HKUST's ODE evolves RL prompts from agent rollouts, boosting multimodal search accuracy
机器之心 · wechat · 2026-11-14
HKUST's MMSenseLab introduced a Visual-Native Agent Harness plus On-policy Data Evolution (ODE), accepted at NeurIPS 2026, for training multimodal deep-search agents.
The harness assigns every image (originals, search results, crops) an ID in a shared image bank, so downstream tools can reuse earlier images — e.g. crop a ridge for reverse image search, then zoom the returned photo, as the demo did to geolocate a mountain photo. ODE lets a data synthesizer revise question-generation rules based on the target model's rollouts, validated over up to five rounds.
With ODE data, an 8B model's average accuracy across eight benchmarks rose from 24.9% to 39.0%, and a 30B model from 30.6% to 41.5% — near Gemini 2.5 Pro's 42.21% under the same harness. Ablations show both pieces matter: disabling image reuse drops MMBC accuracy to 7.59% vs 12.50% with full reuse, and equal-sized ODE data beats fixed-rule data across all RL comparisons.
More from Research
- CoRL 2026 researcher on scientific taste: consensus should come after research, not before — abhishekunique7 · 2026-11-14
- Vauss claims 5.4x faster Gaussian Splat training than gsplat, scene in 106s — ravisparikh · 2026-11-14
- Artificial Analysis launches Decision Index v2.1 benchmarking models across six evaluations — ArtificialAnlys · 2026-11-14
- Gradio launches agent challenge: reproduce NeurIPS 2026 papers on Hugging Face, Nov 20–Dec 5 — Gradio · 2026-11-14
- ICLR 2027 Rebuttal Radar tracks which review scores move during the rebuttal window — 3scorciav · 2026-11-14
- Paper argues 'The End of Harness Engineering': loops rewrite agent harnesses, humans just write the exam — blaizedsouza · 2026-11-14