cs.NINov 5, 2026

When Does Blockwise Additivity Break? Diagnosing and Repairing the Additive Assumption in Distillation-Based Architecture Search

Authors: Yuefeng Xu, Jun Yu

Organizations: Niigata University

Abstract

Blockwise distillation methods for architecture search, such as DNA, DONNA, LANA and Puzzle, score each candidate replacement block in isolation and combine the scores additively. We ask at which budgets this additivity assumption fails, whether fine-tuning repairs the failure, what diagnoses it, and how it can be repaired cheaply. On DeiT-S, DeiT-B, Swin-T and Qwen2.5-0.5B with seven latency budgets each, an additive integer program trails a pairwise interaction-aware program after fine-tuning by at most 1.6 top-1 points on ImageNet-1k at the mildest budgets, but by up to 17.5 points, and a factor of 1.9 in perplexity, at aggressive ones. After 10,000 fine-tuning steps 3.3-9.9 points remain on DeiT, whereas on Swin-T the gap shrinks to at most about 1.8 points within 500 steps. In a learning-rate sweep with the standard 2,000 steps, 6.0-10.7 points remain on DeiT when each method's rate is chosen by its evaluation-set accuracy. An intervention that keeps the block library and only forbids more than two removals in any three adjacent layers shrinks the DeiT gap from 5.8-17.5 to 0.7-5.2 points, mostly by improving the additive solution: adjacent removals are a major failure pattern. Splitting the additivity error into a metric and a functional part, relative to additive logit changes, indicates where a cheap logit-sum predictor can be trusted. A hybrid search that ranks moves with this predictor and measures only the best needs no pair measurements and, after Holm adjustment, is significantly better than the pairwise program in 10 of 14 ImageNet-1k settings, with no statistically significant losses. An ablation shows that its ranking helps except at the most aggressive DeiT-B budget, where the predictor misleads and random ranking or ten-times-costlier evolution does better.

Figures & tables

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

May 16, 2026

HEED: Density-Weighted Residual Alignment for Hybrid Vision-Language Model Distillation

Distilling vision-language models into faster hybrid architectures, such as 3:1 Mamba-2/attention mixes, is now standard practice for making inference efficient. Aggregate benchmarks suggest that this works but they hide selective failures. When we distill Qwen3-VL-8B-Instruct into a 3:1 Mamba-2/attention hybrid, student model stays within 2 points of the teacher across visual reasoning benchmarks like MMStar, MMBench, and MMMU-Pro, while dropping 13 points on optical-character-recognition and document tasks. The student can still understand the scene but loses the fine-grained text needed to answer. We localize much of the failure to a specific kind of position. In a high-resolution image, most patches are sky, wall, or smooth texture, while a small fraction carries text, edges, object boundaries, or other local details. In a token-level diagnostic, the top 10% highest-density patches have 3.6x larger residual drift than the bottom 10% lowest-density patches and 3.5x larger teacher-masking answer contribution. Uniform weighting devotes many loss terms to low-information background patches, whereas sparse answer bearing patches receive no special protection. The required intervention is minimal: we replace uniform residual alignment with density-weighted residual alignment, using patch self-dissimilarity as a training-free proxy for position importance. We call this HEED. Compared with normal end-to-end distillation, HEED increases performance by 8.7 points on OCRBench v2 and 5.13 points on a 10-benchmark average. The gain is realized on different teacher models. After standard post-training, the student reaches teacher-level performance on the 10-benchmark average. HEED adds no parameters and no inference-time cost. Hence, the student maintains the hybrid architecture's 4.12x throughput and 68% memory saving at 128k context. Code is available at https://github.com/jha-lab/HEED.
Nov 2, 2026

Efficient Hybrid Distillation: Synergizing Score and Adversarial Objectives for One-Step Diffusion

Numerous distillation methods have been developed to accelerate diffusion models. Recent research indicates that hybrid strategies---blending different paradigms---yield superior results compared to single-method approaches. However, current works predominantly rely on a simple combination of objectives, leaving the question of how to effectively synergize them largely underexplored. In this paper, we specifically investigate the synergy between Score Distillation and Adversarial Distillation. We reveal that their operating zones are distinct and complementary: Score Distillation is effective in high-noise regimes for capturing the global distribution yet lacks fine-grained supervision in low-noise regions. Conversely, Adversarial Distillation is effective in low-noise regions but is prone to instability and suffers from discriminator failure when applied to high-noise intervals. Leveraging this insight, we propose Efficient Adversarial Score Distillation (EASD), a framework that synergizes these objectives via a regime-adaptive sampling strategy, directing each objective to focus on its effective interval. We further apply this paradigm to DiT-based Flow Matching models. Empirically, our method achieves a promising one-step FID of 1.39 on ImageNet 256×\times256 using the SiT-XL/2+REPA model. This work provides a concrete guideline for maximizing the potential of hybrid distillation. Our code and model are available at https://github.com/pf0607/EASD .
Jul 3, 2026

A Decomposable Probe for Few-Step Diffusion Models: Prompt, Latent, and Score Selectivity across Backbone Families and Distillation Paradigms

Few-step distilled diffusion students cut text-to-image inference from ~50 to 1-8 network evaluations, but the quality gap is usually summarised by a single FID/CLIP scalar that cannot say which axis of the conditioning response changed, nor whether a behaviour comes from the architecture, the distillation objective, or simply from being a diffusion model. We replace the scalar with a decomposable probe that injects controlled perturbations along three layers (prompt encoder, denoiser input, denoiser output) under three modes (mean, variance, scale) and six strengths, reporting a bootstrap-median Bures W2^2 selectivity ratio on Inception features. Under a single matched estimator across 23 models -- five teachers and 18 distilled students spanning five backbone families (SDXL, SD1.5, SD3.5, PixArt-alpha, FLUX), three architecture classes (UNet, DiT, MMDiT), and five distillation paradigms -- the three layers read three empirically separable factors: the prompt layer is a universal prompt-mean response (a sanity channel, not a discriminator), the latent layer reads the prediction type, and the score layer reads the distillation objective. Our main result: within this sweep, the latent layer is a near-binary detector of rectified-flow backbones. Its ratio exceeds 1 across a sustained low-to-mid band only for rectified-flow models (SD3.5, FLUX); no epsilon-prediction model qualifies. A matched epsilon-prediction control (PixArt-alpha) rules out wide-T5 conditioning, and the fingerprint survives adversarial (ADD) distillation as both teacher and student. Two secondary score-layer findings hold under narrower scopes: a canonical 4-step ADD-vs-rest contrast on the UNet families with a non-ADD baseline, and a CI-separated trajectory-rollout early-strength score spike on both UNet and DiT. All ratios are CI-citable under one estimator; we release the per-cell tables and the estimator.