Ethan Li

A rigorous contribution to the theory of diffusion models, with potential impact well beyond ML benchmarks. This paper establishes sample complexity lower bounds and explores generalization limits in generative modeling, core challenges when using these models to discover novel antibiotics and tackle resistance which could saves lives.

16

NVIDIA's research flipping conventional wisdom on SFT and RL training. Math-only RL substantially improving code reasoning. No code training during RL, yet the model solves previously impossible coding problems. Cross-domain transfer at work. 🙃 The results: 72.6% AIME 2024, 64.8% AIME 2025, 52.1% LiveCodeBench v6. State-of-the-art for 7B models. Sometimes taking a step backwards is the fastest way forward. Check out the paper here 🤓 - https://lnkd.in/ebSzkUe3 #reinforcementlearning #llm #reasoning

12

2 Comments

Wittgenstein has an argument that we don't really know what +2 means. IIRC the argument is that for very large numbers we've never empirically tried adding 2, and never had the experience of being corrected if we made a mistake. As Wittgenstein grounds the meaning of words in our ability to learn to use them as tools "correctly", and correctness is determined by community feedback, adding 2 to some very large number is kind of like asserting things about a place no one has been to. Who knows what the rules of addition are like out there? You can argue that we just need to follow some inductive formula. But you can't be sure your inductive formula is the right one either. With neither the result nor the formula being certain, you can't use one to calibrate the other. So you can't know how adding 2 really works up in the 10 to the 3 million neighborhood. (Whereas you know that 2+2=5 is wrong because there's a lot of conventional usage of small number arithmetic, and a lot of people to tell you that this is an incorrect application of the +) I think of this whenever it's revealed that a particular configuration of transistors on a chip gives the "wrong" answer to simple arithmetic operations. "Wrong" according to who? Really there are just several, subtly different, definitions of arithmetic : the ones that are defined by the various IEEE standards for chips, the ones defined by 8-bit adders that roll over at numbers higher than 256, the one defined by .. ahem .. Excel. And the one defined by mathematicians using inductive formulas that can't be fully systemized because of Gödel. They all give the expected results for small numbers. So it's impossible to tell which of these versions of arithmetic you were really thinking of when you learned your times tables back in school. You could have been learning arithmetic1 or arithmetic2 or arithmetic3. But not even you can tell which. You really don't know what +2 means. Now you might think that Wittgenstein is just wrong here. That there is a fact of the matter as to what +2 means. And maths is definitely not "relativist" with multiple alternative versions you can choose between. The thing to remember, though, is that Wittgenstein didn't come to this position - that what "grounds" our "symbols" is social convention - lightly. He came to it after spending half his philosophical career trying to make sense of the opposite theory : that what grounds symbols is the isomorphism between language and the world. And deciding it didn't work. The idea that what gives symbols their meaning, is nothing but the "language game", ie a normative set of rules that a community adopts, is very uncomfortable. But it's also harder than you might think to find a principled alternative that holds up to scrutiny.

4

I’m excited to share a new extension I built for GitHub Spec Kit: Spec Kit Retrospective! If you’re exploring Spec-Driven Development (SDD) with AI agents, you’ve probably run into this exact issue: you go through /specify, /plan, /tasks, and /implement, but the final codebase inevitably diverges from your original spec. What happens to that lost context? Usually, the specs become outdated, and the AI agent repeats the same mistakes on the next feature. Or, even better, when the agent ends the task, It doesn't update the specs! I created Spec Kit Retrospective to close that feedback loop. This extension adds a crucial final phase to the workflow, allowing your AI agent to: - Analyze the gap between the initial plan and the final implemented code. - Capture learnings and automatically re-adjust specs after divergence. - Feed insights back into your project's memory and constitution to mitigate future spec-drift. Instead of losing the "why" behind those last-minute code changes, this tool helps your AI learn from the implementation phase so your next feature build is even sharper. Check out the repository, feel free to give it a try (or a star!), and I’d love to hear your feedback! https://lnkd.in/epYxu6Hz

10

🚀 Code Release: Phys-Instruct is now open-source! Following our recent paper, "Ultra Fast PDE Solving via Physics Guided Few-step Diffusion", we are excited to announce that our official implementation is now publicly available! We built this repository not just to share our method, but to serve as a comprehensive toolkit for diffusion-based PDE solving. It provides a unified framework for benchmarking, experimentation, and reducing the barrier to entry for physics-guided generative AI. 🔗 Links: 💻 Code: https://lnkd.in/g-3himQC 📄 Paper: https://lnkd.in/gWBQfsHe 📂 Pretrained Models: https://lnkd.in/gVJdxhDc 👥 Authors: Xiangrui Kong, Yueqi Wang, Haoyang Zheng, 罗维俭 (Weijian Luo), and Guang Lin ✨ What’s included in the release? ✅ Unified Framework: A single codebase supporting 5 different baselines and approaches, including Standard Diffusion (EDM), Physics-Informed Training (PIDM), and Inference-Time Guidance (DiffusionPDE, CoCoGen). ✅ Multi-PDE Support: validated on 5 benchmark PDEs: Darcy Flow, Burgers’, Helmholtz, Poisson, and Navier-Stokes. ✅ Production-Ready: Includes single & multi-GPU distributed training support, step-by-step instructions, and data generation scripts. ✅ Pretrained Models. 💡 Why does this matter? This release allows researchers and practitioners to: 🔹 Reproduce results end-to-end with ease. 🔹 Benchmark different physics-guided diffusion strategies under one roof. 🔹 Deploy few-step diffusion solvers as fast, reusable priors for downstream tasks. We hope this codebase accelerates research on diffusion models, scientific machine learning, and fast PDE solvers. Feel free to reach out if you have questions, ideas, or would like to collaborate! #AI #ScientificComputing #DiffusionModels #PhysicsInformedML #PDE #MachineLearning #GenerativeAI #OpenSource #Research #Python #Pytorch #AIforScience #ComputationalPhysics

44

We're very happy to share our three preprints. One research program. From biology to adaptive computation. We set out to answer a question that spans biology and computation: how do systems specialize and preserve function under continuous stress? Paper 1 — Energy Conservation Drives Adaptive Specialization Cells exploit nutrient fluctuations to drive tunable transcriptional organizations that lead to durable adaptations. We identified evolutionarily tuned genomic hubs that coordinate stress responses while preserving plasticity—and what happens when that coordination breaks down. Paper 2 — Convergent Signals Encode Persistent Memory of Resilient States Building on that: how do cells lock in beneficial states long-term? We found that convergent signals—cooperative transcriptional regulators, innate immune mediators, and metabolic cofactors—encode epigenetic memory of adaptive phenotypes. Engineering cells with this multimodal programming preserves therapeutic function in metabolic aging models. Paper 3 — Rectifying Stress into Structural Memory The bridge to computation. We showed that asymmetric fluctuations act as tunable drives that, when phase-aligned, get rectified into durable specialization — in both cells and machines. We formalized this using operadic algebra as a topological ratchet, and deployed a self-organizing ML system that discovers hierarchy and tracks its own reorganization costs. One sentence across all three: adaptive systems—biological or computational —convert internal stress into structural memory that encodes function. Why this matters for what we're building at Daice: These aren't academic exercises. They're the foundation for our hybrid AI architecture — continuous adaptive systems that self-organize, encode long-term information through hierarchical nesting, and track thermodynamic costs. The same principles inform our techbio pipeline, where we build hybrid AI systems for automatic discovery of convergent signals to target cell reprogramming. Adaptive specialization. From cells to machines. From research to product. 📄 Paper 1: [https://lnkd.in/ePSkyNEP] 📄 Paper 2: [https://lnkd.in/exBpcnqC] 📄 Paper 3: [https://lnkd.in/erubpAWR] 🔗 Summaries: daicelabs.com/research Grateful to all our collaborators Manolis Kellis (MIT CSAIL), Teresa Femenia (Miguel Hernández University), Lydia Lynch (Princeton), Kevin Grove (Novo Nordisk), Remy Tuyeras (MIT), and the full team across MIT, Broad Institute, and Daice Labs.

7

Really impressed by Alexander’s new paper on Aligned Manifold Property and Topology Point Clouds, describing a sharp and innovative take on improving molecular property prediction through geometry-aware learning. A solid blend of theory and application that pushes the boundaries in molecular ML.

6