paper_notes.md

SPLIT-VLM — Reproduction Notes

Paper: SPLIT-VLM: Salience-Guided Partitioning towards Local Coverage for Importance-Aware Token Dropping in Vision-Language Models Authors: Seungil Lee, Gilha Lee, Hyun Kim (Seoul National University of Science and Technology; KETI) Venue: ICML 2026 (Poster). OpenReview id Elm4TdaXi0 — https://openreview.net/forum?id=Elm4TdaXi0 Public arXiv/GitHub: none. PDF is OpenReview-challenge-gated; text extracted via authenticated browser + pdf.js.

Claims to verify

1. On LLaVA-1.5-7B image understanding, SPLIT preserves >99% avg relative accuracy at 192 vision tokens and ~92.8% at 64 tokens.
2. SPLIT consistently outperforms SOTA token-dropping methods on image & video benchmarks.
3. Token importance is estimated via temporal shifts of hidden states across ViT layers, eschewing attention scores/biases.

Method (Algorithm 1) — training-free, batch size 1

Input: visual token set X={x_j}, partition into K disjoint regions X_k, total budget B, layer set L.

1. Temporal-shift importance (Sec 4.1)

For a token x with ViT hidden state h_ℓ(x) at layer ℓ:

• Relative change rate: Δ_ℓ(x) = ‖h_ℓ(x) − h_{ℓ−1}(x)‖₂ / ‖h_ℓ(x)‖₂ (Eq. 5, normalized so high-norm tokens are not overemphasized)
• Token importance: I(x) = (1/|L|) Σ_{ℓ∈L} Δ_ℓ(x) (Eq. 6)
• Region importance: I(X_k) = (1/|X_k|) Σ_{x∈X_k} I(x)
• Best L = all ViT layers (ablation Fig 6b / Table 16). Middle layers best among single groups; layer 0 hurts.

2. Adaptive region budget (Eq. 7) — hybrid uniform + importance

• B̃_k = B/K + B · I(X_k) / Σ_j I(X_j), rounded to nearest integer.
• K = 16 regions for LLaVA-1.5-7B (24×24 patch grid → 4×4 regions of 6×6 patches). Ablation: 16 most stable; 6×6 helps only at large budgets, hurts at 64.

3. Diversity-score selection (Sec 4.2)

• Normalize each token embedding x_i; self-similarity S_ij = ⟨x_i/‖x_i‖, x_j/‖x_j‖⟩ (Eq. 8, cosine over all N tokens).
• μ_i = mean_j S_ij, σ_i = std_j S_ij (Eq. 9).
• Diversity score D(i) = λ·σ_i − μ_i, λ = 0.5 (fixed, no tuning) (Eq. 10).
• Within each region k, select the B̃_k tokens with largest D.
• Final reduced set R̂ = ∪_k R̂_k.

Note: importance/budget use ViT hidden states (temporal shift); the diversity self-similarity is computed on the token embeddings (post-encoder patch tokens fed to projector). Two complementary signals — ablation Table 7 confirms Temporal-Shift-for-allocation + Diversity-for-selection is best.

Config for LLaVA-1.5-7B

• Vision encoder: CLIP ViT-L/14-336 → 336×336 image → 24×24 = 576 patch tokens.
• Token budgets B ∈ {192 (↓66.7%), 128 (↓77.8%), 64 (↓88.9%)}; baseline = 576.
• Regions K=16 (4×4). Temporal shift over all 24 ViT layers. λ=0.5.
• Pruning applied at the ViT output / before the LLM (ViT-side), batch size 1, no training.
• Env in paper: H100, Python 3.12, PyTorch 2.7.0, CUDA 12.2.

Target: Table 1 (LLaVA-1.5-7B, 10 image tasks). Avg = mean over tasks of score_pruned/score_vanilla.

Vanilla 576 (100%): GQA 61.22 | MMB 63.14 | MMB-CN 54.81 | MME 1477.65 | POPE 85.41 | SQA 68.12 | VQAtext 48.67 | VQAv2 77.41 | VizWiz 53.81 | OCRBench 205

Method	budget	GQA	MMB	MME	POPE	SQA	VQAtext	VQAv2	Avg
SPLIT	192	59.5	62.6	1479.5	85.6	68.1	46.9	76.1	99.3%
DART	192	59.2	62.7	1477.8	84.2	67.9	46.7	76.2	99.0%
SPLIT	128	58.8	62.5	1455.2	85.5	68.4	45.8	75.5	97.8%
SPLIT	64	59.0	60.1	1375.5	84.5	68.6	41.5	73.2	92.8%
DivPrune	64	58.4	58.6	1362.1	83.7	67.3	40.8	72.4	91.7%
FastV	64	53.1	50.1	1292.8	78.5	65.5	39.5	70.3	86.6%

Baselines compared: MustDrop, FastV, PDrop, HiRED, SparseVLM, DivPrune, DART, GreedyPrune (+ VisPruner/VisionZip/CDPruner in Table 6 ablation).

Reproduction strategy (given constraints)

• Full faithful implementation of SPLIT on llava-hf/llava-1.5-7b-hf.
• Evaluate relative accuracy at 576/192/128/64 on standard benchmarks; compare SPLIT vs baselines (random drop, FastV-style attention drop) — verifies Claims 1 & 3, partial Claim 2.
• Substantive run intended for HF GPU Job; HF Jobs blocked by 402 (no credits) as of run — fall back to scaled local run and document scale honestly.