Abe Hou

1/4 Excited to share ProteoCraft😀, our work at Stanford on tree search for hallucination-based protein design. 29/35 designed PD-L1 miniproteins bound experimentally (82.9%), with a best measured SPR KD of 0.879 nM. The question that got us started: what should we try next?

🎉 Five papers accepted at NeurIPS 2026! Excited to share our work across protein design, RNA design and evaluation, efficient LLM agents, and multimodal safet Here’s what we explored 👇 🧬 1 | Grounding protein generation in physics <Equivariant Force Field Calibration for Flow-based Protein Design> Generative models can produce protein structures quickly—but physical plausibility remains a critical challenge. This work combines differentiable force-field guidance with consistency-based training to calibrate flow models toward low-energy, sterically valid conformations. Across full-atom generation and peptide binder modeling, we improve physical plausibility and geometric stability while enabling sampling with very few steps. 🧪 2 | Learning RNA design preferences from physical feedback <Pareto Preference Optimization for Structure- and Stability-Aware RNA Inverse Folding> RNA design requires satisfying several objectives together: structural fidelity, local accuracy, and thermodynamic stability. Optimizing one metric can create shortcuts that undermine the others. We introduce RiboPO, a preference optimization framework that constructs preferences through Pareto comparisons, uses structural gates to guard against compositional reward hacking, and fine-tunes an RNA inverse-folding model using DPO. 📄 Earlier preprint: arxiv.org/abs/2510.21161 💻 Code (anonymous review repository): anonymous.4open.science/r/ri… 🔍 3 | Testing whether RNA design scores deserve our trust <Multi-Oracle Agreement Reveals the Limits of Self-Consistency Evaluation in RNA Design> What happens when the predictor used to optimize a sequence is also used to evaluate it? Across approximately 11,000 RNA designs, we find that 78.1% of designs from single-objective optimization pass one predictor yet fail another. Ensemble-based methods reduce this disagreement substantially. We introduce ACCORD, an evaluation workflow that combines independent predictors and calibrates their disagreement against experimental chemical-mapping data. 📄 Paper: minghao-sun-sc.github.io/pap… 💻 Code (anonymous review repository): anonymous.4open.science/r/AC… ⚡ 4 | Making LLM agents more efficient by changing how they act <Latent Action Reparameterization for Efficient Agent Inference> Agent inference cost depends on both the cost of each decision and the number of decisions required. LAR learns compact, executable latent actions from recurring patterns in agent trajectories. This shortens the effective decision horizon, reducing action tokens and inference time while maintaining or improving task success across the evaluated benchmarks. 📄 Paper: arxiv.org/abs/2605.18597 💻 Code: github.com/EZ-hwh/LAR 🛡️ 5 | Understanding the safety implications of thinking with images <When Think-with-Image Meets Safety: What Determines Multimodal Jailbreak Robustness?> As vision-language models increasingly reason through visual operations and external tools, their inference pipelines become part of the safety question. Across the evaluated models, explicit image-tool interaction reduces jailbreak success by around 30% relative on average. Representation analyses and activation interventions support an account in which tool invocation shifts hidden states toward a safety-relevant direction. 📄 Paper: arxiv.org/abs/2605.27932 Huge thanks to my coauthors, advisors, and collaborators for making this work possible! #NeurIPS2026 #AIforScience #ProteinDesign #RNADesign #LLMAgents

𝗖𝗮𝗻 𝗮 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗺𝗼𝗱𝗲𝗹 𝗯𝗲𝗰𝗼𝗺𝗲 𝗮 𝗯𝗲𝘁𝘁𝗲𝗿 𝗶𝗺𝗮𝗴𝗲 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗼𝗿 𝗯𝘆 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗳𝗲𝗲𝗱𝗯𝗮𝗰𝗸? 🚀 Excited to share 𝗨𝗻𝗶𝗘𝘃𝗼-𝗩𝗟, our new work on on-policy self-distillation for multimodal model self-improvement! The motivation is simple: as generation and understanding come together in multimodal systems, models can provide feedback on their own outputs. Can that feedback become a useful training signal—turning their ability to evaluate into improvements in generation? 📈 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁𝘀 Building on Qwen-image-2512, UniEvo-VL improves: • GenEval: 𝟬.𝟳𝟰𝟳 → 𝟬.𝟴𝟬𝟴 • GenEval2 Soft-TIFA: 𝟯𝟮.𝟵𝟳 → 𝟯𝟱.𝟱𝟯 These gains show the potential of learning from self-generated feedback. But the boundaries are just as interesting. 🔍 𝗪𝗵𝗮𝘁 𝘀𝗲𝘁𝘀 𝘁𝗵𝗲 𝗰𝗲𝗶𝗹𝗶𝗻𝗴? Experiments with stronger external critics suggest room for further improvement. This raises a central challenge: producing feedback is only the beginning. That feedback must be reliable enough to guide learning. Understanding what makes a good critic matters for building a better generator. 🔍 𝗗𝗼𝗲𝘀 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 𝘁𝗿𝗮𝗻𝘀𝗳𝗲𝗿 𝗮𝗰𝗿𝗼𝘀𝘀 𝘁𝗮𝘀𝗸𝘀? Not uniformly. Text-rendering results remain mixed, even as the overall generation benchmarks improve. Progress on one benchmark does not mean every capability improves alongside it—a useful boundary to understand when studying self-improving systems. For me, the broader research question is: 𝗪𝗵𝗲𝗻 𝗶𝘀 𝗮 𝗺𝗼𝗱𝗲𝗹’𝘀 𝗼𝘄𝗻 𝗳𝗲𝗲𝗱𝗯𝗮𝗰𝗸 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗲𝗻𝗼𝘂𝗴𝗵 𝘁𝗼 𝗵𝗲𝗹𝗽 𝗶𝘁 𝗶𝗺𝗽𝗿𝗼𝘃𝗲? UniEvo-VL takes a step toward answering this question for multimodal generation, exploring both the opportunities and the limitations. Grateful to all my collaborators for making this work possible! If you’re working on self-distillation, multimodal post-training, or self-improving models, I’d love to hear your perspective. 📄 Paper: arxiv.org/abs/2609.38721 🤗 Hugging Face: huggingface.co/papers/2609.3… If you find the work interesting, a repost or an upvote on Hugging Face would help more researchers discover it. Thanks for your support! 🙏 #UniEvoVL #SelfDistillation #MultimodalAI #ImageGeneration #PostTraining