acceptodds
Under review as a conference paper at ICLR 2027

SGT-3D: 3D-Guided Self-Distillation and Geometric Verification for Spatial Reasoning

Abstract

Spatial reasoning from images often involves estimating object geometry and using it to answer spatial questions. Some spatial reasoning approaches expose intermediate 3D attributes, but making them visible does not ensure accurate prediction or correct use. These two goals motivate complementary supervision: final-answer rewards do not directly assess attribute accuracy, while training on reference responses may leave reasoning from the model's own imperfect predictions insufficiently supervised. The resulting challenge is to improve object-level spatial understanding and turn better estimates into more accurate answers. We present , which couples learning to predict object centers, dimensions and orientations with learning to reason from the states the student actually generates. A training-only 3D expert guides self-distillation, with continuation supervision following the student's stated attributes within a fixed-state interface, and subsequent reinforcement learning supplies numerical attribute feedback alongside answer rewards. Inference requires only RGB images and questions. SGT-3D achieves QA accuracy of 69.2% on 3DSRBench-real and 96.2% on CV-Bench-3D, with 75.0% simultaneous center/dimension/orientation pass rate on the separate SUN-G400 audit using native SUN RGB-D annotations and fixed object-region prompts.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Spatial Reasoning, Reinforced Learning, On-Policy Distillation

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

SGT-3D trains an RGB-only vision-language model to generate object-level 3D attributes and reason from those attributes. The method contains two main components. CS-OPSD distills supervision using continuations conditioned on the student’s own committed intermediate attributes, addressing the mismatch between ideal reference states during training and imperfect generated states during inference. GF-RL introduces geometry-aware rewards with field-specific normalization and temporal credit assignment.

The paper reports 69.2% accuracy on 3DSRBench-real and 96.2% on CV-Bench-3D. It also provides ablations against answer-only reinforcement learning, attribute cross-entropy, masking variants, different training orders, and alternative reward-credit schemes. The method and results are promising, but the statistical reliability, leakage analysis, and independent annotation protocol are not documented sufficiently for the strength of the claims.

Critical strengths
  1. CS-OPSD addresses a meaningful training-inference mismatch. Supervising continuations from the student’s own committed attributes is better aligned with deployment than always conditioning on correct reference attributes.

  2. The ablation design is unusually comprehensive. Comparisons against answer-only RL, attribute CE, teacher-conditioning variants, alternate training orders, and crossed reward controls make the reported improvement easier to attribute to the proposed components.

  3. The independent geometry evaluation and joint geometry-answer metrics provide evidence beyond final-answer accuracy. This is important for verifying whether the model actually uses the proposed 3D representation.

Critical weaknesses
  1. The method depends heavily on privileged 3D labels, regenerated traces, and a frozen expert during training. This dependency limits scalability and makes the quality and provenance of those labels central to the method, but the paper does not discuss it in sufficient depth.

  2. The proposed object representation may be too narrow for some of the reasoning claims. Center, size, and orientation do not explicitly represent gravity, visibility, occlusion, or geometric uncertainty. The paper notes that camera-up is not necessarily world-up but does not explain how world-relative height questions are resolved.

  3. Several comparisons combine differences in objectives, supervision, training history, and inference procedure. They show that the complete SGT-3D pipeline performs well, but do not always isolate which methodological component is responsible.

Questions for authors
  1. How are gravity-relative questions answered when camera-up does not correspond to world-up?
  2. What information about occlusion and geometric uncertainty is preserved in the object representation?
  3. Which baselines use the same initial checkpoint and supervision resources as SGT-3D?
3 comments
AuthorsOpenReview ↗

Response to Reviewer PGhx [Part 1/2]

We thank the reviewer for the assessment that "CS-OPSD addresses a meaningful training-inference mismatch" and that the ablation design is "unusually comprehensive".

W1. Scalability, replaceability and label provenance of the training-time expert

The dependency on the 3D expert is computationally cheap, so it does not limit scalability, and SGT-3D does not require the expert labels to be exact.

  1. Dependency and scalability. On 8$\times$H200 GPUs, the frozen 3D expert labels one image in 0.19 s. Labeling the whole training set and generating all reasoning traces takes 79 minutes, about 1% of the total training time. The method uses the expert only through the reference ZRZ^R (centers, dimensions, orientations), so any model that outputs these fields can replace it.

  2. Provenance and quality of the labels. The labels come from one automatic pipeline, and SGT-3D remains accurate when they are imperfect.

    • Provenance. Frozen WildDet3D, prompted with 2D boxes from GroundingDINO, produces centers, dimensions and quaternions for SpatialReasoner's training images. A reference is admitted only if prompt-box IoU ≥ 0.5, confidence ≥ 0.3 (2D) and ≥ 0.1 (3D), dimensions and depth are finite and positive, and the quaternion is valid. Each trace uses the operation that its question record specifies, and its arithmetic is re-checked after serialization (Sec. 3.1, App. A.1, A.5).
    • Quality. The raw expert gives the correct depth relation for 92% of the I100 questions (Table 17A). When the expert reference is wrong and the student is right, the CS-OPSD teacher stays factually correct in 95% of cases, versus 24% for expert priority (Table 14).

W2 / Q1 / Q2. Gravity, occlusion and geometric uncertainty in the object representation

The proposed object representation is general, and it handles both cases that the reviewer describes.

  1. Gravity (Q1). Gravity-relative questions are answered along an estimated world-up, not along the camera axis. Think estimates world-up ucu_c from scene cues, such as upright people, vertical edges and support surfaces, and compares Δh=uc⊤(cA−cB)\Delta h = u_c^\top (c_A - c_B) (App. A.1, Eq. 5). The camera translation cancels in Δh\Delta h, so only the up direction is needed. In Fig. 6, where the image is rotated by 180°, this gives the correct order (+0.550), whereas camera-up gives the opposite order (−0.538). In Fig. 7, an unrotated photo, SGT-3D correctly answers that a cat on a bench is higher than a boat on the water, although the cat appears lower in the image. SGT-3D also has the best Height accuracy in Table 2 (60.6), +8.1 over SpatialReasoner.

  2. Visibility, occlusion and geometric uncertainty (W2, Q2). The representation preserves the hidden extent of each object: Gaze predicts a full-object (amodal) 3D box, so its center and dimensions include occluded parts. The hidden height moves the box center downward along gravity (Eq. 6), and Think compares distances between these complete centers. In Fig. 8, a couch hides the lower stem and base of a floor lamp, and the question asks which lamp is closer to the couch. SGT-3D places the complete center of the floor lamp below its visible part and correctly answers “floor lamp”. On D1000, SGT-3D meets the center, dimension and orientation tolerances for all queried objects in 85.8% of the questions, measured against full-object references (Table 3).

AuthorsOpenReview ↗

Response to Reviewer PGhx [Part 2/2]

W3. Isolating the methodological component responsible for each gain

Our ablations isolate each methodological component: every variant starts from the same checkpoint, uses the same supervision resources and changes one component at a time. Variants A–G and every ablation row start from Qwen2.5-VL-7B-Instruct. They use SpatialReasoner's training data, the same regenerated references and matched update budgets (App. A.5). The committed-state teacher is +4.9 QA over plain conditioning and +5.1 over expert priority (Table 5). The GF-RL objective is +3.8 QA over answer-only RL and +2.6 over answer RL with attribute cross-entropy (Table 3). Boundary credit is +3.0 QA over segment credit (Table 6).

Q3. Baselines with the same initial checkpoint and supervision resources as SGT-3D

All models in Table 2 use the same Qwen2.5-VL-7B backbone as SGT-3D, and SpatialReasoner is the only external baseline that also shares our training-data source. The revised Table 2 includes only Qwen2.5-VL-7B-based models, so that the backbone is held fixed across all comparisons. The other baselines use different supervision resources: VST-7B uses VST-P (4.1M samples) and VST-R (135K samples), 3DThinker-7B uses 237K SFT and 57K RL examples, and VERDICT adds no training and uses frozen verifiers at inference. PolyV and Mini-o3 originally use other settings, so we align them with the dense, RGB-only, tool-free inference setting of Table 2: PolyV-dense is our reproduction on the Qwen2.5-VL-7B backbone without MoE experts, and Mini-o3 keeps its entropy-regularized training but disables tools at inference (App. C.1).

With the same backbone and the same data source, SGT-3D exceeds SpatialReasoner by 8.9 points (69.2 vs. 60.3). It also exceeds 3DThinker-7B (66.3) and VST-7B-RL (60.0), although it uses only 25,200 spatial training records, far fewer than either.

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer PGhx,

Thanks again for your thoughtful feedback.

We would greatly appreciate any further discussion of our rebuttal responses. We would be happy to provide further clarification and address remaining concerns.

Thank you so much for your time!

Best regards,

Authors

Reviewer_Qvn9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes SGT-3D, a framework for RGB-only spatial reasoning that combines structured SFT, committed-state on-policy self-distillation (CS-OPSD), and geometric-feedback reinforcement learning (GF-RL). It aims to improve both 3D attribute prediction and reasoning consistency. The paper reports strong results on 3DSRBench and CV-Bench-3D, supported by several ablation studies.

Critical strengths
  1. The paper addresses the important distinction between predicting accurate 3D attributes and reasoning consistently from those predictions.

  2. The reported ablations examine teacher supervision, reward normalization, and credit assignment, potentially providing useful insights into the proposed components.

Critical weaknesses
  1. Full CS-OPSD improves over Answer-masked supervision by only 0.3–0.4 percentage points, raising questions about the necessity of its complete supervision strategy.

  2. Insufficient details on independent 3D annotations. The construction and provenance of D1000 annotations are unclear, and some orientation-related thresholds are unspecified, limiting reproducibility.

  3. Performance degrades under stronger reference noise, where CS-OPSD even underperforms conflict masking, raising concerns about robustness to imperfect geometric supervision.

Questions for authors
  1. Table 19 shows only a 0.3–0.4 percentage-point improvement over Answer-masked supervision. Is this improvement consistent across multiple random seeds?

  2. Could the authors clarify how the D1000 annotations were collected and validated, and specify the orientation thresholds used?

  3. Table 6 shows that CS-OPSD underperforms conflict masking at higher noise levels. How sensitive is the method to errors in real-world 3D expert predictions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Qvn9

We thank the reviewer for recognizing that the paper "addresses the important distinction between predicting accurate 3D attributes and reasoning consistently from those predictions". The reviewer's Tables 19 and 6 are now Tables 20 and 7 in the revised PDF.

W1 / Q1. Consistency of the gain across seeds and necessity of full CS-OPSD

Full CS-OPSD is necessary because (1) it performs consistently better than both Answer-masked variants (ablation variants of CS-OPSD, not baselines) on all three training seeds, and (2) it is the simpler design, because it does not need the conflict indicator χ\chi. We added a three-seed experiment to support this (Table 21).

  1. Consistency across seeds (Q1). Full CS-OPSD is above both Answer-masked variants on each of the three seeds (seeds 29, 17 and 43; Table 21): | Variant | Seed 1 (reported) | Seed 2 | Seed 3 | Mean ± SD | |---|---|---|---|---| | Full CS-OPSD | 69.2 | 68.9 | 69.4 | 69.16 ± 0.24 | | Answer-only mask | 68.8 | 68.7 | 69.0 | 68.83 ± 0.18 | | Rescaled Answer-only mask | 68.9 | 68.8 | 69.1 | 68.91 ± 0.16 |

  2. Necessity of the complete strategy (W1). The Answer-masked variants must first detect conflicting states with χ\chi. This needs parsed states and a decidability test, and χ\chi is defined only for supported binary depth questions (App. B.2). Full CS-OPSD applies one loss to every token of every response (Eq. 2). It is the simpler design. It is also more accurate: it is higher than both Answer-masked variants on all four metrics (full-benchmark QA, D1000 QA, attribute correctness and joint correctness; Table 20).

W2 / Q2. Construction and validation of D1000 references, and orientation thresholds

D1000 references come from an automatic annotation pipeline that is separate from the training expert and the student (App. C.1). The two orientation-related values are ηR=π/12\eta_R = \pi/12 rad (15°), the pass threshold, and τR=π/12\tau_R = \pi/12 rad (15°), the reward scale (App. A.2, C.2). We will release the D1000 references.

  1. How the references were collected (construction and provenance). App. C.1 now describes the full procedure. The questions are 1,000 original, unflipped 3DSRBench-real questions (100 Height, 325 Location including all 175 camera-depth questions, 150 Orientation and 425 Multi-Object). We fixed them before any student output existed. To annotate them, we adapt the automated segmentation, depth and pose-proposal stages and the filtering of SpatialReasoner, together with the object-to-scene correspondence and geometric alignment of GS-Reasoner. The pipeline associates each referent with an instance mask and estimates an oriented, full-object 3D box in the camera frame. No reference copies a WildDet3D output or a student output.
  2. How they were validated. Automatic filters check referent correspondence, finite box parameters, positive dimensions, valid rotations, and the consistency of projected corners with registered visible surfaces. Eligibility is fixed per field before student evaluation.
  3. The orientation thresholds. A predicted orientation passes when its symmetry-aware rotation error is at most ηR=π/12\eta_R = \pi/12 rad (15°). In GF-RL, the orientation reward is rori=exp⁡(−dori/τR)r_{\mathrm{ori}} = \exp(-d_{\mathrm{ori}}/\tau_R) with τR=π/12\tau_R = \pi/12 rad (15°). This scale follows the convention of the center and size rewards, whose scales equal their tolerances (τc=0.1\tau_c = 0.1, τs=0.2\tau_s = 0.2). We fixed both values before evaluation (App. A.2, C.2).

W3 / Q3. Robustness to imperfect geometric supervision and to errors in real-world 3D expert predictions

SGT-3D is robust to both imperfect geometric supervision and errors in real-world 3D expert predictions, and it is more robust than SpatialReasoner.

  1. Imperfect geometric supervision (W3). Under added noise of SD 0.05, SGT-3D loses only 0.9 QA (69.2 → 68.3). In a new comparison experiment, SpatialReasoner-SFT loses 3.2 QA (58.3 → 55.1) under the same noise. SGT-3D also stays +2.8 QA above conflict masking (Table 7). Conflict masking is ahead only at SD 0.10, a stress test whose added noise per coordinate (0.50 m at a 5 m median depth) equals our tolerance for a correct center.
  2. Errors in real-world 3D expert predictions (Q3). All main results already include these errors: SGT-3D is trained only on raw WildDet3D predictions and reaches 69.2 on 3DSRBench-real and 96.2 on CV-Bench-3D. In this real-world setting (SD 0 in Table 7), full CS-OPSD is +3.2 QA above conflict masking.

Clarification of the outdated PDF metadata

The discrepancy comes from outdated metadata. All reported results are measured. The metadata came from an earlier compact-layout test. That test checked whether the inserted content would push the paper over the page limit. We did not update it before the final compile. The revised PDF corrects it.

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer Qvn9,

Thanks again for your thoughtful feedback.

We would greatly appreciate any further discussion of our rebuttal responses. We would be happy to provide further clarification and address remaining concerns.

Thank you so much for your time!

Best regards,

Authors

Thank you for the detailed rebuttal. Several concerns have been addressed, including the PDF metadata issue, for which I have removed the ethics flag. However, the modest incremental gains, annotation reliability, and robustness under stronger noise remain concerns. I therefore maintain my original rating.

Thank you for your feedback and for the further discussion. We answer the three remaining concerns directly.

  1. Modest incremental gains.

    (1) The 0.3–0.4-point difference compares two variants that both retain our core mechanism. Answer-only masking and full CS-OPSD are both introduced in this work and share committed-state supervision on Think tokens. Their difference therefore measures only whether Answer tokens on conflicting prefixes are also supervised, rather than the contribution of committed-state supervision itself.

    (2) The ablations directly demonstrate the value of the core mechanism. The committed-state teacher improves QA accuracy by 4.9 points over plain conditioning and 5.1 points over expert priority (Table 5). On conflicting prefixes, retaining committed-state supervision on Think tokens yields +2.8 points with Answer-only masking and +3.2 points with full CS-OPSD, compared with dropping all supervision on those prefixes (Table 20).

    (3) The variant that leads at SD 0.10 is also ours. Conflict masking, which is 0.6 QA ahead of full CS-OPSD at SD 0.10, is also a CS-OPSD design of this paper (Table 7). Full CS-OPSD therefore does not lead in every noise regime; we select it for its higher accuracy with the actual expert outputs (+3.2 QA at SD 0; Table 7) and its simpler implementation.

  2. Annotation reliability.

    (1) The D1000 references agree with independent human labels. The 2,100 question-answer pairs of 3DSRBench-real were annotated by ten human annotators and reviewed in two stages, and samples without annotator consensus were discarded (Ma et al., 2025a). Answers computed from our references reach 98.6% agreement with these human answer keys. This agreement validates our references against independent human labels.

    (2) The references are independent of what they evaluate. Annotation excludes student predictions, training-expert outputs and QA answer keys (App. C.1).

    (3) The main QA results do not depend on these references. All QA results (69.2 on 3DSRBench-real and 96.2 on CV-Bench-3D) use the benchmarks' own answer keys.

  3. Robustness under stronger noise. Since your concern focuses on stronger noise, we add the comparison with SpatialReasoner, the most closely related baseline, at the stronger-noise setting SD 0.10, which our previous response did not include.

    (1) SGT-3D degrades less than SpatialReasoner-SFT at both noise levels. SpatialReasoner-SFT shares our base model and training-data source. SGT-3D loses 0.9 and 6.1 QA at SD 0.05 and 0.10, whereas SpatialReasoner-SFT loses 3.2 and 7.3 QA.

    (2) Under the strongest noise, SGT-3D still outperforms baselines trained without added noise. At SD 0.10, SGT-3D reaches 63.1 QA, 12.1 points above SpatialReasoner-SFT (51.0), and higher than SpatialReasoner (60.3), VST-7B-RL (60.0), PolyV-dense (61.8) and Mini-o3 (62.5) without added noise (Table 2).

    (3) The variant that leads at SD 0.10 is also ours. Conflict masking, which is 0.6 QA ahead of full CS-OPSD at SD 0.10, is also a CS-OPSD design of this paper (Table 7). We acknowledge that conflict masking performs better at SD 0.10 by 0.6 points. Full CS-OPSD therefore does not dominate every noise regime, we select it for its clean-setting performance and simpler implementation.

If you have any other concerns, we would be glad to discuss them further.

Reviewer_58xdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents SGT-3D, a training framework designed to improve 3D spatial reasoning by enhancing both attribute prediction and attribute use. The framework consists of two main components: Committed-State On-Policy Self-Distillation (CS-OPSD) and Geometric Feedback Reinforcement Learning (GF-RL). CS-OPSD employs a shared-weight self-teacher conditioned on 3D reference attributes provided by an external expert. The teacher guides the student's attribute predictions while supervising consistent reasoning based on the student's already-generated attributes, even when they are inaccurate. GF-RL further improves attribute prediction through geometry-aware rewards that evaluate predicted 3D centers, dimensions, and orientations, alongside a final-answer reward. Together, these components improve question-answering accuracy on 3DSRBench-real and CV-Bench-3D, with experimental results also demonstrating improved 3D attribute prediction accuracy.

Critical strengths
  1. The paper provides a clear motivation for improving the use of predicted 3D attributes in spatial reasoning tasks. The proposed training framework demonstrates improvements in question-answering accuracy on both evaluated benchmarks, 3DSRBench-real and CV-Bench-3D.
  2. The ablation studies are comprehensive and provide convincing evidence for the effectiveness of both CS-OPSD and GF-RL.
Critical weaknesses
  • The connection between the method design and its two objectives could be clearer. While the paper aims to improve both 3D attribute prediction and attribute use, I find the connection between the specific design choices and these two objectives somewhat unclear. Although the proposed method appears to address both aspects, it is not always clear how each step is designed to improve attribute prediction or attribute use. For example, in CS-OPSD, the teacher uses expert references to guide attribute prediction at earlier token positions, while supervising subsequent reasoning based on the student's already-generated attributes, even when they conflict with the expert references. Although this mechanism is described in Section 3.2, a clearer explanation of how these different supervision signals contribute to improving attribute prediction and attribute use, respectively, would help readers better understand the method.

  • The mathematical notation in Section 3 could be clearer. For example, Equation (2) uses several symbols, such as (\theta), (q_t^{(n)}), (\operatorname{stopgrad}), and (\hat{y}_{<t}), without explicitly explaining their meanings and roles in the surrounding text. Providing clear definitions when these symbols first appear would improve the readability of the method.

Questions for authors
  1. From my understanding, the teacher is supposed to provide reliable reasoning supervision even when the student's predicted attributes conflict with the expert reference. In the example where the expert depths are (z_A=2, z_B=5), but the student has generated (z_A=5, z_B=2), the teacher is expected to supervise the continuation "B is nearer." However, since this is a self-distillation and committed-state priority is implemented through an instruction, what ensures that the teacher correctly reasons over the student's committed attributes rather than making the same reasoning errors as the student? Although Table 13 provides evidence of improved arithmetic consistency, further clarification of how reliable reasoning supervision is obtained in such conflicting states would be helpful.

  2. While the ablation studies are impressive and provide convincing evidence of the effectiveness of CS-OPSD, I think the presentation of the teacher policy and its role in guiding attribute use could be clearer. Although Section 3.2 and Appendix B.1 describe the committed-state priority rule, it would be helpful to provide a more intuitive, step-by-step illustration of how the teacher guides the student's attribute prediction and subsequent reasoning. For example, a figure showing how the teacher uses expert references to supervise unfinished attribute fields, while following the student's committed attributes to guide Think and Answer tokens, particularly when the two states conflict, would make the mechanism easier to understand. Such an illustration would complement the existing ablation studies and help readers better understand how the proposed teacher policy improves attribute use.

  3. In Figure 4(c), the two methods use different reasoning strategies given the same 3D attributes: Mask + GF-RL considers only horizontal distances, while SGT-3D uses full 3D Euclidean distances. Could the authors elaborate on why this difference occurs? Does it strengthen the claim of improved attribute use? It would be interesting to see whether the method helps the model better use or combine relevant 3D attributes to answer specific spatial questions.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 58xd [Part 1/2]

We thank the reviewer for the careful reading, and for the assessment that the ablations "provide convincing evidence for the effectiveness of both CS-OPSD and GF-RL". Table and figure numbers below refer to the revised PDF: the reviewer's Table 13 and Fig. 4(c) are now Table 14 and Fig. 5(c).

W1. Clarification of the connections between design choices, objectives and supervision signals

We clarify that each design choice in SGT-3D is tied to one objective through the tokens it supervises, so attribute prediction and attribute use receive separate and complementary supervision signals. In CS-OPSD, expert guidance at the attribute tokens improves attribute prediction, and committed-state supervision at the reasoning tokens improves attribute use. In GF-RL, the attribute reward improves attribute prediction, and the answer reward optimizes the final answer. The revised paper adds this mapping as Table 1.

DesignsSupervision signalSupervised tokensImprovesEvidence
CS-OPSDExpert guidanceScan, GazeAttribute predictionD1000 attribute correctness 58.0 → 70.0 (A→B, Table 3)
CS-OPSDCommitted-state supervisionThink, AnswerAttribute use+2.8 QA (Table 20); state agreement 96% vs. 87% (Table 17A)
GF-RLAttribute rewardUp to each field's endAttribute prediction+13.8 attribute, +14.9 joint over answer-only RL (Table 3)
GF-RLAnswer rewardWhole responseFinal answerQA 63.0 → 65.4 (B→D, Table 3)

Together, the two steps raise QA from 53.8 to 69.2 and joint attribute-and-answer correctness from 48.0 to 78.4 (A→E, Table 3).

W2. Definitions and roles of the symbols in Eq. (2)

The four symbols are defined below. The revised Section 3.2 defines them at their first use, together with all other symbols of Eq. (2).

  • θ\theta: the trainable student parameters (LoRA adapters on Qwen2.5-VL-7B-Instruct). Role: the only quantity that Eq. (2) updates. θn\theta_n is its value at iteration n.

  • qt(n)(v)q_t^{(n)}(v): the self-teacher's probability of token v at position t. Role: the distillation target. The teacher uses the same weights θn\theta_n, but its context adds the expert reference and the teacher instruction.

  • stopgrad: blocks the gradient through the teacher. Role: the target stays fixed, so the update changes only the student.

  • y^<t\hat{y}_{<t}: the first t−1 tokens of the student's sampled response y^\hat{y}. Role: the shared prefix. Student and teacher both predict token t after this same student-written prefix.

AuthorsOpenReview ↗

Response to Reviewer 58xd [Part 2/2]

Q1. Why the teacher reasons reliably over the student's committed attributes

Reliable reasoning supervision in conflicting states comes from splitting the work into two separate steps: one step corrects the attribute values, and the other computes the answer from the values that the student has already written. The teacher is never asked to be more accurate than the student. It only maps the values that the student has already printed to an answer.

  1. Why the teacher does not repeat the student's error. In the reviewer's example (expert zA=2,zB=5z_A = 2, z_B = 5; student zA=5,zB=2z_A = 5, z_B = 2), the student's error is a geometric error. Geometric errors are corrected where they occur: at the attribute tokens, where the teacher follows the expert, and by the GF-RL attribute reward. Think and Answer have one task, to map printed values to an answer: 2<52 < 5, so "B is nearer". The shared weights learned this mapping from validated SFT traces (App. A.3), and it needs no privileged information.

  2. Why the teacher follows the student's values. If the teacher answered from the expert's values, the student's error would be hidden inside the response, and the student would learn answers that contradict its own values. With committed-state supervision, the answer always follows from the predicted attributes. Better attributes then give better answers: the GF-RL attribute reward raises QA by +3.8 after CS-OPSD, versus +2.4 after continued SFT (Table 4).

  3. Measured reliability. In wrong, conflicting states, the teacher's arithmetic/trace consistency is 94%/95% with the committed-state instruction, 35%/38% with plain conditioning and 7%/9% with expert priority (Table 14). After identical GF-RL, the final QA is +4.9 and +5.1 above these two policies (Table 5).

Q2. Step-by-step illustration of teacher guidance in conflicting states

The teacher guides the student in two phases. For attribute values, it always follows the expert. After the values are committed, it supervises the reasoning from those values, because they are the only values that the student has at inference. The revised paper adds this illustration as Fig. 3, with the reviewer's example (expert zA=2,zB=5z_A = 2, z_B = 5; student zA=5,zB=2z_A = 5, z_B = 2):

  1. Unfinished attribute field. While the student is still writing zAz_A, the teacher target follows the expert reference (zA=2z_A = 2). This step supervises attribute prediction. It is where the student's error is corrected.

  2. Committed field. After the student has printed zA=5z_A = 5 and zB=2z_B = 2, these values are the state that the reasoning must use.

  3. Think and Answer. The teacher target computes from the committed values: 2<52 < 5, so "B is nearer". This step supervises attribute use: the answer must follow from the stated values.

  4. Contrast. An expert-priority teacher supervises "A is nearer", which contradicts the printed values. It teaches the student to ignore its own state; its trace consistency in such states is 9% (Table 14).

Q3. Source of the gain over Mask + GF-RL and evidence of improved attribute use

The difference occurs because only SGT-3D keeps the teacher's supervision of the reasoning step on conflicting states. It strengthens the claim of improved attribute use, and the method does help the model combine attributes.

  1. Why the difference occurs. On conflicting prefixes, Mask + GF-RL removes the teacher's supervision of both Think and Answer tokens; SGT-3D keeps it. Restoring only the Think supervision recovers +2.8 of the +3.2 QA gap (Table 20), so the gain comes mainly from supervising the computation. Without that supervision, only the answer reward shapes the computation in those states, and the answer reward does not specify which operator to apply.

  2. Does it strengthen the claim of improved attribute use? Yes. With identical oracle attributes, the CS-OPSD student reaches 98% QA, versus 94% for the mask student and 90% for the plain student (Table 17A). The attributes are identical, so the whole gain comes from attribute use.

  3. Does the method help the model better use or combine relevant 3D attributes? Yes. At every Think position, the teacher writes the complete computation on the student's own values (operation, arguments, numbers), so the student learns to use all relevant coordinates. On questions that combine attributes, SGT-3D has the best accuracy in Table 2: 64.6 on Multi-Object (versus 63.2 for the next best model) and 95.7 on CV-Bench-3D Distance (versus 94.7).

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer 58xd,

Thanks again for your thoughtful feedback.

We would greatly appreciate any further discussion of our rebuttal responses. We would be happy to provide further clarification and address remaining concerns.

Thank you so much for your time!

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.