ai-sherpa/repro-vlmrobustbench-bundle

VLM-RobustBench reproduction bundle (arXiv 2603.06148)

Independent reproduction of "VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models" (EdinburghNLP; OpenReview HwXyyvK7ZJ) for the ICML-2026 Agent Reproducibility Challenge. The official repo (github.com/saxenarohit/vlm_robustbench) is an empty stub, so everything here is implemented from the paper description.

What is reproduced

The paper's central, falsifiable claim (Claim 2): visual severity is a weak predictor of difficulty — low-severity glass_blur (a photometric corruption) reduces MMBench accuracy by ~8pp, while geometric / resampling distortions (e.g. elastic_transform, upsample) drive the largest drops, up to ~34pp.

We evaluate open VLMs from two of the paper's four families (Qwen, InternVL) on a seeded subset of MMBench (lmms-lab/MMBench_EN, dev split) under clean vs. corrupted images, and measure the percentage-point accuracy drop per corruption.

Layout

run_eval.py                  self-contained UV script (inline deps); runs on
                             HF GPU Jobs and CPU-smoke-tests unchanged
tests/test_corruptions.py    local CPU tests: corruption pipeline + answer parsing
results/                     results.json + per_condition_accuracy.csv per model

Corruptions

glass_blur and elastic_transform are reimplemented verbatim from the standard imagecorruptions library (Michaelis et al. 2019) — see the inlined _glass_blur_np / _elastic_np in run_eval.py. They are inlined (not import imagecorruptions) only because that package's module-level import cv2 pulls non-headless OpenCV (needs libGL, absent in the Jobs container). The computation is identical (skimage gaussian + scipy map_coordinates). rotate_30 and upsample_x4 are additional geometric/resampling controls.

All images are uniformly downscaled (aspect preserved, MAX_SIDE=336) before every condition including clean, so the only difference between conditions is the corruption, and the pixel-swapping glass_blur stays cheap.

Reproduce

Local logic tests (CPU, seconds):

python -m pytest tests/test_corruptions.py -v

Full eval on a Hugging Face GPU Job (L4, ~$0.5-0.8 per model):

hf jobs uv run --flavor l4x1 --timeout 60m -d --secrets HF_TOKEN \
  -e MODEL_ID=Qwen/Qwen2.5-VL-7B-Instruct -e N_ITEMS=300 \
  -e OUT_REPO=ai-sherpa/repro-vlmrobustbench-bundle \
  run_eval.py

The same script CPU-smoke-tests with a tiny model (MODEL_ID=Qwen/Qwen2-VL-2B-Instruct, N_ITEMS=6). Config is entirely env-var driven (MODEL_ID, N_ITEMS, SEED, MAX_SIDE, MAX_PIXELS, CONDITIONS, OUT_REPO).