Participated in public demo 2603 with Visual Blind Spots

The problem [00:00] Saurabh had observed that in his work as a machine learning engineer, most of his time wasn't spent fixing a model's mistakes but figuring out why it made them, and today's evaluation process (train, score against metrics like precision/recall, deploy) does nothing to surface a model's inherent weaknesses before it ships. Failures typically only turn up after deployment, as a manual, one-off case review that doesn't scale: one engineer working through a handful of examples cannot systematically surface the dozens of failure patterns a model may actually have. What they built [01:30] Saurabh built an agentic harness that acts as a forensic auditor for a vision-language model, starting cold with no hints about what might be wrong. Given a video and its ground truth, the agent runs a five-tool diagnostic toolkit, frame shuffling, attention-map probing, frame captioning, language-prior probing, and temporal-segment probing, to test whether the model's answer holds up under different perturbations and views into its internal state. It compiles this evidence into a per-video hypothesis, then updates that hypothesis as it works through more videos, converging on a failure taxonomy without ever being told which failure modes to look for, which tools to call, or how many patterns exist. What happened [03:20] Run cold-start on 100 videos from the NextQA dataset against Qwen2-VL 7B, the agent surfaced a taxonomy of 8 failure patterns (4 primary drivers, 4 co-occurring amplifiers), a taxonomy that had already emerged by video 3 and held up through video 100, for a total cost of about $28 across roughly 5 hours, 592 API calls, and 3.5M tokens. [05:05] A follow-up literature review cross-referenced these against published VideoQA/video-LLM research: most patterns were directly or partially confirmed by prior work (e.g. U-shaped attention, temporal-order effects), and a few, including a primacy-bias variant, appear to be novel, unquantified findings, with wording different enough from the cited papers to rule out the agent having simply read and repeated them. The ask Saurabh is open to hearing from anyone evaluating ML models or working on similar problems.