{"id":"7e1bd2be-2ef0-4584-b066-59dae5fee341","arxiv_id":"2506.17221","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VLN-R1 shows that a Qwen2-VL model trained with supervised fine-tuning plus GRPO-based reinforcement fine-tuning can navigate continuous indoor environments from egocentric video, improving success rate on VLN-CE R2R and RxR benchmarks over its own SFT baselines.","lead":"The authors trained a vision-language model to navigate rooms from first-person camera video, using a two-stage recipe of imitation learning followed by reinforcement learning with a time-weighted reward. It tests whether large multimodal models can become embodied agents that follow natural-language instructions without maps or extra sensors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The dense six-action labels in VLN-Ego are never shown to be correct or unique; since both SFT (Eq. 8) and RFT/TDR (Eq. 9) optimize against them, label generation is the load-bearing gap.","rationale":"The reader's weakest assumption is also the most load-bearing one. The paper's headline comparisons are meaningful only if the six-action supervision in VLN-Ego is a faithful, deterministic expert signal. Since Section 3.2 omits the generation algorithm and no artifacts are released, this is a genuine gap in the central argument rather than a stylistic issue. Baseline coverage, error bars, and test-server evaluation are important secondary concerns, but they concern interpretation and reproducibility; the label-generation problem undermines both training stages and the reward function itself. A focused dataset-generation audit would settle the question. The reader's CONDITIONAL verdict remains appropriate because the issue is addressable by releasing code/data and demonstrating label consistency; it does not by itself show the method is wrong. Hence no change to the verdict is recommended.","tokens_in":13393,"tokens_out":6122,"duration_ms":75636,"concrete_test":"Obtain or reconstruct the VLN-Ego label-generation code. For 100 randomly sampled R2R training timestamps, from the exact saved pose/heading: (a) execute the recorded next-six-action sequence in Habitat and measure the distance to the next reference viewpoint; (b) regenerate labels with an independent shortest-path/controller using randomized tie-breaking and compare. If more than ~5% of executions fail to follow the reference path, or if the regenerated sequences disagree, the dense action supervision is not a reliable reward for either SFT or RFT; if the labels are reproducible and path-consistent, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central result—that GRPO-based RFT improves navigation and that VLN-R1 reaches SOTA—rests on correctness of the next-six-action ground-truth labels in VLN-Ego. Section 3.2 states only that annotations are 'acquired action-by-action' and that 'the ground truth will include the next 6 actions'; it never specifies how R2R/RxR reference paths are converted into fixed atomic sequences (FORWARD 25 cm, TURN 30 deg), nor does it report any quality or uniqueness validation. This matters because Eq. 8 supervises SFT on the full action text and Eq. 9 rewards exact matches to the same labels. In Habitat, a reference waypoint path does not determine a unique low-level action sequence: the mapping depends on the controller, discretization, action ordering, and tie-breaking. If the labels were generated with a shortest-path planner or a hand-tuned controller, the TDR reward is not a verifiable navigation outcome but a proxy for the generator's choices, and the reported RFT gains (30.2 vs 24.9 SR in Table 1) could reflect better imitation of that generator rather than improved navigation. No code, dataset, or checkpoints are released, so this assumption is not checkable from the manuscript. If the labels are noisy, non-unique, or generated under assumptions that differ from the evaluation policy, both training stages inherit the error and the headline RFT conclusion is uninterpretable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VLN-R1, an end-to-end vision-language navigation framework that feeds egocentric video streams to a large vision-language model (Qwen2-VL) and outputs low-level navigation actions (FORWARD, TURN-LEFT, TURN-RIGHT, STOP). Training proceeds in two stages: supervised fine-tuning (SFT) on a newly constructed VLN-Ego dataset of instruction-video-action samples, followed by reinforcement fine-tuning (RFT) with Group Relative Policy Optimization (GRPO) and a proposed Time-Decayed Reward (TDR) that weights exact matches to ground-truth future actions with an exponential temporal decay. The authors evaluate on VLN-CE R2R and RxR val-unseen, reporting that RFT improves success rate (SR) from 24.9 to 30.2 on R2R for the 7B model and from 21.2 to 25.6 for the 2B model, with similar gains on RxR. The paper claims state-of-the-art performance and that RFT enables smaller models to match larger SFT-only models.","tokens_in":13724,"tokens_out":3093,"duration_ms":30978,"significance":"If the reported results are reliable, the paper makes a useful empirical contribution: it demonstrates that GRPO-based reinforcement fine-tuning improves an LVLM's low-level navigation ability beyond SFT, and that the proposed TDR reward, together with long-short memory sampling, is an effective design choice. The internal ablations (action space size, history sampling, number of GRPO generations, reward type) are coherent and support the main RFT conclusion. However, the significance is tempered by two issues: the 'state-of-the-art' claim is not supported by the baselines actually compared, and the ground-truth action labels on which both SFT and RFT rely are not described or validated. The paper also does not release code, data, or checkpoints, so the central assumption about label correctness is not independently checkable. With the label-generation process clarified and comparisons against recent LVLM-based navigation agents added, the work would be a solid contribution to the VLN-CE community.","major_comments":[{"comment":"The construction of the VLN-Ego ground-truth action sequences is not described. Section 3.2 states only that annotations are 'acquired action-by-action' and that 'the ground truth will include the next 6 actions', but does not specify how R2R/RxR reference paths are converted into fixed atomic actions (FORWARD 25 cm, TURN 30 degrees). This is load-bearing because the SFT loss in Eq. (8) and the TDR reward in Eq. (9) both supervise against exactly these labels. In Habitat, a reference waypoint path does not determine a unique low-level action sequence: the conversion depends on the controller, discretization, action ordering, and tie-breaking. If the labels were produced by a shortest-path planner or a hand-tuned controller, the reported RFT gains (e.g., 30.2 vs. 24.9 SR in Table 1) could reflect better imitation of that generator rather than improved navigation ability. Please provide the generation algorithm, discuss non-uniqueness, and report a quality/agreement analysis, or the headline conclusion is uninterpretable.","section":"Section 3.2 and Eqs. (8)-(9)"},{"comment":"The claim of state-of-the-art performance is not supported by the presented baselines. All task-specific baselines in Table 1 and the first two tables are from 2021-2022, while the paper's own related work section (§2) identifies Navid and Uni-Navi as the most relevant LVLM-based navigation agents; neither is benchmarked. The abstract and conclusion both state that VLN-R1 'achieves state-of-the-art navigation performance', but no comparison against these recent approaches is provided. Please include results for Navid and Uni-Navi (and, if feasible, other recent VLN-CE methods) on the same val-unseen splits, or revise the claim to 'strong performance among the evaluated methods'.","section":"Section 4.2, Tables 1 and 2"},{"comment":"The inference-time control loop is unspecified. The paper says the agent outputs a sequence of n=6 future actions and that a navigation episode is successful if the agent stops within a threshold distance, but it does not describe how the re-planning frequency, the STOP action, and episode termination are handled when the model outputs a multi-action sequence. Specifically, does the agent execute all six actions before re-querying the model, and can an intermediate STOP action end the episode? These details directly affect SR and SPL numbers and are needed for reproducibility.","section":"Section 4.1 and Section 3.3"}],"minor_comments":[{"comment":"The figure contains a typo: 'Syetem Message' should be 'System Message'.","section":"Figure 2"},{"comment":"The sentence 'We presents our experimental setup and results here' should read 'We present our experimental setup and results here'.","section":"Section 4.1"},{"comment":"In the qualitative example, 'poast' should be 'past'.","section":"Figure 4"},{"comment":"The table headers use 'Odom.', 'Depth', and 'S.RGB' without defining these abbreviations in the table captions; please add a note explaining the observation modalities.","section":"Tables 1 and 2"},{"comment":"The dataset sizes (630K R2R, 1.2M RxR) would benefit from a brief explanation of how these counts are derived from the 7,189 R2R and 42,023 RxR trajectories, since this affects the reader's understanding of the data engine.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's core empirical claim is plausible and internally consistent, but the missing description of the action-label generation is a serious reproducibility gap because both training stages depend on those labels. The SOTA claim also needs to be benchmarked against the LVLM-based methods the paper itself cites. I recommend major revision rather than rejection, as both issues are addressable with additional description and experiments. I would also flag that the paper's assertion of being the 'first application of RFT for continuous navigation' may be worth softening given the rapidly moving literature on LVLMs for embodied tasks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core result is believable: two-stage SFT then GRPO-based RFT with a time-decayed reward does improve navigation on VLN-CE, and the internal controlled comparison (SFT-only 24.9 SR vs SFT+RFT 30.2 SR for Qwen2-VL-7B) is consistent. The paper also shows the 2B model roughly matching the 7B SFT result after RFT, which is a nice finding and aligns with the DeepSeek-R1 story. The long-short memory sampling and the TDR ablation are sensible contributions. This is a legitimate data-efficient recipe, and the VLN-Ego dataset is a real artifact.\n\nThe biggest soft spot is exactly where the stress-test lands: the paper never explains how the next-six-action ground-truth labels in VLN-Ego are generated from R2R/RxR reference paths. Section 3.2 says annotations are 'acquired action-by-action' and that the ground truth includes the next 6 actions, but there is no description of the action discretization, the controller, tie-breaking, or any validation that the resulting sequences are correct or unique. Since Eq. 8 supervises SFT on those labels and Eq. 9 rewards exact matches to them, the RFT gains could simply reflect better imitation of the label generator rather than genuine navigation ability. This is a load-bearing gap, not a minor omission. Without code, dataset, or a precise description of the label-generation pipeline, the headline result is unverifiable.\n\nSecond, the 'state-of-the-art' claim is not supported by the baselines shown. The comparisons are mostly 2021–2022 task-specific methods; Navid and Uni-Navi are cited but not benchmarked, even though they are the closest LVLM-based continuous navigation works. That is an easy fix, but it needs doing. Also, there are no error bars or test-server results, and the ablation split for hyperparameter selection is ambiguous—it looks like the same Val-Unseen split is used for both development and final reporting. These are addressable, but they matter.\n\nWho is this for? Researchers working on LVLM-based embodied agents and RL post-training for vision-language tasks. The paper deserves a serious referee because the recipe is new and the internal evidence is coherent, but the referee should push hard for label-generation details and a fuller baseline comparison. I would not cite this yet, and I would not bring it to reading group until the label issue is cleared up.\n\nRecommendation: send to peer review, with the expectation of heavy revision rather than acceptance as-is.","headline":"A plausible empirical recipe for turning an LVLM into a continuous navigation policy via GRPO-based RFT, but the paper under-specifies the ground-truth action labels that both training stages depend on, so the headline SOTA claim is not yet verifiable.","tokens_in":14304,"tokens_out":1628,"would_cite":false,"duration_ms":17361,"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":"The paper claims a vision-language model trained with supervised fine-tuning and reinforcement fine-tuning can translate egocentric video plus a natural-language instruction directly into low-level navigation actions, reaching…","keywords":["vision-language navigation","egocentric video","continuous control","reinforcement fine-tuning","group relative policy optimization","time-decayed reward","VLN-CE","large vision-language models"],"falsifier":"A direct test is to rerun the reinforcement stage with the ground-truth action labels in the reward shifted by one time step; if success rates still improve, the reward is teaching the model to mimic label statistics rather than to time its actions correctly, and the navigation claim is weakened.","tokens_in":13165,"feed_emoji":"🧭","tokens_out":9299,"duration_ms":88502,"temperature":0.7,"pith_summary":"VLN-R1 sets out to show that a single large vision-language model can act as an end-to-end navigation agent, converting an egocentric RGB video stream and a natural-language instruction directly into a sequence of low-level actions (forward, turn left, turn right, stop) in a continuous 3D environment. The authors build a large training corpus of instruction-video-future-action samples, first align the model with supervised fine-tuning, then apply reinforcement fine-tuning with a time-decayed reward that weights earlier correct actions more heavily. On the VLN-CE benchmark, this pushes a 7-billion-parameter model to 30.2% success rate on unseen R2R scenes (from 24.9% after supervised fine-tuning alone), and lets a 2-billion-parameter model reach 25.6% after reinforcement fine-tuning, close to the larger model's supervised result. The broader claim is that reward-driven post-training, previously applied to math and code reasoning, transfers to embodied sequential decision-making, so the paper asks readers to see navigation as a downstream skill of general vision-language models.","feed_headline":"RL fine-tuning lifts a 7B vision-language model to 30.2% success","feed_subtitle":"Time-decayed action rewards let a video-trained vision-language model beat graph-based planners on unseen scenes.","key_machinery":"The load-bearing component is the Time-Decayed Reward (TDR): for a predicted block of $n$ future actions, it sums $\\gamma^k$ over exactly those positions where the predicted action matches the ground-truth action, so the first action contributes weight $1$, the second $\\gamma$, and so on. This turns action prediction into a softly graded, position-aware optimization target rather than an all-or-nothing text match. It is paired with group-relative advantage normalization, which compares each sampled completion against its group, and with a frame-sampling scheme (dense sampling of recent frames plus sparse sampling of older frames) that feeds both immediate context and long-horizon memory into the vision-language model.","core_discovery":"The central discovery, stated on the paper's own terms, is that low-level continuous navigation can be reduced to a text-generation problem over a compact action vocabulary, and that this behavior is improved by reinforcement fine-tuning with a temporally weighted reward. The model predicts the next six actions as text; supervised fine-tuning first imitates expert demonstrations, and then group-relative policy optimization is run with the Time-Decayed Reward $$R_{\\text{nav}}=\\sum_{k=0}^{n-1}\\gamma^k\\,\\mathbb{1}(\\alpha_{t+k}=\\$\\alpha$^*_{t+k}),$$ which rewards correct matches but discounts later steps geometrically. The reported numbers on the unseen validation split are the evidence: SFT alone gives 24.9 SR on R2R for the 7B model, and adding RFT raises this to 30.2 SR; on RxR, reinforcement fine-tuning with only 10K cross-domain samples outperforms training on the full dataset. The authors read these results as proof that large vision-language models can drive embodied navigation and that verifiable-reward reinforcement post-training is a data-efficient way to teach them the skill.","pith_inferences":["The authors leave implicit that the time-decayed reward is a generic temporal-credit-assignment trick; the same functional form could be lifted to manipulation or driving benchmarks where early action correctness dominates task success.","Because the six-action ground truths are simulator-derived, a natural next test is to vary the expert that produces them; if the reward's gains vanish with a different expert, the method is learning one expert's action style, not navigation generally.","The four-action discrete vocabulary is the main limit on fine-grained control; a natural extension, not tested here, is to let the model emit continuous parameters such as distance and angle in the same text format.","If the 'small model plus RFT matches large model plus SFT' pattern generalizes, it suggests the efficient frontier for embodied agents may involve spending compute at post-training time rather than at pretraining time."],"forward_implications":["Egocentric RGB video plus language is enough for a vision-language model to produce low-level navigation commands, so map, depth, and odometry modules are not required at inference time.","Reinforcement fine-tuning with a temporally weighted reward improves navigation success beyond supervised fine-tuning alone, so RL post-training can be a data-efficient second stage for embodied tasks.","A 2-billion-parameter model after reinforcement fine-tuning matches a 7-billion-parameter model trained only with supervision, suggesting model scale can be traded for post-training compute.","On the RxR benchmark, fine-tuning on 10K cross-domain samples after R2R pretraining outperforms training on the full dataset, indicating fast cross-domain transfer through reinforcement post-training.","Predicting a six-step action sequence outperforms single-action prediction, so multi-step temporal supervision is itself a source of navigation performance."],"supporting_citations":[{"why":"Defines the R2R vision-and-language navigation task and provides the instruction and path data used to build the training corpus.","marker":"[2]"},{"why":"Introduces the VLN-CE continuous-environment benchmark and the SR, OS, SPL, NE, and TL metrics used in all evaluations.","marker":"[23]"},{"why":"Provides the Room-Across-Room instructions and trajectories that form the second training and evaluation dataset.","marker":"[24]"},{"why":"Supplies the verifiable-reward reinforcement learning methodology and the observation that RL post-training can boost smaller models.","marker":"[14]"},{"why":"Introduces the group-relative policy optimization used as the reinforcement fine-tuning optimizer.","marker":"[49]"},{"why":"Defines the base vision-language model architecture that the navigation policy is built on.","marker":"[55]"},{"why":"Provides the waypoint-based continuous navigation baseline and the contrast for frame-history sampling strategies.","marker":"[21]"},{"why":"Describes a prior video-based VLM planner that selects waypoints rather than low-level actions, against which VLN-R1 positions itself.","marker":"[60]"}],"fun_headline_variants":["VLN-R1: RL fine-tuning lifts VLM navigation to 30.2%","Time-decayed rewards teach VLMs continuous navigation","7B VLM navigates unseen scenes via RL fine-tuning","GRPO with time-decayed rewards boosts SR to 30.2","Video-driven VLM beats graph planners via RL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the six-action ground-truth sequences generated from simulator trajectories are correct, unique, and noise-free expert demonstrations, even though the paper does not describe how those action sequences are produced or validate their quality; if the labels are wrong or non-unique, both the supervised loss and the reinforcement reward inherit the error, and the reported gains measure label alignment rather than navigation ability.","fun_headline_variants_meta":{"raw":{"variants":["VLN-R1: RL fine-tuning lifts VLM navigation to 30.2%","Time-decayed rewards teach VLMs continuous navigation","7B VLM navigates unseen scenes via RL fine-tuning","GRPO with time-decayed rewards boosts SR to 30.2","Video-driven VLM beats graph planners via RL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1334,"prompt_tokens":1007,"completion_tokens":327,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":239}},"tokens_in":623,"tokens_out":327,"duration_ms":3794,"temperature":1.0,"reasoning_tokens":239,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:09:34.928837+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to rerun the reinforcement stage with the ground-truth action labels in the reward shifted by one time step; if success rates still improve, the reward is teaching the model to mimic label statistics rather than to time its actions correctly, and the navigation claim is weakened.","supporting_citations":[{"cited_title":"Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments","cited_arxiv_id":null,"evidence_quote":"Defines the R2R vision-and-language navigation task and provides the instruction and path data used to build the training corpus."},{"cited_title":"Beyond the nav-graph: Vision and language navigation in continuous environments","cited_arxiv_id":null,"evidence_quote":"Introduces the VLN-CE continuous-environment benchmark and the SR, OS, SPL, NE, and TL metrics used in all evaluations."},{"cited_title":"Room-across- room: Multilingual vision-and-language navigation with dense spatiotemporal grounding","cited_arxiv_id":null,"evidence_quote":"Provides the Room-Across-Room instructions and trajectories that form the second training and evaluation dataset."},{"cited_title":"Waypoint models for instruction-guided navigation in continuous environments","cited_arxiv_id":null,"evidence_quote":"Provides the waypoint-based continuous navigation baseline and the contrast for frame-history sampling strategies."},{"cited_title":"Navid: Video-based vlm plans the next step for vision-and-language navigation","cited_arxiv_id":null,"evidence_quote":"Describes a prior video-based VLM planner that selects waypoints rather than low-level actions, against which VLN-R1 positions itself."}],"review_version":2}