Tag
This paper introduces the Expectation Consistency Loss (ECL), a theoretically grounded loss function for calibrating classifier confidence under covariate shift, derived from a necessary and sufficient condition called the Expectation Consistency Condition.
TILT introduces a novel objective for unsupervised domain adaptation under covariate shift that penalizes an auxiliary component on unlabeled target data, implicitly achieving self-localized importance weighting with bounded estimands. Theoretical guarantees and experiments on shifted CIFAR-100 show improved target performance over baselines.
This paper revisits Dataset Aggregation (DAgger) for training long-horizon LLM agents, demonstrating that turn-level teacher-student policy interpolation mitigates covariate shift and outperforms existing methods on software engineering benchmarks like SWE-bench Verified.