Lightning Talks!
summary
A rapid-fire set of concrete results: an online load‑balancing scheme learns from a 1% sample and achieves a 1+ε load guarantee (using fractional relaxations, convex dual weights, and multiplicative weights); a boosting-style method turns multiple probabilistic models into a safer sampler that trades abstain vs safety; RNN-trained next‑token models provably yield next‑k-token indistinguishability against bounded RNN distinguishers; and a proof shows Kikuchi lifts satisfy an eigenvalue bound roughly M + k, resolving a recent conjecture.
AI-written summary · about
chapters (13)
- 0:00Welcome to Lightning Talks
- 0:23Online algorithms with predictions
- 6:57AI's role in theory
- 11:09Multi-armed bandits improvement
- 14:45Generative models evaluation
- 20:47Robust ML across distributions
- 26:27Conditional sampling with LLMs
- 31:58Token prediction and proof generation
- 36:36Contradiction graphs and VC dimension
- 40:37Consensus sampling risks
- 46:38Credit assignment via games
- 51:41Kikuchi lifts in graph theory
- 55:41Closing and banquet
related talks
49mIntroduction and Research Fellows Lightning Talks
21mFireside Chat
58mLightning Talks
01h 01mRobustness, Stochastics, Uncertainty 1
10mLightning Talk: Not All Tokens Are Equal: Semantic KV-Cache for Agen...
39mLightning Talks
29mEnhancing Performance Test Results with ML & LLMs: Communication Beyond the Numbers
01h 01mRobustness, Stochastics, Uncertainty 2