Kai Ye

Large language models need reliable test-time control of hallucinations. We propose Conditional Factuality Control (CFC), a post-hoc conformal framework that uses feature-conditional thresholds to produce set-valued outputs with conditional coverage guarantees. Compared with marginal conformal methods, CFC adapts to prompt difficulty, improving coverage on hard prompts while reducing prediction-set size. We also introduce CFC-PAC, which provides finite-sample risk certificates. Experiments on synthetic, QA/reasoning, and VLM benchmarks show near-target coverage with smaller sets than baselines.

The Evidential Regression Network (ERN) combines deep learning with Dempster-Shafer theory to jointly predict a target and quantify its uncertainty. However, the non-negativity that the theory enforces through specific activation functions prevents the model from learning effectively over part of the sample space, which limits performance. This paper characterizes that limited-learning region, analyzes the ERN to explain why the constraint arises, and introduces a regularization term that lets the network learn from the entire training set. Extensive experiments support the theoretical analysis and demonstrate the effectiveness of the proposed method.

Deep learning models of the interaction between brain structure and function have proven effective at identifying biomarkers for clinical phenotypes and brain diseases, but most prior work maps in only one direction—structure to function or function to structure—ignoring the intrinsic unity of the two modalities and introducing directional bias. To address this, we propose Bidirectional Mapping with Contrastive Learning (BMCL), which reduces the bias between the two one-way mappings through ROI-level contrastive learning. Evaluated on the HCP and OASIS datasets for clinical phenotype and neurodegenerative disease prediction, BMCL outperforms several state-of-the-art methods.