{"id":"be8cd465-3e8a-49ad-a7ef-45db1d6f116b","arxiv_id":"2412.05552","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"One navigation model with state-adaptive mixture-of-experts routing matches or exceeds task-specific agents on several of seven navigation benchmarks.","lead":"Researchers trained one navigation model to handle seven different language-guided indoor navigation tasks at once, using a mixture-of-experts design that switches specialist modules based on what the agent currently sees and hears. The model matches or beats separately trained specialists on several benchmarks, suggesting a single agent can cover multiple instruction styles.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ObjectNav evidence rests on a filtered, unreleased discrete projection of standard MP3D validation episodes; without retained-episode counts or a full-set evaluation, the seven-task and task-specific-comparison claims are not yet controlled.","rationale":"The reader's weakest assumption identifies the same area I would attack: the unvalidated discrete projection of continuous ObjectNav data. I agree and would sharpen the point: the filtering is applied not only to training data but also to the MP3D validation split, so the evaluation set itself may be a strict subset of the standard benchmark. The paper gives no episode counts before and after filtering, and the average step compression from 243 to 20 indicates a major distributional shift. This is more consequential than the other candidate concerns. The RxR-EN and REVERIE gaps versus task-specific agents are real but are mostly about the vague phrase 'highly comparable' and can be addressed by calibration or rewording; the single-run ablations are an experimental-rigor issue, but the main MoE-vs-baseline improvements are large and consistent. In contrast, a filtered ObjectNav evaluation set would invalidate one of the seven benchmark comparisons and the claim of beating Habitat-Web specifically. The proposed check is cheap and decisive: evaluate on the full standard validation split and report the filtering dropout rate. The paper otherwise has a controlled multi-task baseline (ScaleVLN-dagger) and a clear architecture ablation, so a conditional verdict is appropriate until this check is run.","tokens_in":23630,"tokens_out":8887,"duration_ms":80830,"concrete_test":"Recompute the MP3D ObjectNav row of Table 5 on the full, unfiltered standard validation split: run SAME plus the waypoint predictor on all original continuous episodes in Habitat, and report both the number of validation episodes removed by the 0.5m endpoint filter and the SR/SPL on the complete set. If the filtered subset drops more than a few percent of episodes, or if full-set SR is materially lower than 43, the ObjectNav claim and the seven-task headline need revision; if the numbers are essentially unchanged, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The ObjectNav result is load-bearing because it is one of the seven tasks in the headline claim, and the only cross-method comparison is Table 5, where SAME is compared to continuous-space methods (SemExp, PONI, Habitat-Web). Section 2.2 describes converting continuous Habitat-Web demonstrations to the discrete connectivity graph by Euclidean nearest-node matching, merging repeated nodes, and removing disconnected paths and paths whose endpoint is more than 0.5m from the original endpoint. It then states: 'Similarly, we transfer the data from the MP3D validation split to evaluate model performance in discrete environments.' The paper never reports how many standard MP3D validation episodes survive this filtering, nor how the dropped episodes distribute across scenes and goal categories. Because the original demonstrations average 243 steps while the projected trajectories average 20, the projection heavily compresses exploration behavior. If the 0.5m endpoint filter preferentially removes long, hard episodes or shortens them to near-target paths, the evaluation set is easier than the standard ObjectNav benchmark. Consequently, SAME's 76.3 SR / 42.7 SPL in Table 4 and 43 SR / 21 SPL in Table 5 may not be measuring the same skill as the continuous-space baselines, which are evaluated on unfiltered standard episodes. The Table 4 discrete ObjectNav row has no direct task-specific competitor, so the entire ObjectNav superiority claim rests on Table 5's comparability. This is a testable validity issue rather than a disagreement with consensus, and it can be resolved with a small amount of additional reporting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAME, a State-Adaptive Mixture of Experts model for unifying seven language-guided navigation tasks into a single policy. The model is built on the DUET architecture, initializes from ScaleVLN, and routes among experts based on a learned combination of the current visual observation and the language instruction rather than on task identity or individual tokens. The authors report experiments on R2R, RxR-EN, REVERIE, CVDN, SOON, ObjectNav-MP3D, and R2R-CE, and claim that one set of weights outperforms or is highly comparable to task-specific agents. The main evidence is a controlled comparison against a same-initialization multi-task DUET baseline (ScaleVLN†), plus a series of ablations on routing signal, expert placement, and training schema.","tokens_in":23983,"tokens_out":6378,"duration_ms":57837,"significance":"If the claims hold, the paper makes a useful contribution: it demonstrates that a single navigation policy can handle instruction granularities from fine-grained VLN to zero-grained ObjectNav, and the idea of routing experts based on the agent's multimodal state is well motivated. The experimental design comparing SAME against a multi-task DUET baseline initialized from the same ScaleVLN weights is a strength, because it isolates the effect of the MoE mechanism. The routing ablations in Tables 2 and 3 are informative and go beyond simply reporting final numbers. However, the seven-task claim is currently not fully supported: the ObjectNav evidence rests on an unvalidated discrete projection of continuous demonstrations, the REVERIE and SOON gaps against specialists contradict the 'highly comparable' wording, and all key comparisons are single runs without variance estimates.","major_comments":[{"comment":"The ObjectGoal Navigation evidence rests on an unvalidated projection of continuous Habitat-Web demonstrations onto the discrete MP3D connectivity graph. The text states that each viewpoint is matched to the nearest node, repeated nodes are merged, and 'disconnected paths and paths with an ending position more than 0.5m away from the original endpoint are removed,' yielding 58,803 training trajectories with average length 20 steps from demonstrations of average length 243 steps. Crucially, the paper does not report how many standard MP3D validation episodes survive the same filtering, nor how the dropped episodes are distributed across scenes and goal categories. If the filter preferentially removes long or hard episodes, SAME's Table 4 ObjectNav row (76.3 SR/42.7 SPL) and the Table 5 comparison (43 SR/21 SPL vs SemExp 28/11, PONI 32/12, Habitat-Web 35/10) are not measuring the same skill as the continuous-space baselines evaluated on unfiltered standard episodes. The authors should report retained-episode statistics, evaluate on the full MP3D validation split without endpoint filtering, or provide a continuous-space evaluation of SAME before the ObjectNav part of the seven-task claim can be accepted.","section":"Section 2.2, Tables 4 and 5"},{"comment":"The headline claim 'outperforms or achieves highly comparable performance to task-specific agents' is contradicted by the REVERIE and SOON rows of Table 4. On REVERIE Val Unseen, SAME obtains 46.4 SR/36.1 SPL versus 57.0/41.8 for ScaleVLN and 56.0/39.7 for VER; on SOON Val Unseen, SAME obtains 36.1 SR versus 40.4 for GOAT and 57.0 for ScaleVLN. These are gaps of 10-21 SR points, not 'highly comparable.' The sentence 'SAME performs at the same level as VER and GOAT comparing the SPL on R2R and REVERIE' refers only to SPL and does not cover SR. The claim should be restricted to the tasks where parity actually holds (CVDN, R2R in part), or the abstract and introduction should be revised to describe SAME as a competitive unified agent with task-specific trade-offs rather than as one that outperforms or matches specialists across all seven tasks.","section":"Section 4.1 and Abstract"},{"comment":"The R2R-CE result depends on an external waypoint predictor: the text says, 'We follow Hong et al. [37] and deploy the waypoint predictor to bridge the gap between discrete and continuous,' but the paper does not specify which checkpoint was used, how the predictor was trained, or whether it is fixed for all compared methods. Moreover, SAME's R2R-CE numbers (47 SR/38 SPL) are below ScaleVLN (55/51) and BEVBert (59/50), so R2R-CE cannot be cited as evidence of outperformance. The paper does state that this is zero-shot inference, but the abstract-level seven-task claim does not carry this caveat. Please either remove R2R-CE from the 'outperforms or highly comparable' statement, provide a properly trained R2R-CE variant, or report full details of the waypoint module and clarify that the R2R-CE numbers are zero-shot with an external component.","section":"Section 4.1, Table 5"},{"comment":"All ablations and the main SAME-vs-ScaleVLN† comparison are single runs with no variance estimates. Since the claimed advantage over the same-initialization baseline averages about 3 SR points and individual task differences are sometimes zero (R2R Val Unseen is 76 vs 76), run-to-run noise could change the conclusions. Please report mean±std over at least three seeds for the key comparisons (SAME vs ScaleVLN† in Table 4, and the routing/position ablations in Tables 2 and 3), or provide statistical significance tests for the observed differences.","section":"Tables 2-7 and Table 4"},{"comment":"The paper never states the number of experts N or the top-k value used in the MoE layers, nor does it report the total parameter count after replacing visual-query projections with MoE layers. These values are required to reproduce the method and to interpret the routing ablations, since the router behavior and the load-balancing loss both depend on N and top-k. Please add these implementation details explicitly.","section":"Section 4, Implementation Details"}],"minor_comments":[{"comment":"Grammar and typos: 'SAME achieve State-of-the-Art performance' and 'SAME peform significantly better' should be corrected, and 'Dagger' should be written consistently as 'DAgger'.","section":"Section 4.1"},{"comment":"The load-balancing loss notation is inconsistent: Equation (8) sums over N experts but Equation (7) multiplies by N, and the index ranges are not fully defined. Please align the notation and define the batch size K explicitly.","section":"Equations (7)-(9)"},{"comment":"The table is titled 'Full results of SAME on all VLN benchmarks' but does not include R2R-CE or ObjectNav-MP3D rows; please add those rows or rename the table to avoid misleading the reader.","section":"Supplementary Table 8"},{"comment":"A sentence about decoupling vision-language-action alignment and improving VLN-CE success rates by approximately 20% appears twice in consecutive paragraphs; please remove the duplicate.","section":"Supplementary Section A"},{"comment":"The spacing in 'O BJECT NAV' is unusual and appears inconsistently; please use 'ObjectNav' consistently across the paper.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The self-citation of ScaleVLN is disclosed and the comparison is controlled by using the same initialization, so I do not see a circularity problem. The main risk is the ObjectNav evaluation protocol: if the filtered discrete validation set cannot be justified, the seven-task claim should be narrowed. I also recommend that the editor ask the authors for the missing N/top-k values and the waypoint-predictor details before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real idea here is routing experts on the agent's multimodal state (mean visual feature plus text CLS) instead of by task or token, and it earns its keep: the controlled comparison against their own ScaleVLN multi-task reimplementation, same initialization, shows a consistent ~3% SR gain across all seven tasks. The ablations on routing type and expert placement are well designed, and putting MoE on visual queries in cross-attention rather than on the FFN is a sensible, empirically supported choice. That part of the paper is solid and citable.\n\nThe headline claim in the abstract, \"outperforms or achieves highly comparable performance to task-specific agents,\" is overstated. In their own Table 4, SAME trails the best single-task models on R2R (76 vs 79 val unseen) and clearly on REVERIE (46.4 vs 57.0 val unseen). \"Highly comparable\" is doing a lot of work. It is still a good unified model, but the gap should be stated plainly.\n\nThe load-bearing soft spot is ObjectNav. Section 2.2 describes converting Habitat-Web's continuous 243-step demonstrations into discrete 20-step trajectories by nearest-node matching, dropping disconnected paths and any path whose endpoint is more than 0.5m from the original. That might be fine for training data, but the same filter is applied to the MP3D validation split, and the paper never reports how many standard episodes survive, how the dropped episodes distribute across scenes and goal categories, or a full-set evaluation. Since the Table 4 ObjectNav row has no discrete competitor and Table 5 compares only against continuous-space methods on this filtered set, the seven-task claim rests on an unvalidated evaluation set. This is fixable with a few numbers, but as it stands the ObjectNav comparison is not controlled.\n\nMinor points: all ablations are single runs with no variance, the R2R-CE zero-shot result leans on an external waypoint predictor, and lambda=0.8 plus the data sampling ratio are tuned hyperparameters. None of these are fatal.\n\nWho this is for: people working on multi-task or generalist embodied agents, and anyone designing MoE routing for sequential decision-making. It deserves a serious referee; the routing study is a genuine contribution even if the ObjectNav numbers need more work. Send it out, and ask for the ObjectNav retention counts, a full-set evaluation, and multi-seed results.","headline":"A genuinely new state-adaptive MoE routing for multi-task VLN with a solid controlled baseline study, but the seven-task claim is weakened by an unvalidated ObjectNav data projection and an overstated comparison to specialists.","tokens_in":24502,"tokens_out":2314,"would_cite":true,"duration_ms":21022,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One model with one set of weights can follow step-by-step instructions, dialogue, and object-goal searches on seven navigation benchmarks, matching or beating task-specific agents.","keywords":["vision-and-language navigation","mixture of experts","multi-task learning","instruction following","object goal navigation","embodied AI","state-adaptive routing","navigation benchmarks"],"falsifier":"For the MP3D validation split, compute the fraction of projected episodes in which the projected endpoint is a different object than the human demonstrator reached, or in which the target object is never visible along the projected path; if that fraction is substantial, the reported ObjectNav numbers measure a shortcut-friendlier task than continuous ObjectNav, and the seven-task claim would need re-evaluation.","tokens_in":23457,"feed_emoji":"🧭","tokens_out":12617,"duration_ms":112912,"temperature":0.7,"pith_summary":"The paper tries to establish that all forms of language-guided visual navigation fall on one spectrum—from step-by-step instructions, through dialogue and referring expressions, to bare object-category goals—and that a single agent, trained once, can cover the whole spectrum with one set of weights. Its model, SAME, uses a Mixture of Experts where a router picks different expert sub-networks at every timestep based on the agent's current visual-linguistic state. Trained on seven benchmarks simultaneously (R2R, RxR-EN, REVERIE, CVDN, SOON, ObjectNav-MP3D, R2R-CE), SAME is reported to match or beat specialists on six of the seven and to beat a multi-task DUET baseline on all of them. If correct, this would mean navigation agents no longer need per-benchmark fine-tuning, and that contrasting training data can be mixed productively once the model can specialize its skills.","feed_headline":"One model runs seven navigation tasks with one set of weights","feed_subtitle":"A state-adaptive router picks experts per step, so one agent follows commands, dialogue, and object-goal searches.","key_machinery":"The load-bearing mechanism is the State-Adaptive Mixture of Experts (SAME) layer—a sparse MoE placed on the visual query of every cross-attention layer in a DUET-style transformer. At each timestep, a linear router reads the concatenation of the instruction's [CLS] token and the mean of the current panoramic view features, selects the top-k experts, and combines the selected experts' outputs with router probabilities; a load-balancing loss (weight λ = 0.8) prevents a few experts from dominating all routes. The paper shows that this state-based routing outperforms token-wise routing, task-wise routing, and routing with task embeddings appended, and that placing the experts on the visual query beats placing them in the feed-forward network or on the textual key/value projections, because cross-modal attention is the point at which navigation behaviors differentiate.","core_discovery":"The central claim is that navigation task conflicts sit not in the shared visual-semantic backbone but in how decisions are derived from it: the paper shows that training one DUET transformer on mixed navigation data degrades coarse-grained target search (REVERIE success drops 6–7% when ObjectNav data is added), while a model that routes through a sparse mixture of experts on the visual-query projection of cross-attention, with the router fed by the mean panoramic feature plus the instruction's [CLS] token, recovers and exceeds that performance. With this design, SAME is reported to reach 7.07 goal progress on CVDN test, 76.3% success on discrete ObjectNav-MP3D validation, and 73.9% success on R2R test, closely matching the best single-task models on R2R and REVERIE and clearly outperforming the multi-task-tuned baseline on every one of the seven tasks. The same checkpoint transfers zero-shot to continuous environments through a waypoint predictor, reporting 47% success on R2R-CE and 43% success on continuous ObjectNav-MP3D.","pith_inferences":["The routing signal (mean view plus [CLS] token) discards spatial structure; a natural extension is to route on a spatial summary of the 36 views and test whether experts specialize even more cleanly.","Because no task embedding is used, the agent could discover skill groupings that cut across the seven labeled tasks; inspecting which experts fire during exploration-heavy versus instruction-heavy timesteps would test whether the learned specializations match the paper's exploration-versus-instruction-following story.","The discrete projection of Habitat-Web data is the main unvalidated link; re-training SAME on a denser connectivity graph or on continuous actions would show whether the reported ObjectNav numbers are an artifact of the 0.5-meter endpoint filter."],"forward_implications":["A single checkpoint can serve seven navigation tasks at once, so downstream systems could replace seven fine-tuned navigation models with one model trained on mixed data, cutting storage and adaptation cost.","Coarse-grained and zero-grained navigation (REVERIE, SOON, ObjectNav) inherit better visual-semantic grounding from fine-grained instruction data, so future data collection can prioritize instruction-trajectory pairs even when the deployment task is pure object search.","The router conditions only on the current observation and language, not on task labels, so the same architecture can absorb a new task by adding its dataset to training without changing the routing inputs.","Discrete-to-continuous transfer via waypoint prediction works for a multi-task agent as well as for single-task ones, with zero-shot continuous success of 47% on R2R-CE and 43% on ObjectNav-MP3D."],"supporting_citations":[{"why":"Supplies the DUET dual-scale graph transformer that SAME modifies, and the multi-task-tuned baseline used for the conflict analysis and comparisons.","marker":"[16]"},{"why":"Provides the ScaleVLN pretrained weights that initialization experiments show are responsible for large multi-task gains (roughly 15% success on R2R).","marker":"[106]"},{"why":"Supplies the Habitat-Web human demonstrations that are discretized into the 58,803 ObjectNav training trajectories and used as a continuous evaluation baseline.","marker":"[87]"},{"why":"Provides the environment discretization into a connectivity graph and the waypoint predictor that bridge discrete and continuous evaluations; the ObjectNav data conversion follows this work.","marker":"[37]"},{"why":"Defines the R2R fine-grained VLN task, dataset, and evaluation protocol that anchor the instruction-following side of the seven tasks.","marker":"[6]"},{"why":"Defines the REVERIE coarse-grained referring-expression navigation task that supplies evidence that plain data mixing hurts target-oriented performance.","marker":"[80]"},{"why":"Defines the ObjectGoal Navigation task and evaluation, the zero-grained end of the task spectrum used for the ObjectNav experiments.","marker":"[8]"},{"why":"Provides the Habitat simulator and continuous environment used for the zero-shot continuous evaluations of SAME.","marker":"[92]"}],"fun_headline_variants":["Seven navigation tasks, one model: SAME adapts per step","State-adaptive mixture of experts masters 7 navigation tasks","One SAME model, seven tasks: state-adaptive routing wins","SAME: one agent that navigates seven tasks in one go","State-adaptive experts let one model beat seven specialists"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that snapping the continuous human object-search demonstrations onto a discrete grid of navigable points—and throwing out any demonstration whose end point moves more than half a meter—still leaves training data that teaches the same exploration skill that continuous ObjectNav agents are judged on.","fun_headline_variants_meta":{"raw":{"variants":["Seven navigation tasks, one model: SAME adapts per step","State-adaptive mixture of experts masters 7 navigation tasks","One SAME model, seven tasks: state-adaptive routing wins","SAME: one agent that navigates seven tasks in one go","State-adaptive experts let one model beat seven specialists"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000799,"raw_usage":{"total_tokens":3508,"prompt_tokens":932,"completion_tokens":2576,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2500}},"tokens_in":548,"tokens_out":2576,"duration_ms":18837,"temperature":1.0,"reasoning_tokens":2500,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:35:32.231007+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the MP3D validation split, compute the fraction of projected episodes in which the projected endpoint is a different object than the human demonstrator reached, or in which the target object is never visible along the projected path; if that fraction is substantial, the reported ObjectNav numbers measure a shortcut-friendlier task than continuous ObjectNav, and the seven-task claim would need re-evaluation.","supporting_citations":[{"cited_title":"Scal- ing data generation in vision-and-language navigation","cited_arxiv_id":null,"evidence_quote":"Provides the ScaleVLN pretrained weights that initialization experiments show are responsible for large multi-task gains (roughly 15% success on R2R)."},{"cited_title":"Habitat-web: Learning embodied object-search strategies from human demonstrations at scale","cited_arxiv_id":null,"evidence_quote":"Supplies the Habitat-Web human demonstrations that are discretized into the 58,803 ObjectNav training trajectories and used as a continuous evaluation baseline."},{"cited_title":"Reverie: Remote embodied visual referring ex- pression in real indoor environments","cited_arxiv_id":null,"evidence_quote":"Defines the REVERIE coarse-grained referring-expression navigation task that supplies evidence that plain data mixing hurts target-oriented performance."},{"cited_title":"Habitat: A platform for embodied ai research","cited_arxiv_id":null,"evidence_quote":"Provides the Habitat simulator and continuous environment used for the zero-shot continuous evaluations of SAME."}],"review_version":1}