prediction. We train the suggested variant to generate Scene, Progress, and Subgoal at reasoning positions, while supplying only Subgoal for subsequent action prediction. All variants share the same trajectories, action labels, reasoning-update positions, teacher targets, and training settings.

Table S8. Combined reasoning supervision with plan-only action context.

ConditionR2R-CE SR (%) ↑R2R-CE SPL (%) ↑RxR-CE SR (%) ↑RxR-CE SPL (%) ↑
Combined Context50.445.453.045.2
Combined reasoning, plan-only action context50.145.455.046.9
Subgoal-only reasoning and action context52.146.556.347.9

Restricting action-time context to planning improves RxR-CE SR by 2.0 points over Combined Context, but remains 1.3 points below Subgoal-only training. On R2R-CE, the suggested variant remains 2.0 points below Subgoal-only training.

These results demonstrate that the benefits of planning-focused context arise from both reasoning supervision and action conditioning, rather than simply filtering the generated context at inference.

Comments on the paper

1 comment

Dear Reviewers and Area Chairs,

We thank all reviewers for their constructive feedback and recognition of the important research question (Zkvk, kb89), clean controlled experiments (6Hsm, Zkvk), intuitive method (6Hsm, Zkvk, kb89), competitive performance across benchmarks (Zkvk, kb89), and clear presentation (Zkvk, kb89).

To address the shared concerns, we provide additional experiments and clarifications summarized below.

1. Validate the robustness of the main finding (kb89, Zkvk)

Across environments, prompts, and backbones, the finding that more reasoning leads to worse navigation still holds. Combined Context, which contains scene description, progress assessment, and subgoal planning, underperforms subgoal planning alone on both R2R-CE and RxR-CE, evaluated in 11 unseen buildings, and still has the lowest SR on RxR-CE with both an alternative prompt and a Qwen2.5-VL-7B backbone. This degradation cannot be explained by context length, relevance, or reasoning correctness. Our final-goal replanning method, Context-Think, uses a context of nearly the same length but achieves 5.6 points higher SR. In manual annotation, Combined Context is 99.2% relevant and more often correct than progress assessment or subgoal planning alone (83.0% vs. 64.0%/60.2%), yet its SR is lower.

2. Validate training reliability (6Hsm, Zkvk, kb89)

Both this degradation and the gain of final-goal replanning hold across training seeds and with the full training data. Across three seeds, subgoal planning outperforms Combined Context by 1.74/3.58 SR points on R2R-CE/RxR-CE, and final-goal replanning adds another 0.97/2.42 points, improving in every seed. With the full training data including DAgger, which the paper's ablations exclude, the same ordering holds, and final-goal replanning outperforms subgoal planning by 2.3/2.6 SR points.

3. Clarify the distinction from existing planning methods (Zkvk)

Context-Think differs from existing planning methods in planning horizon, retained context, and plan source, which allow future route requirements to inform current action selection while adapting to execution progress. Since the best way to complete a subgoal can depend on what follows it, Context-Think plans the entire remaining route and keeps only this plan. NavGPT and MapGPT instead plan subgoals or map nodes zero-shot while keeping reasoning traces or an online map, and AwareVLN plans only the next navigation phase while keeping Combined Context.

4. Clarify the ablation setup (6Hsm)

The performance differences come only from the reasoning context, as shown by controlled experiments with identical teacher annotations. Combined Context simply concatenates the single-component annotations, and all ablations share the same trajectories and action labels. The same control is kept with the alternative prompt, the Qwen2.5-VL-7B backbone, and all three training seeds.

After the discussion, Reviewer kb89 raised the rating to weak accept, and Reviewer Zkvk stated that our response has largely addressed the concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.