Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

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.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 06:57 UTC pith:AKS7URUN

load-bearing objection 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. the 4 major comments →

arxiv 2601.21409 v3 pith:AKS7URUN submitted 2026-01-29 cs.RO

DSCD-Nav: Dual-Stance Cooperative Debate for Object Navigation

Weitao An , Qi Liu , Chenghao Xu , Jiayi Chai , Xu Yang , Kun Wei , Cheng Deng This is my paper
classification cs.RO
keywords object navigationzero-shot navigationmulti-agent debatevision-language modelaction selectionarbitrationmicro-probingexploration redundancy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

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.

Core claim

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

What carries the argument

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

Load-bearing premise

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.

What would settle it

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.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • 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.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [§3.1; §4.2.2; Table 1; Table 3(c)] 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
  2. [§3.2, Eqs. (3)-(4)] 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.
  3. [Appendix A.2 vs. Appendix A.3] 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.
  4. [Tables 1-3; §4.2.2] 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.
minor comments (5)
  1. [§3.3, Eq. (9)] 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.
  2. [Appendix B.3] The text 'foundation-model–drivenmicro-probing' appears to be missing a space or hyphen after 'driven'; the sentence is otherwise unclear.
  3. [Table 6] 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.
  4. [§4.3(c)] 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.
  5. [§4.2.2] 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.

Circularity Check

0 steps flagged

No significant circularity; empirical benchmark comparison with external baselines.

full rationale

The paper's core claim is an empirical comparison on standard benchmarks. The method is defined independently of the results: TSU and SIB stances are constructed from task objectives, NCA arbitration is a specified procedure, and two-mode execution is a defined policy. No parameter is fitted to the target metric and then renamed a prediction. The ablations compare configurations of the method itself, including single-stance baselines, which provide a within-method control. There are no self-citations. Any concern that the candidate generator or backbone settings are not identical across published baselines concerns experimental validity/attribution, not circularity, since the comparison is against externally reported numbers rather than derivations. The paper does claim 'we keep the candidate generator and perception settings identical to the baselines' (Sec. 4.1), but this is an unverified fairness claim rather than a reduction of the result to an input. The derivation chain is self-contained: the method's equations (e.g., Eqs. 8-13) define an execution policy but do not presuppose the benchmark outcomes. Thus no circular step is identified.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 3 invented entities

The central claim is empirical and rests on many hand-set dialogue/execution parameters, on the behavior of a proprietary VLM, and on an uncharacterized candidate generator. The paper provides no formal derivation or external benchmark of its belief-update mechanism.

free parameters (6)
  • β_r (micro-probing forward scale) = 0.50 (default; tested 0.33–0.67)
    Eq. (13) and Appendix A.3. Hand-chosen; sensitivity is mild on HM3Dv1 with a 20-step budget, but no sensitivity is reported on the main HM3Dv2/MP3D results.
  • β_θ (micro-probing yaw interpolation) = 1/3 (default; tested 0.25–0.50)
    Eq. (13). Set by hand; no principled derivation. Sensitivity is only reported on one benchmark.
  • α (soft-compromise angular threshold) = 60° implied by Appendix A.2 '∆θmax=60°'; not explicitly connected to α in main text
    Controls when Mode B triggers. The value is not ablated and is only given in the appendix.
  • K (debate rounds) = 3
    Table 3(b): rounds=3 gives 73.0 SR and rounds=4 gives 73.1; K=3 is chosen for cost. This is a validation-set tuning choice.
  • AORI weights w_c, w_d = 0.8 / 0.2
    Eq. (19). The redundancy metric used to support 'reduced exploration redundancy' depends on these hand-set weights.
  • AORI auxiliary constants τ, λ, η, voxel/ray size = not specified
    Eqs. (16)-(18). The AORI metric is under-specified; different choices would change the reported redundancy numbers.
axioms (4)
  • domain assumption The upstream 'Geometry-Aware' candidate generator supplies a fixed, executable candidate set independent of the debate policy.
    Sec. 3.1, Eq. (1)-(2). The controlled comparison and the claim of decision-layer-only gains depend on this premise.
  • domain assumption Gemini-2.5-Flash-Lite, CLIP ViT-B/32, and GroundingDINO reliably ground egocentric images and candidate descriptions.
    Appendix A.2. The entire system's evidence quality depends on proprietary or unspecified model behavior that is not independently verified.
  • domain assumption Natural-language debate traces correspond to true belief updating.
    Sec. 3.2, Eqs. (3)-(4): 'Update(·) is an abstract operator implemented implicitly via next-round prompting, not an explicit analytical rule.' The paper assumes LLM prompting can reliably revise beliefs.
  • domain assumption The standard ObjectNav protocol (Batra et al., 2020) is the correct evaluation setting for zero-shot object navigation.
    Sec. 3 task definition and Appendix A.2. Success is defined as STOP within 0.5m of the target with visual confirmation of the object.
invented entities (3)
  • TSU-Agent (Task-Scene Understanding) no independent evidence
    purpose: Proposes the goal-progress action using layout cues and room-object priors.
    A prompt role introduced by the paper. Its necessity is supported only by this paper's ablations, not by independent external evidence.
  • SIB-Agent (Safety-Information Balancing) no independent evidence
    purpose: Audits the proposal from a task-agnostic safety and information-gain perspective.
    A prompt role introduced by the paper. No external validation exists outside the reported ablations.
  • NCA-Agent (Navigation Consensus Arbitration) no independent evidence
    purpose: Arbitrates between TSU and SIB using the full debate trace.
    A prompt role introduced by the paper. Its arbitration quality is not separately validated against ground-truth optimal actions.

pith-pipeline@v1.3.0-alltime-deepseek · 18962 in / 18170 out tokens · 173187 ms · 2026-08-03T06:57:20.391711+00:00 · methodology

0 comments
read the original abstract

Adaptive navigation in unfamiliar indoor environments is crucial for household service robots. Despite advances in zero-shot perception and reasoning from vision-language models, existing navigation systems still rely on single-pass scoring at the decision layer, leading to overconfident long-horizon errors and redundant exploration. To tackle these problems, we propose Dual-Stance Cooperative Debate Navigation (DSCD-Nav), a decision mechanism that replaces one-shot scoring with stance-based cross-checking and evidence-aware arbitration to improve action reliability under partial observability. Specifically, given the same observation and candidate action set, we explicitly construct two stances by conditioning the evaluation on diverse and complementary objectives: a Task-Scene Understanding (TSU) stance that prioritizes goal progress from scene-layout cues, and a Safety-Information Balancing (SIB) stance that emphasizes risk and information value. The stances conduct a cooperative debate and make policy by cross-checking their top candidates with cue-grounded arguments. Then, a Navigation Consensus Arbitration (NCA) agent is employed to consolidate both sides' reasons and evidence, optionally triggering lightweight micro-probing to verify uncertain choices, preserving NCA's primary intent while disambiguating. Experiments on HM3Dv1, HM3Dv2, and MP3D demonstrate consistent improvements in success and path efficiency while reducing exploration redundancy.

Figures

Figures reproduced from arXiv: 2601.21409 by Cheng Deng, Chenghao Xu, Jiayi Chai, Kun Wei, Qi Liu, Weitao An, Xu Yang.

Figure 1
Figure 1. Figure 1: (a) Map-based navigation is brittle to map errors. (b) VLM-based navigation is overly confident, making greedy Top-1 choices. (c) Our DSCD-Nav uses collaborative debate and arbitra￾tion for more reliable decisions. object navigation, an embodied agent must move from a random start to a target object in an unseen scene using only egocentric observations, without prior maps or task-specific training (Batra e… view at source ↗
Figure 2
Figure 2. Figure 2: DSCD-Nav overview. (I) Candidate and context construction packages the goal, optional memory, and evidence cues, and a geometry-pruned set of executable polar action candidates with language descriptions into debate-ready inputs. (II) TSU and SIB conduct multi-round cooperative debate over the shared candidates, exchanging preferences and cue-grounded evidence while updating beliefs. (III) The NCA judge ar… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative case study on HM3Dv2 ObjectNav. We overlay the executed trajectory on the reconstructed 3D scene and annotate several decision points with the corresponding TSU and SIB preferences and the NCA arbitration result. The trace highlights DSCD-Nav’s ability to resolve disagreement at visually ambiguous viewpoints and, when needed, trigger a conservative micro-probing step before committing to the ex… view at source ↗
Figure 4
Figure 4. Figure 4: An 8-step, debate-guided HM3Dv1 ObjectNav episode from the agent’s first-person view, illustrating sequential observations, executed actions, and successful termination at the target. where θ ⋆ denotes the yaw angle of the NCA-arbitrated ac￾tion, θ alt denotes the yaw angle of the other stance’s final preferred candidate, and ∆θ = wrap(θ alt − θ ⋆ ) ∈ (−π, π] is the signed shortest angular difference (wrap… view at source ↗
Figure 5
Figure 5. Figure 5: Top-down trajectory visualization. The floor-plan view overlays the agent’s navigation trace within the scene. Colored arrows denote the executed motions and a small set of salient alternative directions at key junctions, while red crosses mark branches that are discarded or lead to failures or inefficient detours. B.3. Detailed Description of Baselines To assess the performance of DSCD, we compare against… view at source ↗
Figure 6
Figure 6. Figure 6: Navigation case and debate trace. Each row corresponds to one decision step along the trajectory, with the executed action highlighted in green. The debate trace summarizes TSU’s goal-progress proposal, SIB’s safety and visibility audit, and the NCA judge’s final arbitration grounded in the current scene evidence. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Uni-LaViRA: Language-Vision-Robot Actions Translation for Unified Embodied Navigation

    cs.RO 2026-05 unverdicted novelty 6.0

    A zero-shot unified agent for VLN-CE, ObjectNav, EQA and Aerial-VLN on wheeled, quadruped, humanoid and UAV platforms that translates language and vision inputs into actions via MLLMs plus TDM and SCB mechanisms, matc...

  2. HAM-VLN: Harnessing Hierarchical Agentic Memory for Zero-Shot Vision-and-Language Navigation

    cs.RO 2026-07 conditional novelty 5.0

    A decision-coupled, hierarchical agentic memory, stored in a depth-grounded world graph, lets a zero-shot multimodal LLM navigator reach higher success rates with fewer context tokens.

Reference graph

Works this paper leans on

13 extracted references · 5 linked inside Pith · cited by 2 Pith papers

  1. [3]

    VL-Nav: Real-time vision-language navigation with spatial reasoning.arXiv preprint arXiv:2502.00931,

    Du, Y ., Fu, T., Chen, Z., Li, B., Su, S., Zhao, Z., and Wang, C. VL-Nav: Real-time vision-language navigation with spatial reasoning.arXiv preprint arXiv:2502.00931,

    Pith/arXiv arXiv

  2. [7]

    OpenFMNav: Towards open-set zero-shot object navigation via vision-language foundation models

    Kuang, Y ., Lin, H., and Jiang, M. OpenFMNav: Towards open-set zero-shot object navigation via vision-language foundation models. InFindings of the Association for Computational Linguistics: NAACL 2024, pp. 338–351,

    2024

  3. [8]

    Encouraging divergent thinking in large language models through multi-agent debate

    9 DSCD-Nav: Dual-Stance Cooperative Debate for Object Navigation Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y ., Wang, R., Yang, Y ., Shi, S., and Tu, Z. Encouraging divergent thinking in large language models through multi-agent debate. InProceedings of the 2024 Conference on Em- pirical Methods in Natural Language Processing, pp. 17889–17904,

    2024

  4. [9]

    and Hooi, B

    Lin, Z. and Hooi, B. Enhancing multi-agent debate system performance via confidence expression. InFindings of the Association for Computational Linguistics: EMNLP 2025, pp. 6453–6471,

    2025

  5. [10]

    and Bai, Q

    Ling, L. and Bai, Q. Endowing embodied agents with spatial reasoning capabilities for vision-and-language navigation. arXiv preprint arXiv:2504.08806,

    arXiv

  6. [11]

    Multi-agent LLM actor-critic framework for social robot navigation.arXiv preprint arXiv:2503.09758, 2025a

    Wang, W., Obi, I., and Min, B.-C. Multi-agent LLM actor-critic framework for social robot navigation.arXiv preprint arXiv:2503.09758, 2025a. Wang, X., Wang, J., Wang, Y ., Dang, P., Cao, S., and Zhang, C. MARS: toward more efficient multi- agent collaboration for LLM reasoning.arXiv preprint arXiv:2509.20502, 2025b. 10 DSCD-Nav: Dual-Stance Cooperative De...

    Pith/arXiv arXiv

  7. [12]

    JanusVLN: De- coupling semantics and spatiality with dual implicit memory for vision-language navigation.arXiv preprint arXiv:2509.22548,

    Zeng, S., Qi, D., Chang, X., Xiong, F., Xie, S., Wu, X., Liang, S., Xu, M., and Wei, X. JanusVLN: De- coupling semantics and spatiality with dual implicit memory for vision-language navigation.arXiv preprint arXiv:2509.22548,

    arXiv

  8. [13]

    NaVid: Video- based VLM plans the next step for vision-and-language navigation.arXiv preprint arXiv:2402.15852, 2024a

    Zhang, J., Wang, K., Xu, R., Zhou, G., Hong, Y ., Fang, X., Wu, Q., Zhang, Z., and Wang, H. NaVid: Video- based VLM plans the next step for vision-and-language navigation.arXiv preprint arXiv:2402.15852, 2024a. Zhang, L., Zhang, Q., Wang, H., Xiao, E., Jiang, Z., Chen, H., and Xu, R. TriHelper: Zero-shot object navigation with dynamic assistance. InProcee...

    Pith/arXiv arXiv 2025

  9. [2020]

    B., Bauer, N., Wahle, J

    Becker, J., Kaesberg, L. B., Bauer, N., Wahle, J. P., Ruas, T., and Gipp, B. MALLM: Multi-agent large language mod- els framework. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pp. 418–439,

    2025

  10. [2022]

    ObjectNav Revisited: On evaluation of embodied agents navigating to objects.arXiv preprint arXiv:2006.13171,

    Batra, D., Gokaslan, A., Kembhavi, A., Maksymets, O., Mottaghi, R., Savva, M., Toshev, A., and Wijmans, E. ObjectNav Revisited: On evaluation of embodied agents navigating to objects.arXiv preprint arXiv:2006.13171,

    Pith/arXiv arXiv 2006

  11. [2023]

    G., and Loquercio, A

    Goetting, D., Singh, H. G., and Loquercio, A. End-to-end navigation with vision language models: Transforming spatial reasoning into question-answering.arXiv preprint arXiv:2411.05755,

    Pith/arXiv arXiv

  12. [2024]

    DORAEMON: Decentralized ontology-aware reliable agent with enhanced memory oriented navigation.arXiv preprint arXiv:2505.21969,

    Gu, T., Li, L., Wang, X., Gong, C., Gong, J., Zhang, Z., Xie, Y ., Ma, L., and Tan, X. DORAEMON: Decentralized ontology-aware reliable agent with enhanced memory oriented navigation.arXiv preprint arXiv:2505.21969,

    arXiv

  13. [2025]

    B., Becker, J., Wahle, J

    Kaesberg, L. B., Becker, J., Wahle, J. P., Ruas, T., and Gipp, B. V oting or consensus? decision-making in multi-agent debate. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 11640–11671,

    2025