acceptodds
Under review as a conference paper at ICLR 2027

GSSR: Graph-Spectral Control of Text-Induced Logit Updates for 3D Part Segmentation

Abstract

Meaning-preserving prompts for the same 3D part can produce different pointwise masks. We measure the graph-frequency structure of paraphrase-induced logit changes and find that normalized Dirichlet energy adds modest held-out predictive value for soft-mask disagreement across two fusion heads. We then test a selective placement of a standard graph filter: GSSR filters the final query-dependent logit residual while leaving the query-independent base unfiltered. For the same frozen fields and filter, residual-only and full-logit filtering act identically on prompt-pair logit differences; the placement question is whether preserving the base benefits segmentation. In a PTv3-based controlled model, GSSR lowers soft disagreement () from to and raises worst-prompt mIoU from to relative to the unfiltered Direct control. Its primary 3DCoMPaT-Coarse mIoU is versus (five-seed means). Frozen and retrained placement controls show higher observed region and boundary scores than full-logit filtering; a second fusion head replicates the Direct/GSSR direction. These findings support selective filter placement in the tested model family.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
text-prompted 3D part segmentation, graph signal processing, multimodal fusion, prompt robustness, counterfactual analysis

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_zctyOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the problem of spatial drift induced by semantically-equivalent prompts in language-guided 3D part segmentation, and proposes GSSR, which performs graph filtering only on query-related logit residuals while preserving the query-agnostic base field. The motivation is somewhat appealing; nevertheless, the paper contains substantial flaws in several key aspects.

Critical strengths

(1)This paper achieves sota performance on the part segmentation task.

(2)This paper organizes in a good shape.

Critical weaknesses

(1)The paper claim that “observe that graph energy of paraphrase-induced logit differences adds modest held-out predictive value for soft-mask disagreement in two fusion heads.”. But, there is no evidence that this happens in the part segmentation task. Building the whole paper on the motivation that does not validate is no acceptable.

(2)The core mechanism of retaining the unfiltered base field and filtering only residuals has known precedents in graph neural networks, time-series denoising, and Transformer adaptation [1,2]. The papers’ main contribution is the observation. But the observation itself does not validate. Thus, this paper is a application rather than a research paper.

[1] Adaptive Initial Residual Connections for GNNs with Theoretical Guarantees, AAAI 2026

[2]Double-Diffusion: Balancing Speed, Accuracy, and Uncertainty inProbabilistic Forecasting for Urban Sensor Networks, Arxiv

(3)The Assumption of the paper is weird, test-independent backbone produces is separated into bϕb_{\phi} and aθa_{\theta}. Then, the most central part of this paper comes with Eq. 6. This is ok. But, the whole part of the derivation is built on the separation. We even do not know why this separation works. Why a simple separation could produce prompt-independent base field. This is a simple assumption but rather than a real problem.

(4)The motivation centers on prompt sensitivity, yet the primary controlled target metric is standard part-segmentation mIoU, rather than prompt-robustness metrics. Prompt-robustness results only appear in the top panel of Table 3. Even in this section, the paper fails to establish a mechanistic link between residual filtering and reduced prompt sensitivity. It does not verify whether filtered components correspond to prompt-introduced noise instead of authentic boundary signals, nor does it measure whether the discrepancy field (IN−H)b(I_N - H)b localizes at true boundaries. Consequently, the paper only demonstrates overall performance improvements but cannot explain why the gains occur, and lacks direct experimental validation supporting its core motivation.

Questions for authors

(1)This paper give the claim that even has not been validated, like “Filtering the complete logit can attenuate prompt-induced variation”, why? Does the logit really reflect the prompt-induced variation?

(2)The paper adopts normalized Dirichlet energy as a diagnostic metric without justifying why this statistic is appropriate, nor does it compare it against alternative graph-signal metrics. Meanwhile, it remains unclear whether the proposed framework is specifically tailored for part-segmentation tasks or enjoys general applicability. The idea is presented, yet its underlying rationale is not elaborated.

(3)This paper changes the symbol used in the paper through the paper. XX is defined as the point cloud. In the stage, it’s meaning is changed into the text promt xx. This leads to a very bad reading experience. And the similar happens in the whole paper, like aia_i and x+x^+. The meaning of this symbols needs to be guessed. This paper, from the very beginning contains many confused symbols.

2 comments
AuthorsOpenReview ↗

Response to Reviewer zcty

Purple highlighting in the revised PDF marks the construction of the shared base (Sections 3.2–3.3) and the boundary experiment (Section 4.5).

Critical Weaknesses

W1. Evidence for the observation in 3D part segmentation. We tested the diagnostic on pointwise outputs of two trained 3D part-segmentation heads. Each seed was evaluated on 1,200 equivalent-prompt pairs from 100 held-out objects. All pairs for an object stayed in the same regression fold. Adding normalized Dirichlet energy to predictors including change magnitude and confidence increased out-of-fold R² by 0.0618 for the pooled-prompt head and 0.0579 for the token–point head (Section 3.1, Table 9). This is a modest association with soft-mask disagreement. Q2 compares alternative diagnostics.

W2. Relationship to prior residual and diffusion methods. Section 2 now discusses both suggested papers. Adaptive Initial Residual Connections preserve initial node features during graph message passing. Double-Diffusion combines a graph forecasting prior with residual refinement and uses a spectral diagnostic to guide its design. Our contribution is to test a specific placement of an established filter: after text fusion, should it act on the entire pointwise prediction or only on the update for the requested part? Equivalent-prompt pairs measure wording sensitivity, and frozen placement controls isolate the effect of retaining the shared base.

W3. Why does the decomposition produce a query-independent base? The architecture specifies two heads trained jointly for segmentation. The base head takes only point features, so its output b is identical for every prompt on a fixed object. The query head also takes text. Subtracting its output for the fixed empty reference defines the query residual (Section 3.2). The empty reference sets the residual to zero for that input and cancels when two prompts are compared. Whether the learned base contains useful boundary detail is an empirical question, tested in Section 4.5.

W4. Prompt robustness as an outcome and evidence for the mechanism. The top panel of Table 2 directly evaluates the held-out paraphrases. Across five paired training seeds, GSSR reduces soft disagreement from 3.20 to 2.25 on the 100esoft100e_{\mathrm{soft}} scale and the point-flip rate from 2.85% to 1.95%. Worst-prompt mIoU increases from 44.60 to 48.80 (+4.20 percentage points). Coarse mIoU is the primary endpoint for the separate segmentation evaluation.

For identical frozen fields and the same filter H, residual-only and full-logit outputs for one prompt differ by (I − H)b. This restored component has a mean boundary enrichment of 1.51. Adding it back reduces error by 2.40 percentage points near ground-truth boundaries and by 0.20 elsewhere (Appendix C.3, Table 11). This identifies useful boundary information in the shared base under the frozen comparison.

Questions for Authors

Q1. Why do logit differences measure wording sensitivity, and how can full-logit filtering attenuate it? A logit is the pointwise score before the sigmoid. We fix the object, model, and target part and change only the wording, so subtracting the two logit fields measures the change in the model’s scores under that rewording. For the full-logit output H[b + r(x)], the shared base cancels under subtraction, giving H[r(x) − r(x′)]. Residual-only filtering gives the same difference for fixed fields and H (Eqs. 9–10). The low-pass filter attenuates components that vary rapidly on the graph. Probability disagreement after the sigmoid is measured separately in Table 2.

Q2. Why use normalized Dirichlet energy, how does it compare with other metrics, and what is the scope? Normalization lets us ask how sharply a score change varies between neighboring points relative to its magnitude. Under the same object-grouped evaluation, Table 9 gives the following increments over the regression using nonspectral covariates.

Increase in out-of-fold R²Pooled-prompt headToken–point head
Normalized Dirichlet energy0.06180.0579
Total variation0.03970.0385
High-pass summary0.04620.0442
Shuffled graph−0.0005−0.0001

Dirichlet energy has the largest point estimate in both heads, and shuffling the graph brings the increment close to zero. These results concern text-prompted 3D part segmentation with two PTv3 fusion heads.

Q3. Which symbols denote the point cloud, point attributes, and text? Section 3.1 uses X\mathcal{X} for the point cloud, pip_i for the coordinates of point ii, and cic_i for its input attributes. The text prompt is xx, and x+x^+ is another wording that names the same part and shares its target mask. Section 3.2 uses aθ(F,x)a_\theta(F,x) for the query-head output. Appendix A.2 uses t(x)t(x) for the pooled text embedding.

Thank you to the authors for their reply and for the reminder regarding AC. From the authors' response, I still find the overall readability of the paper to be quite poor: logically, the motivation in the earlier sections depends on experiments presented later, which seriously undermines readability. The novelty is also incremental. Furthermore, the late reply made it difficult for the reviewer to properly reassess the overall quality of the paper. I therefore maintain my score.

Reviewer_SV76OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies mask drift in 3D part segmentation under semantically equivalent prompts. GSSR decomposes the final logits into a query-independent base and an empty-string-anchored query residual, then applies a fixed graph low-pass filter only to the residual. A held-out diagnostic finds that normalized Dirichlet energy provides a modest additional prediction of prompt disagreement. Across five paired seeds, GSSR reduces soft disagreement from 3.20 to 2.25, improves worst-prompt mIoU from 44.60 to 48.80, and raises 3DCoMPaT-Coarse mIoU from 47.80 to 51.17. Frozen, retrained, and alternative-head controls examine the filter-placement choice.

Critical strengths
  • Spatial drift under semantically equivalent prompts is a clear and practically meaningful reliability problem.
  • The control suite is extensive, including Direct, frozen same-residual, norm-matched scalar, high-pass, shuffled-graph, full-logit, Laplacian-regularization, and alternative-fusion-head controls.
  • Comparisons with external methods are explicitly labeled descriptive rather than incorrectly interpreted as paired effects.
Critical weaknesses

Here I have several main concerns:

  1. The central full-logit control is not sufficiently fair. Full logits and residuals have different scales and spectra, so there is no reason to expect the same k/eta/m configuration to be optimal for both. The two placements require independent searches with identical tuning budgets.
  2. The frozen control involves an unauditable conditional selection. Panel A reports only the common intersection for which all norm-matching denominators are valid and scale <= 10, but does not report how many tuples were excluded. This leaves open the possibility of selection bias against difficult or high-frequency examples.
  3. The prompt-robustness dataset is insufficiently documented. The paper does not report the numbers of objects, categories, prompts per group, the source of prompt generation, or human validation of semantic equivalence in the confirmation split.

Beisdes, I hope the belowing issues can be also solved:

  • Part of the reduction in Dirichlet energy follows directly from the low-pass operator's design; changes in accuracy and boundary quality are the genuinely diagnostic outcomes.
  • The increase in BIoU16 from 7.60 to 16.11 is large, but no seed-level variance or paired interval is provided. Defined-case counts are also missing for the external methods.
  • The paper does not directly test whether (I-H)b is concentrated near ground-truth boundaries. Thus, the claim that base preservation retains boundaries remains a mechanistic hypothesis.
  • Absent prompts are not used during training. The appendix reports only the fusion head's FPR and does not provide a paired Direct-versus-GSSR comparison.
  • The second experiment still uses the same PTv3 backbone with a different fusion head, so generalization across backbones remains unknown.
Questions for authors

A few key experiments and clarifications would help me re-evaluate the paper:

Fairness in baseline comparison: Residuals and full logits behave quite differently. Could you run an independent search for k/eta/m on full-logit filtering using the same tuning budget, and report paired results across seeds?

Prompt benchmark transparency: The confirmation split needs more documentation. What are the counts for objects, categories, part groups, and prompts/pairs? Also, how were they created and cross-checked for semantic equivalence? Showing this dataset is solid would greatly support the robustness claims.

Selection bias in Panel A: How many tuples were actually dropped when forming the common-valid subset? Please report numbers for both the full data and the intersection. If the vast majority of samples are retained, my worry about selection bias will be mostly cleared up.

Open-vocabulary evaluation: How does GSSR affect the absent-query FPR in a paired setup, and does the performance stay consistent with different anchors?

Sanity check on the mechanism: To back up the claim that base preservation keeps boundary information intact, is there direct evidence showing that (I-H)b actually aligns with ground-truth boundaries?

1 comment
AuthorsOpenReview ↗

Response to Reviewer SV76

In the revised PDF, orange highlights the full-set comparison (Table 3) and the independently tuned filters (Table 4).

Filter placement under two controls. Giving full-logit filtering its own search changes its preferred setting. With 27 configurations per placement, three search seeds per configuration, and 40,000 updates per run, the design set selects (k, η, m) = (16, 0.10, 4) for full-logit filtering and (16, 0.25, 4) for GSSR (Appendix B.5). The reported segmentation and prompt-consistency scores favor residual-only filtering.

After independent tuning (Table 4)Full-logitResidual-only
Coarse mIoU ↑49.7351.17
Boundary F1 ↑42.7443.80
Strict boundary IoU ↑12.4916.11
Soft disagreement ↓2.662.25
Worst-prompt mIoU (%) ↑47.0048.80

Table 4 uses the main-study result for the selected residual configuration. The differences of means are 1.44 percentage points for Coarse mIoU and 1.80 percentage points for worst-prompt mIoU (48.80 versus 47.00). Seed pairing for the final Full–Residual comparison is unverified. For the separate five-seed GSSR–Direct comparison, the paired gain is 3.37 points (descriptive 95% interval 2.57 to 4.17).

Does the frozen result depend on exclusions? Table 3 evaluates all 1,000 valid queries per seed. With the same five frozen checkpoints and the same filter, residual-only improves on full-logit by 1.41 mIoU points and 2.96 boundary-F1 points on the full set. The exclusion rule applies to the auxiliary norm-matched controls. They retain 4,839 of 5,000 seed–query observations (96.78%) and exclude 161. Median residual energy is 0.18 among excluded observations and 0.30 among retained observations. Tables 6–8 give the counts and reasons by seed.

Prompt construction and review. The confirmation split contains 100 objects from five categories, with no object shared across splits. Two reviewers independently checked all 1,000 distinct candidate prompt records drawn from manual templates. Table 5 records 80 initial disagreements and 190 semantic rejections. The final 200 object–part–level groups each contain four accepted prompts, yielding 800 records and 1,200 within-group pairs.

Does the preserved base help at boundaries? Restoring the base component removed by full-logit filtering corrects 3.70% of points in the ground-truth boundary band and introduces errors on 1.30%. Boundary error falls from 31.20% to 28.80%, compared with a 0.20-point reduction outside the band (Table 11). The band contains 1.51 times the share of component energy expected from its share of points. These measurements track errors against ground truth as well as the spatial effect of smoothing. For BIoU₁₆, Table 2 reports means of 7.60 and 16.11, and Appendix B.6 reports 920 eligible queries and predicted-empty counts per seed. The seed-level BIoU variance, paired interval, and external methods’ defined-case counts are unavailable.

Absent queries, anchors, and backbone. GSSR’s point FPR is 9.38% on 360 absent queries from 120 objects at threshold 0.5. This set has no paired Direct estimate. In the frozen anchor test, changing the empty reference to “object” or “shape” increases mIoU by 0.33 or 0.28 points and lowers FPR by 0.35 or 0.53 points, respectively (Table 12). The second fusion head also improves on Direct, by 0.9 points in four-dataset macro mIoU (Table 13). Both heads use PTv3, which sets the scope of the replication.

Reviewer_dttxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper aims to address the difference between 3D part segmentation predictions induced by prompts that describe the same part in different ways. It is found that normalized Dirichlet energy can serve as a relatively useful indicator. Based on this, the authors propose to use graph filters to process the final query-dependent logit residual, which leads to consistent improvement.

Critical strengths
  1. The research question is clear and real. The research is conducted following the motivation.
  2. The experimental results show consistent improvement.
Critical weaknesses
  1. The writing is hard to understand. I have to admit that I am not very familiar with 3D part segmentation, but this also gives me a chance to evaluate if this paper is accessible to newcomers to this field. From my perspective, the notations in this manuscript are only partially explained, which appears to be very wierd, and there are many unexplained terminologies. Furthermore, this paper focuses much on details and especially favors mathematical formulation, but this is done without enough high-level interpretation. Therefore, I am concerned if a potential reader can learn much from this paper.
  2. The ideal organization would be to first show a key observation and then base the proposed method on it. However, the authors choose to place the observation in a later section, together with the experiment that demonstrates the effectiveness of the method. This breaks the flow.
Questions for authors
  1. Please provide some high-level interpretation of the story: given the problem that segmentation prediction varies when the prompt is paraphrased, what is observed to be the root of this phenomenon? Then, how is the method naturally derived from this observation? Also show how such interpretation is able to be derived directly from the manuscript.
1 comment
AuthorsOpenReview ↗

Response to Reviewer dttx

In the revised PDF, blue highlights the observation and its link to GSSR (Section 3.1).

Why does rewording change the mask? Take two prompts for the same eyeglass frame, “temple of eyeglasses” and “eyeglasses temple.” Both refer to the same points. Yet changing the wording can raise the predicted score at some points and lower it at others. Points that cross the decision threshold then enter or leave the mask. Our observation is that where the scores change matters. Even after accounting for the size of the change, its variation across neighboring points helps predict mask disagreement on held-out objects in both fusion heads (Section 3.1).

Why smooth only the query update? Smoothing can reduce sharp changes caused by rewording, but neighboring points can also lie on opposite sides of a real part boundary. GSSR therefore applies smoothing to the update produced by the prompt and keeps the score shared by all prompts for that object. Across five paired training seeds, soft disagreement falls from 3.20 to 2.25, a reduction of about 30%, while worst-prompt mIoU rises from 44.60 to 48.80, a gain of 4.20 percentage points (Table 2). The masks agree more across descriptions, and even the least successful description gives a more accurate mask on average.

The revision presents the observation in Section 3.1 before the design in Sections 3.2–3.3. Figure 1 follows the same sequence. We define the point attributes, prompt pair, and pointwise score where they first appear and give the spectral derivation in Appendix A.3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.