Search.
Filter by area, type, author, year, or venue — or just type a few words.
Search the corpus
Optimistic Online Newton with Sublinear Regret on Strongly Curved Losses
Leila Aydın, Maya Singh, Renat Ostrovsky
A second-order online learner that matches the regret of the best Newton step in hindsight, while remaining one-pass with O(d²) memory.
Learned Cardinality with Distribution-Free Confidence
Jihoon Park, Aylin Demir, Sasha Volkov
Calibrated cardinality estimation under workload drift — confidence intervals that hold without distributional assumptions.
Adaptive Mirror Descent for Heavy-Tailed Bandits
Leila Aydın, Maya Singh, Renat Ostrovsky
A mirror descent variant whose potential adapts to the heaviness of the loss tail, achieving instance-optimal regret without prior tail knowledge.
On the Price of Adaptivity in Online Convex Optimization
Adaptive regret bounds come at a constant — but unavoidable — multiplicative cost. We pin down the exact frontier and exhibit matching algorithms.
oco-bench: A Reproducible Suite for Online Convex Optimization
14 environments, 23 algorithms, one CLI. Reproducible regret curves with seed-level controls and standard plotting harnesses.
On the Statistical Cost of Open-Vocabulary Decision-Making
Leila Aydın, Jihoon Park, Sasha Volkov, Renat Ostrovsky
How much harder is bandit learning when the action set is unbounded? Surprisingly, only by a logarithmic factor — provided the loss is well-conditioned.
Adaptive Buffer Pools via Bandit Feedback
Aylin Demir, Jihoon Park, Sasha Volkov
A buffer manager that treats page replacement as a contextual bandit, with regret guarantees against the best fixed policy in hindsight.
Logarithmic Regret for Convex–Concave Saddle Problems
Maya Singh, Renat Ostrovsky, Leila Aydın
When both players play strongly convex losses, an extragradient method attains O(log T) duality gap with high probability.
Adaptive Algorithms in Adversarial Worlds (invited talk)
A 45-minute survey of recent progress on parameter-free and adaptive online learning, given as a plenary at COLT 2025.
Online Query Plan Selection with Sublinear Regret
A plan-selection bandit that achieves sublinear regret against the best plan per query template, even under heavy-tailed latency.
Learned Indexes Under Distribution Shift
A self-correcting learned index whose error stays bounded as the underlying key distribution drifts, without retraining from scratch.
Stability and Generalization in Online Stochastic Gradient Methods
A unified stability framework that recovers and tightens generalization bounds for SGD under online and i.i.d. settings.