- 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
-
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?
-
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.
-
Equation (2) is underspecified. The authors should explain how and are computed and normalized, and specify the value of .
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 (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).
| Benchmark | Metric | Full | Scene | Progress | Subgoal | Final-goal |
|---|---|---|---|---|---|---|
| R2R-CE | SR (%) ↑ | 50.44 ± 0.19 | 50.34 ± 0.16 | 51.47 ± 0.22 | 52.18 ± 0.30 | 53.15 ± 0.70 |
| R2R-CE | SPL (%) ↑ | 45.39 ± 0.25 | 45.51 ± 0.15 | 46.03 ± 0.12 | 46.55 ± 0.18 | 47.62 ± 0.45 |
| RxR-CE | SR (%) ↑ | 52.90 ± 0.42 | 55.92 ± 0.17 | 56.19 ± 0.28 | 56.48 ± 0.23 | 58.90 ± 0.31 |
| RxR-CE | SPL (%) ↑ | 45.15 ± 0.54 | 47.40 ± 0.16 | 47.87 ± 0.31 | 48.01 ± 0.22 | 49.82 ± 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 throughout. Let and denote the action and reasoning queries, respectively, and let be the query input:
Here, and 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.
| Method | SR (%) ↑ | SPL (%) ↑ | NE (m) ↓ |
|---|---|---|---|
| GPT-6-Astra (zero-shot) | 54.00 | 50.53 | 5.50 |
| Context-Think | 72.00 | 63.59 | 4.07 |
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.
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.
Table S2. Qwen2.5-VL-7B backbone on RxR-CE Val-Unseen.
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.