ICML
2026
REPRO
SPLIT-VLM: Attention-Free Vision-Token Dropping [ A Reproduction ]
Reproducing salience-guided partitioning for importance-aware token dropping in VLMs (OpenReview Elm4TdaXi0)
Artifacts & Code
1The Paper & SPLIT
Prior token dropping trades salience (attention — position-biased) against diversity (coverage-unstable). SPLIT is training-free and unifies both:
1. Temporal-shift importance (attention-free)
$$I(x)=\tfrac{1}{|L|}\!\sum_{\ell\in L}\frac{\lVert h_\ell(x)-h_{\ell-1}(x)\rVert_2}{\lVert h_\ell(x)\rVert_2}$$
2. Adaptive region budget (K=16 regions)
$$\tilde B_k=\tfrac{B}{K}+B\cdot\frac{I(X_k)}{\sum_j I(X_j)}$$
3. Diversity selection ($\lambda{=}0.5$), keep top-$\tilde B_k$/region
$$D(i)=\lambda\,\sigma_i-\mu_i,\quad S_{ij}=\langle\hat x_i,\hat x_j\rangle$$
2Claim 3 ✓ Reproduced ★ KEY
Temporal-shift importance is spread; CLS attention collapses onto a few low-row tokens (position bias). SPLIT covers all 16 regions.
Attention-free & unbiased: center-of-mass row 11.5 (=center) vs attention 12.7–13.6; Spearman(imp, attn) ≈ −0.7.
3Claim 1 ≈ Trend reproduced ★
LLaVA-1.5-7B (MPS); POPE 150, TextVQA 60, ScienceQA 60; 576→192/128/64.
Left: SPLIT degrades gracefully (~92→88%), above random. Right: on OCR TextVQA, random collapses at 64 tokens; SPLIT holds.
| Avg rel. acc. (%) | 192 | 128 | 64 |
|---|---|---|---|
| SPLIT (paper, 10 bench) | 99.3 | 97.8 | 92.8 |
| SPLIT (ours, 3 bench) | 92.4 | 89.3 | 87.9 |
| random (ours) | 88.3 | 82.7 | 71.5 |
SPLIT ≫ random; Claim 2's sub-2% SOTA margins sit below subset noise.
5Takeaways ★
- Claim 3 ✓ attention-free, unbiased, coverage-preserving.
- Claim 1 ≈ graceful degradation reproduced; absolute lower.
- Claim 2 – not resolvable at this scale; SPLIT clearly beats random dropping.
- Cost. Apple M1 Pro (MPS), ~1.5 h, $0 — HF Jobs was blocked (402).
In one sentence: the attention-free mechanism reproduces cleanly; the headline accuracy needs the full 10-benchmark GPU suite.
LAB · INSTITUTION