- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes RCSR-D, a corrective post-training method for Vision-and-Language Navigatio designed to extract local supervision from a weak recovery policy without deploying it. When evaluated as a full navigator on a reserved 128-episode cohort, a recovery-tuned Qwen3.5-9B policy fails severely (achieving only 6.25% Success Rate vs. 33.59% for the original policy). RCSR-D bridges this gap by decoupling supervision utility from trajectory execution: an offline margin-based routing rule identifies confident corrective turn proposals, a student model initialized from the original navigator trains via cross-entropy on these turns while matching the original policy via top-k KL divergence on all other states, and parameter interpolation (α=0.5) moderates the final weight update. On the 1,839-episode R2R-CE val-unseen benchmark, RCSR-D improves the 9B model's SR from 37.25% to 46.93% (+9.68 pp) and SPL from 29.36% to 36.28% (+6.92 pp), deploying only a single student policy at runtime.
Critical strengths
-
Valuable Conceptual Insight: Decouples the utility of a policy's action proposals from its closed-loop navigation competence, establishing that a checkpoint with catastrophic full-trajectory failure can still serve as an effective local teacher.
-
Clean Single-Agent Deployment: Unlike runtime mixture-of-experts or multi-policy cascading schemes, the offline selection rule and recovery policy are discarded after distillation, resulting in a single deployed student model using the original inference interface without extra communication or model switching.
-
Statistically Substantial Gains: On the complete 1,839-episode R2R-CE val-unseen split, the method yields statistically robust gains across all metrics (SR +9.68 pp, SPL +6.92 pp, Navigation Error -1.73 m), with 95% bootstrap intervals safely excluding zero and consistent improvements across all 11 evaluation scenes.
-
Rigorously Controlled Baselines: Validates the distillation strategy against direct weight interpolation (θ_0+0.5(θ_r-θ_0 )), demonstrating that RCSR-D significantly outperforms direct parameter blending (+19.34 pp SR on a 512-episode cohort).
Critical weaknesses
-
Inconclusive Ablations for Core Algorithmic Components: The margin-based routing rule (g_τ) fails to show a statistically distinguishable SR advantage over stratified random delegation (+0.78 pp difference, with a 97.5% CI of [-9.38,+10.94]pp).
Soft KL retention shows no statistically significant SR improvement over standard hard-label CE preservation (+5.47 pp difference with a 97.5% CI of [-3.91,+14.84]pp on the 9B model, and slightly negative point estimates in the StreamVLN experiments).
-
Heuristic and Fragile Calibration Pipeline: Thresholds (τ=2.25, α=0.5) and tolerances (e.g., 5-point preservation drops, 3% false Stop limits) are tuned via multi-step heuristics on training and development slices; as shown in Appendix D, identical thresholds failed to meet constraints when applied to StreamVLN.
-
Inconsistent Generalization Across Backbones: Gains on native-action architectures are noticeably smaller or statistically ambiguous (e.g., anchored NaVIDA achieves only a +0.65 pp SR gain over the baseline, with a 95% CI crossing zero at [-1.36,+2.61]pp).
Questions for authors
-
Given that margin-based delegation does not statistically outperform stratified random delegation on the 128-episode cohort (Δ"SR"=+0.78 pp, 97.5% CI includes zero), is the specific margin-thresholding mechanism strictly necessary, or does the benefit derive primarily from exposing the student to any balanced mix of corrective turns alongside reference retention?
-
In both the 9B cohort (Table 4) and the StreamVLN experiments, the SR differences between soft KL and hard-label CE cross zero (with KL yielding lower point estimates in StreamVLN). In light of this, what is the empirical justification for preferring top-k KL divergence over standard, computationally simpler cross-entropy for reference preservation?
-
Parameter interpolation is fixed at α=0.5 based on development checks. How sensitive is the final closed-loop navigation performance on R2R-CE to the choice of α(e.g., across a sweep of α∈{0.0,0.25,0.5,0.75,1.0})?
-
Why does the anchored selective-distillation approach exhibit marginal and inconclusive improvements on NaVIDA (+0.65 pp SR, CI crossing zero) compared to the substantial +9.68 pp gain observed on the Qwen3.5-9B model?
2 comments
Thanks for the rebuttal. I was expecting some numerical evidence, especially for Q1-4. Based on the current rebuttal. I choose to stay with my current rating.
Thank you for the careful review and for highlighting the distinction between useful local proposals and poor navigation performance. We address your questions and the calibration concern below.
The current comparison does not show that it is necessary. On the matched 128 episodes, the primary student achieves 42.97% SR and stratified random delegation achieves 42.19%. The difference is +0.78 points, with a 97.5% paired interval of [-9.38, 10.94] (Table 4). The random control matches supervision counts, source roles, candidate actions, and semantic action-change status. This makes it a comparison of selection rules, but the interval is too wide to establish either superiority or equivalence.
The margin rule selects proposals offline according to empirical recovery and retention criteria (Section 3.2). Those checks do not by themselves prove a closed-loop SR advantage. We also cannot conclude from this comparison that any balanced mix of corrective turns would work equally well.
The motivation in Section 3.3 is to match the reference distribution rather than keep sharpening a single selected target. Specifically, the KL approximation compares distributions renormalized on the reference's top-128 token support. This explains the objective, but is not evidence that KL gives better navigation performance or lower computational cost.
On the 9B cohort, KL minus hard preservation gives an SR difference of +5.47 points, with a 97.5% interval of [-3.91, 14.84], and an NE difference of -0.97 m [-1.91, -0.05] (Section 5.4). The evidence favors KL for NE in that comparison, not for SR. On the two matched StreamVLN seeds, the KL-minus-CE SR differences are -1.09 [-3.15, 0.92] and -0.22 [-2.18, 1.69] points, with 97.5% intervals (Section 5.6). We therefore do not claim that KL is generally a better preservation loss than CE.
As described in Section 3.4, alpha=0.5 was selected during development, fixed before the reported unseen evaluations, and reused across seeds. Those results show performance at that coefficient; they do not answer how closed-loop performance changes across the requested values {0, 0.25, 0.5, 0.75, 1.0}.
The retention constraints and development checks in Sections 3.2-3.4 test behavior on fixed states. They do not guarantee behavior on future trajectories or show that the same thresholds can be used unchanged on another backbone. We can therefore claim performance under the specified calibration settings, not robustness to changes in those settings.
The anchored NaVIDA variant gains +0.65 SR points, with a 95% interval of [-1.36, 2.61] (Section 5.6). Since this interval includes zero, the experiment does not establish a gain. NaVIDA and the 9B model also use different native action interfaces and training recipes, so the comparison cannot isolate why their gains differ. An explanation based on saturation or another mechanism would be speculative.
The 9B full-split gain is +9.68 SR points, with a paired 95% interval of [7.40, 11.96], for one fixed primary training seed (Table 1). That result supports the procedure in the 9B setting, but does not imply a similar gain on NaVIDA.