| --- |
| license: mit |
| tags: |
| - icml2026-repro |
| - paper-Elm4TdaXi0 |
| - vision-language-models |
| - token-pruning |
| --- |
| |
| # SPLIT-VLM reproduction bundle |
|
|
| Reproduction of **SPLIT-VLM** (ICML 2026, OpenReview `Elm4TdaXi0`) — *Salience-Guided |
| Partitioning towards Local Coverage for Importance-Aware Token Dropping in Vision-Language Models*. |
|
|
| Contents: |
| - `scripts/split_prune.py` — faithful SPLIT implementation (temporal-shift importance, adaptive region budgets, diversity selection; Algorithm 1). |
| - `scripts/validate_mechanism.py` — Claim 3 mechanism check on the CLIP vision tower. |
| - `scripts/llava_split_eval.py` — LLaVA-1.5-7B token-dropping eval (POPE/TextVQA/ScienceQA) with SPLIT vs random/attention baselines. |
| - `scripts/aggregate.py`, `make_figures.py`, `make_results_fig.py` — aggregation + figures. |
| - `outputs/` — result JSONs, aggregate.{json,csv}, figures, mechanism validation. |
| - `paper_notes.md` — recovered method + target tables. |
|
|
| **Compute:** local Apple M1 Pro (MPS, fp16); reduced-scale subsets (POPE 150, TextVQA 60, ScienceQA 60). |
| HF Jobs (GPU) was blocked by lack of credits (402). See the Trackio logbook for the full write-up. |
|
|
| Reproduce: `python scripts/llava_split_eval.py --task pope --n 150 --budgets 192,128,64 --methods vanilla,split,random,attn` |
|
|
