DeepMind's Ian Osband to argue policy gradient 'can be a terrible way to learn' in Oxford talk
IanOsband · x · 2026-10-12
Ian Osband, who leads the Science of Post-Training team at Google DeepMind, gives an Oxford seminar talk titled 'The Price of Learning' (Oct 19).
- Core claim: policy gradient sits at the heart of modern RL and LLM post-training, yet on image classification—where the gradient of accuracy is exact—it loses to cross-entropy even on accuracy
- Framing: learning is a sequential resource allocation problem; data, compute and capacity are scarce and each update sets where the next starts, but policy gradient spends greedily
- Proposal: simple methods that 'price learning,' improving robustness, compute efficiency and accuracy
Osband previously spent two years at OpenAI on post-training for ChatGPT and reasoning models. The series also features Zhanxing Zhu on token distinguishability and safety alignment.
More from Research
- Graph algorithm researchers shrug at OpenAI's general matching math result — thegautamkamath · 2026-10-13
- STOKE: Stanford's superoptimizer finds optimal assembly through randomness with early pruning — lauriewired · 2026-10-13
- Muse Wonder Full beats its scaling trend on out-of-distribution tasks, Meta says — AIatMeta · 2026-10-13
- ReplaySSM ships in GLM-5.3-Flash and Kimi-K3, speeding decoding up to 2.1x — tri_dao · 2026-10-13
- DeepRSI benchmark pushes agents to 48-hour ML tasks to measure recursive self-improvement — garrytan · 2026-10-13
- Datalab launches OmniParseBench, a game-resistant OCR benchmark where it doesn't rank first — VikParuchuri · 2026-10-13