In briefWeFair: Fidelity-aware human-feedback alignment for personalized query rewriting in short-video search
SummaryThis paper proposes WeFair, a method for personalized query rewriting in WeChat short-video search. It mines real user reformulations from search sessions, filters unsupported transitions with LLM screening and contextual verification, and trains the rewriter through supervised curriculum learning, direct preference optimization, and GRPO with a strict verification reward. Online A/B testing reports a 1.576% relative increase in effective clicked-query rate (VV>10s) and a 1.731% relative reduction in query reformulation rate.
In briefComparing generative AI, machine learning, and manual review for eDiscovery
SummaryThis paper examines a Redgrave LLP case study that reported 88% recall for Relativity's generative AI tool aiR and 64% for its machine-learning tool RAL. The authors argue the two results are not directly comparable due to differing objectives, inter-assessor disagreement, and different recall-precision and cost-benefit trade-offs. They conclude that, once these differences are considered, aiR and manual review are comparable. They also report that established machine-learning models using Continuous Active Learning and Scalable Continuous Active Learning perform at state-of-the-art levels when trained on only 2,000 first-pass reviewer responsiveness determinations.
In briefProving Shampoo is minimax optimal for tracking moving optima in nonstationary optimization
SummaryThis paper theoretically analyzes the Matrix Shampoo optimizer when the optimum it must track changes over time. It derives finite-horizon high-probability error bounds that account for momentum, noise, and adaptive preconditioning. Under a squared path-variation constraint, the authors prove a matching minimax lower bound, showing Shampoo is information-theoretically optimal among causal procedures. The theoretical findings are validated on synthetic and real-world problems.
In briefTime window optimization for motor imagery brain-computer interface classification using structure-aware reinforcement learning
SummaryMotor imagery brain-computer interfaces usually decode EEG with fixed sliding windows, which may not align with neural events. This paper proposes TSE-RL, a reinforcement learning framework that embeds knowledge of motor imagery event structure, assigning each window a Discard, Onset, Sustain, or Offset role. A finite-state machine masks invalid actions to preserve temporal order, and a Transformer-based history encoder models dependencies across windows. The authors report consistent decoding improvements in cross-session and cross-subject settings on multiple benchmark datasets.
In briefML Engineering Agents: A Data Management Perspective
SummaryMachine learning engineering (MLE) agents use large language models to generate code and predictive performance as feedback to automate end-to-end ML pipeline development. This tutorial introduces their foundations and current evaluation limits. It highlights database research opportunities, including efficient pipeline execution, data discovery over heterogeneous sources, reliability, and fairness.
