Google's Agile Deliberation surfaces borderline cases where your labeling rules break down
nukebuddy · reddit · 2026-11-05
Google Research's Nov 4 write-up introduces Agile Deliberation, tackling a common failure mode: a classifier can follow your rule perfectly and still disappoint you, because the rule itself was under-specified.
- How it works: you start with a category and its exclusions; the system finds borderline examples (e.g. a smoothie in a "healthy food" task, exposing disagreements about sugar, processing, or what counts as a meal) for you to judge, then refines the instructions automatically.
- Evaluation: 18 live sessions of 90 minutes each, across food classification and misleading reward imagery. Google reports better classification performance with less human effort and stress than manual approaches.
- Caveat: a small study in two domains can't settle how it generalizes across communities or moderation policies.
The most compelling product idea: instead of another empty prompt box, show people exactly which cases their instructions don't cover.
Related event: Google Open-Sources Agile Deliberation for Ambiguous Classification(4 posts)→
More from Research
- Diffusion-Qwen3.8-27B: Qwen3 adapted into a masked diffusion LM with bidirectional attention — GabGarrett · 2026-11-06
- Researcher slams Inkling 2 Small time-series demo as a practically pointless task — mariofilhoml · 2026-11-06
- Wavefunction Bench Tests Coding Agents on Quantum Chemistry: Only 3 of 9 Systems Passed — CatAstro_Piyush · 2026-11-06
- Stable reinforce handles 8-minute rollouts with FP4 weights in open RL training — simonguozirui · 2026-11-06
- Biggest open RL run yet: 2,048 GPUs, 9 days, 17M rollouts, 3T+ tokens — LiTianleli · 2026-11-06
- Solv3R reconstructs 128-photo scenes with under 1.5GB of VRAM — taiyasaki · 2026-11-06