Native image reading beats OCR for deep research agents, cutting tool calls
RulinShao · x · 2026-11-04
New work extends the "no OCR, read text natively from images" approach to deep research agents: post-training open-weight models to work with page images boosts performance and reduces tool calls, likely because such models were never trained on page images. Led by Hughie Hu.
More from coding & agent
- SurrealDB 4 launches with SurrealDB Earth: serverless agent database on Cloudflare — ritakozlov · 2026-11-04
- Karpathy's "vibe coding" post hailed as a shoo-in for the test-of-time award — MillionInt · 2026-11-04
- OpenAI SDK ships official WebSocket helpers for Responses, GPT-Live and Realtime APIs — craigsdennis · 2026-11-04
- Pydantic AI adds ConversationStore to prevent concurrent chat saves from overwriting each other — samuelcolvin · 2026-11-04
- Gemini API CLI skill lets agents call image, TTS and video models with one command — _philschmid · 2026-11-04
- Anthropic silently pushed a server-side rule to Claude Code banning recreations of specific games — alex_verem · 2026-11-04