{"id":"9d6da666-caa7-4d6a-a6cb-17e454a1a518","arxiv_id":"2601.21409","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A dual-stance debate between a goal-focused and a safety-focused VLM, plus arbitration and optional micro-probing, improves zero-shot object navigation success and path efficiency on HM3Dv1, HM3Dv2, MP3D, and GOAT.","lead":"DSCD-Nav replaces the usual one-pass scoring of possible robot moves with a 'debate': one agent argues for the move that best advances the goal, another argues for safety and information gain, and a judge picks the final action. The authors report large success-rate and path-efficiency gains on standard indoor object-navigation benchmarks, making the paper a test of whether multi-agent debate can improve embodied decision-making without retraining.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Attribution of the SR/SPL gains to the debate mechanism is not yet supported: the candidate generator and prompts are undisclosed, and §4.3(c) shows backbone choice alone moves SR by >14 points; a same-candidate, same-VLM one-shot baseline is missing.","rationale":"The paper proposes a plausible, modular decision-layer mechanism and provides component ablations that internally support the mechanism. The strongest internal evidence is Table 3(a), where TSU-only (a likely single-pass scoring version) reaches 62.4 SR versus DSCD's 73.0 on HM3Dv2; this suggests the debate may genuinely add value. However, the central cross-benchmark claim—'DSCD-Nav achieves higher SR and SPL across all three benchmarks'—requires holding the candidate generator and VLM fixed. That condition cannot be verified from the manuscript: no generator identity, candidate counts, prompt text, or baseline re-implementation details are given, and some baselines (VLMNav, SSSD) are undefined or uncited. Table 3(c) aggravates the concern because VLM choice alone spans 58.7–80.1 SR, larger than many headline gaps. Thus the main risk is not that the method is internally inconsistent, but that the reported margins could be confounded by upstream candidate quality or backbone strength. The reader's verdict of CONDITIONAL is therefore appropriate; my stress test does not move it. The proposed concrete test—an exact same-backbone, same-candidate one-shot baseline across multiple benchmarks—would directly settle whether the debate layer is the cause of the gains.","tokens_in":19380,"tokens_out":6367,"duration_ms":69318,"concrete_test":"Release the candidate-generation module and prompt templates, then run a controlled single-pass scoring baseline on the same episodes: identical Gemini-2.5-Flash-Lite backbone, identical Geometry-Aware candidate set, the TSU prompt (or the full TSU+SIB scoring prompt) called exactly once, argmax action, and no debate, NCA, or micro-probing. If the TSU-only ablation is already this baseline, reproduce it and extend it to HM3Dv1, HM3Dv2, MP3D, and GOAT with 3 seeds or bootstrap CIs. The debate claim survives only if DSCD beats this controlled baseline by a margin exceeding the cross-seed/backbone noise on at least two benchmarks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central claim is that only the decision layer changes across the comparison. §3.1 and §4.1 assert that the candidate generator and perception settings are 'identical to the baselines,' but the generator is never named: Figure 2 labels it 'Geometry-Aware Action Proposal and Pruning,' yet the paper gives no candidate count, pruning rule, or interface details, and no baseline re-implementation is specified. Instead, Table 1 compares against published numbers (DORAEMON, VLMNav, etc.) that use their own candidate generators and backbones. The sensitivity in Table 3(c) is decisive for the confound: with everything else fixed, swapping the VLM moves SR from 58.7 to 80.1, a spread larger than the 62.0→73.0 gap used to headline the debate gain. So the observed margins are within the range that could be produced by upstream candidate quality or backbone strength alone. The internal TSU-only ablation on HM3Dv2 (62.4→73.0) is evidence that a single-pass same-backbone baseline exists, but it is reported for one benchmark only and the configuration is under-specified; it does not establish the 'consistent across all three benchmarks' claim, nor does it let a reviewer verify that the prompt-only baselines (VLMNav, SSSD) used the same generator. The central attribution is therefore currently unverified rather than falsified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DSCD-Nav, a training-free decision-layer mechanism for zero-shot 3D object navigation. Instead of scoring candidate actions once, a Task-Scene Understanding agent and a Safety-Information Balancing agent debate over the same candidate set for multiple rounds, a Navigation Consensus Arbitration agent judges the debate trace, and a two-mode execution policy either executes the consensus action or applies a soft-compromise micro-probing step when disagreement persists. Experiments on HM3Dv1, HM3Dv2, MP3D, and GOAT report improved Success Rate, SPL, and lower AORI over prior prompt-based and training-free baselines, with ablations of the main components, debate rounds, and VLM backbone.","tokens_in":19754,"tokens_out":6311,"duration_ms":69029,"significance":"If the attribution claim is verified, the contribution is practically useful: a plug-and-play decision-layer wrapper that can be added to existing VLM-based navigation pipelines without modifying perception or candidate generation, while providing interpretable per-step rationales. The paper contains several strengths: Table 3(a) gives internal evidence that the TSU, SIB, NCA, and micro-probing components each contribute; Table 3(b) shows saturation with debate rounds; Appendix A.3 reports sensitivity of the micro-probing hyperparameters; and the qualitative debate traces in Figure 6 help explain the mechanism. However, the central attribution of the performance gains to the debate mechanism is not yet established because of the undisclosed candidate generator and the backbone confound documented below. The paper is therefore promising but currently under-specified for the strength of its claims.","major_comments":[{"comment":"The central claim that DSCD-Nav's gains come from the debate mechanism is not supported by the current comparison. Equation (1) only says an 'upstream candidate generator' provides At, and §4.1 says the generator is 'identical to the baselines,' but no generator identity, candidate count, pruning rule, or baseline re-implementation is given. Table 1 compares against published numbers from methods with their own candidate generators and backbones. Table 3(c) is the decisive confound: with everything else fixed, changing only the VLM moves SR from 58.7 to 80.1 on HM3Dv2, a spread larger than the 62.0→73.0 improvement over DORAEMON reported in Table 2(a). The TSU-only ablation (62.4→73.0) is suggestive, but it is reported for one benchmark and its configuration is under-specified. Please disclose the candidate generator, re-implement the prompt-based baselines with the same generator and sa","section":"§3.1; §4.2.2; Table 1; Table 3(c)"},{"comment":"The formal belief-update mechanism is not operational. Eqs. (3)-(4) define Update as an 'abstract operator implemented implicitly via next-round prompting, not an explicit analytical rule.' Since no prompt templates, API decoding parameters, or concrete update logic are provided, the claim that TSU and SIB actually revise their beliefs in response to each other's evidence cannot be checked. The qualitative trace in Figure 6 shows that convergence sometimes occurs, but it does not verify that the modeled Update is the cause. Please provide the exact prompts used for TSU, SIB, and NCA, at least one representative prompt per agent, and an analysis showing that agreement/counter behavior changes with evidence as the model assumes.","section":"§3.2, Eqs. (3)-(4)"},{"comment":"There is an inconsistency in the episode budget that affects the reproducibility of the headline numbers. Appendix A.2 states the maximum navigation steps per episode is 40, while Appendix A.3 says the micro-probing sensitivity study uses 'a stricter episode budget' with maximum 20 steps. Yet Table 5, default row, reports SR=75.6 and SPL=40.1 on HM3Dv1, exactly matching Table 1 for DSCD on HM3Dv1. Either the main results use a 20-step budget (contradicting A.2) or the sensitivity table duplicates the main numbers. This must be corrected and clarified.","section":"Appendix A.2 vs. Appendix A.3"},{"comment":"No error bars, seeds, or run-to-run variability are reported for any experiment. This is consequential for the central 'consistent improvements across all three benchmarks' claim. For example, Table 3(b) shows rounds=3 at SR 73.0 and rounds=4 at 73.1, and Table 1 shows MP3D DSCD at 47.8 vs. CogNav at 46.6; these differences could plausibly be within the noise of a stochastic VLM API. Please report multiple seeds (at least 3) with means and standard deviations for the main comparison and the ablations, or otherwise justify that the reported single runs are stable.","section":"Tables 1-3; §4.2.2"}],"minor_comments":[{"comment":"The symbol θ⋆ in Eq. (9) is used before it is defined in Eq. (12). Define θ⋆ and θalt just before Eq. (8) or restructure the presentation.","section":"§3.3, Eq. (9)"},{"comment":"The text 'foundation-model–drivenmicro-probing' appears to be missing a space or hyphen after 'driven'; the sentence is otherwise unclear.","section":"Appendix B.3"},{"comment":"The caption says 'HM3D ObjectNav' without specifying the split. The numbers (75.6/40.1) match Table 1's HM3Dv1 row; please state this explicitly.","section":"Table 6"},{"comment":"The sentence 'This indicates that the improvements mainly stem from the decision structure rather than a specific backbone model' is an interpretation, not a result. As written it is undermined by the backbone-only spread in the same table; rephrase as an interpretation pending the controlled comparison requested in the major comments.","section":"§4.3(c)"},{"comment":"The phrase 'consistent with the conclusions on ObjectNav' should cite Table 2(a) and Table 2(b) directly so the reader can map the claim to the evidence.","section":"§4.2.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the internal ablations suggest the mechanism has substance, but the paper currently cannot support its central attribution claim because the candidate generator is undisclosed and the backbone sensitivity is large. This is fixable within the manuscript's scope: the authors can provide the generator identity, a same-candidate/same-VLM one-shot baseline, error bars, and prompt details. I would not accept the paper in its current form, but I would welcome a revised submission that resolves these verification issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"New here is putting dual-stance debate and arbitration on the candidate-action layer itself rather than on text reasoning or high-level plans. TSU and SIB see the same candidate set from goal-progress and safety/information stances, exchange evidence, and NCA either picks a consensus or triggers a small geometric compromise. That is a clean, training-free, plug-in idea, and the qualitative traces show real interpretability value. Give credit: the component ablations in Table 3(a), the debate-round saturation in 3(b), and the micro-probing sensitivity analysis in A.3 are genuine internal checks. The fact that four rounds buys only +0.1 SR suggests the effect is not just \"more prompting.\" The paper is coherent on its own terms.\n\nSoft spot is verification of the headline comparison, not the mechanism. The stress-test note is right: Table 1 compares against published numbers from systems with different candidate generators and backbones, so the 73.0 vs 62.0 gap is not attributable to debate unless the upstream is held fixed. The paper states that it keeps the generator and perception settings identical, but never names the generator, gives no candidate count, no prompts, no seeds, and no error bars. Table 3(c) hurts: swapping the VLM moves SR from 58.7 to 80.1, a spread larger than the debate gain claimed, so backbone or candidate quality could explain the margin. The internal TSU-only vs DSCD gap on HM3Dv2 (62.4 vs 73.0) is reassuring, but it is one benchmark and under-specified. The absent baseline is a same-backbone, same-candidate one-shot comparison in the main table. Also, the \"Bayesian-style\" update in Eqs. (3)-(4) is an abstract operator implemented by prompting, so the formalism is decorative; not a fatal flaw, but there is no mathematical content to inspect. Some baselines in Table 2 and Table 4 (VLMNav, SSSD) are never defined or cited.\n\nBottom line: the central claim is unverified, not falsified. The paper deserves a serious referee because the idea is new, modular, and the internal evidence looks honest. Ask for code, exact prompts, generator details, and a controlled same-VLM one-shot baseline. If they deliver, this becomes a solid contribution.","headline":"A plausible, well-ablated plug-in debate mechanism for zero-shot ObjectNav, but the headline gains are not yet attributable to it until the authors release the candidate generator, prompts, and a same-backbone one-shot baseline.","tokens_in":20278,"tokens_out":2331,"would_cite":false,"duration_ms":25434,"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":"This paper claims that replacing one-shot action scoring with a two-stance cooperative debate and an arbitration judge makes zero-shot object navigation consistently more successful and more efficient.","keywords":["object navigation","zero-shot navigation","multi-agent debate","vision-language model","action selection","arbitration","micro-probing","exploration redundancy"],"falsifier":"Keep the upstream candidate generator and the vision-language backbone fixed, then run a controlled comparison where the only difference is whether final action selection uses the full two-stance debate plus arbitration or a single-pass scoring pass over the same candidates. If success and path efficiency do not improve with the debate, the central claim is false. A second check: feed the same recorded debate traces to the judge but scramble the round order; if arbitration outcomes are unchanged, the exchange of reasons itself is not what matters.","tokens_in":19232,"feed_emoji":"🤖","tokens_out":4585,"duration_ms":52746,"temperature":0.7,"pith_summary":"The paper sets out to fix a specific failure in zero-shot indoor object navigation: at each step, current systems score candidate actions once and commit to the top one, which can lock in overconfident errors under partial visibility. Its proposal is to replace that single pass with a structured debate between two complementary evaluators — one arguing from goal progress and scene layout, the other auditing from safety and information gain — and to let a third arbitration agent resolve the disagreement into an executable action. When the two stances cannot agree but their preferred headings are close, the agent takes a short, slightly biased probing step instead of committing. The paper reports consistent gains in success, path efficiency, and lower exploration redundancy across three indoor benchmarks and a lifelong navigation benchmark, with the candidate generator and perception modules unchanged. If true, the finding points to the decision layer itself, not better perception or better candidates, as a profitable place to invest in zero-shot navigation.","feed_headline":"Two-stance debate lifts zero-shot navigation success","feed_subtitle":"A decision-layer debate between goal progress and safety cuts redundant exploration and improves path efficiency.","key_machinery":"The central mechanism is the debate-and-arbitration loop over a shared polar candidate set. Each candidate is packaged as a card with an identity, distance, yaw, and direction text; a goal-oriented stance and a safety/information stance exchange top picks with cue-grounded reasons over multiple rounds. The belief-update operation that revises each stance is not an explicit rule but is implemented implicitly by prompting each side with the other's arguments. A judge then reads the full trace and selects an action with a rationale. The two-mode execution policy is the second load-bearing piece: consensus executes directly; persistent near-angle disagreement triggers a soft compromise — a short","core_discovery":"The central claim is that two deliberately opposing evaluation stances, forced to exchange reasons and evidence over the same candidate set, produce more reliable navigation decisions than any single scoring function. The Task-Scene Understanding stance argues for the action most likely to advance the goal using layout priors and visual cues; the Safety-Information Balancing stance independently counters with risk and visibility concerns. Their disagreement is not averaged away: a Navigation Consensus Arbitration agent reads the full debate trace and issues a final choice, and if the stances remain split over directions that are geometrically near, the agent executes a conservative micro-pro","pith_inferences":["Beyond the paper's experiments, the same two-stance pattern could apply to other partially observable embodied tasks where one objective is progress and another is risk, such as manipulation or search-and-rescue; the paper only tests navigation.","Because the debate trace is logged and the final rationale is readable, the framework could serve as a debugging tool: operators could inspect whether repeated safety-stance counters flag a genuinely dangerous region or an over-cautious prompt, a use the paper does not develop.","A testable extension would be to make the compromise threshold and probe coefficients adaptive — for example, smaller probes in tight spaces — rather than fixed; the sensitivity analysis shows the current defaults are safe, not optimal.","The implicit belief update is prompt-only; replacing it with an explicit evidence-logging mechanism would test whether the debate actually revises beliefs or merely converges on the judge's prior."],"forward_implications":["If the mechanism is correct, any vision-language navigation pipeline could gain success and efficiency simply by wrapping its existing candidate set with the debate-and-arbitration layer, with no retraining.","With perception and candidate generation held fixed, dual-stance debate reduces exploration redundancy, so the reported effect is about decision quality rather than better sensors or maps.","The number of debate rounds matters: going from one round to three raises success substantially, while a fourth round adds almost nothing, indicating a practical saturation point.","The headline gains transfer from single-goal navigation to a multi-subgoal lifelong navigation benchmark, suggesting the decision mechanism is not tuned to one benchmark."],"fun_headline_variants":["Debate between safety and progress cuts robot detours","Two stances argue to improve robot navigation success","Cooperative debate reduces redundant exploration in robots","Stance-based arbitration boosts path efficiency in navigation","Robot navigation gets a debate-based decision upgrade"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The main load-bearing premise is that the compared baselines share the same upstream candidate generator and perception settings, so the reported gains come from the debate layer rather than from better candidates or a stronger vision-language model; the paper asserts this but does not identify the generator or show re-implementation details.","fun_headline_variants_meta":{"raw":{"variants":["Debate between safety and progress cuts robot detours","Two stances argue to improve robot navigation success","Cooperative debate reduces redundant exploration in robots","Stance-based arbitration boosts path efficiency in navigation","Robot navigation gets a debate-based decision upgrade"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1277,"prompt_tokens":737,"completion_tokens":540,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":470}},"tokens_in":481,"tokens_out":540,"duration_ms":6724,"temperature":1.0,"reasoning_tokens":470,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T06:57:20.391711+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep the upstream candidate generator and the vision-language backbone fixed, then run a controlled comparison where the only difference is whether final action selection uses the full two-stance debate plus arbitration or a single-pass scoring pass over the same candidates. If success and path efficiency do not improve with the debate, the central claim is false. A second check: feed the same recorded debate traces to the judge but scramble the round order; if arbitration outcomes are unchanged, the exchange of reasons itself is not what matters.","supporting_citations":[],"review_version":1}