{"id":"7ba1eff9-d950-4cf1-9cbf-37a8722e220c","arxiv_id":"2507.12911","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"LaViPlan uses RLVR with GRPO and ADE/FDE rewards to fine-tune a 2B VLM for trajectory prediction, improving ADE/FDE on ROADWork and a normalized safety score on CODA-LM over supervised fine-tuning.","lead":"LaViPlan fine-tunes a vision-language model with reinforcement learning that rewards accurate future-driving trajectories, improving path prediction on construction-zone and corner-case datasets. The approach is a step toward making language-guided driving models output safe, low-level actions, but the reported gains are small and lack statistical guarantees.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The out-of-domain generalization claim is unsupported: the CODA-LM safety score (Eq. 7) is a relative min-max rank over only three models with no raw F/C/P values or ground-truth trajectories, so it cannot show that LaViPlan plans better on OOD scenes.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern. The in-domain half of the central claim is moderately supported: Table 5 shows consistent improvements from reinforcement fine-tuning over SFT on a held-out-style validation split, and since the RLVR reward is literally ADE/FDE, the direction of the effect is expected. However, no error bars are reported and the validation split's disjointness from the 5,430-sample training pool is only asserted indirectly in Appendix B. The more severe problem is the OOD evaluation. Equation 7 normalizes three metrics over k = {Baseline, SFT(5k), LaViPlan}, so the resulting safety score is a relative ranking rather than a calibrated safety measure; raw F/C/P values are never given. CODA-LM has no ground-truth trajectories, and bounding-box-line intersection in the image plane is an unvalidated proxy for drivability or collision. Appendix D provides only qualitative examples. A secondary contribution claim about 'fewer training samples' is also not supported by the setup: LaViPlan uses 4,344 SFT + 1,086 RFT = 5,430 samples, the same total as the SFT baseline, so saying 'RLVR requires fewer samples' ignores the prerequisite Phase 1 SFT. These issues are addressable with reporting and additional evaluation, so a conditional acceptance remains appropriate rather than a rejection.","tokens_in":14135,"tokens_out":8107,"duration_ms":99000,"concrete_test":"Report raw Fail Rate, Collision Count, and Penetration Length (with scene count and bootstrap 95% CIs) for Baseline, SFT(5k), and LaViPlan on the full CODA-LM split, and recompute the safety scores after adding at least Senna and DriveLM to the min-max normalization set. If LaViPlan's raw metrics are not strictly better than SFT's, or if its rank under Eq. 7 flips when the normalization set changes, the OOD generalization claim is an artifact of the three-model relative score.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LaViPlan improves planning performance out-of-domain is carried entirely by Table 4, computed from Eq. 7. Each safety score is a weighted min-max normalization of Fail Rate, Collision Count, and Penetration Length across the models k in the evaluation set. With only three models (Baseline, SFT(5k), LaViPlan), this is essentially a rank transformation, not a calibrated safety measure: a model can receive a high normalized score while remaining unsafe in absolute terms if all three models collide frequently, and small absolute differences are inflated by the min-max denominator. The paper never reports raw F, C, or P for any model, nor the number of CODA-LM scenes used, nor confidence intervals. Because CODA-LM has no ground-truth trajectories, the score also does not establish that LaViPlan's outputs are drivable or realistic; only qualitative examples in Appendix D are offered. A related weakness is that the 'fewer training samples' contribution is not supported by the experimental design: LaViPlan consumes 4,344 SFT samples plus 1,086 RFT samples, the same total as the SFT baseline, so the comparison cannot demonstrate data efficiency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LaViPlan, a two-stage framework for vision-language path planning. Stage 1 is supervised fine-tuning (SFT) of a VLM on image-instruction-trajectory pairs; Stage 2 is reinforcement fine-tuning with verifiable rewards (RLVR) using GRPO, where the reward combines a format reward and a planning reward based on ADE and FDE. The method is evaluated on the ROADWork dataset in-domain and on CODA-LM as a zero-shot, out-of-domain benchmark. The reported results show improved ADE/FDE on ROADWork relative to SFT baselines, higher relative safety scores on CODA-LM, and a trade-off in which linguistic fidelity, measured by BERTScore and NLI, decreases while planning-oriented reasoning is claimed to improve. The paper also includes ablations on easy-to-hard sampling ratios and on the presence of explicit reasoning in the two fine-tuning phases.","tokens_in":14408,"tokens_out":4366,"duration_ms":46211,"significance":"If the central claims were fully established, the contribution would be a useful and timely application of RLVR to low-level trajectory planning in autonomous driving, complementing the more common high-level decision use of VLMs. The training recipe is simple and largely reproduces standard GRPO machinery, which is a strength in terms of reproducibility, and the paper uses a publicly available benchmark for the in-domain evaluation. However, the current evidence does not support the full strength of the abstract's claim of out-of-domain improvement, and the 'fewer training samples' contribution is not supported by the experimental comparison. The paper also reports no variance or multiple-run statistics, so the small in-domain gains are not yet statistically grounded.","major_comments":[{"comment":"The out-of-domain safety score is a relative min-max normalization computed across only three models (Baseline, SFT (5k), LaViPlan), and the raw values of Fail Rate, Collision Count, and Penetration Length are never reported. With three models, Eq. (7) is essentially a rank transform, so the claim that LaViPlan 'achieves the highest safety scores' only means that it ranks first among these three; it does not establish that any of the trajectories are safe or even drivable in an absolute sense. The paper also does not state the number of CODA-LM scenes used, does not report confidence intervals, and CODA-LM has no ground-truth trajectories, so the score cannot measure trajectory accuracy or realism. The qualitative examples in Appendix D are suggestive but cannot replace a quantitative, absolute, or calibrated evaluation.","section":"§4.2.1, Eq. (7), Table 4"},{"comment":"The contribution statement that RLVR 'requires significantly fewer training samples compared to supervised fine-tuning' is not supported by the experimental design. The paper states that the SFT baselines are trained on the full 5K samples, while LaViPlan uses 4,344 SFT samples plus 1,086 RFT samples, so the total number of samples is the same as (or slightly greater than) the full dataset. Thus, the comparison cannot demonstrate data efficiency; it demonstrates that adding an RFT phase on top of an SFT phase can improve results, which is a different claim. The authors should either compare LaViPlan with an SFT baseline trained on 4,344 samples only (which appears in Table 5 but is not the headline comparison) or rephrase the claim to be about the relative contribution of the RFT phase rather than about sample efficiency.","section":"§4, Table 2 and Table 4"},{"comment":"The in-domain improvements are reported as single numbers without error bars, multiple seeds, or significance tests. The gains are often small in absolute terms (for example, ADE Easy 4.12 for SFT (4k) versus 3.62 for LaViPlan in Table 5, and ADE Hard 5.31 versus 4.83), and the comparison in Table 5 is confounded because LaViPlan receives an additional 1,086 RFT samples that the SFT (4k) model does not. Without repeated runs or at least a clear statement of variance, the central claim that RLVR 'improves planning performance' beyond SFT is not yet statistically established. The authors should provide at least three seeds with means and standard deviations for the main ADE/FDE experiments.","section":"§4.1 and §4.3, Tables 2, 5, 6"},{"comment":"The out-of-domain ablation in Table 7 is used to draw the conclusion that 'out-of-domain robustness requires greater emphasis on hard cases.' However, the evaluation is again the relative min-max safety score from Eq. (7), so the differences between the 9:1, 7:3, and 6:4 ratios are relative comparisons across the same small set of models. The conclusion that a particular easy-to-hard ratio is best for OOD generalization would require an absolute safety measure, raw metric values, and ideally more than one run or a statistical comparison. As written, the domain-sensitivity claim in the text is stronger than the evidence in Table 7.","section":"§4.3.2, Tables 6 and 7"}],"minor_comments":[{"comment":"There is a typo: 'Qwne2VL-2B-Instruct' should be 'Qwen2VL-2B-Instruct', and the model name should be consistent with Table 2 and the rest of the paper.","section":"Figure 2 caption"},{"comment":"The captions of Figures 4 and 5 appear contradictory: Figure 4 says the trajectories in Deasy exhibit lower x-variance than Dhard, while Figure 5 says the trajectories exhibit lower x-variance than Deasy. The captions should be corrected to state the intended ordering.","section":"Figures 4 and 5 captions"},{"comment":"Table 7 has formatting problems: the rows for different models are run together without line breaks, making the table difficult to read. The table should be reformatted with each model on a separate row.","section":"Table 7"},{"comment":"The GRPO objective in Eq. (4) writes the KL divergence term outside the token-level expectation without explicitly defining the expectation over the group and sequence. Clarifying the notation would improve the presentation.","section":"Eq. (4)"},{"comment":"The set of models k used in the min-max normalization is not explicitly defined. The paper should state that k ranges over the three compared models and specify whether the normalization is performed per evaluation run or pooled across runs.","section":"Eq. (7)"},{"comment":"The random sampling steps in Algorithm 1 and the construction of Dhard use randomness, but no random seed is reported. Adding the seed would improve reproducibility.","section":"Appendix B, Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a plausible and timely application of RLVR to trajectory planning, and the in-domain results may well be correct. However, the two central headline claims, out-of-domain generalization and data efficiency, are not supported by the current experimental evidence: the OOD evaluation is a relative rank score over three models, and the training-sample comparison does not actually use fewer samples. The lack of error bars is also a recurring issue. These are load-bearing but fixable within the paper's scope, so I recommend major revision rather than rejection. I would encourage the editor to require raw F/C/P values on CODA-LM, a corrected data-efficiency comparison, and multi-seed statistics for the main tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent but incremental application of RLVR/GRPO to VLM trajectory prediction. The in-domain numbers are probably real; the OOD safety claim is not supported by the evidence as written.\n\nWhat's actually new: applying ADE/FDE as verifiable rewards to a VLM trajectory predictor, and the observation that RLVR trades linguistic fidelity for planning accuracy. The easy-vs-hard sample split for the RL stage is a reasonable extension of Visual-RFT/AlphaDrive, not a new method.\n\nWhat the paper does well: the in-domain setup is straightforward and the gains over SFT are consistent — roughly 10% ADE/FDE reductions on both easy and hard splits. The ablations on reasoning and on easy/hard ratios are useful. The discussion section is honest about sparse rewards and about GRPO depending on a strong SFT initialization. Credit for not overselling the linguistic fidelity — the BERTScore/NLI drop is reported plainly.\n\nSoft spots, in proportion:\n\n(1) The OOD claim is load-bearing and weak. Table 4 uses Eq. 7, a min-max normalization over only three models. That is a rank transform, not a safety metric. No raw fail rates, collision counts, or penetration lengths are reported, no scene count, no confidence intervals. With three models, a model can score high while still colliding in every scene. The paper cannot claim \"improves planning performance across OOD datasets\" on this evidence. The qualitative appendix is nice but does not carry that weight.\n\n(2) \"Fewer training samples\" is not supported by the experimental design. LaViPlan uses 4,344 SFT + 1,086 RFT samples, which equals the 5,430 the SFT baseline gets. If the claim is that RLVR needs fewer samples than SFT, the comparison should hold total data fixed and vary the split, or compare against an SFT baseline trained on only 4,344 samples.\n\n(3) Minor but real: no error bars, single seed, and the main results use the 6:4 easy/hard ratio but this is only clear in the ablation table. The paper's own tables are otherwise clean.\n\nI'd send this to review. The in-domain result is plausible and worth checking, and the OOD evaluation is fixable if the raw numbers are reported. But I would not accept it as-is. The core recipe is not novel enough to excuse the missing evidence.","headline":"Competent, incremental RLVR-for-trajectory-planning paper with plausible in-domain gains, but the OOD safety claim is built on a relative min-max score over three models and the data-efficiency claim is not supported by the experimental design.","tokens_in":14893,"tokens_out":2529,"would_cite":false,"duration_ms":28628,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Reinforcement rewards, not imitation, sharpen VLM driving plans","keywords":["reinforcement learning with verifiable rewards","vision-language models","trajectory planning","autonomous driving","out-of-distribution generalization","GRPO","average displacement error","final displacement error"],"falsifier":"Run LaViPlan's predicted trajectories through a closed-loop simulator with a rule-based safety checker, or compare them against human-labeled drivable paths on CODA-LM, and measure absolute collision and off-road rates without min-max normalization; if these absolute rates are no better than supervised fine-tuning's, the claimed out-of-domain safety advantage disappears.","tokens_in":13953,"feed_emoji":"🚗","tokens_out":3899,"duration_ms":41902,"temperature":0.7,"pith_summary":"LaViPlan claims that fine-tuning a vision-language model with reinforcement learning whose reward is trajectory error (ADE/FDE) improves trajectory planning for autonomous driving, both on the construction-zone dataset it trained on and on an unseen corner-case dataset. The paper reports that LaViPlan beats supervised fine-tuning on in-domain ADE/FDE and achieves the highest normalized safety scores on out-of-domain CODA-LM, while using only about one-fifth of the data for the reinforcement stage. It argues this closes the gap between a VLM's language reasoning and the low-level paths it outputs. The authors also document a drop in linguistic fidelity after reinforcement fine-tuning, interpreting it as a shift from verbose scene description toward hazard-focused reasoning that better supports planning.","feed_headline":"Reward-tuned VLM beats supervised fine-tuning on driving trajectories","feed_subtitle":"Planning-error rewards improve path predictions in-domain and on unseen corner cases with fewer training samples.","key_machinery":"The load-bearing mechanism is Group Relative Policy Optimization (GRPO) applied to a vision-language policy, with a compound verifiable reward $R = R_{\\text{format}} + R_{\\text{planning}}$. $R_{\\text{planning}}$ is the negative logarithm of $(1 + \\text{ADE})$ plus $(1 + \\text{FDE})$, making trajectory error a differentiable signal in the token-level policy update, while $R_{\\text{format}}$ checks the required <think>/<answer> structure. Advantages are computed from group statistics over $G$ sampled trajectories, and the update is clipped and KL-regularized against the supervised reference model. This turns a scalar planning error into a preference signal among candidate trajectories, which the paper argues gives richer supervision than a single ground-truth imitation target.","core_discovery":"The central claim is that planning-oriented metrics, used as verifiable rewards in GRPO-style reinforcement fine-tuning, realign a VLM's trajectory outputs with its visual reasoning and improve generalization. Given an image and a trajectory-prediction prompt, the model must emit reasoning inside <think> tags and 20 image-plane coordinates inside <answer> tags; the reward adds a format component to a log-smoothed ADE/FDE trajectory-match term, and the policy is kept near a supervised reference model by KL regularization. On ROADWork, LaViPlan achieves ADE 3.62/4.83 and FDE 3.85/6.09 on the easy/hard splits, outperforming every supervised baseline, and on CODA-LM it scores highest under balanced, safety-focused, and equal weighting schemes, with only a fifth of the training budget allocated to reinforcement. The paper treats this as evidence that RLVR is a viable post-training paradigm for aligning language reasoning with action-level driving plans.","pith_inferences":["The out-of-domain safety score on CODA-LM is a relative ranking among only three models, not an absolute measure of drivability, so the OOD advantage should be read as suggestive until absolute collision or closed-loop metrics are reported.","Because the reward is sparse, computed only after the full trajectory is generated, the method would likely benefit from dense step-wise rewards or auxiliary collision penalties, a direction the paper itself flags in its discussion.","The same recipe could transfer to other structured trajectory-output tasks, such as pedestrian motion forecasting or robot arm path planning, wherever a trajectory error can be computed at the end of a formatted response.","Since the policy is KL-regularized to a supervised backbone, the ceiling of the approach is partly set by the quality of the supervised fine-tuning stage; a stronger base VLM or better SFT phase would probably shift final trajectory quality as well."],"forward_implications":["If RLVR with planning rewards works as claimed, VLM planners can be improved without collecting many new labeled trajectories: only 1,086 reinforcement samples were needed on top of 4,344 supervised ones.","The in-domain gains, roughly 9 to 13 percent ADE/FDE reductions over supervised fine-tuning, suggest that preference among candidate trajectories provides supervision beyond simple imitation.","The easy-to-hard sampling ratio becomes a tunable lever: a 6:4 ratio favors in-domain accuracy while a 7:3 ratio favors out-of-domain safety, implying training-set composition should be chosen according to the target deployment.","The drop in linguistic fidelity alongside better trajectories implies that safety-critical evaluation should weight functional relevance of reasoning over textual similarity to human annotations."],"supporting_citations":[{"why":"Supplies the GRPO algorithm that the policy update in LaViPlan is built on.","marker":"[10]"},{"why":"Establishes the RLVR approach of using rule-based verifiable rewards for vision-language models, which LaViPlan adapts to planning.","marker":"[20]"},{"why":"Preceding work applying reinforcement learning and reasoning to VLM autonomous driving, motivating the setting.","marker":"[15]"},{"why":"Provides the ROADWork dataset used for both the supervised and reinforcement fine-tuning phases.","marker":"[9]"},{"why":"Supplies the CODA-LM corner-case dataset on which out-of-distribution generalization is evaluated.","marker":"[2]"},{"why":"Source of the min-max normalized safety-score aggregation procedure used for the out-of-domain evaluation.","marker":"[13]"},{"why":"Cited theoretical analysis of GRPO dynamics that justifies why RLVR amplifies success and converges, supporting the method's premise.","marker":"[21]"}],"fun_headline_variants":["Reward-tuned VLM outdrives supervised fine-tuning","Planning-error rewards align VLM reasoning with trajectories","RLVR fine-tuning boosts VLM generalization on unseen scenes","VLM path planning improved by RL with planning metrics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The out-of-distribution safety result rests on a min-max normalized safety score computed from bounding-box intersections on CODA-LM, which has no ground-truth trajectories and only three models in the normalization set, so the score is relative rather than an absolute measure of drivability.","fun_headline_variants_meta":{"raw":{"variants":["Reward-tuned VLM outdrives supervised fine-tuning","Planning-error rewards align VLM reasoning with trajectories","RLVR fine-tuning boosts VLM generalization on unseen scenes","VLM path planning improved by RL with planning metrics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1451,"prompt_tokens":927,"completion_tokens":524,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":459}},"tokens_in":543,"tokens_out":524,"duration_ms":6350,"temperature":1.0,"reasoning_tokens":459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:34:54.508316+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run LaViPlan's predicted trajectories through a closed-loop simulator with a rule-based safety checker, or compare them against human-labeled drivable paths on CODA-LM, and measure absolute collision and off-road rates without min-max normalization; if these absolute rates are no better than supervised fine-tuning's, the claimed out-of-domain safety advantage disappears.","supporting_citations":[{"cited_title":"Roadwork: A dataset and benchmark for learn- ing to recognize, observe, analyze and drive through work zones","cited_arxiv_id":null,"evidence_quote":"Provides the ROADWork dataset used for both the supervised and reinforcement fine-tuning phases."},{"cited_title":"Au- tomated evaluation of large vision-language models on self- driving corner cases","cited_arxiv_id":null,"evidence_quote":"Supplies the CODA-LM corner-case dataset on which out-of-distribution generalization is evaluated."},{"cited_title":"Bench2drive: Towards multi-ability bench- marking of closed-loop end-to-end autonomous driving","cited_arxiv_id":null,"evidence_quote":"Source of the min-max normalized safety-score aggregation procedure used for the out-of-domain evaluation."}],"review_version":1}