Protein inverse folding models conditioned on structure achieve high sequence recovery but often fail to preserve biological function due to the lack of functional supervision. We propose a function-aware preference alignment framework that improves functional preservation by fine-tuning models to favor function-preserving sequences over function-disrupting alternatives, avoiding the need for explicit function optimization. Our approach constructs reliable preference pairs in silico using hypothesis-driven perturbations of critical residues and model-consistent likelihood constraints, enabling scalable supervision without additional wet-lab measurements. The resulting framework guides protein sequence design models toward generating sequences that better preserve functional integrity, while remaining compatible with existing inverse folding pipelines such as ProteinMPNN and ESM-IF. Extensive experiments on protein design benchmarks and enzyme datasets with established wet-lab validation show that our fine-tuned models consistently outperform pretrained counterparts in preserving functional integrity during protein sequence design. The code is available at https://github.com/EvaFlower/Function-aware-Protein-Inverse-Folding
Nilufer Tamatgar, Soobin Park, Yinghua Yao et al.· Proceedings of the 32nd ACM...· 0 citations
Chain-of-thought (CoT) reasoning is widely used to improve both the performance and interpretability of large language models (LLMs), yet the generated reasoning may not faithfully support the final answer. We study this problem from a causal perspective, where a faithful CoT process should follow the chain $Z\rightarrow X\rightarrow Y$, with $Z$, $X$, and $Y$ denoting the instruction, reasoning chain, and final answer, respectively. In this process, the instruction should affect the answer only through the reasoning chain. However, conventional autoregressive LLMs condition answer generation on both the instruction and the CoT, which still allows a direct instruction-to-answer shortcut. To address this issue, we propose CASE, a framework that combines training-time causal alignment and inference-time structural enforcement. During training, CASE builds counterfactual-CoT, biased-instruction, and empty-instruction datasets, and applies selective-loss fine-tuning to strengthen CoT-to-answer dependence while suppressing instruction shortcuts. During inference, CASE masks direct attention from instruction tokens to answer tokens, preventing the model from bypassing the generated CoT. We provide an information-theoretic analysis showing how these components promote faithful chains. Experiments on three models and four benchmarks show that CASE achieves a 37\% average per-setting relative improvement in overall CoT faithfulness over the strongest baselines, exhibits stronger cross-dataset faithfulness transfer, and maintains competitive average accuracy. Code is available at https://github.com/oddwang/CASE.
Ziming Wang, Yinghua Yao, Changwu Huang et al.· 0 citations
This work forms the specialization of pretrained virtual screening models to individual pockets as a test-time adaptation problem and proposes PETA, a parameter-efficient framework that directly adapts pretrained model at test time and outperforms both pretrained and fully retrained baselines while updating only the LayerNorm parameters.
Jia-Qi Lin, Yinghua Yao, Changran Wang et al.· 0 citations