acceptodds
Under review as a conference paper at ICLR 2027

Learning Multi-Step 3D Reasoning for Embodied Question Answering

Abstract

Embodied Question Answering (EQA) requires embodied agents to explore a 3D environment to find evidence and answer questions related to the environment. Some methods employ off-the-shelf vision-language models (VLMs) to solve EQA tasks, which exhibit limited 3D reasoning capabilities, making it difficult to accurately retrieve evidence and resulting in inefficient exploration and incorrect responses. In this paper, we propose a chain-of-exploration EQA agent that unifies exploration, spatial evidence acquisition, and question answering through iterative reasoning with tool use. Specifically, our agent reasons over both current observations and historical information, determines what spatial information is required for solving the task, selects appropriate tools to acquire geometric evidence, and decides where to explore next. To improve the reasoning and exploration behaviors of the agent, we further design an EQA data generation pipeline that automatically constructs EQA tasks with 3D reasoning trajectories. Based on the pipeline, we collect the EQA-RT dataset, which contains 18K tasks and is divided into a training set, EQA-RT-Train, and two test sets, EQA-RT-Seen (scenes overlapping with the training set) and EQA-RT-Unseen (unseen scenes). We perform supervised fine-tuning on EQA-RT-Train to initialize the model's tool-use capability, followed by evidence-grounded reinforcement learning to enhance task-relevant spatial evidence acquisition. Across public benchmarks, our method improves LLM-Match by 0.69 points and grounding-aware success by 2.93 points, while reducing exploration distance by up to 61.4%. On our EQA-RT, it improves target-object recall by 0.06–0.12 and LLM-Match by 4.19–5.63 points.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Question Answering, Embodied Agent, 3D Reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GZktOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a chain-of-exploration agent for embodied question answering that integrates planning, spatial memory, and iterative tool use to acquire evidence during exploration. It introduces EQA-RT, a dataset of approximately 18K question–answer pairs with reasoning and tool-use trajectories, constructed through an automated generation and verification pipeline. The controller is trained through supervised fine-tuning followed by evidence-grounded reinforcement learning, which jointly considers evidence coverage, answer quality, and interaction cost. Experiments demonstrate improvements in answer quality and exploration efficiency on EQA-RT, together with competitive results across three public benchmarks.

Critical strengths
  1. EQA-RT provides a useful resource for studying embodied evidence acquisition through question–answer pairs paired with intermediate reasoning and tool-use trajectories, including evaluation in both seen and unseen scenes.
  2. Controlled training and reward ablations show that evidence-grounded reinforcement learning improves answer quality, target retrieval, and exploration efficiency over supervised fine-tuning alone within the proposed framework.
Critical weaknesses
  1. The experiments report overall answer quality, target recall, and exploration distance, but do not show where the proposed multi-step 3D reasoning provides a clear advantage. Results are not broken down by question type or reasoning complexity, and the ablations examine training stages and reward terms without isolating explicit reasoning, spatial memory, or geometric tool use. Consequently, it remains unclear whether the improvements demonstrate stronger spatial reasoning or primarily reflect better tool selection and task-specific training, which weakens the experimental support for the paper’s central claim.
  2. The evidence reward depends on coverage of predefined facts, but the main text does not sufficiently explain how the correctness and sufficiency of the acquired evidence are established. This matters because collecting the expected information does not necessarily mean that it is accurate or correctly used to derive an answer. Table 4 further shows that removing evidence coverage from the answer reward improves LLM-Match on both splits, despite reducing coverage. Although the paper acknowledges this trade-off, independent evaluation of evidence correctness and its connection to the final answer is needed to establish that the proposed reward improves grounding reliability.
Questions for authors
  1. Were all synthetic trajectory observations obtained or replaced through actual tool execution, and what checks ensure that intermediate reasoning does not depend on oracle information unavailable at inference?
  2. How well does the evidence-coverage metric agree with independently verified target identity, geometric accuracy, and visual-answer correctness?
4 comments

We sincerely thank the reviewer for the insightful and constructive feedback, which has been highly valuable in helping us strengthen and improve our manuscript.

Weakness 1 (Attribution of Multi-Step 3D Reasoning Improvements)

The experiments report overall answer quality, target recall, and exploration distance, but do not show where the proposed multi-step 3D reasoning provides a clear advantage. Results are not broken down by question type or reasoning complexity, and the ablations examine training stages and reward terms without isolating explicit reasoning, spatial memory, or geometric tool use. Consequently, it remains unclear whether the improvements demonstrate stronger spatial reasoning or primarily reflect better tool selection and task-specific training, which weakens the experimental support for the paper’s central claim.

Response:

Our improvements arise from both tool selection and spatial reasoning, with a larger reduction in spatial reasoning errors. To distinguish their contributions, we conduct a Bad Case Analysis. We classify failures as Tool Selection Errors when necessary tools are not invoked despite sufficient observations, and Spatial Reasoning Errors when the agent fails to identify the prerequisites for acquiring spatial evidence or incorrectly integrates acquired geometric information. As shown in the following table, SFT+RL reduces Tool Selection Errors by 7.3 percentage points and Spatial Reasoning Errors by 11.4 points compared with None, indicating that the improvements extend beyond tool selection.

Failure TypeNonSFT-only (%)SFT+RL (%)
Tool Selection Error22.619.815.3
Spatial Reasoning Error33.229.521.8

Both Spatial Memory and Planner contribute to multi-step spatial reasoning. As shown in the following table, removing either component reduces performance on EQA-RT-Unseen, while removing both causes a larger decline. The full model achieves 48.16% LLM-Match with a trajectory length of 23.19 m, compared with 40.37% and 49.99 m when both modules are removed.

SettingLLM-Match (%) ↑Epath@10 ↑Recall@10 ↑Traj. Length (m) ↓
w/o Spatial Memory42.910.150.2845.11
w/o Planner45.690.170.3238.68
w/o Both40.370.130.2249.99
Full (Ours)48.160.210.3723.19

Our method also achieves substantial gains on tasks requiring geometric measurements and multi-object evidence integration. As shown in the following table, SFT+RL consistently outperforms SFT-only across question types, particularly on Attribute-Size and Distance, with gains of 13.90/14.30 points on Seen and 13.69/14.28 points on Unseen, respectively.

Question TypeSFT-only SeenSFT+RL SeenSFT-only UnseenSFT+RL Unseen
Attribute-Size39.1853.0837.4051.09
Attribute-Color46.6850.9844.5148.79
Attribute-Special44.0848.6841.8146.50
Distance35.9850.2834.1148.39
Counting37.1950.0834.8147.09
Location-Location41.1848.1839.0245.83
Location-Special42.4848.3840.9146.69
Relationship41.9349.3639.5346.96
Status46.2050.4844.5849.06
Overall41.2550.3239.8348.16

Our method also improves performance on tasks requiring multi-object evidence integration. As shown in the following table, SFT+RL consistently outperforms SFT-only across different numbers of target objects, supporting its effectiveness in tasks requiring evidence acquisition and integration across multiple objects.

Target ObjectsSFT-only SeenSFT+RL SeenSFT-only UnseenSFT+RL Unseen
144.5151.2142.5249.21
237.0149.2836.2747.04
≥336.5848.0936.1344.46
Overall41.2550.3239.8348.16

We have added a Spatial Reasoning Analysis to Section 4.4, with the complete component and error analyses and per-category performance breakdowns in Appendix D.

Weakness 2 (Evidence Correctness, Sufficiency, and Reward Design)

The evidence reward depends on coverage of predefined facts, but the main text does not sufficiently explain how the correctness and sufficiency of the acquired evidence are established. This matters because collecting the expected information does not necessarily mean that it is accurate or correctly used to derive an answer. Table 4 further shows that removing evidence coverage from the answer reward improves LLM-Match on both splits, despite reducing coverage. Although the paper acknowledges this trade-off, independent evaluation of evidence correctness and its connection to the final answer is needed to establish that the proposed reward improves grounding reliability.

Response:

Acquired evidence is validated through type-specific checks rather than being credited unconditionally, and its sufficiency is measured against the required GT Evidence Facts. Specifically, evidence validation includes target association, tool-output validity, and spatial constraints. For instance, 3D Position requires agreement with the GT position within 2 m, while Object Size and VisualQA are checked for output validity. Evidence Coverage then measures the proportion of required GT Evidence Facts acquired by the agent. Thus, it primarily evaluates task-relevant evidence acquisition rather than the complete accuracy of individual tool predictions.

Removing Coverage Scaling slightly improves LLM-Match but reduces evidence acquisition. Without Coverage Scaling, the Answer Reward no longer depends on evidence sufficiency, allowing optimization to focus more directly on final-answer scores. Our design instead encourages the agent to acquire sufficient evidence and penalizes incorrect answers more strongly when sufficient evidence has been collected, promoting evidence utilization. Meanwhile, the separate Evidence Reward encourages evidence acquisition. As shown in Table 4 of the main paper, removing Coverage Scaling improves LLM-Match by 0.76 and 0.46 percentage points on Seen and Unseen, respectively, but reduces Evidence Coverage by 4.82 and 2.68 points, with Recall@10 also declining.

Evidence Coverage and final-answer quality improve together under SFT+RL, although they measure different aspects of agent performance. Evidence Coverage evaluates the acquisition of required evidence, whereas LLM-Match assesses final-answer quality. As shown in Table 3 of the main paper, SFT+RL improves Evidence Coverage from 17.72% to 41.09% on Seen and from 15.50% to 38.50% on Unseen, while LLM-Match increases from 41.25% to 50.32% and from 39.83% to 48.16%, respectively. These results demonstrate simultaneous improvements in evidence acquisition and answer quality, supporting the objective of jointly optimizing both.

Section 3.3 now clarifies the Evidence Coverage validation rules and reward design, while Appendix D includes the independent evidence-correctness evaluation and reward-scaling comparisons.

Question 1 (Synthetic Observations and Oracle Information)

Were all synthetic trajectory observations obtained or replaced through actual tool execution, and what checks ensure that intermediate reasoning does not depend on oracle information unavailable at inference?

Response:

Our synthetic trajectory observations are constructed from pre-extracted scene evidence and mapped to the outputs of the corresponding tools, rather than obtained through actual tool execution. Compared with actual tool outputs, ours observations retain similar information while avoiding potential execution errors and instability, thereby providing more accurate and reliable supervision. We first use the simulator to generate reference exploration trajectories and their corresponding visual Observations. For geometric perception tools, Observations are constructed from scene GT or automatically generated object annotations and formatted to match the corresponding tool outputs. Navigation Observations are generated from the observations and actions recorded in the reference trajectories.

Oracle information is used for offline trajectory generation, while intermediate reasoning is constrained to previously available Observations. Navigation actions are predetermined by the reference trajectories; found and all_found identify target-discovery states and key tool calls, while expected_answer constrains the final answer. During generation, the Prompt explicitly prohibits intermediate Thoughts from using geometric information that has not yet been acquired, and geometric computations in the final Thought strictly use positions and dimensions already contained in historical Observations. We additionally apply rule-based checks and VLM consistency verification to assess tool-call validity and reasoning–observation consistency. At inference time, the agent receives no oracle information, reference trajectories, or GT object information. Instead, it follows the Planner-generated Plan, obtains Observations through actual tool execution, and autonomously explores and answers using interaction history and Spatial Memory.

Section 3.2 now specifies how synthetic Observations are constructed and how intermediate reasoning is constrained, and Appendix G documents the offline oracle fields and their absence at inference.

Question 2 (Independent Validation of Evidence Coverage)

How well does the evidence-coverage metric agree with independently verified target identity, geometric accuracy, and visual-answer correctness?

Response:

Evidence Coverage is positively associated with independently verified target identity, geometric accuracy, and VisualQA correctness, supporting the validity of our metric. To verify this relationship, we evaluate 300 EQA-RT-Unseen samples, divided into three Evidence Coverage ranges with 100 samples each. We independently assess Target Identity Accuracy, Geometric Accuracy, and VisualQA Accuracy. As shown in the following table, all three metrics consistently improve with increasing Evidence Coverage.

Evidence Coverage# SamplesTarget Identity Acc. (%) ↑Geometric Accuracy (%) ↑VisualQA Accuracy (%) ↑
Low (0-15%)100635269
Medium (15-40%)100706675
High (40-100%)100797987

We have added the full 300-sample evaluation to Appendix D (Independent Evidence Correctness Evaluation) and summarized its findings in Section 4.4.

Thank you to the authors for their detailed response. My concerns have been largely addressed, and I have therefore raised my score.

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

The paper targets embodied question answering (EQA) for questions that require geometric evidence, such as sizes, distances, and counts. It proposes a tool-using agent in which a Qwen3-VL-8B controller emits a thought and code at each step to navigate, detect, localize in 3D and query a VQA model. The authors build EQA-RT, about 18K tasks synthesized on HM3D with GPT-4o-generated trajectories along A* paths. The controller is trained with SFT followed by GDPO-based RL, using an "evidence-grounded" reward that scales the answer reward by evidence coverage and penalizes interaction cost. On EQA-RT the method improves LLM-Match over Pred-EQA by 4–6 points and shortens paths by about 30%; gains on HM-EQA, OpenEQA and EXPRESS-Bench are small or mixed.

Critical strengths

S1. The evidence-grounded RL addresses a real failure mode: answering geometric questions without measuring. RL accounts for most of the gains (Table 3: 41.25→50.32 on seen and 39.83→48.16 on unseen scenes). The paired tool-use analysis (Table 8) and the candid reward ablations (Table 4) show concretely what behavior changes.

S2. EQA-RT, together with the full prompts, the evidence-validation rules, and the promised release, is a potentially useful resource for training tool-using embodied agents.

Critical weaknesses

W1. The gains are largely in-distribution. The test sets come from the same pipeline as the training data. The paper does not report training any baseline on EQA-RT, and does not specify the backbones used for the baselines on EQA-RT. The untrained agent scores below Pred-EQA (38.32 / 35.17 vs. 44.69 / 43.97). On public benchmarks, results are mixed: HM-EQA accuracy is 65.77 vs. 69.2 for FAST-EQA, and OpenEQA LLM-SPL and EXPRESS dTd_T are both slightly worse than Pred-EQA. The RL answer reward appears to use the same Qwen3-VL-8B scorer as the reported LLM-Match (App. C), and no seed variance or confidence intervals are reported.

W2. The labels and trajectories do not reliably encode what the paper claims. Geometric answers are produced by a 3D detection model applied to the whole scene (l.203–210), which is not named; Table 11 specifies DetAny3D only for the inference-time ObjectLocation3D tool. The paper does not report checking these answers against ground-truth meshes or semantic annotations. The trajectory prompt (App. G) further gives GPT-4o oracle found flags and forces it to output the expected answer, which is at odds with the paper's own requirement (l.085–087) that each trajectory should only use information that the agent could actually observe. Consistent with this, SFT adds only 2.9–4.7 LLM-Match points over the untrained controller (Table 3).

W3. The reward design is flawed, and the "3D reasoning" claim is not isolated. For submitted answers, Rans(τ)=C(τ)⋅s(τ)R_{\text{ans}}(\tau) = C(\tau) \cdot s(\tau) (Eq. 4a), so when s(τ)<0s(\tau) < 0, the penalty grows as more evidence is gathered (e.g., a wrong closed-vocabulary answer with s(τ)=−1s(\tau) = -1 receives Rans=0R_{\text{ans}} = 0 at C(τ)=0C(\tau) = 0, but −1-1 at C(τ)=1C(\tau) = 1). This may explain why removing C(τ)C(\tau) from RansR_{\text{ans}} raises LLM-Match (Table 4). In the coverage computation, grounding requires only a name-matched detection without agreement to the annotated position; size and VQA facts are credited without accuracy checks; positions use a 2 m tolerance; and the same coverage quantity is also reported as an evaluation metric. After RL, 3D calls that add position evidence barely change (Table 8), and no per-question-type results are reported.

Questions for authors

Q1. Could you re-run the strongest baselines with Qwen3-VL-8B under the same tools and budget, and report None / SFT-only / SFT+RL on the public benchmarks? Do the EQA-RT training scenes overlap with those benchmarks' test scenes? If the gains hold, I would raise my score.

Q2. Which 3D detector produced the geometric labels (l.203), and is it the DetAny3D used by ObjectLocation3D (Table 11)? Please re-score size, distance and counting questions against HM3D meshes or semantic annotations on a subset. Are the SFT observations executed tool outputs or GPT-4o's "expected" results?

Q3. Why are negative answer scores scaled by C(τ)C(\tau)? Please report a variant that scales only s(τ)>0s(\tau) > 0, and re-evaluate with a judge independent of the RL reward or on a human-rated subset.

Q4. Please report per-question-type results and mean ± std over at least three seeds. How is the "Evidence checklist" in Spatial Memory (l.1555) generated at inference, and does it use question-type labels or target annotations?

9 comments

We sincerely thank the reviewer for the insightful and constructive feedback, which has been highly valuable in helping us strengthen and improve our manuscript.

Weakness 1 & Question 4 (Generalization, Baseline Fairness, and Evaluation Reliability)

The gains are largely in-distribution. The test sets come from the same pipeline as the training data. The paper does not report training any baseline on EQA-RT, and does not specify the backbones used for the baselines on EQA-RT. The untrained agent scores below Pred-EQA (38.32 / 35.17 vs. 44.69 / 43.97). On public benchmarks, results are mixed: HM-EQA accuracy is 65.77 vs. 69.2 for FAST-EQA, and OpenEQA LLM-SPL and EXPRESS dTd_T are both slightly worse than Pred-EQA. The RL answer reward appears to use the same Qwen3-VL-8B scorer as the reported LLM-Match (App. C), and no seed variance or confidence intervals are reported.

Please report per-question-type results and mean ± std over at least three seeds. How is the "Evidence checklist" in Spatial Memory (l.1555) generated at inference, and does it use question-type labels or target annotations?

Response:

The training gains are not confined to the EQA-RT generation pipeline and can transfer to different task distributions and unseen scenes. As shown in the table below, we additionally evaluated Zero-shot, SFT-only, and SFT+RL on HM-EQA, OpenEQA, and EXPRESS-Bench, where SFT+RL outperforms the other two settings on all reported metrics.

BenchmarkMetric0-shotSFT-onlySFT+RL
HM-EQAAccuracy ↑63.8164.5365.77
HM-EQAPath Length (m) ↓27.6424.8320.93
OpenEQA (A-EQA)LLM-Match ↑50.8752.1653.99
OpenEQA (A-EQA)LLM-SPL ↑43.7645.3447.21
EXPRESS-BenchSucc. ↑47.8350.2655.51
EXPRESS-BenchSucc.* ↑65.7468.2372.61
EXPRESS-BenchEpath ↑38.9242.6148.37
EXPRESS-BenchdT ↓7.186.475.89

The existing baselines primarily employ training-free exploration and reasoning frameworks, so we evaluate them using their original configurations. The backbone configurations are listed in the table below. Because these methods have architectures different from ours, applying Thought–Code–Observation supervision directly would require additional training interfaces; using Question–Answer supervision alone would not directly optimize active exploration and evidence acquisition and could introduce dataset shortcuts. To isolate the contribution of training, Table 3 of the main paper presents controlled training ablations with the backbone, agent architecture, and tools held fixed.

MethodBackbone
Explore-EQAPrism-DINOSigLIP 7B
FineEQAPrism-DINOSigLIP 7B
3D-MemGPT-4o
GraphEQAGPT-4o
FAST-EQAGPT-4o
Pred-EQAQwen3-VL-8B-Instruct
OursQwen3-VL-8B-Instruct

Our results demonstrate the effectiveness of task-specific training. Controlled ablations isolate the training contribution (main-paper Table 3). Although Zero-shot underperforms Pred-EQA, SFT+RL improves LLM-Match by 12.00/12.99 points on Seen/Unseen, surpassing Pred-EQA by 5.63/4.19 points with the same backbone. On public benchmarks, our method outperforms Pred-EQA on several metrics, while SFT+RL consistently improves upon Zero-shot and SFT-only across all reported metrics (see the above Table).

The Answer Reward during RL training uses Qwen3-VL-8B-Instruct, whereas the final LLM-Match evaluation uses Qwen3-VL-8B-Thinking. We will clarify this setup in the revised manuscript.

As shown in the table below, we repeated inference with three different random seeds and report the mean and standard deviation. The results show relatively stable answering and exploration performance on both Seen and Unseen.

DatasetModelRecall@10 ↑Epath@10 ↑Traj. Length (m) ↓LLM-Match (%) ↑
SeenOurs (SFT+RL)0.40 ± 0.020.21 ± 0.0121.98 ± 4.8650.13 ± 0.48
UnseenOurs (SFT+RL)0.37 ± 0.010.21 ± 0.0123.81 ± 2.9348.66 ± 0.86

The Evidence Checklist is generated from the Planner's predictions without using GT annotations at inference time. Specifically, the Planner determines the question type, target objects, and required evidence types from the input question and passes them to Spatial Memory through the Plan. Spatial Memory then initializes the Checklist and dynamically updates evidence-acquisition status based on actual tool outputs. Neither GT question-type labels nor target or evidence annotations are used.

The revised manuscript clarifies Evidence Checklist construction in Section 3.1 and Appendix A and distinguishes the RL reward scorer from the evaluation judge in Appendix C. The public-benchmark ablations, baseline backbones, per-question-type results, and three-seed inference statistics are now reported in Appendix D.

Weakness 2 & Question 2 (Geometric Annotation and Trajectory Reliability)

The labels and trajectories do not reliably encode what the paper claims. Geometric answers are produced by a 3D detection model applied to the whole scene (l.203–210), which is not named; Table 11 specifies DetAny3D only for the inference-time ObjectLocation3D tool. The paper does not report checking these answers against ground-truth meshes or semantic annotations. The trajectory prompt (App. G) further gives GPT-4o oracle found flags and forces it to output the expected answer, which is at odds with the paper's own requirement (l.085–087) that each trajectory should only use information that the agent could actually observe. Consistent with this, SFT adds only 2.9–4.7 LLM-Match points over the untrained controller (Table 3).

Which 3D detector produced the geometric labels (l.203), and is it the DetAny3D used by ObjectLocation3D (Table 11)? Please re-score size, distance and counting questions against HM3D meshes or semantic annotations on a subset. Are the SFT observations executed tool outputs or GPT-4o's "expected" results? Response:

Response:

**Our geometric annotations are derived from official HM3D annotations or a validated automatic annotation pipeline, and reference answers are computed using deterministic rules. **For scenes with official semantic annotations, we use them directly; otherwise, we construct 3D object annotations using Grounded-SAM with multi-view RGB-D observations and camera poses, rather than the inference-time DetAny3D. The automatic pipeline achieves 40.5% [email protected] against official annotations. Answers to size, distance, and counting questions are then computed from geometric annotations and simulator-provided GT information, rather than generated by GPT-4o.

Oracle information is used only for offline trajectory generation, while intermediate reasoning is constrained to information available in the observation history. Navigation actions follow precomputed reference trajectories, while found and all_found identify target-discovery states and expected_answer constrains the final answer. Intermediate Thoughts cannot use geometric information before it appears in Observations, and rule-based checks and VLM verification assess reasoning–observation consistency. Synthetic Observations are constructed from scene annotations in the corresponding tool-output formats, rather than necessarily obtained through actual tool execution. At inference, the agent receives no oracle information and acquires Observations through actual tool calls.

We have clarified geometric annotations, synthetic Observations, and trajectory verification in Section 3.2, added annotation and reference-answer details to Appendix B, and explained the offline oracle inputs in Appendix G.

Weakness 3 (Reward Design and 3D Reasoning)

The reward design is flawed, and the "3D reasoning" claim is not isolated. For submitted answers, Rans(τ)=C(τ)⋅s(τ)R_{ans}(\tau)=C(\tau)\cdot s(\tau)(Eq. 4a), so when s(τ)<0s(\tau)<0, the penalty grows as more evidence is gathered (e.g., a wrong closed-vocabulary answer with s(τ)=−1s(\tau)=-1 receives Rans=0R_{ans}=0 at C(τ)=0C(\tau)=0, but -1 at C(τ)=1C(\tau)=1). This may explain why removing C(τ)C(\tau) from RansR_{ans} raises LLM-Match (Table 4). In the coverage computation, grounding requires only a name-matched detection without agreement to the annotated position; size and VQA facts are credited without accuracy checks; positions use a 2 m tolerance; and the same coverage quantity is also reported as an evaluation metric. After RL, 3D calls that add position evidence barely change (Table 8), and no per-question-type results are reported.

Response:

Scaling negative answer scores by C(τ)C(\tau) is intentional and encourages better utilization of acquired evidence. When the agent answers incorrectly despite having sufficient evidence, a stronger penalty reinforces learning from such errors. Meanwhile, the separate Evidence Reward encourages evidence acquisition and discourages premature answering. As shown in Table 4 of the main paper, removing Coverage Scaling increases LLM-Match by 0.76/0.46 points on Seen/Unseen but reduces Evidence Coverage by 4.82/2.68 points, reflecting a trade-off between answer quality and evidence acquisition.

Evidence Coverage measures the acquisition of required evidence rather than the complete accuracy of individual tool predictions. Evidence matching requires target association and valid tool outputs. In particular, 3D Position must be within 2 m of the GT position, whereas Object Size and VisualQA are checked for output validity rather than numerical or semantic accuracy. We use the same Coverage definition for training and evaluation to consistently assess task-relevant evidence acquisition, while final-answer correctness is evaluated separately using LLM-Match.

The improvements in multi-step 3D reasoning are reflected in stronger performance on geometry-dependent tasks, rather than simply increased tool usage. The limited change in 3D calls adding position evidence (Table 8) does not fully capture the agent's ability to identify missing evidence and integrate geometric information across exploration steps. Tables 3 and 8 of the main paper show improvements in target recall, Evidence Coverage, and exploration efficiency. As shown in the table below, our additional per-question-type results show consistent gains from SFT+RL, particularly on Attribute-Size and Distance, which improve by 13.90/14.30 points on Seen and 13.69/14.28 points on Unseen, respectively.

Question TypeSFT-only SeenSFT+RL SeenSFT-only UnseenSFT+RL Unseen
Attribute-Size39.1853.0837.4051.09
Attribute-Color46.6850.9844.5148.79
Attribute-Special44.0848.6841.8146.50
Distance35.9850.2834.1148.39
Counting37.1950.0834.8147.09
Location-Location41.1848.1839.0245.83
Location-Special42.4848.3840.9146.69
Relationship41.9349.3639.5346.96
Status46.2050.4844.5849.06
Overall41.2550.3239.8348.16

Section 3.3 now explains the evidence-validation criteria and the rationale for negative-score scaling, while Appendix D provides the full per-category results and additional reward-scaling comparisons.

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

The paper proposes a multi-step reasoning 3D (embodied) QA agent. The claim is that VLMs are not accurate enough for answering questions that deal with object geometries and 3D properties and this is the motivation to introduce a 3D reasoning agent. The proposed agent uses spatial history along with the current observations to answer questions in a 3d scene. The contribution consists of a 3D-trajectory based dataset generation pipeline, a multi-step 3D reasoning agent based on iterative tool-usage, EQA dataset containing 18k tasks, and a SFT+RL recipe to train the EQA agent. They evaluate on HM-EQA, OpenEQA, EXPRESS-Bench, and their own EQA-RT-Unseen dataset. The authors claim that this agent beats Pred-EQA which is the strongest available baseline for the LLM-match metric.

Critical strengths

Applying RL to improve agents is a well motivated direction for research. The reward designing is clearly explained in the paper. The ablation study on the training regimes proves how important the evidence-grounded RL is (Table 3). Claims on the results in Table 1-4 are impressive.

Critical weaknesses
  1. The ablations in Table 3 are run for the EQA-RT-Seen and EQA-RT-Unseen datasets. Were training-regime ablation tried on other benchmarks?
  2. The introduction would benefit from a more concrete discussion of why 3D spatial reasoning is necessary for specific tasks, e.g., examples of questions that 2D or frame-based approaches fail on.
  3. The paper does not cite or compare against ToolEQA (Zhai et al., ICLR 2026), which proposes a closely related tool-augmented, multi-step reasoning agent for EQA trained with SFT and evaluated on the same benchmarks (HM-EQA, OpenEQA, EXPRESS-Bench). Without this comparison, it is difficult to assess the contribution of the proposed RL training beyond existing work. It is possible that the SFT-only ablation reflects a ToolEQA-like setup; if so, the authors should state this explicitly.

Minor issues: 1) Typo in Introduction last sentence. Line 123, Page 3: "effective and effective"

Questions for authors
  1. How does the method compare to a strong frontier VLM (like the recent GPT/Claude/Gemini models) used zero-shot as an agent?
  2. How does performance change if the Qwen3 based controller is replaced with a frontier model, i.e., how much of the gain comes from the training method versus the base model ?
  3. Training-regime ablations in Table 3 appear to be run only on the new benchmark. Do the same trends hold on the public benchmarks?'
  4. How does the proposed method differ from ToolEQA? Could the authors include it as a baseline, or clarify whether the SFT-only ablation corresponds to a ToolEQA-like setup?
3 comments

We sincerely thank the reviewer for the insightful and constructive feedback, which has been highly valuable in helping us strengthen and improve our manuscript.

Weakness 1 & Question 3 (Cross-Benchmark Training-Regime Ablations)

The ablations in Table 3 are run for the EQA-RT-Seen and EQA-RT-Unseen datasets. Were training-regime ablation tried on other benchmarks?

Training-regime ablations in Table 3 appear to be run only on the new benchmark. Do the same trends hold on the public benchmarks?'

Response: The same training-regime improvement trend holds on HM-EQA, OpenEQA, and EXPRESS-Bench. As shown in the table below, SFT+RL outperforms both Zero-shot and SFT-only on all reported metrics across these three public benchmarks, demonstrating that the benefits of our training strategy extend to benchmarks with different question distributions and evaluation protocols.

BenchmarkMetric0-shotSFT-onlySFT+RL
HM-EQAAccuracy ↑63.8164.5365.77
HM-EQAPath Length (m) ↓27.6424.8320.93
OpenEQA (A-EQA)LLM-Match ↑50.8752.1653.99
OpenEQA (A-EQA)LLM-SPL ↑43.7645.3447.21
EXPRESS-BenchSucc. ↑47.8350.2655.51
EXPRESS-BenchSucc.* ↑65.7468.2372.61
EXPRESS-BenchEpath ↑38.9242.6148.37
EXPRESS-BenchdT ↓7.186.475.89

We have incorporated the full cross-benchmark training-regime results into Appendix D (Cross-Benchmark Training-Regime Ablations).

Weakness 2 (Necessity of 3D Spatial Reasoning)

The introduction would benefit from a more concrete discussion of why 3D spatial reasoning is necessary for specific tasks, e.g., examples of questions that 2D or frame-based approaches fail on.

Response: 3D spatial reasoning is necessary for geometric reasoning across observations acquired from different viewpoints and locations, because understanding and reasoning about 3D spatial relationships between objects requires integrating observations acquired from different viewpoints and locations into a unified spatial representation, rather than relying solely on independent frame-by-frame visual understanding. For example, determining the spatial distance between two objects in different rooms requires the agent to acquire observations at multiple locations, localize the targets in a common 3D coordinate system, and compute their distance from their positions. Similarly, comparing the sizes of objects in different rooms requires acquiring and integrating geometric measurements across exploration steps. These tasks require the agent not only to interpret the current observation but also to actively acquire, maintain, and integrate spatial evidence across steps. We will add these concrete examples to the Introduction.

We have added these cross-room distance and size-comparison examples to Section 1 (Introduction).

Weakness 3 & Question 4 (Comparison with ToolEQA)

The paper does not cite or compare against ToolEQA (Zhai et al., ICLR 2026), which proposes a closely related tool-augmented, multi-step reasoning agent for EQA trained with SFT and evaluated on the same benchmarks (HM-EQA, OpenEQA, EXPRESS-Bench). Without this comparison, it is difficult to assess the contribution of the proposed RL training beyond existing work. It is possible that the SFT-only ablation reflects a ToolEQA-like setup; if so, the authors should state this explicitly.

How does the proposed method differ from ToolEQA? Could the authors include it as a baseline, or clarify whether the SFT-only ablation corresponds to a ToolEQA-like setup?

Response: Our method differs from ToolEQA in two respects. Architecturally, we introduce Spatial Memory to maintain and retrieve spatial evidence across exploration steps. In terms of training, we introduce Evidence-Grounded RL to further learn how to actively acquire and use spatial evidence. Our SFT-only setting is not equivalent to ToolEQA: it uses the current architecture, including Spatial Memory, and applies supervised fine-tuning on EQA-RT. We also wish to clarify that ToolEQA was not accepted to ICLR 2026.

The architectural role of Spatial Memory is described in Section 3.1 and Appendix A, and the SFT-only setting is specified in Section 4.3.

Question 1 (Zero-Shot Frontier VLM Comparison)

How does the method compare to a strong frontier VLM (like the recent GPT/Claude/Gemini models) used zero-shot as an agent?

Response: We outperform the Codex + GPT-6 Astra zero-shot agent in both answer accuracy and exploration efficiency. We evaluate Codex + GPT-6 Astra as a Zero-shot Agent on EQA-RT-Seen and EQA-RT-Unseen. As shown in the table below, our method achieves LLM-Match scores that are 4.65 and 2.25 percentage points higher on Seen and Unseen, respectively, while reducing exploration length by 61.8% and 62.8%. These results indicate that, despite their strong general reasoning and tool-use capabilities, frontier models used as zero-shot agents do not necessarily acquire spatial evidence efficiently. In contrast, task-specific SFT+RL training improves answer quality while substantially reducing exploration cost, demonstrating its value for multi-step spatial evidence acquisition in EQA.

ModelSeen LLM-Match ↑Seen Length (m) ↓Unseen LLM-Match ↑Unseen Length (m) ↓
Codex + GPT-6 Astra45.6757.9245.9162.38
Ours50.3222.1548.1623.19

The additional experiment is now included in Appendix D (Frontier Model Comparisons).

Question 2 (Frontier Controller Replacement)

How does performance change if the Qwen3 based controller is replaced with a frontier model, i.e., how much of the gain comes from the training method versus the base model ?

Response: A stronger frontier controller improves zero-shot performance, but our SFT+RL model achieves comparable or better accuracy with shorter exploration trajectories. We replace the Qwen3-VL-8B Controller with GPT-6 Astra while keeping the other agent components, tool suite, and interaction budget unchanged. GPT-6 Astra achieves LLM-Match scores of 48.19% and 48.63% on Seen and Unseen, respectively, showing that a stronger base model can improve zero-shot EQA performance. However, our Qwen3-VL-8B trained with SFT+RL achieves a higher LLM-Match on Seen (50.32% vs. 48.19%) and comparable performance on Unseen (48.16% vs. 48.63%), while reducing exploration length by 38.3% and 41.2%, respectively. These results show that both base-model capability and task-specific training contribute to performance, and that training for multi-step spatial evidence acquisition enables a smaller model to achieve answer quality comparable to a stronger frontier model with greater exploration efficiency.

ModelEQA-RT-Seen LLM-Match ↑Seen Length (m) ↓EQA-RT-Unseen LLM-Match ↑Unseen Length (m) ↓
GPT-6 Astra Controller48.1935.8748.6339.44
Ours (Zero-shot)38.3243.6535.1755.37
Ours (SFT+RL)50.3222.1548.1623.19

We have added the GPT-6 Astra controller experiment to Appendix D (Frontier Model Comparisons).

Dear Reviewer ZKsK,

Thanks for your time and effort in reviewing this paper. Your comments are valuable to improve the quality of this paper. As the author-reviewer discussion period ends in 48 hours, we kindly request that you review the rebuttal and confirm whether your concerns have been addressed. Thanks a lot.

Regards from the authors.

Comments on the paper

1 comment

Summary of Revisions: We sincerely thank the reviewers for their constructive feedback and valuable suggestions. Through the rebuttal and additional experiments, we have addressed the major concerns regarding generalization, spatial reasoning, data reliability, and reward design. Following our discussions, the two reviewers who participated in the discussion acknowledged that their concerns had been addressed. We have incorporated the rebuttal updates into the revised manuscript and appendix to further strengthen the experimental evidence and clarify the methodology. Specifically, we have:

  • Expanded the experimental comparisons, including zero-shot frontier models, frontier controller replacement, cross-benchmark training-regime ablations, and the transfer of Evidence-Grounded RL to Pred-EQA.
  • Added comprehensive spatial reasoning analyses, including per-question-type and multi-target evaluations, failure analysis, and component ablations to distinguish spatial reasoning improvements from tool selection.
  • Strengthened the evaluation reliability, including reporting the mean and variance of performance across repeated experiments, independent evidence-correctness verification, and human evaluation of answer quality under different reward settings.
  • Extended the reward analysis, including additional Coverage Scaling ablations and a detailed discussion of the relationship between evidence acquisition, evidence correctness, and final-answer quality.
  • Clarified the data generation and training procedures, including geometric annotation sources, synthetic trajectory construction, oracle information usage, Evidence Checklist generation, and benchmark scene overlap.
  • Refined the manuscript and appendix to improve clarity, reproducibility, and the motivation for multi-step 3D spatial reasoning.

We sincerely thank the Area Chair and reviewers again for their time and consideration. We hope that the additional evidence and revisions presented throughout the discussion will contribute to a comprehensive final assessment of our work.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.