distribution-shift

This paper identifies distribution shift and scale constraints as critical failure modes for statistical contamination detection methods in LLM benchmark auditing. Evaluating three paradigms across 27 models reveals only 199 correct outcomes out of 335 evaluations, indicating a systematic reliability gap that prevents these methods from replacing transparent data provenance.

This paper introduces a theoretical framework for quantifying deployment risk when training and deployment distributions differ due to latent regime dynamics modeled as a Markov-switching process, providing exact decomposition and finite-sample bounds.

Introduces TASER, a training-time regularization framework derived from Langevin Stein operators that encourages geometric compatibility between predictors and data density, improving adversarial robustness and stability on CIFAR-10 without significant clean accuracy degradation.

This paper theoretically identifies and mitigates context distribution shift in multi-turn dialogue RL, proposing Calibrated Interactive RL that couples interactive RL with simulator alignment to reduce the sim-to-real gap and achieve state-of-the-art performance.

MARGIN is a runtime confidence calibration method for multi-agent foundation model systems that learns per-agent calibration factors online, improving pairwise resolution from below random to 70-89% on hard benchmarks, requiring no held-out data or retraining.

This paper develops a PAC-Bayesian framework for test-time adaptation that uses MMD-balls as credal sets, providing formal generalization bounds and separating epistemic from aleatoric uncertainty under distribution shift.

This paper proposes Physics-Informed Multi-Scale Mamba (PIMSM), a state-space architecture that aligns model memory with physical timescales to improve robustness under distribution shift in scientific time series, demonstrating improvements on fMRI and weather forecasting tasks.

This paper introduces ICRL, a framework that jointly trains a solver and critic with reinforcement learning to internalize critique guidance, enabling the solver to improve without external critique. It uses distribution calibration and role-wise group advantage estimation, achieving 6-7 point gains over GRPO on agentic and mathematical reasoning tasks.

This paper investigates how informal text (slang, emoji, Gen-Z filler tokens) degrades NLI accuracy in ELECTRA-small and RoBERTa-large models, identifying two distinct failure mechanisms—tokenization failure (emoji mapped to [UNK]) and distribution shift (out-of-domain noise tokens)—and proposes targeted mitigations that recover accuracy without harming clean-text performance.

This paper proposes a conformal prediction framework for LLMs that leverages internal representations rather than output-level statistics, introducing Layer-Wise Information (LI) scores as nonconformity measures to improve validity-efficiency trade-offs under distribution shift. The method demonstrates stronger robustness to calibration-deployment mismatch compared to text-level baselines across QA benchmarks.