Lost in Interpolation: Why Predictive Feedback Fails in Diffusion Language Models
Gaurav Kumar Nayak, Lavanya Nigam, Ishaan Bansal, Aryan Sood, Vidit Aggarwal
dlm-study
Research metadataShow detailsHide details
- Affiliations
- Not available
- Published
- 2026-08-06
- Processed
- 8/30/2026, 11:59:09 AM
- Analysis model
- Not available
- Analysis status
- partial
- Local PDF artifact
- papers/pdf/2026/lost-in-interpolation-why-predictive-feedback-fails-in-diffusion-language-models.pdf
Abstract
Soft-masking accelerates the convergence of Masked Diffusion Language Models (MDLMs). Existing formulations build this blend with linear interpolation (LERP) in the raw embedding space, which implicitly treats that space as Euclidean. We analyze the embedding space of MDLMs and find that the mask and predicted-token embeddings maintain a near-constant angle of (\approx 73^\circ) throughout training, while embedding norms remain essentially flat across vocabulary-frequency rank. These indicate a hyperspherical geometry, for which LERP is the wrong interpolation primitive. We introduce Spherical Soft-Masking (S-SM), a drop-in replacement that aggregates the top-(k) predictions with a Fr'echet mean on the hypersphere and blends this mean with the mask direction using spherical linear interpolation (SLERP), then restores the native mask norm. We evaluate S-SM on continued pre-training of a released 169M-parameter MDLM checkpoint across a wide range of inference-time step budgets, SLERP feedback avoids the training degradation that LERP feedback induces and delivers MAUVE gains of up to 2x over the vanilla MDLM baseline and 27.5-56.1% over TopK/LERP at various sampling budgets, alongside consistently lower generative perplexity (16.9-19.6% over the baseline), while leaving output entropy and convergence essentially unchanged.
Summary
이 논문은 MDLM(Masked Diffusion Language Model)의 soft-masking이 왜 때로는 잘 안 되는지, 원인을 “임베딩 공간의 기하”에서 찾습니다. 기존 soft-masking은 [MASK] 임베딩과 예측 토큰 임베딩을 직선으로 섞는 LERP를 씁니다. 그런데 저자들은 MDLM 임베딩이 평평한 유클리드 공간이라기보다 거의 같은 반지름의 초구면 위에 놓여 있다고 봅니다.