Mocked benchmark shows DeepSeek's document chart grounding badly misaligned
burkov · x · 2026-10-18
ML practitioner Andriy Burkov mocks how "the new faster DeepSeek 4.1 Flash" sees illustrations in documents. Screenshots from an algorithms textbook show the model's predicted regions (red) badly misaligned with ground-truth figure boxes (green) — offset, overlapping body text, and wrong on all three sample pages.
Note: the model name appears sarcastic; the substance is a demonstration of poor document layout understanding / chart grounding.
More from Models
- One 3D avatar generation burned 15% of a $100 Claude Opus 5.5 plan, user reports — mallow610 · 2026-10-18
- Dev complains Claude's habit of saying "not done" is his biggest pain point — jasonkneen · 2026-10-18
- Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18
- Claude refuses bulk opcode dump for user's own hardware, admits its rule is "crude" — epitaxial_layer · 2026-10-18
- Paying Google AI subscriber blocked: custom-domain Workspace accounts can't access AI Studio — ChrisRackauckas · 2026-10-18
- Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18