DeepMind paper: without direct image viewing, Gemini's defect-spotting falls from 86.7% to 14.3%
nukebuddy · reddit · 2026-11-22
A Reddit post digs into DeepMind's Multimodal Knowledge Agent tech report, using a road-inspection case to show why agents must actually open images. Given direct image viewing, Gemini 4 Carbon spotted a pothole in every run; without it, the same model and setup fell back to OCR, OpenCV, even ASCII renderings of photos—and missed it every time. Scores on harder criteria dropped from 86.7% to 14.3% while the no-viewing variant used 2.2× more output tokens; full-benchmark drops were smaller (6 points) and GDPval barely moved. Practical takeaway: when auditing agent work on scans or drawings, ask it to point to the actual evidence on the page—a convincing spreadsheet can be built from an incomplete reading.
More from coding & agent
- Redditor argues AI agents need decision expiration dates, not just bigger context windows — Elvantime · 2026-11-22
- ianand's 20-min AI Engineer World's Fair talk covers a paper and how it can go wrong — js_craft_hq · 2026-11-22
- Agent pipeline silently skipped a midnight run due to a UTC vs local time mismatch — fantomkoenig · 2026-11-22
- llms.txt Protocol Goes Mainstream as Google Antigravity Adopts It Two Years On — jeremyphoward · 2026-11-22
- He built a feed only agents can read: no UI, semantic matching, agent-rated posts — imdogefather · 2026-11-22
- Claude Code Debuts Experimental Code Mode as AI-Driven Tooling Gains Momentum — lxfater · 2026-11-22