acceptodds
Under review as a conference paper at ICLR 2027

ARIADNE: Disambiguation-Guided Active Perception for Zero-Shot 3D Visual Grounding

Abstract

Zero-shot 3D visual grounding aims to locate an object in a 3D scene from a natural-language expression without task-specific training. Object-centric local views preserve photographic color and texture detail from captured RGB images, but may omit the anchor objects needed to judge spatial relations. Rendered global perspective views provide wider context, but small object scale, occlusion, and reconstruction artifacts can obscure the details needed for candidate disambiguation. We introduce ARIADNE, a disambiguation-guided active perception framework that adapts its observations to the ambiguity remaining among candidates without fine-tuning or test-time training. We first decompose the expression into visual claims about target attributes and relations to anchor objects and track which claims remain unresolved for each candidate. Our Relation-Centric Observation Construction (RCOC) provides complementary captured, stitched, and rendered views for comparing candidates and anchor objects in shared visual context. When several views are available for a comparison, Online Evidence Transport (OET) jointly matches the candidates' evidence needs to these views and selects the next observation. OET balances visibility and visual quality within a transport objective, with graph and history regularization coordinating view selection across candidates and successive comparisons. Together, these components form an adaptive observation and decision loop, with each new view guiding the next comparison toward the distinctions still needed to identify the target. Experiments on ScanRefer and Nr3D show gains of 3.5 percentage points in overall [email protected] and 4.2 percentage points in overall accuracy, respectively, over the current state-of-the-art zero-shot baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D visual grounding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_g4cXOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ARIADNE, a disambiguation-guided active perception framework for zero-shot 3D visual grounding. The method first decomposes natural-language referring expressions into visual claims and continuously tracks the unresolved claims for each candidate to identify missing evidence. Subsequently, its Relation-Centric Observation Construction (RCOC) module establishes a shared visual context for directly comparing candidates against anchor objects by providing three complementary views: local, stitched, and rendered. Next, the Online Evidence Transport (OET) module integrates graph and history regularization to model view selection as a transport problem, jointly matching and selecting the optimal next observation view for competing candidates. Experiments on the ScanRefer and Nr3D datasets demonstrate that this approach surpasses existing zero-shot visual grounding baselines in overall accuracy.

Critical strengths

(1) Zero-shot 3D grounding is an important research question with great potential. The paper identifies a critical bottleneck in current zero-shot 3D visual grounding: static object-centric or global perspective views often fail to capture the shared visual context and spatial relations needed to distinguish similar candidates. The motivation to use "unresolved visual claims" to actively guide the perception loop directly addresses this candidate disambiguation problem, making the approach intuitive and valuable.

(2) The proposed framework forms a complete and logically sound perception loop. Specifically, formulating the joint view selection for multiple competing candidates as an optimal transport problem (Online Evidence Transport) is insightful.

(3) The paper achieves leading results on two major benchmarks (ScanRefer and Nr3D) compared with other zero-shot baselines. The ablation studies systematically verify the individual contributions of different view sources (Local, Stitch, Render), the OET module, and its specific regularization terms.

Critical weaknesses

(1) While the paper acknowledges that the active perception loop increases inference time and token usage, it lacks concrete quantitative comparisons of computational overhead against baselines like SPAZER or SeqVLM. This makes it difficult to evaluate the framework's efficiency for real-world deployment.

(2) The disambiguation loop completely relies on pre-generated object proposals. If the target object is missed or incorrectly merged during the initial generation or filtering stages, the system cannot rediscover it during active exploration, limiting its fault tolerance.

(3) The experiments primarily rely on Seed 2.0 Lite (with a brief mention of DeepSeek in the appendix). The paper lacks a comprehensive evaluation using a broader, diverse set of widely adopted and smaller open-source MLLMs (e.g., Qwen series from 9B-27B-35BA3B). This leaves the framework's generalizability and robustness across different MLLM backbones insufficiently validated.

(4) Lack of discussion and comparison of recent and highly relevant works [1]

[1] Think, Act, Build: An Agentic Framework with Vision Language Models for Zero-Shot 3D Visual Grounding, 2026 EMNLP

Questions for authors

Please see the weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer g4cX

Thank you for your careful, constructive review and for recognizing our unresolved-claim active perception loop, Online Evidence Transport formulation, and systematic ablations. Our General Response also summarizes six studies on equal-budget planning, score-matched OET, runtime/token efficiency, extended-set evaluation and five-run stability, results with open-weight Qwen MLLMs, and proposal fault tolerance and TAB. We address your questions and concerns below.

W1: How does ARIADNE's computational overhead compare quantitatively with SPAZER and SeqVLM?

ARIADNE's median single-query time is the lowest among the compared methods (10.2 s vs. 37.9 s for SPAZER, 57.1 s for SeqVLM, and 34.1 s for SeeGround), because its extra MLLM calls (6.5 vs. 1.9 on average) are spent only on the 34% of queries still ambiguous after initial screening. Below we report single-query time (one query at a time), MLLM calls, and tokens on ScanRefer under the same MLLM; for TAB, times and proposal-free accuracy come from its paper, while calls and tokens were measured on 20 random queries using its released code and MLLM (to be added to the revised manuscript):

Method (ScanRefer, 250 queries)View / perception prep. (mean / med.)MLLM time (mean / med.)Total 1-query time (mean / med.)Calls (mean / med.)Tokens (in / out)[email protected] / @0.5
SeeGround0.4 s / 0.3 s45.9 s / 33.7 s46.3 s / 34.1 s2.0 / 2.06.0k / 1.6k53.2 / 47.6
SeqVLM0.3 s / 0.2 s136.6 s / 56.8 s136.9 s / 57.1 s1.9 / 1.07.4k / 13.3k57.2 / 52.0
SPAZER (same MLLM)4.2 s / 3.5 s37.1 s / 36.1 s41.3 s / 37.9 s1.9 / 2.06.8k / 2.3k60.8 / 52.8
SPAZER (GPT-4o, as reported in its paper)——23.5 s / ———57.2 / 48.8
TAB24.2 s / —59.5 s / —83.7 s / —295 / 260685k / 5.0k71.2 / 46.4
ARIADNE (Ours)15.1 s / 1.7 s33.2 s / 8.7 s48.3 s / 10.2 s6.5 / 3.041.9k / 2.3k62.4 / 55.6 (72.6 / 64.0 with TAB's inputs; SAM3 masks and refined queries; please see the response to W4)
  1. 3.3×–5.6× lower median time: stopping once one candidate remains finishes 66% of queries in zero verification rounds (further observing and re-judging remaining candidates; 1.3 rounds per query on average), whereas SPAZER and SeqVLM output long chain-of-thought traces per step.
  2. Competitive mean time (48.3 s vs. 41.3 s SPAZER, 46.3 s SeeGround, 83.7 s TAB, 136.9 s SeqVLM): ARIADNE has the lowest mean MLLM time (33.2 s) and needs about 45× fewer MLLM calls and 16× fewer input tokens per query than TAB, while OET (Online Evidence Transport, view selection) takes 0.5 ms/step on CPU (negligible cost).
  3. Further reducible: a new early-stopping variant (halting if a round eliminates no candidate) cuts mean rounds to 0.5 and input tokens by 30.3% at 61.6 / 55.2 (vs. 62.4 / 55.6 full budget).

W2: The loop depends on pre-generated proposals: can a missed or merged target be recovered (fault tolerance)?

We agree: in the paper setting, a target missed or merged during proposal generation or filtering cannot be rediscovered by the disambiguation loop. This was a deliberate compromise for fairness: like prior zero-shot 3DVG work, ARIADNE uses pre-generated Mask3D proposals and prior class-name filters (SeqVLM's on ScanRefer, SPAZER's on Nr3D) without modifying proposals or recovering lost objects, so Tables 1–2 and 6 isolate our active disambiguation workflow rather than better perception. To add fault tolerance rather than a benchmark-specific design, we tested two simple, generic fallbacks (reported here; the limitation in Appendix C will move to the main text); prior pipelines already use similar fallbacks (SPAZER screens all proposals on zero label match; SeqVLM's code falls back to ZSVG3D on VLM failure), whereas our paper used none. (i) Zero-match fallback: if no candidate passes the target category, screen the remaining scene proposals. (ii) 2D open-vocabulary masks (SAM3, as in TAB): also keep proposals supported by 2D masks of the target class. Each alone recovers lost targets:

Initial candidates (ScanRefer, 250 original queries)Queries whose target is among the candidates (@0.25 / @0.5)[email protected] / @0.5
Paper setting: Mask3D proposals + category filter, no fallback83.2% / 74.0%62.4 / 55.6
(i) Zero-match fallback only88.4% / 78.0%66.8 / 58.8
(ii) SAM3 2D masks only90.8% / 81.6%64.8 / 56.8

The same fallbacks also help on Nr3D (250 queries). In the paper setting, the target is among the candidates for 84.0% of queries and accuracy is 68.0% (Table 2 of the paper). With (i), accuracy rises to 69.2%; with (i) + (ii), the target is among the candidates for 98.0% of queries and accuracy rises to 72.4% (+4.4% over the paper setting). These gains show that ARIADNE's workflow improves directly with better perception inputs, further supporting its generality.

AuthorsOpenReview ↗

Response to Reviewer g4cX (continued)

W3: Does ARIADNE generalize to smaller open-source MLLM backbones such as the Qwen series?

With fixed prompts and no backbone-specific tuning, ARIADNE generalizes to open-weight Qwen backbones (9B–27B; reported here and to be added to the main text) and to DeepSeek-V4.1-Flash (Table 8 in the Appendix of the paper). First, under the same backbone, ARIADNE is the best of the four frameworks:

MLLM backbone (ScanRefer [email protected] / @0.5, 250 queries)SeeGroundSeqVLMSPAZERARIADNE (Ours)
Seed 2.0 Lite53.2 / 47.657.2 / 52.060.8 / 52.862.4 / 55.6
Qwen3.8-27B52.8 / 45.653.6 / 47.658.4 / 50.861.6 / 54.4
Qwen3.6-27B51.6 / 45.252.0 / 46.457.6 / 49.259.6 / 52.4

On Qwen3.8-27B, ARIADNE beats SPAZER (+3.2% / +3.6%), SeqVLM (+8.0% / +6.8%, p=0.005p = 0.005 at [email protected] / p=0.010p = 0.010 at [email protected]), and SeeGround (+8.8% / +8.8%, p=0.005p = 0.005 / p=0.002p = 0.002); on Qwen3.6-27B, it beats SPAZER (+2.0% / +3.2%), SeqVLM (+7.6% / +6.0%, p=0.005p = 0.005 / p=0.017p = 0.017), and SeeGround (+8.0% / +7.2%, p=0.027p = 0.027 / p=0.033p = 0.033). Second, ARIADNE itself stays strong across backbones:

MLLM backboneARIADNE on ScanRefer ([email protected] / @0.5, 250 queries)
Seed 2.0 Lite62.4 / 55.6
DeepSeek-V4.1-Flash61.2 / 54.0
Qwen3.8-27B61.6 / 54.4
Qwen3.6-27B59.6 / 52.4
Qwen3.5-27B60.0 / 53.2
Qwen3.5-9B51.6 / 45.6

It is worth noting that a 9B model is not a common setting: prior zero-shot 3DVG work uses backbones of about 30B parameters or larger, or proprietary models such as GPT-4o. In this common setting (Seed 2.0 Lite, DeepSeek-V4.1-Flash, and three 27B Qwen models), ARIADNE consistently reaches 59.6–62.4 / 52.4–55.6, showing that its agentic workflow transfers across MLLMs without any tuning.

W4: Missing discussion of and comparison with the recent TAB (EMNLP 2026)

Thank you for pointing to TAB (EMNLP 2026), an insightful concurrent agentic framework for proposal-free 3DVG that combines 2D detection/segmentation (Grounding DINO + SAM3) with VLM reasoning and is a useful point of comparison; we will cite, discuss, and compare with TAB in Sections 2 and 4 of the revised manuscript. However, TAB targets a complementary bottleneck with different trade-offs: its 13-step skill commits to the target on one static reference frame from raw video before tracking 2D masks, whereas ARIADNE performs dynamic multi-round disambiguation across competing 3D candidates driven by unresolved claims (parts of the query not yet verified for a candidate) and synthesizes views beyond raw video frames (stitched panoramas and targeted renders). For a fair comparison, we give ARIADNE SAM3 2D masks and the refined queries introduced by TAB (manually rewritten queries):

Method (250 queries per benchmark)ScanRefer [email protected] / @0.5Nr3D Acc.
TAB (proposal-free)71.2 / 46.468.0
TAB (with Mask3D proposals)71.6 / 61.667.2
ARIADNE (Ours, with SAM3 2D masks and refined queries)72.6 / 64.0 (mean of two runs)77.2

TAB's numbers are computed from its officially released inference results. With these inputs, ARIADNE beats TAB on both benchmarks: 72.6 / 64.0 vs. 71.6 / 61.6 on ScanRefer and 77.2% vs. 67.2% on Nr3D (+10.0%). Because the two methods start from different perception inputs, we also compare them on the 208 ScanRefer queries where our proposals contain the target, so that both are judged only on disambiguation: there ARIADNE is correct on 75.0% vs. 71.6% for TAB. Given the same perception inputs as TAB (the table above), ARIADNE still grounds the target more accurately, which confirms that its advantage comes from the multi-round disambiguation workflow itself rather than from its perception inputs.

AuthorsOpenReview ↗

Brief summary of our responses

Dear Reviewer g4cX,

Thank you again for your thoughtful review. We appreciate your time during this busy period, and as the discussion comes to a close, we briefly summarize our main responses below.

We added stage-split runtime, call, and token comparisons (10.2 s median single-query time, 3.3×–5.6× faster than prior baselines and 45× fewer calls than TAB) and evaluated simple proposal-recovery fallbacks (lifting ScanRefer from 62.4 / 55.6 to 66.8 / 58.8, and Nr3D target coverage from 84.0% to 98.0% and accuracy from 68.0% to 72.4%). Across open-weight Qwen MLLMs (9B–27B), ARIADNE remains the strongest framework (61.6 / 54.4 on Qwen3.8-27B, 59.6 / 52.4 on Qwen3.6-27B), and with SAM3 masks and TAB's refined queries it surpasses concurrent TAB on both ScanRefer (72.6 / 64.0 vs. 71.6 / 61.6) and Nr3D (77.2% vs. 67.2%).

We welcome any feedback on whether our responses address your concerns, and we would be glad to clarify any remaining questions.

Reviewer_c4jsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ARIADNE for zero-shot 3D visual grounding. The model decomposes referring expressions into visual claims, continuously tracks whether candidate objects support, contradict or remain undecided for each claim, and actively selects the next relation to verify accordingly. RCOC constructs shared local, stitched and rendered views, and OET selects observations via a transport objective with history and graph regularization. The authors report better results than existing zero-shot methods on ScanRefer and Nr3D.

Critical strengths
  1. This paper identifies a genuine and important failure mode in zero-shot 3DVG: object-centric views capture fine details but may lack anchors, while global/rendered views provide context yet can lose candidate-level details. Leveraging a candidate–claim ledger to drive relation-specific shared observations, the method has clear intuition and forms a closed loop with the final grounding decision.

  2. The experiments do not merely report end-point SOTA results. Tables 4–6 analyze observation sources and history/graph regularization, and multiple frameworks are re-run under the unified Seed 2.0 Lite setting. Such controls make the results more credible than direct comparisons against numbers from original papers, which is the main reason I do not lean toward clear rejection.

Critical weaknesses
  1. The disambiguation-guided active planner has not been tested under equal budget. Section 3.2’s core idea picks the next claim or context from candidate–claim judgements. But Table 3’s one-pass baseline uses fewer observation rounds, and Random/Greedy/OET comparisons all run after this planner selects what to evaluate. The 7.6/4.4 gain only proves iterative observation helps, not that picking observations based on unresolved claims beats simple exploration with the same number of rounds. To validate this core contribution, the authors should compare their full planner against random/fixed-order claim scheduling on the same 250 ScanRefer samples with a 5-round cap and identical candidate grouping, RCOC, OET and MLLM budget. Only the claim selection logic should change, and [email protected]/0.5 should be reported.

  2. The standalone benefit and reliability of OET’s small gains are unclear. Table 3 shows OET beats Greedy by only 1.6/1.2 points, yet Greedy only uses visibility ratio while OET’s cost mixes fidelity, projected size and observability, changing both scoring and optimization. Table 6’s advantage over SPAZER is also just 1.6/2.8 points on a fixed 250-sample set with no repeated runs, standard deviation or confidence intervals. The authors should add a score-matched baseline a∗=arg⁡max⁡a∑eq^t(e,a)a^*=\arg\max_a\sum_e\hat q_t(e,a) with identical proposals, planner and round limits to isolate the OET objective. They also need mean±std or bootstrap confidence intervals for OET vs baseline and ARIADNE vs SPAZER, plus details on temperature, seeds and checkpoints/API settings.

Questions for authors

1.Do the authors have results or logs for the baseline with the same 5-round budget that selects the next claim without using the candidate–claim ledger? If available, what is the performance gap between this baseline and the full planner?

2.Has the author tried directly aggregating q^t(e,a)\hat q_t(e,a) across candidates from Eq. (2) and greedily selecting views? How does its performance compare with full OET?

3.Is Seed 2.0 Lite inference deterministic? Are there multiple decoding seeds or repeated runs for Tables 3, 5 and 6, or are per-sample results available to compute paired confidence intervals?

4.How were the fidelity priors and hyperparameters (ϵ,λ,η,γ,ρ)(\epsilon,\lambda,\eta,\gamma,\rho) chosen in Table 7? Were these values fixed before inspecting the 250-example evaluation subset, and were they tuned on separate validation or development samples?

4 comments
AuthorsOpenReview ↗

Response to Reviewer c4js

Thank you for your careful, constructive review and for recognizing the local-vs.-global view bottleneck, our closed-loop candidate–claim design, and our same-MLLM ablations. Our General Response also summarizes six studies on equal-budget planning, score-matched OET, runtime/token efficiency, extended-set evaluation, five-run and hyperparameter stability, results with open-weight Qwen MLLMs, and a concurrent method (TAB). We address your questions and concerns below.

W1 / Q1: Under the same five-round budget, does judgment-guided claim selection beat random or fixed-order selection?

Under the suggested equal-budget design on the paper's 250 ScanRefer queries (five-round cap; identical candidate grouping, view construction (RCOC), joint view selection (OET), and MLLM; varying only which claim is selected), Judgment-guided planning beats Fixed order by +2.8% / +3.2% (62.4 / 55.6 vs. 59.6 / 52.4) and Random order by +2.0% / +2.8% (vs. 60.4 / 52.8), because tracking candidate–claim judgments avoids wasting rounds on claims already satisfied by remaining candidates (rechecking a satisfied claim, unable to eliminate any candidate; 4.1%–6.2% is the five-run range):

Claim-selection policy (ScanRefer, 250 queries, 5-round cap)Steps rechecking satisfied claimsMLLM calls / queryScanRefer [email protected] / @0.5
Fixed order38.8%7.059.6 / 52.4
Random order34.7%6.960.4 / 52.8
Judgment-guided (Ours)4.1%–6.2%6.562.4 / 55.6

Because claim 1 checks coarse category (confirmed initially), fixed order spends round 1 rechecking category on 100% of queries entering verification, whereas judgment-guided planning skips supported claims on 88% of round-1 steps. It cuts redundant rechecks 6×–9×, eliminates 55% more distractors than fixed order, and beats fixed order (95% CI [+1.2, +5.6], p=0.008p=0.008 at [email protected]) and random order ([+0.8, +5.2], p=0.039p=0.039 at [email protected]) with 15%–19% fewer verification tokens; all five repeated judgment-guided runs beat both baselines (95% CIs are paired bootstrap intervals; pp-values are from two-sided exact McNemar tests). This confirms that the gain comes from choosing what to verify from the unresolved claims, not merely from iterating for the same number of rounds.

W2 / Q2: Does OET still help against a score-matched greedy baseline on Eq. (2), and how reliable are the gains (CIs, temperature, seed, API)?

Against the suggested score-matched baseline, which, unlike the visibility-only Greedy of Table 3 in the paper, aggregates the view score of Eq. (2) across candidates and picks the top view without transport (Score-Matched Greedy, keeping proposals, planner, five-round cap, and the requirement that candidates and anchors (the reference objects named in the query) appear together identical), OET still gains +3.2% / +2.8% on ScanRefer, as Eq. (3)'s per-candidate demand prevents easy candidates from dominating the sum. We answer the three parts of this question in turn: the isolated effect of OET, the statistical reliability of the gains, and the temperature, seed, and API settings.

  • Isolated OET effect. OET (62.4 / 55.6) beats Score-Matched Greedy (59.2 / 52.8) by +3.2% / +2.8% (paired 95% CI [+0.4, +6.0] / [+0.4, +5.6]) and Eq. (2) greedy without shared views (58.8 / 52.4) by +3.6% / +3.2% ([+0.4, +6.8], p=0.049p=0.049 at [email protected]).
  • Statistical reliability. The 250-query subset is the standard evaluation protocol of zero-shot 3DVG, but on 250 queries paired bootstrap 95% CIs are inherently wide (about ±5–6 points), so a comparison on this subset alone may not fully resolve gains of a few points. We therefore verified the gains further, on more data and across MLLMs. Extending the evaluation from the 250 queries to 750 ScanRefer queries, ARIADNE reaches 62.1 / 55.5 vs. SeqVLM's 52.3 / 46.3 under SeqVLM's category filter (Table 7) and 64.9 / 57.6 vs. SPAZER's 62.4 / 54.9 (+2.5 [-0.9, +5.9] / +2.7 [-0.5, +5.9]) with SPAZER's zero-match fallback rule (retaining scene proposals when no initial candidate passes the target category). Across 3 MLLMs × 2 metrics, ARIADNE beats SPAZER 6/6 times (Seed 2.0 Lite +1.6 [-4.4, +7.6] / +2.8 [-2.4, +8.4], Qwen3.8-27B +3.2 / +3.6, Qwen3.6-27B +2.0 / +3.2) and significantly beats SeqVLM and SeeGround on each MLLM (e.g., +9.2% [+2.8, +16.0], p=0.009p=0.009 / +8.0% [+2.0, +14.0], p=0.013p=0.013 vs. SeeGround on Seed 2.0 Lite); please also refer to the table in our response to Rg4cX (W3). The five-run standard deviation (please see the response to Q3) further confirms stability.
  • Temperature, random seed, API. Seed 2.0 Lite (checkpoint doubao-seed-2-0-lite-260215) uses the commercial API at its default sampling settings (temperature 1.0, top-p 0.95; not overridden), seed = 0 for grouping and category filter, and structured tool output; we report these here and will add API details to Appendix B.

Q3: Is Seed 2.0 Lite inference deterministic, and do Tables 3, 5, and 6 rest on repeated runs or per-sample results that allow paired confidence intervals?

Hosted LLM APIs such as Seed 2.0 Lite are not fully deterministic, as documented by the providers themselves and by prior studies [1–5]; this comes from the hosted backend, not from ARIADNE. In our setting the effect is small: over five repeated runs of the paper setting, accuracy varies by a standard deviation of only 0.6 / 0.9 points ([email protected] / [email protected]), and lowering the temperature to 0 further raises cross-run answer agreement from 73%–74% to 77%. Locally served open-weight models are more deterministic: we found empirically that with greedy decoding, sending the same request six times to Qwen3.8-27B or Qwen3.6-27B returns six identical outputs.

The reported gains, including those of the ablations in Tables 3, 5, and 6, are therefore not the product of a favorable run. Even the worst of the five repeated runs (61.6 / 54.2) still outperforms all prior zero-shot methods (Table 1 of the paper) as well as the single-pass and no-regularization variants (Tables 3 and 5); every run beats the fixed- and random-order planners and Score-Matched Greedy (please see the response to W1 and W2); and ARIADNE beats SPAZER, the strongest same-MLLM baseline of Table 6, on all three MLLMs tested (please see the response to W2). Per-query results of all runs, from which the paired CIs in the response to W2 are computed, will be released on publication.

References

  • [1] OpenAI, "How to make your completions outputs reproducible with the new seed parameter," OpenAI Cookbook. https://developers.openai.com/cookbook/examples/reproducible_outputs_with_the_seed_parameter
  • [2] Anthropic, "Glossary: Temperature," Claude API Documentation. https://platform.claude.com/docs/en/about-claude/glossary
  • [3] S. Ouyang, J. M. Zhang, M. Harman, and M. Wang, "An Empirical Study of the Non-determinism of ChatGPT in Code Generation," ACM Transactions on Software Engineering and Methodology, vol. 34, no. 2, 2025.
  • [4] B. Atıl, S. Aykent, A. Chittams, et al., "Non-Determinism of 'Deterministic' LLM System Settings in Hosted Environments," in Proc. 5th Workshop on Evaluation and Comparison of NLP Systems (Eval4NLP), 2025.
  • [5] J. Yuan, H. Li, X. Ding, et al., "Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference," in Advances in Neural Information Processing Systems (NeurIPS), 2025.

Q4: How were the fidelity priors and hyperparameters in Table 7 chosen, and was a separate validation set used?

ARIADNE is not sensitive to these hyperparameters. The Table 7 values were set once from simple a-priori principles and not tuned on a specific set: (i) the fidelity priors only encode an order of trust in the view types (raw captured frames 1.0 > stitched 0.9 > rendered 0.75 > warp-flagged 0.7); (ii) the OET weights are generic defaults (unit demand weight λ=1.0\lambda = 1.0; small entropy and history weights ϵ=η=0.15\epsilon = \eta = 0.15 with a 10% history refresh ρ=0.1\rho = 0.1; graph weight γ=0.25\gamma = 0.25). They were fixed without reference to the 250-query evaluation subset, without grid search, and without a separate validation split (following SPAZER and SeqVLM). We re-tested them here in two ways. Extending the evaluation from the paper's 250 ScanRefer queries to 750, ARIADNE scores 62.1 / 55.5 vs. 62.4 / 55.6 on the 250 (-0.3% / -0.1%), so the result does not depend on the 250-query subset. And varying each hyperparameter around its Table 7 value barely changes which view is selected:

Hyperparameter (Table 7 value)Range around the Table 7 valueSelected views unchanged
Entropy weight ϵ\epsilon (0.15)0.05 – 5≥ 98%
Demand weight λ\lambda (1.0)0.25 – 4≥ 99%
History weight η\eta (0.15)0.05 – 1.5≥ 95%
History refresh ρ\rho (0.1)0.05 – 0.3≥ 97%
Graph weight γ\gamma (0.25)0.1 – 10≥ 99%
Fidelity priors of stitched / rendered views (0.9 / 0.75)± 0.05≥ 96%

Within the ranges above, fewer than 5% of the selected views change for any OET weight, so further tuning would change almost nothing. For the fidelity priors, what matters is their order rather than their exact values: shifting them by ±0.05\pm 0.05 while keeping the order changes at most 3.1% of the selections, whereas reversing the order changes 15.5%. This is why they follow a simple fixed order (raw captured frames over stitched over rendered views) rather than values fitted to the benchmark.

AuthorsOpenReview ↗

Brief summary of our responses

Dear Reviewer c4js,

Thank you again for your thoughtful review. We appreciate your time during this busy period, and as the discussion comes to a close, we briefly summarize our main responses below.

Under the suggested five-round equal-budget control and Eq. (2) Score-Matched Greedy baseline, judgment-guided planning beats fixed and random orders (62.4 / 55.6 vs. 59.6 / 52.4 and 60.4 / 52.8; 6×–9× fewer redundant rechecks) and OET gains +3.2 / +2.8 (95% CI [+0.4, +6.0] / [+0.4, +5.6]). We also reported the checkpoint and API settings, 3× extended evaluation (from the standard 250-query ScanRefer subset to 750 queries) and 3-MLLM comparisons, five-run stability (standard deviation 0.6 / 0.9), and sensitivity checks across all five OET weights (≥ 95% unchanged views) and fidelity priors (≥ 96%).

We welcome any feedback on whether our responses address your concerns, and we would be glad to clarify any remaining questions.

Dear Authors, thank you for the extensive rebuttal and carefully designed control experiments. The equal-budget comparison against fixed/random claim scheduling, the score-matched greedy baseline for OET, repeated runs with confidence intervals, extended evaluation on 750 ScanRefer samples, and hyperparameter sensitivity analysis have addressed my main concerns. The results clearly disentangle the benefits brought by the judgment-guided planner and the OET objective.

One minor point remains: the small-magnitude gains of OET still rely on a hosted LLM API with inherent non-determinism, even though you have demonstrated stability across five repeated runs. I suggest the final manuscript clearly mark this limitation in discussion.

I adjust my score to 3. I also encourage you to submit rebuttal materials earlier in future cycles to allow more time for discussion.

Reviewer_71GFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ARIADNE, a training-free framework for zero-shot 3D visual grounding. The method uses a multimodal large language model to break down referring expressions into visual claims and identify which conditions remain unclear for different object candidates. It then constructs different types of views, including local, stitched, and rendered images, to provide additional visual evidence. An Online Evidence Transport (OET) module selects the next view for comparison. Experiments on ScanRefer and Nr3D show improvements over existing zero-shot methods.

Critical strengths
  1. The paper addresses a meaningful problem in zero-shot 3D visual grounding. Actively selecting views to distinguish similar objects is a reasonable and useful idea.

  2. The proposed framework is easy to understand, and its components work together in a clear way.

  3. The method achieves competitive results on ScanRefer and Nr3D. The paper also provides ablation studies and comparisons using the same MLLM, which help support the effectiveness of the framework.

Critical weaknesses
  1. Limited methodological novelty. Although the overall framework is reasonable, the main components are largely based on existing techniques, including language-based condition decomposition, multi-view observation, and transport-based optimization. The method mainly combines these components into an iterative framework, and the technical novelty appears relatively limited.

  2. The necessity of OET is not sufficiently demonstrated. OET introduces a relatively complex optimization objective with graph and history regularization. However, compared with a simple greedy view-selection strategy, it only improves [email protected] by 1.6 percentage points and [email protected] by 1.2 percentage points on ScanRefer. It is unclear whether this additional complexity is justified by the performance gains.

  3. Limited analysis of inference efficiency. The proposed method requires multiple rounds of MLLM reasoning, view construction, and optimization. However, the paper does not provide sufficient analysis of end-to-end inference time, the number of MLLM calls, or token consumption. These results are important for understanding the practical cost of the proposed active perception framework.

  4. The evaluation scale is relatively small. The experiments use fixed subsets of 250 examples from ScanRefer and Nr3D. Given the relatively modest improvements over some baselines, evaluation on larger subsets or additional analysis of statistical stability would make the results more convincing.

Questions for authors
  1. Could the authors further clarify the main technical novelty of ARIADNE compared with existing active perception and multi-view reasoning frameworks?

  2. Could the authors compare OET with simpler view-selection strategies that consider the joint visibility of candidate and reference objects, without transport optimization?

  3. What is the average inference time, number of MLLM calls, and token consumption per example? How does grounding accuracy change with different observation budgets?

  4. Could the authors evaluate the method on larger test subsets or provide statistical analysis to show that the reported improvements are stable?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 71GF

Thank you for your time and for the careful, constructive review, as well as for recognizing the value of active view selection for candidate disambiguation, the clarity of our framework, and our controlled ablations. Please also see our General Response, which summarizes six additional studies covering equal-budget planning, score-matched OET, runtime/token efficiency, extended-set evaluation and five-run stability, results with open-weight Qwen MLLMs, and concurrent TAB. We address your questions and concerns below.

W1 / Q1: What is ARIADNE's main technical novelty over existing active-perception and multi-view frameworks?

Claim decomposition, multi-view observation, and transport optimization do exist individually; ARIADNE's main technical novelty over existing multi-view and active perception frameworks is not their combination but closing the loop between per-candidate unresolved claims (query conditions not yet verified for a candidate) and comparison-specific joint view synthesis and selection. This loop is what lets ARIADNE resolve the cases where predetermined views and judgment-agnostic schedules fail: they either miss the relation between a candidate and its anchor (the reference object named in the query) or waste rounds re-observing settled attributes. Specifically, ARIADNE differs in three ways:

  1. Candidate–claim judgments drive observation planning. SeeGround, SeqVLM, and SPAZER run fixed stages over predetermined views; View-on-Graph traverses a static scene graph; and the concurrent method TAB (EMNLP 2026) runs a prescribed 13-step skill that commits to the target on one reference frame before tracking its 2D mask. None maintains a per-candidate, per-claim judgment record (supported, contradicted, unresolved). In an equal-budget ScanRefer ablation (varying only claim choice under a five-round cap), replacing judgment-guided selection with a fixed or random claim order drops accuracy from 62.4 / 55.6 to 59.6 / 52.4 (p=0.008p = 0.008 at [email protected]; all pp-values are from two-sided exact McNemar tests on paired per-query results) and 60.4 / 52.8 (p=0.039p = 0.039 at [email protected]) while increasing redundant rechecks (steps inspecting a claim already supported for all remaining candidates, unable to eliminate any candidate) by 6×–9×, because state-independent orders fail to introduce new informative observations.
  2. Views are synthesized on demand for the unresolved relation. Because existing frames or global renders often miss fine candidate detail or the reference anchor, RCOC (Relation-Centric Observation Construction) constructs depth-reprojected panoramas (Stitch) and targeted 3D camera views (Render) aimed at the current candidates and anchors, lifting accuracy from 56.8 / 52.8 with captured Local views alone to 62.4 / 55.6 (Table 4 of the paper).
  3. Views are selected jointly across competing candidates. Because a close-up favoring one candidate can hide rival candidates or the anchor, OET (Online Evidence Transport) coordinates view selection across all compared candidates rather than scoring proposals independently.

We report the following comparison table here and will add it to Section 2 of the revised manuscript:

FrameworkWhat drives the next observationViews shown to the MLLMHow views are selected
SeeGroundFixed single pass1 mesh render oriented toward anchorDeterministic rule
SeqVLMFixed elimination tournamentObject-centric sequences of captured RGB framesPer-proposal top-kk
SPAZERFixed 2-stage pipeline5 pre-rendered global mesh views + captured RGB framesPer-stage / per-candidate
VLM-Grounder2D grounding feedbackCaptured RGB framesPer-query retrieval
View-on-GraphScene-graph traversalCaptured RGB framesGraph neighbor step
Concurrent TAB (EMNLP'26)Prescribed 13-step ReAct skillCaptured RGB frames; target chosen on 1 reference frameGrounding DINO detection + SAM3 segmentation + VLM scoring
ARIADNE (Ours)Unresolved candidate–claim ledgerCaptured RGB frames (Local), depth-reprojected Stitch, and on-demand candidate+anchor 3D mesh RenderJoint co-visibility + OET across competing candidates
AuthorsOpenReview ↗

Response to Reviewer 71GF (continued)

W2 / Q2: Is OET necessary compared with simpler joint-visibility view selection without transport?

OET is worth its added complexity: it adds negligible overhead (0.5 ms per step on one CPU core, <0.01% of query runtime, zero extra MLLM calls) and outperforms simpler joint-visibility rules: summing visibility scores lets one easily visible candidate dominate view choice and leaves harder rival candidates poorly observed, which OET's joint allocation prevents. As suggested in Q2, to test whether joint-visibility scoring alone suffices without transport, we compared OET on ScanRefer against three new non-transport greedy policies, with the proposals, claims, and planner kept identical:

  • Score-Matched Greedy on Eq. (2): retains views where compared candidates and anchors can be shown together, sums the exact same Eq. (2) joint-visibility score q^t(e,a)=ft(a)min⁡j∈Mt(e)vt(j,a)\widehat q_t(e,a)=f_t(a)\min_{j\in\mathcal M_t(e)}v_t(j,a) across candidates, and picks the maximum without transport.
  • Joint-Visibility Greedy without the shared-view requirement: maximizes the summed Eq. (2) score over all candidate views without requiring candidates and anchors to appear together.
  • Target-Visibility Greedy: ranks all candidate views by target visible fraction alone.

All four policies are evaluated on the 250-query ScanRefer subset used in the paper:

View-selection policy (ScanRefer, 250 queries)Joint comparisons showing candidates & anchors together[email protected] / @0.5
Target-Visibility Greedy (no shared-view requirement)24.3%59.2 / 52.4
Joint-Visibility Greedy on Eq. (2) (no shared-view requirement)20.3%58.8 / 52.4
Score-Matched Greedy on Eq. (2)63.8%59.2 / 52.8
OET (Ours)66.9%62.4 / 55.6

First, omitting the shared-view requirement drops joint coverage to 20.3%–24.3%, lowering ScanRefer accuracy (p=0.049p = 0.049 at [email protected], p=0.022p = 0.022 at [email protected]), whereas OET shows all candidates and anchors together in 66.9% of steps (100% of physically feasible cases; in the remaining ~one-third, objects sit across room regions that no single camera or panorama can capture together). Second, even under the same shared-view requirement and Eq. (2) scores, Score-Matched Greedy simply sums scores across candidates, so a view showing most candidates clearly but one rival poorly can outscore a balanced view showing all of them clearly; failing to eliminate that poorly observed rival leaves larger candidate sets in later rounds (harder to fit into one shared view, reducing joint coverage to 63.8% and requiring 15% more verification images) and lowers accuracy by 3.2% / 2.8% overall (+3.8% / +3.3% for OET on queries with same-class distractors, 55.4 / 47.8 vs. 51.6 / 44.6), whereas OET's per-candidate demand in Eq. (3) and history/graph regularizers (+4.8%, Table 5 of the paper) enforce balanced evidence across all competing candidates (all five repeated OET runs beat 59.2 / 52.8).

AuthorsOpenReview ↗

Response to Reviewer 71GF (continued)

W3 / Q3: What are the inference time, MLLM calls, and tokens per query, and how does accuracy change with the observation budget?

On ScanRefer, ARIADNE averages 6.5 MLLM calls (median 3.0), 41.9k input / 2.3k output tokens, and 48.3 s mean / 10.2 s median end-to-end single-query runtime (including view construction, OET, and MLLM calls). This is 3.7×–5.6× faster in median time than SPAZER (37.9 s) and SeqVLM (57.1 s), because verification rounds (further rounds of observing and re-judging the remaining candidates) run only when several candidates survive the initial screening (66% of queries finish in zero verification rounds). To evaluate practical cost under a common MLLM, we measured per-query calls, tokens, and single-query runtime (one query at a time) for SeeGround, SeqVLM, and SPAZER, the baselines of Table 6 in the paper, alongside TAB:

Method (ScanRefer, 250 queries)MLLM calls / query (mean / med.)Tokens / query (in / out)Single-query time, 1 query at a time (mean / med.)ScanRefer [email protected] / @0.5
SeeGround2.0 / 2.06.0k / 1.6k46.3 s / 34.1 s53.2 / 47.6
SeqVLM1.9 / 1.07.4k / 13.3k136.9 s / 57.1 s57.2 / 52.0
SPAZER (same MLLM)1.9 / 2.06.8k / 2.3k41.3 s / 37.9 s60.8 / 52.8
SPAZER (GPT-4o, as reported in its paper)——23.5 s / —57.2 / 48.8
ARIADNE (Ours)6.5 / 3.041.9k / 2.3k48.3 s / 10.2 s62.4 / 55.6 (72.6 / 64.0, mean of two runs, with TAB's inputs; SAM3 masks and TAB's manually refined queries)
TAB295 / 260685k / 5.0k83.7 s / —71.2 / 46.4

TAB's time and accuracy are those reported in its paper; its calls and tokens were measured by us on 20 randomly sampled queries with its released code and model (tokens are counted by each method's own MLLM). TAB thus needs about 45× more MLLM calls and 16× more input tokens per query than ARIADNE.

First, computation scales with ambiguity: stopping once one candidate remains means 66% of queries use zero verification rounds (1.3 mean, median 0). ARIADNE's 10.2 s median time is therefore 3.3×–5.6× lower than the baselines', whose MLLM calls each produce a long reasoning trace before answering (about 0.8k–7k output tokens per call, vs. 0.35k for ARIADNE), while its mean time stays on the same scale as SPAZER and SeeGround (48.3 s vs. 41.3 s and 46.3 s) and far below SeqVLM (136.9 s).

Second, across observation budgets, accuracy improves from 54.8 / 51.2 under single-pass observation (Table 3 of the paper) to 62.4 / 55.6 (up to five rounds), with eliminations concentrated in early rounds (45%, 29%, 26%, 17%, and 13% of queries running rounds 1–5 eliminate a candidate). A new early-stopping variant that halts verification once a round eliminates no candidate (averaged over all 250 queries) cuts input tokens by 30.3% at a cost of only 0.8 / 0.4 points:

Observation-budget setting (ScanRefer, 250 queries)Mean verification rounds / queryInput tokens (relative)ScanRefer [email protected] / @0.5
Full budget (up to 5 rounds, paper setting)1.3100%62.4 / 55.6
Stop when a round eliminates no candidate (new variant)0.569.7%61.6 / 55.2

W4 / Q4: Do the results hold on a larger test set, and are the reported gains statistically stable?

The results hold on a larger test subset and are stable across repeated runs. To verify this and rule out overfitting to the 250-query subset used in the paper (a pre-established evaluation subset introduced by VLM-Grounder and followed by SeqVLM, SPAZER, and TAB to bound API cost, which we simply follow), we extended, as suggested, the evaluation from the 250-query subset to 750 ScanRefer validation queries and ran five repeated runs of the paper setting. Under the paper's SeqVLM category filter (Table 7), ARIADNE achieves 62.1 / 55.5 on all 750 queries, matching 62.4 / 55.6 on the 250-query subset and beating SeqVLM (52.3 / 46.3) by +9.9 / +9.2 (p<0.001p < 0.001 at both thresholds); with SPAZER's zero-match fallback rule (retaining scene proposals when no initial candidate passes the target category), ARIADNE reaches 64.9 / 57.6, surpassing SPAZER (62.4 / 54.9). Across these five repeated runs, accuracy varies by a standard deviation of only 0.6 / 0.9 points ([email protected] / [email protected]), so the reported results depend neither on the 250-query subset nor on a single run.

AuthorsOpenReview ↗

Brief summary of our responses

Dear Reviewer 71GF,

Thank you again for your thoughtful review. We appreciate your time during this busy period, and as the discussion comes to a close, we briefly summarize our main responses below.

We clarified ARIADNE's closed-loop novelty (with a 7-method comparison and an equal-budget planner ablation, 62.4 / 55.6 vs. 59.6 / 52.4 and 60.4 / 52.8) and showed that OET (0.5 ms/step) outperforms Score-Matched Greedy (59.2 / 52.8) and simpler joint-visibility rules (58.8–59.2 / 52.4). We also added end-to-end cost and budget analyses (10.2 s median runtime, 3.7×–5.6× faster than SPAZER and SeqVLM; 30.3% fewer input tokens at 61.6 / 55.2 with early stopping) and verified stability by extending the standard 250-query ScanRefer subset 3× to 750 queries (62.1 / 55.5, and 64.9 / 57.6 vs. SPAZER's 62.4 / 54.9 under its fallback rule) and across five repeated runs (standard deviation 0.6 / 0.9).

We welcome any feedback on whether our responses address your concerns, and we would be glad to clarify any remaining questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all three reviewers (R71GF, Rc4js, and Rg4cX) for their constructive, detailed feedback and for recognizing the significance of candidate disambiguation in zero-shot 3D visual grounding (R71GF-S1, Rc4js-S1, Rg4cX-S1), the clarity of ARIADNE's closed-loop active perception design (R71GF-S2, Rg4cX-S2), and the value of our ablations and controlled same-MLLM comparisons (R71GF-S3, Rc4js-S2, Rg4cX-S3).

To address raised questions and concerns from reviewers directly, we further ran six controlled studies, which will be added to the revised manuscript (all pp-values below are from two-sided exact McNemar tests on paired per-query results):

  • Equal-budget visual claim planner ablation (Rc4js W1/Q1, R71GF W1/Q1). To test whether ARIADNE's gain comes from which visual claim (an attribute or relation stated in the query) it verifies next, and not merely from running more rounds, we followed the equal-budget design proposed by Rc4js: only this choice changes, while the input proposals, view construction (RCOC), joint view selection (OET), MLLM, and five-round cap stay identical. Choosing the next claim from the current per-candidate judgments reaches 62.4% / 55.6% on ScanRefer, against 59.6% / 52.4% when the claims are verified in a fixed order (p=0.008p = 0.008 at [email protected]) and 60.4% / 52.8% in a random order (p=0.039p = 0.039 at [email protected]), with 6×–9× fewer redundant rechecks (steps spent on a claim already supported for all remaining candidates, which cannot eliminate any candidate). This empirically demonstrates that the gain comes from what the planner chooses to verify, not from the extra rounds.
  • Score-matched OET isolation (R71GF W2/Q2, Rc4js W2/Q2). Both reviewers ask whether OET is worth its complexity over simple greedy view selection. As Rc4js suggested, we built Score-Matched Greedy: unlike the visibility-only Greedy of Table 3 in the paper, it uses exactly OET's view score (Eq. (2) of the paper) over the same views, and only replaces transport by picking the top-scoring view. On ScanRefer, OET reaches 62.4% / 55.6% against 59.2% / 52.8% for Score-Matched Greedy, and it also outperforms the two simpler visibility-based greedy rules requested by R71GF (details in our response to R71GF, W2/Q2). With the score held identical, this shows that the gain comes from selecting the view jointly for all candidates, and it is free in practice: OET takes 0.5 ms per step on CPU (<0.01% of query time) and adds no MLLM calls.
  • Inference efficiency and observation budget (R71GF W3/Q3, Rg4cX W1). Both reviewers ask what the active loop costs. It does not make ARIADNE slow, because extra observations are spent only on ambiguous queries: verification stops as soon as one candidate remains, so 66% of ScanRefer queries need no verification round at all. In single-query runtime (one query at a time) under the same MLLM, ARIADNE's 10.2 s median time is 3.7× faster than SPAZER (37.9 s) and 5.6× faster than SeqVLM (57.1 s). The budget can be reduced further at almost no cost in accuracy: a new early-stopping variant (also stopping when a round eliminates no candidate) cuts input tokens by 30.3% and still reaches 61.6% / 55.2%, against 62.4% / 55.6% with the full budget. The full per-query time, MLLM-call, and token tables are in our responses to R71GF (W3/Q3) and Rg4cX (W1).
  • Larger test subset, repeated runs, and hyperparameter sensitivity (R71GF W4/Q4, Rc4js W2/Q3/Q4). As R71GF suggested, to check that our results do not depend on the 250-query ScanRefer subset used in the paper (the standard subset of prior zero-shot work), we extended the evaluation from these 250 queries to 750 ScanRefer queries. Under the paper's SeqVLM category filter, ARIADNE reaches 62.1% / 55.5% on all 750, matching its 62.4% / 55.6% on the original 250 and beating SeqVLM (52.3% / 46.3%) by +9.9% / +9.2%; with SPAZER's zero-match fallback rule (last bullet), ARIADNE reaches 64.9% / 57.6%, surpassing SPAZER (62.4% / 54.9%). The results are also stable across five repeated runs (standard deviation of 0.6 / 0.9 points) and not sensitive to hyperparameters: varying each of the five OET weights over a wide range (e.g., the entropy weight from 0.05 to 5; all ranges in our response to Rc4js, Q4) keeps at least 95% of view selections unchanged, and shifting the fidelity priors of the view types by ±0.05\pm 0.05 keeps at least 96%. Together, these checks confirm that the reported accuracy is not an artifact of the 250-query subset, a single run, or the exact hyperparameter values.
AuthorsOpenReview ↗

General Response (continued)

  • Results with open-weight Qwen MLLM backbones (Rg4cX W3). Rg4cX asks whether ARIADNE generalizes to smaller open-source MLLM backbones such as the Qwen series. With unchanged prompts, ARIADNE remains the best of the four frameworks under the same backbone: 61.6% [email protected] on Qwen3.8-27B (SPAZER 58.4%, SeqVLM 53.6%, SeeGround 52.8%) and 59.6% on Qwen3.6-27B (SPAZER 57.6%, SeqVLM 52.0%, SeeGround 51.6%); the full results are in our response to Rg4cX (W3). Its own accuracy is also similar across the 27B Qwen models and the two commercial models, Seed 2.0 Lite and DeepSeek-V4.1-Flash (59.6%–62.4% [email protected]). This reveals that ARIADNE's advantage holds across MLLM backbones and is not tied to a specific one.
  • Proposal fault tolerance and comparison with TAB (Rg4cX W2/W4). Rg4cX notes that a target missed by the pre-generated proposals or by the initial filtering cannot be recovered. This was deliberate: every method relying on such proposals loses these targets, and we chose not to compensate, because our contribution is the agentic disambiguation workflow, not better perception; keeping exactly the proposals and class-name filter of prior work ensures that the reported gains come from the workflow alone. Fault tolerance is nevertheless simple to add: a zero-match fallback of the kind SPAZER already uses (screening the remaining scene proposals when no initial candidate passes the target category) lifts ARIADNE from 62.4% / 55.6% to 66.8% / 58.8% on ScanRefer, a clear gain from such a simple procedure. Unlike TAB (EMNLP 2026), which picks its target on one static video frame, ARIADNE disambiguates candidates across rounds via stitched panoramas (Stitch) and targeted renders (Render). Under matched inputs (2D SAM3 masks and the refined queries introduced by TAB, with the fallback above), ARIADNE reaches 72.6% / 64.0% on ScanRefer (mean of two runs; vs. TAB's 71.6% / 61.6%) and 77.2% on Nr3D (vs. TAB's 67.2%). This illustrates that ARIADNE's workflow improves directly with better perception inputs, further supporting its generality.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.