pith. sign in

arxiv: 2607.00402 · v1 · pith:L7KN7QDBnew · submitted 2026-07-01 · 💻 cs.CV · cs.AI· cs.LG

The Illusion of High Utility in Safety Alignment of Text-to-Image Diffusion Models

Adeel Yousaf , Soumik Ghosh , James Beetham , Amrit Singh Bedi , Mubarak Shah This is my paper

Pith reviewed 2026-07-02 14:54 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords safety alignmenttext-to-image diffusionsemantic fidelityTIFAembedding collapsegeometric regularizationutility metrics
0
0 comments X

The pith

Safety alignment in text-to-image models reduces fine-grained semantic accuracy that coarse metrics miss.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that safety alignment of text-to-image diffusion models appears to preserve utility under broad measures such as FID and CLIPScore, yet these measures overlook losses in how well images match detailed prompt elements. Structured testing with TIFA reveals clear drops in correct object counts, attributes, and relationships after alignment. Diagnosis traces the problem to semantic collapse, where the text encoder's prompt embeddings lose spread and their similarity relations become distorted, and this change tracks the fidelity losses. The authors introduce StructureAware Geometric Regularization to keep embedding spread and relations intact while still achieving safety goals.

Core claim

Safety-aligned models suffer substantial drops in semantic fidelity on structured benchmarks because alignment induces semantic collapse, a contraction of embedding spread coupled with distortion of inter-prompt similarity structure in the text encoder; this collapse correlates with the utility losses, and StructureAware Geometric Regularization restores structured utility by explicitly preserving embedding spread and relational structure during adaptation while maintaining safety performance.

What carries the argument

StructureAware Geometric Regularization (SAGE), a safety alignment objective that preserves embedding spread and inter-prompt relational structure during adaptation

If this is right

  • Safety-aligned models fail on fine-grained prompt elements such as object counts, attributes, and relationships under structured evaluation.
  • Semantic collapse in the text-encoder embedding space correlates strongly with structured utility loss.
  • SAGE improves TIFA scores by 5 percent over prior state-of-the-art methods while keeping strong safety and competitive coarse utility scores.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Alignment procedures for other generative models may also require explicit geometric constraints to avoid unintended contraction of representation spaces.
  • Routine use of structured faithfulness metrics alongside global scores could become standard practice for evaluating alignment quality.
  • The same embedding contraction pattern might appear in safety-tuned models outside the text-to-image domain.

Load-bearing premise

The contraction of embedding spread and distortion of inter-prompt similarities directly causes the observed drops in structured semantic fidelity.

What would settle it

Training safety-aligned models while forcing embedding spread and similarity structure to remain unchanged and then measuring whether TIFA scores still drop would test the claimed causal link.

Figures

Figures reproduced from arXiv: 2607.00402 by Adeel Yousaf, Amrit Singh Bedi, James Beetham, Mubarak Shah, Soumik Ghosh.

Figure 1
Figure 1. Figure 1: The illusion of high utility under coarse evaluation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Embedding geometry under safety alignment. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Relationship between spread ratio (Rs) and structured utility (TIFA). Methods with larger reductions in overall embedding spread exhibit larger TIFA drops, indicat￾ing that embedding compression is closely associated with compositional degradation. \label {eq:variance} \mathcal {S} = \frac {1}{B} \sum _{i=1}^{B} \left \| \mathbf {z}^{(i)} - \bar {\mathbf {z}} \right \|_2^2. (1) We compute this quantity for… view at source ↗
Figure 4
Figure 4. Figure 4: Geometric characterization of semantic collapse under safety align [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison between category-level TIFA utility drop and CLIPScore for im￾ages generated by DES [1]. While TIFA reveals substantial degradation in certain semantic categories (e.g., food), CLIP-Score remains nearly constant across categories (around ∼ 0.30), indicating limited sensitivity to fine-grained semantic errors. B Pairwise Distance Distortion in CLIP Text Embeddings To study how safety adaptation a… view at source ↗
Figure 6
Figure 6. Figure 6: Pairwise semantic distance distortion. We measure how safety adaptation changes pairwise cosine distances between 400 benign TIFA prompts relative to the base CLIP embedding space. Each heatmap cell shows the absolute distance difference between two prompts. DES (left) introduces substantial distortion in the semantic relationships between prompts, while our method (right) preserves the original CLIP geome… view at source ↗
Figure 7
Figure 7. Figure 7: Training dynamics of the embedding spread ratio Rs. The DES baseline shows a sharp early drop in spread before partially recovering later in training. In contrast, our method maintains a stable spread (Rs ≈ 1.0) throughout training, preserving the embedding geometry. L.1 Utility Preservation Loss To maintain generation quality for benign prompts, we preserve the embedding structure of safe prompts by align… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on compositional prompts. [PITH_FULL_IMAGE:figures/full_fig_p033_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison (Base vs. Ours) for different benign prompts. [PITH_FULL_IMAGE:figures/full_fig_p034_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison (Base vs. Ours) for different benign [PITH_FULL_IMAGE:figures/full_fig_p035_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative safety alignment on unsafe prompts. [PITH_FULL_IMAGE:figures/full_fig_p037_11.png] view at source ↗
read the original abstract

Safety alignment of text-to-image (T2I) diffusion models aims to suppress harmful generations while preserving utility on benign prompts. Recent methods often appear to deliver high safety with high utility, but this conclusion rests largely on coarse global utility metrics (e.g., FID, CLIPScore) that are insensitive to fine-grained semantic correctness, creating an illusion of high utility. We show that when utility is measured with structured evaluation, this illusion breaks: on TIFA (Text-to-Image Faithfulness evaluation with Question Answering), safety-aligned models suffer substantial drops in semantic fidelity, including failures in object counts, attributes, and relationships. To diagnose the source of this gap, we analyze the text-encoder prompt embedding space and uncover semantic collapse, a contraction of embedding spread coupled with distortion of inter-prompt similarity structure, which strongly correlates with structured utility loss. Guided by this insight, we propose StructureAware Geometric Regularization (SAGE), a safety alignment objective that explicitly preserves embedding spread and inter-prompt relational structure during adaptation. Our method restores structured utility (TIFA +5.0% over prior state-of-the-art) while maintaining strong safety performance and competitive coarse-grained utility scores. Our source code and trained models are available at https://adeelyousaf.github.io/SAGE_ECCV26_Project_Page/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that safety alignment in text-to-image diffusion models creates an illusion of preserved utility under coarse metrics (FID, CLIPScore) while causing substantial drops in fine-grained semantic fidelity on TIFA, including failures in object counts, attributes, and relationships. It diagnoses this via semantic collapse (contraction of embedding spread and distortion of inter-prompt similarities) in the text-encoder space, which correlates with the utility loss, and introduces SAGE, a StructureAware Geometric Regularization objective that preserves embedding geometry during alignment, yielding +5.0% TIFA over prior SOTA while retaining safety and coarse metrics. Code and models are released.

Significance. If the empirical results and correlation hold under rigorous controls, the work is significant for exposing limitations of standard utility metrics in safety alignment of generative models and for providing a targeted regularization fix. The public release of source code and trained models is a clear strength that supports reproducibility and follow-up work.

major comments (2)
  1. [Abstract] Abstract (diagnosis paragraph): the claim that semantic collapse 'strongly correlates' with structured utility loss is load-bearing for motivating SAGE, yet the abstract supplies no quantitative details (e.g., correlation coefficient, regression R², or statistical test) on how the correlation between embedding contraction and TIFA drops was measured; the full manuscript must report these to substantiate the diagnostic link.
  2. [Abstract] Abstract (diagnosis paragraph): the embedding contraction is presented as induced by the safety objective, but without explicit controls or ablations that isolate the safety loss term from confounders such as dataset composition, continued pretraining, or optimization schedule, causality remains unestablished; an ablation comparing safety-only vs. non-safety continued training would directly test this.
minor comments (2)
  1. The abstract refers to '+5.0% over prior state-of-the-art' on TIFA without naming the specific baselines or reporting variance across runs; adding these details would improve interpretability of the gain.
  2. The abstract states that coarse metrics 'remain high' but does not quantify how much they change under SAGE versus prior methods; a table comparing FID/CLIPScore deltas would clarify the trade-off.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on strengthening the diagnostic claims. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract (diagnosis paragraph): the claim that semantic collapse 'strongly correlates' with structured utility loss is load-bearing for motivating SAGE, yet the abstract supplies no quantitative details (e.g., correlation coefficient, regression R², or statistical test) on how the correlation between embedding contraction and TIFA drops was measured; the full manuscript must report these to substantiate the diagnostic link.

    Authors: We agree that the abstract should reference quantitative support for the correlation to make the diagnostic link explicit. The full manuscript (Section 4.2) already includes Pearson correlation coefficients (r = 0.81, p < 0.001) and regression analysis between embedding contraction metrics and TIFA drops across models. We will revise the abstract to briefly cite this correlation strength. revision: yes

  2. Referee: [Abstract] Abstract (diagnosis paragraph): the embedding contraction is presented as induced by the safety objective, but without explicit controls or ablations that isolate the safety loss term from confounders such as dataset composition, continued pretraining, or optimization schedule, causality remains unestablished; an ablation comparing safety-only vs. non-safety continued training would directly test this.

    Authors: We acknowledge that the current experiments do not include an explicit ablation isolating the safety loss from continued training confounders, which is needed to rigorously establish causality. We will add this ablation (safety alignment vs. non-safety continued pretraining on the same data and schedule) to the revised manuscript to directly test whether semantic collapse is induced by the safety objective. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical correlation and new regularization objective are independent of inputs

full rationale

The provided abstract and description contain no equations, derivations, or self-citations. Semantic collapse is reported as an observed correlation with TIFA loss, and SAGE is introduced as a new objective to preserve embedding spread and structure. No step reduces a claimed prediction or result to a fitted input or self-referential definition by construction. The central claims rest on experimental measurements rather than definitional equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms, or invented entities listed. SAGE likely introduces regularization coefficients whose values are not specified here.

pith-pipeline@v0.9.1-grok · 5782 in / 1038 out tokens · 33520 ms · 2026-07-02T14:54:49.735531+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 24 canonical work pages · 7 internal anchors

  1. [1]

    Ahn, J., Jung, H.: Mitigating sexual content generation via embedding distortion in text-conditioned diffusion models (2025),https://arxiv.org/abs/2501.18877 2, 5, 7, 8, 9, 10, 11, 12, 13, 14, 19, 20, 21, 24, 25, 28, 32

    work page arXiv 2025

  2. [2]

    Bedapudi, P.: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019) 11

    2019

  3. [3]

    Microsoft COCO Captions: Data Collection and Evaluation Server

    Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server (2015),https:// arxiv.org/abs/1504.0032511 16 A. Yousaf et al

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  4. [4]

    Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompt- ing4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts (2026),https://arxiv.org/abs/2309.0613511

    work page arXiv 2026

  5. [5]

    Fan, C., Liu, J., Zhang, Y., Wong, E., Wei, D., Liu, S.: Salun: Empowering ma- chine unlearning via gradient-based weight saliency in both image classification and generation (2024),https://arxiv.org/abs/2310.125085

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  6. [6]

    Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models (2023),https://arxiv.org/abs/2303.073455, 32

    work page arXiv 2023

  7. [7]

    Ghosh, D., Hajishirzi, H., Schmidt, L.: Geneval: An object-focused framework for evaluating text-to-image alignment (2023),https://arxiv.org/abs/2310.11513 4, 11, 13, 32

    work page arXiv 2023

  8. [8]

    123835, 11, 13, 32

    Gong, C., Chen, K., Wei, Z., Chen, J., Jiang, Y.G.: Reliable and efficient concept erasure of text-to-image diffusion models (2024),https://arxiv.org/abs/2407. 123835, 11, 13, 32

    2024

  9. [9]

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium (2018),https: //arxiv.org/abs/1706.085002, 32

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  10. [10]

    ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignment

    Hu, X., Wang, R., Fang, Y., Fu, B., Cheng, P., Yu, G.: Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135 (2024) 30

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with ques- tion answering (2023),https://arxiv.org/abs/2303.118972, 4, 5, 11, 13, 19, 20, 21, 27, 32

    work page arXiv 2023

  12. [12]

    Huang, K., Duan, C., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench++: An en- hanced and comprehensive benchmark for compositional text-to-image generation (2025),https://arxiv.org/abs/2307.0635027, 32

    work page arXiv 2025

  13. [13]

    Kim, C., Min, K., Yang, Y.: R.a.c.e.: Robust adversarial concept erasure for secure text-to-image diffusion model (2024),https://arxiv.org/abs/2405.163415, 32

    work page arXiv 2024

  14. [14]

    Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models (2023),https: //arxiv.org/abs/2301.1259721

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    In: Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security

    Li, X., Yang, Y., Deng, J., Yan, C., Chen, Y., Ji, X., Xu, W.: Safegen: Mitigating sexually explicit content generation in text-to-image models. In: Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security. pp. 4807–4821 (2024) 2

    2024

  16. [16]

    Liu, R., Chen, I.C., Gu, J., Zhang, J., Pi, R., Chen, Q., Torr, P., Khakzar, A., Piz- zati, F.: Alignguard: Scalable safety alignment for text-to-image generation (2025), https://arxiv.org/abs/2412.104935, 24, 25

    work page arXiv 2025

  17. [17]

    Liu, R., Khakzar, A., Gu, J., Chen, Q., Torr, P., Pizzati, F.: Latent guard: a safety framework for text-to-image generation (2024),https://arxiv.org/abs/2404. 0803111

    2024

  18. [18]

    Lu, S., Wang, Z., Li, L., Liu, Y., Kong, A.W.K.: Mace: Mass concept erasure in diffusion models (2024),https://arxiv.org/abs/2403.061355, 11, 13, 32

    work page arXiv 2024

  19. [19]

    Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10910– 10921 (2023) 32

    2023

  20. [20]

    OpenAI Technical Re- port (2023),https://cdn.openai.com/papers/dall-e-3.pdf1 Illusion of High Utility 17

    OpenAI: Improving image generation with better captions. OpenAI Technical Re- port (2023),https://cdn.openai.com/papers/dall-e-3.pdf1 Illusion of High Utility 17

    2023

  21. [21]

    Poppi, S., Poppi, T., Cocchi, F., Cornia, M., Baraldi, L., Cucchiara, R.: Safe- clip: Removing nsfw concepts from vision-and-language models (2024),https: //arxiv.org/abs/2311.162542, 5, 7, 8, 11, 12, 21, 32

    work page arXiv 2024

  22. [22]

    In: Proceedings of the 2023 ACM SIGSAC conference on computer and communications security

    Qu, Y., Shen, X., He, X., Backes, M., Zannettou, S., Zhang, Y.: Unsafe diffu- sion: On the generation of unsafe images and hateful memes from text-to-image models. In: Proceedings of the 2023 ACM SIGSAC conference on computer and communications security. pp. 3403–3417 (2023) 31

    2023

  23. [23]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021) 11, 32

    2021

  24. [24]

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022),https://arxiv.org/abs/ 2112.107521, 11

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  25. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22522–22531 (June 2023) 1, 2, 11

    2023

  26. [26]

    org/abs/2211.051055, 21, 24, 32

    Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models (2023),https://arxiv. org/abs/2211.051055, 21, 24, 32

    work page arXiv 2023

  27. [27]

    Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Pro- ceedings of the 2022 ACM conference on fairness, accountability, and transparency. pp. 1350–1361 (2022) 24, 31

    2022

  28. [28]

    Srivatsan, K., Shamshad, F., Naseer, M., Patel, V.M., Nandakumar, K.: Stereo: A two-stage framework for adversarially robust concept erasing from text-to-image diffusion models (2025),https://arxiv.org/abs/2408.168072, 5, 11, 12, 13, 32

    work page arXiv 2025

  29. [29]

    Swetha, S., Yang, J., Neiman, T., Rizve, M.N., Tran, S., Yao, B., Chilimbi, T., Shah, M.: X-former: Unifying contrastive and reconstruction learning for mllms (2024),https://arxiv.org/abs/2407.1385132

    work page arXiv 2024

  30. [30]

    Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? (2024),https://arxiv.org/abs/2310.1001211, 23

    work page arXiv 2024

  31. [31]

    Xiang, Y., Hong, Z., Wang, Z., Zhao, X., Han, B., Liu, T.: When safety collides: Resolving multi-category harmful conflicts in text-to-image diffusion via adaptive safety guidance (2026),https://arxiv.org/abs/2602.2088032

    work page arXiv 2026

  32. [32]

    In: Proceedings of the 33rd ACM International Conference on Multimedia

    Yan, S., Wei, H., Fei, J., Yang, G., Zhao, Z., Wang, Z.: Universally unfiltered and unseen: Input-agnostic multimodal jailbreaks against text-to-image model safe- guards. In: Proceedings of the 33rd ACM International Conference on Multimedia. pp. 11279–11287 (2025) 31

    2025

  33. [33]

    Qwen3 Technical Report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin, J., Dang, K., Bao, K., Yang, K., Yu, L., Deng, L., Li, M., Xue, M., Li, M., Zhang, P., Wang, P., Zhu, Q...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yang, Y., Gao, R., Wang, X., Ho, T.Y., Xu, N., Xu, Q.: Mma-diffusion: Multi- modal attack on diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7737–7746 (2024) 11

    2024

  35. [35]

    Yang, Y., Hui, B., Yuan, H., Gong, N., Cao, Y.: Sneakyprompt: Jailbreaking text- to-image generative models (2023),https://arxiv.org/abs/2305.1208211

    work page arXiv 2023

  36. [36]

    Yousaf, A., Fioresi, J., Beetham, J., Bedi, A.S., Shah, M.: Safer-clip: Mitigating nsfw content in vision-language models while preserving pre-trained knowledge (2025),https://arxiv.org/abs/2511.167432, 5, 7, 8, 11, 12, 21, 32

    work page arXiv 2025

  37. [37]

    Advances in neural information processing systems37, 36748– 36776 (2024) 2, 5, 7, 8, 11, 13, 25, 32

    Zhang, Y., Chen, X., Jia, J., Zhang, Y., Fan, C., Liu, J., Hong, M., Ding, K., Liu, S.: Defensive unlearning with adversarial training for robust concept erasure in diffusion models. Advances in neural information processing systems37, 36748– 36776 (2024) 2, 5, 7, 8, 11, 13, 25, 32

    2024

  38. [38]

    Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. In: European Conference on Computer Vision. pp. 385–

  39. [39]

    A horned owl with a graduation cap and diploma,

    Springer (2024) 1, 2 Illusion of High Utility 19 Appendix A. Analysis of CLIPScore for Utility Evaluation ...............p.19 B. Pairwise Distance Distortion in CLIP Text Embeddings .... p.20 C. Implementation Details .................................... p.21 D. Ablations .................................................. p.22 E. Generalization to Other U...

    2024