acceptodds
Under review as a conference paper at ICLR 2027

What to Attend, Where to Act: Referent–Affordance Co-Evolution for Embodied Spatial Grounding

Abstract

Embodied spatial grounding refers to interpreting a language instruction and identifying the visual region for intended interaction. Existing methods broadly follow two paradigms: coordinate-generation methods produces discrete coordinate tokens as answers, whereas spatial-decoding methods predicts directly from visual representations to preserve spatial structure. However, both typically ground semantics into spatial predictions in a one-way manner, limiting reciprocal interaction between referent semantics and visual affordance evidence. To this end, we propose RACE, a Referent–Affordance Co-Evolution framework that couples referent semantics specifying what to attend to with an affordance state indicating where to act, jointly refining referent understanding and spatial localization. Specifically, referent semantics guide updates to the affordance state, which in turn feeds back visual evidence to refine referent understanding. To supervise guide this co-evolution, we unify heterogeneous grounding annotations as target distributions and combine distribution matching with a geometry-aware constraint penalizing spatially inconsistent predictions. Our model achieves 54.50% on RefSpatial-Bench and 74.00% on Where2Place while maintaining competitive general spatial-reasoning performance. These results suggest that referent–-affordance co-evolution provides an effective interface between task-relevant semantics and fine-grained spatial grounding, offering a general pathway from task-level reasoning to actionable spatial representations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, Spatial Grounding, Affordance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GAJcOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RACE, which achieves spatial grounding through the co-evolution of referent semantics and spatial predictions. It uses affordance probability maps as an intermediate representation for grounding and introduces an explicit architecture for progressively refining this map.

Critical strengths
  1. The co-evolution of semantics and affordance enables progressive refinement of grounding predictions. The explicit affordance representation supports residual delta updates, allowing the model to localize the target object more accurately through coarse-to-fine refinement. The explicit affordance delta updates in a corarse to fine manner is novel for sptial grounding.
Critical weaknesses
  1. The paper lacks sufficient novelty. Though the explicit affordance used in spatial grounding is novel, its central idea of reciprocal interaction between semantics and spatial predictions has already been explored in FlowSeg. In my view, the two methods do not differ substantially at a conceptual level: RACE uses an explicit affordance representation for grounding, whereas FlowSeg uses visual queries.
  2. The experimental evaluation is insufficient. The paper does not compare against FlowSeg, and the existing experiments do not sufficiently isolate the benefits of using an explicit affordance probability map for grounding. More baseline comparisons and ablation studies are needed, particularly a variant that retains co-evolution without using an explicit affordance representation, to assess whether this representation provides additional benefits.
Questions for authors
  1. Would grounding performance differ substantially if spatial predictions were represented without an explicit affordance map, for example using visual queries as in FlowSeg or another representation? If so, why?If there is no substantial difference, what distinguishes the proposed method from FlowSeg?
5 comments
AuthorsOpenReview ↗

Response to Reviewer GAJc

GAJc-Q1: Would grounding performance differ substantially if spatial predictions were represented without an explicit affordance map, for example using visual queries as in FlowSeg or another representation? If so, why?If there is no substantial difference, what distinguishes the proposed method from FlowSeg?

A1: Thank you for this question. Our comparison with a FlowSeg-style latent co-evolution decoder shows accuracy gains of 3.04–7.80 percentage points across the reported subsets and benchmarks. RACE and FlowSeg share bidirectional semantic–visual interaction, but their differences go beyond performance and task settings:

1.Motivation. FlowSeg focuses on aligning language conditions with candidate masks. RACE aims to preserve the backbone’s semantic and QA capabilities, using a dedicated spatial decoder for task-relevant grounding. This design assigns distinct roles to general semantic understanding and task-relevant spatial inference, while coupling semantic and spatial states through iterative feedback for fine-grained grounding.

2.Mechanism. FlowSeg refines language conditions and visual queries; RACE unifies heterogeneous annotations into distributional targets and explicitly maintains the same spatial distribution for both grounding prediction and feedback evidence selection. Grounding supervision therefore directly constrains feedback weighting, enabling co-evolution between referent semantics and explicit spatial hypotheses.

3.Experimental Comparisons. To evaluate an alternative representation, we constructed a FlowSeg-style latent co-evolution decoder with 16 visual queries propagated across stages through bidirectional interaction with semantic representations. It uses the same backbone initialization, training data, and evaluation protocol as RACE.

MethodLocationPlacementUnseenWhere2PlaceRoboRefIt
Latent co-evolution decoder51.0045.0036.3667.0078.29
RACE58.0051.0044.1674.0081.33

These results support the benefit of an explicit affordance distribution in our grounding setting. The coupling between grounding supervision and feedback weighting may help explain these gains.

Reviewer_GAJcOpenReview ↗

Official Comment by Reviewer GAJc

It remains unclear why RACE places greater emphasis than FlowSeg on preserving the backbone’s semantic and QA capabilities. I do not see how this distinction follows from the proposed mechanism or architectural design. FlowSeg also includes a language modeling loss and uses a dedicated decoder for spatial grounding. Explicitly maintaining a spatial distribution alone does not appear to explain better preservation of the backbone’s semantic and QA capabilities. Could the authors clarify which specific architectural or training choices support this claimed distinction?

AuthorsOpenReview ↗

Clarifying the Differences Between RACE and FlowSeg

We really appreciate your question. We clarify the differences between RACE and FlowSeg in terms of application, motivation, and mechanism.

1. Application. General and embodied grounding/QA in RACE vs. Segmentation in general scenes in FlowSeg.

RACE targets diverse grounding and spatial QA tasks across general and embodied settings within a single model. Our training data and evaluations cover multiple settings and tasks, with results showing strong overall performance. FlowSeg focuses on language-conditioned instance segmentation in general visual scenes, and its reported evaluations focus on segmentation benchmarks.

2. Motivation. Semantic–spatial correlation in RACE vs. Semantic disambiguation in FlowSeg.

RACE uses iterative interaction to jointly refine semantic understanding and spatial grounding: semantics guide spatial distribution updates, while spatial evidence feeds back to refine referent semantics. FlowSeg focuses on aligning language conditions with candidate masks for semantic disambiguation.

3. Mechanism. Explicit spatial distributions in RACE vs. Implicit spatial representations in FlowSeg.

RACE explicitly maintains and updates a spatial distribution across stages. Grounding supervision acts directly on this distribution, which weights the visual evidence fed back for semantic refinement and serves as the final grounding prediction. The same distribution connects supervision, feedback, and prediction, without a separate set of latent visual queries carrying the spatial state. FlowSeg instead introduces an additional set of visual queries and jointly refines these queries and language conditions through bidirectional interaction in latent space.

Reviewer_sDCKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RACE for embodied spatial grounding. Its grounding head jointly updates a referent query and a spatial distribution over four stages. Previous spatial predictions guide visual feature aggregation, which then informs query refinement and subsequent localization. Point, box, and mask annotations are converted into distributional targets, with an additional geometry-aware objective penalizing spatially inconsistent responses. The paper evaluates a 2B model on grounding and spatial-understanding benchmarks and demonstrates integration with an action head in real-world pick-and-place experiments.

Critical strengths
  1. A coherent feedback mechanism with supporting ablations. The spatial distribution has an explicit role in selecting visual evidence for subsequent query refinement. Table 3 shows that the full configuration outperforms variants without feedback or with uniform feedback, supporting the usefulness of spatial selectivity.
  2. Competitive performance across multiple tasks. The 2B model reports strong results on several grounding and spatial-understanding benchmarks, particularly EmbSpatial, SAT, and BLINK Spatial. The mixed-supervision formulation is practically useful, and the robotic experiments demonstrate that the learned representations can be integrated into a manipulation pipeline.
Critical weaknesses
  1. The methodological contribution relative to closely related work is insufficiently established. Bidirectional refinement between language semantics and visual decoding states already appears in FlowSeg. DeRIS also uses iterative perception–cognition interaction and mask-guided visual aggregation. The paper acknowledges these precedents, but does not sufficiently explain or experimentally isolate the additional benefit of RACE’s particular update mechanism. Removing feedback demonstrates that the proposed connection helps, but does not establish its advantage over existing feedback designs. Applying these ideas to embodied grounding may be valuable, but the specific methodological advance needs stronger justification.
  2. The empirical evidence does not yet establish a sufficiently clear advance. Some grounding gains are modest: 1.5 percentage points over RoboRefer-RFT on RefSpatial-Bench, 1.0 over RoboMAP on Where2Place, and 0.12 over Si0 on RefCOCO. Other results are lower than some listed baselines. Moreover, Appendix A notes that source-reported evaluation settings are not necessarily directly comparable, which limits conclusions drawn from these margins. Section 4.4 demonstrates feasibility through 36 successful tasks out of 50 trials, but provides no controlled baseline comparison to establish the contribution of RACE to robotic task success. Reporting inconsistencies further complicate interpretation of the evidence.
Questions for authors
  1. What are the essential methodological differences between RACE and the bidirectional refinement mechanisms in FlowSeg and DeRIS? Could a comparison under a common backbone, training data, and evaluation protocol establish the benefit of RACE’s specific design? I recognize that their original segmentation scores cannot be directly compared with point-grounding accuracy.
  2. Could the robotic evaluation include a controlled baseline using the same action-head training and deployment conditions, such as a configuration without spatial feedback? Reporting target-localization accuracy separately from execution success would also help identify the source of any benefit.
5 comments
AuthorsOpenReview ↗

Response to Reviewer sDCK

sDCK-Q1: What are the essential methodological differences between RACE and the bidirectional refinement mechanisms in FlowSeg and DeRIS? Could a comparison under a common backbone, training data, and evaluation protocol establish the benefit of RACE’s specific design? I recognize that their original segmentation scores cannot be directly compared with point-grounding accuracy.

A1: Thank you for highlighting the need for a clearer methodological comparison and an evaluation under common settings.

1. Motivation. RACE: Spatially constrained grounding from general semantics vs. FlowSeg/DeRIS: Instance segmentation and semantic disambiguation. Beyond differences in tasks and application scenarios, RACE aims to preserve the backbone’s general semantic and QA capabilities while using a dedicated spatial decoder to construct and refine explicit spatial hypotheses for task-relevant, fine-grained grounding.

2. Mechanism. RACE: An evolving spatial distribution unifying supervision, feedback, and prediction vs. FlowSeg/DeRIS: Query-based bidirectional refinement. RACE unifies heterogeneous annotations into distributional targets and explicitly maintains a grounding distribution across stages. The spatial distribution serves both as the basis for selecting visual evidence for semantic feedback and as the grounding prediction, enabling co-evolution between referent semantics and explicit spatial hypotheses.

3. Experimental Comparisons. We constructed a FlowSeg-style latent co-evolution decoder with 16 visual queries propagated across stages, interacting bidirectionally with the semantic information to iteratively refine grounding predictions. This variant uses the same backbone initialization, training data, and evaluation protocol as RACE.

MethodLocationPlacementUnseenWhere2PlaceRoboRefIt
Latent co-evolution decoder51.0045.0036.3667.0078.29
RACE58.0051.0044.1674.0081.33

These results support the benefit of RACE’s explicit distribution-based co-evolution design over the latent-query alternative in our grounding setting. The latent queries may mix regional information and disrupt referent refinement through irrelevant feedback. The gap may also reflect the latent design’s suitability for the grounding paradigm and task settings.

sDCK-Q2: Could the robotic evaluation include a controlled baseline using the same action-head training and deployment conditions, such as a configuration without spatial feedback? Reporting target-localization accuracy separately from execution success would also help identify the source of any benefit.

A2: We added a no-feedback baseline to the real-robot evaluation section, removing only spatial feedback while keeping the action-head training and deployment conditions unchanged. Each task category was tested over 10 trials. The numbers of successful trials are reported below:

ConfigurationSimilar objectsUnseen instancesAbsolute positionRelative positionPartial occlusion
No feedback86864
RACE87876

Within the limited set of scenes covered by this evaluation, both configurations achieved 100% object-level grounding accuracy. The execution success rate conditioned on correct target selection therefore equals the overall success rate. However, selecting the correct target does not guarantee precise localization. Under partial occlusion, for example, removing spatial feedback reduced grounding precision, leading to gripper misalignment and task failures.

sDCK—Weaknesses section: Competitiveness of our results and evaluation methodology

Response: RACE combines grounding and spatial QA across general and embodied settings in a single 2B model. It achieves the best reported results among the listed 2B models on RefSpatial-Bench, Where2Place, and the RefCOCO family, and outperforms larger models on multiple benchmarks, demonstrating broad cross-task competitiveness at a small model size. For example, it achieves 81.33% on SAT vs. 78.67% for MiMo-Embodied (7B).

We selected or re-evaluated baselines using consistent criteria wherever possible. Evaluation protocols are broadly aligned; the differences noted in Appendix A concern only a few models and do not affect the overall conclusions.

Thank you for the detailed response and the additional experiments. I appreciate the comparisons with the simpler decoder and the FlowSeg-style variant, as well as the new real-robot baseline. These results provide useful evidence for the proposed design and address part of my original concerns.

My main remaining concern is the contribution of RACE relative to existing bidirectional semantic–visual refinement methods. The FlowSeg-style variant is a useful comparison, but I still find it difficult to assess how much of the reported improvement comes from RACE’s explicitly supervised spatial distribution and its use for feedback. Could the authors clarify:

  • Why were 16 visual queries chosen for the FlowSeg-style variant, especially given that the official FlowSeg implementation uses 200 queries? Were other query counts considered?
  • Does this variant use the same number of refinement stages as RACE? How do the two configurations compare in parameter count, supervision, and training budget?

I also appreciate the clarification of the RefSpatial-Bench evaluation scope and the corrections in the revised manuscript. However, I am still uncertain about the stage-wise results. The revised Figure 6 reports that accuracy increases from 52.00% to 54.50% between stages 3 and 4, which corresponds to a net gain of 5 correct predictions over 200 examples. Figure 4(b), however, shows 15 corrections and 5 regressions, corresponding to a net gain of 10. Could the authors clarify whether these figures use the same examples and evaluation settings?

Overall, the response has strengthened my view of the paper’s empirical results. However, I remain uncertain about the methodological contribution relative to FlowSeg and how to interpret the stage-wise analysis.

AuthorsOpenReview ↗

Response to Reviewer sDCK’s Follow-up Comments

We really appreciate your follow-up comments and your recognition of the value of our additional experiments. We address the remaining questions below.

Why were 16 visual queries chosen for the FlowSeg-style variant, especially given that the official FlowSeg implementation uses 200 queries? Were other query counts considered?

1. Query-count exploration. We also explored configurations with 32 and 64 visual queries. At matched training steps during the early phase, both achieved lower grounding accuracy than the 16-query configuration (at step 1,000 on the Location subset of RefSpatial-Bench: 16 queries, 48%; 32 queries, 40%; 64 queries, 33%). We therefore retained 16 queries for the reported comparison. A larger query set introduces more intermediate latent states that must be coordinated with referent semantics and mapped to the grounding output, creating additional learning and alignment demands and potentially introducing alignment errors into semantic feedback.

2. Role of visual queries. Our latent-query variant produces an instruction-conditioned spatial distribution for final point prediction. Its visual queries aggregate visual evidence and interact bidirectionally with referent semantics to refine the grounding distribution instead of maintaining a large set of independent mask candidates. The compact 16-query variant provides an alternative way to implement co-evolution within our grounding framework.

Does this variant use the same number of refinement stages as RACE? How do the two configurations compare in parameter count, supervision, and training budget?

Both configurations use four refinement stages and share the same backbone initialization, supervision, and training budget. The RACE decoder contains approximately 2.5M parameters, compared with 3.5M for the latent co-evolution decoder.

Clarification of the stage-wise results.

Thank you for pointing out this inconsistency. Figure 6 had been updated to the 200 Location and Placement expressions, while Figure 4(b) had not been updated accordingly and still reported statistics over 277 entries. We have now recomputed the transition statistics on the same 200-expression set, using the unchanged checkpoint and evaluation settings. The updated counts are consistent with the stage-wise accuracies, as summarized below:

Stage transitionCorrectedRegressedNet gainAccuracy (%)
S1 → S23333026.50 → 41.50
S2 → S32322141.50 → 52.00
S3 → S494552.00 → 54.50
We have updated Figure 4(b) and its caption accordingly.

Further clarification of RACE’s methodological contribution

1. A unified spatial distribution for supervision, prediction, and feedback. RACE explicitly maintains a grounding distribution across refinement stages. This distribution directly receives grounding supervision, represents the current spatial hypothesis, and weights visual evidence for semantic refinement. This unified design requires no additional set of visual latent queries. The latent-decoder comparison and the spatial-feedback ablations provide complementary evidence for the benefit of this design.
2. Fine-grained grounding and multi-task capability. RACE builds on the backbone’s general semantic understanding and uses a dedicated spatial decoder to perform fine-grained grounding through spatial reasoning over the coarse visual-token grid. This supports instruction-conditioned object localization and placement. The same model also supports general and embodied spatial QA tasks. Its results across general and embodied grounding and spatial QA benchmarks demonstrate the effectiveness of this design across multiple tasks.

Reviewer_BVghOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RACE, a framework for embodied spatial grounding that jointly refines a referent representation (“what to attend to”) and a spatial affordance distribution (“where to act”). A four-stage grounding head uses the current spatial distribution to pool visual evidence and update the referent query, which then guides the next spatial prediction. The method also converts point, box, and mask annotations into unified distributional targets, combining stage-wise distribution matching with a geometry-aware false-peak penalty. The framework is evaluated using a 2B vision-language backbone on embodied grounding and general spatial benchmarks, reporting 54.50% on RefSpatial-Bench and 74.00% on Where2Place. Component ablations examine query initialization, persistence, and spatial feedback. An additional action head demonstrates downstream applicability through real-world pick-and-place experiments, completing 36 of 50 trials.

Critical strengths

The strongest aspect is the explicit use of spatial predictions to select visual evidence for subsequent referent refinement. This gives the intermediate grounding distribution a functional role beyond producing the final location. The ablations provide targeted support: on Where2Place, the full model achieves 74% accuracy, compared with 70% without visual feedback and 68% with uniform visual pooling. These comparisons suggest that spatially selective feedback contributes beyond simply propagating the query across stages. This is consequential because it provides empirical support for the paper’s central mechanism, rather than relying solely on aggregate benchmark improvements.

Critical weaknesses
  1. The incremental contribution and necessity of the proposed architecture are insufficiently established. The paper acknowledges prior work on perception–cognition interaction and bidirectional semantic–visual refinement, but does not clearly establish which distinguishing mechanism accounts for its advantages. Table 3 provides useful component ablations, yet a comparison against a simpler spatial decoder with the same backbone, training data, supervision, and comparable optimization budget is still needed. This matters because the current results support the usefulness of particular components without sufficiently establishing that the proposed four-stage co-evolution architecture offers a compelling advance over simpler alternatives.
  2. Inconsistencies in reporting weaken confidence in the central experimental conclusions. The main text reports a final RefSpatial-Bench accuracy of 54.50%, whereas Appendix C reports a final-stage SoftArgmax hit rate of 50.18% over all 277 expressions. Given the stated point-based evaluation protocol, the relationship between these results requires clarification. Additionally, the query-persistence ablation is described as resetting the query at each stage in the discussion, but as keeping it fixed at initialization in the table caption; these are different interventions. These ambiguities affect both the headline performance and the interpretation of a central ablation, making it difficult to assess the evidence reliably.
Questions for authors
  1. Clarification of evaluation results: The main text reports 54.50% on RefSpatial-Bench, while Appendix C reports a final-stage SoftArgmax hit rate of 50.18% over all 277 expressions. Do these results use different checkpoints, model variants, or evaluation settings? Please reconcile the figures and identify which configuration underlies the main comparisons and stage-wise analyses.
  2. Clarification of the query-persistence ablation: Is the query reset to its initial value before each stage and subsequently refined within that stage, or held fixed at its initial value throughout decoding? The discussion and Table 3 caption appear to describe different interventions. Which was implemented?
  3. Evidence supporting architectural necessity: Beyond the component ablations, do you have a matched comparison against a simpler spatial decoder using the same backbone, training data, distributional supervision, and backbone fine-tuning? Such evidence would materially affect my assessment of whether the gains justify the proposed co-evolution architecture, rather than primarily reflecting the training recipe. This question concerns support for the current architectural claim, rather than a suggestion for extending the work.
2 comments
AuthorsOpenReview ↗

Response to Reviewer BVgh

BVgh-Q1: Clarification of evaluation results: The main text reports 54.50% on RefSpatial-Bench, while Appendix C reports a final-stage SoftArgmax hit rate of 50.18% over all 277 expressions. Do these results use different checkpoints, model variants, or evaluation settings? Please reconcile the figures and identify which configuration underlies the main comparisons and stage-wise analyses.

A1: Thank you for pointing out this inconsistency. Both results use the same full RACE checkpoint on RefSpatial-Bench; the discrepancy arises from the evaluation scope. The main-text result of 54.50% was computed over the 200 Location and Placement expressions, whereas the 50.18% in Appendix C was computed over 277 expressions, including 77 Unseen expressions. Since these Unseen expressions are already included in Location and Placement, pooling all three subsets gives them extra weight. We have revised the appendix to report all diagnostics on the same 200-expression set used for the main comparisons. We apologize for the unclear reporting of the evaluation protocol.

BVgh-Q2: Clarification of the query-persistence ablation: Is the query reset to its initial value before each stage and subsequently refined within that stage, or held fixed at its initial value throughout decoding? The discussion and Table 3 caption appear to describe different interventions. Which was implemented?

A2: We implemented the first setting: the input query is reset to q0q_0 at the beginning of each stage and subsequently refined within that stage. Thus, this ablation isolates cross-stage query persistence and supports the importance of carrying refined semantics across stages. We have corrected the caption and clarify the reset operation in the main text.

BVgh-Q3: Evidence supporting architectural necessity: Beyond the component ablations, do you have a matched comparison against a simpler spatial decoder using the same backbone, training data, distributional supervision, and backbone fine-tuning? Such evidence would materially affect my assessment of whether the gains justify the proposed co-evolution architecture, rather than primarily reflecting the training recipe. This question concerns support for the current architectural claim, rather than a suggestion for extending the work.

A3: We added a matched comparison with a four-layer cross-attention spatial decoder (Simple decoder), using the same backbone initialization, training data, distributional supervision, backbone fine-tuning settings, and training/evaluation protocols. Accuracy (%) is reported below. Location, Placement, and Unseen are subsets of RefSpatial-Bench.

MethodLocationPlacementUnseenWhere2PlaceRoboRefIt
Simple decoder51.0041.0033.7771.0072.43
RACE58.0051.0044.1674.0081.33

The results support the benefit of the co-evolution architecture beyond the shared training recipe. The existing query-persistence and spatial-feedback ablations further support the contributions of its key mechanisms.

AuthorsOpenReview ↗

Follow-up on Our Response to Reviewer BVgh

Thank you again for your thoughtful comments. As the discussion period is nearing its end, we would be grateful for any further feedback you may have on our response. In our previous response, we reconciled the evaluation results, clarified the query-persistence ablation, and added a matched comparison with a simpler spatial decoder. We hope these clarifications and additional results help address your concerns, and would be happy to discuss any remaining questions. Thank you for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Additional Evidence and Revisions

Dear Chairs and Reviewers,

We sincerely appreciate the reviewers’ constructive feedback. We have supplemented the manuscript with controlled experiments and clarified the methodological design and evaluation reporting. We summarize the main additions below.

  • Controlled decoder comparisons. We added a four-layer cross-attention decoder and a FlowSeg-style latent co-evolution decoder under common experimental settings. RACE outperformed both alternatives, and we clarified their configurations, supervision, and training budgets.
  • Controlled real-robot evaluation. We added a no-feedback baseline with the same action-head training and deployment conditions. RACE completed 36/50 tasks (72%), compared with 32/50 (64%) without spatial feedback, providing additional evidence for its contribution to downstream manipulation.
  • Clarification of the methodological contribution. RACE directly uses the same explicitly supervised spatial distribution for grounding prediction, cross-stage refinement, and visual-evidence feedback, without introducing a separate latent visual representation for co-evolution. It achieves fine-grained object localization and placement through spatial reasoning over the backbone’s coarse visual-token grid, while supporting general and embodied spatial QA.
  • Consistent evaluation reporting. We clarified the query-reset ablation and standardized the RefSpatial-Bench diagnostic and transition analyses to the same 200-expression set to ensure a consistent basis for quantitative comparisons.

We hope these additions and clarifications help the reviewers assess the contribution and empirical evidence more precisely. Thank you for your time and consideration.

Sincerely,
Authors of Submission #3732

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.