Lijie Hu

Contrastive learning's entanglement problem may be solved: BayesNCL disentangles representations by selectively filtering task-irrelevant features, leading to a 142% boost in semantic consistency.

Standard attention mechanisms inevitably cause intertask interference in in-context continual learning, leading to systematic bias and performance degradation in long prompts.

Applying representation interventions adaptively based on input characteristics dramatically improves alignment without sacrificing general capabilities, a feat previously unmet by uniform intervention methods.

The best LLM to answer a question isn't always the best LLM to *teach* the answer, and matching the "difficulty" of the explanation to the student's current abilities yields better learning.

LLM watermarks can now survive fine-tuning, quantization, and distillation thanks to a new method that embeds them in a stable functional subspace.

By enforcing graph isomorphism across counterfactual inputs, UGID reveals that debiasing LLMs can be achieved by directly manipulating internal representations and attention mechanisms.