Steffi Chern

Should data "evolve"? 🧬 Scaling is not enough. Model performance is bounded by data, but its value is defined by processing depth. We introduce Data Darwinism, a 10-level hierarchy (L0-L9) redefining data as an eternal co-evolutionary process.(1/n) huggingface.co/papers/2602.07…

@o_cube01 yayy congrats Ayoung! 🤩

Accepted to ICLR 2026!🎉 So grateful to my amazing collaborators 🫶 We introduce CLASH to evaluate value reasoning, revealing new failure modes in reasoning models and intriguing steerability results! 📰 Paper: arxiv.org/pdf/2504.10823

@tyao923 🐮！

𝗧𝗵𝗶𝗻𝗸 𝘄𝗶𝗱𝗲𝗿. 𝗧𝗵𝗶𝗻𝗸 𝘀𝗵𝗼𝗿𝘁𝗲𝗿. 🚀 𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗠𝘂𝗹𝘁𝗶𝗽𝗹𝗲𝘅 𝗧𝗵𝗶𝗻𝗸𝗶𝗻𝗴: token-wise branch-and-merge reasoning for LLMs. 💸 Discrete CoT is costly. 🎛️ Existing continuous tokens often clash with 𝗼𝗻-𝗽𝗼𝗹𝗶𝗰𝘆 𝗥𝗟 𝗲𝘅𝗽𝗹𝗼𝗿𝗮𝘁𝗶𝗼𝗻. 🎥 𝗠𝘂𝗹𝘁𝗶𝗽𝗹𝗲𝘅 𝗧𝗵𝗶𝗻𝗸𝗶𝗻𝗴, a sampling-based continuous reasoning paradigm:

LiveTalk Real-Time Multimodal Interactive Video Diffusion via Improved On-Policy Distillation huggingface.co/papers/2512.23…

There are competing views on whether RL can genuinely improve base model's performance (e.g., pass@128). The answer is both yes and no, largely depending on the interplay between pre-training, mid-training, and RL. We trained a few hundreds of GPT-2 scale LMs on synthetic GSM-like reasoning data from scratch. Here are what we found: 🧵

Agents are killing it at coding, deep research, Q&A...But the next frontier? Seamlessly orchestrating multiple apps to solve tasks end2end in real states -- Toolathlon is just for this! So if you want to make agents truly useful in the beautiful mess of real work, don't miss it!

🔥 Announcing our new paper: "SR-Scientist: Scientific Equation Discovery With Agentic AI" Most current work using LLMs for scientific discovery, like AlphaEvolve, follows a rigid "generate → evaluate → refine" loop. We challenge this paradigm for equation discovery. Our work, SR-Scientist, empowers an LLM to act as an autonomous agent, discovering scientific equations through long-horizon, tool-driven data analysis and equation evaluation—much like a human scientist. We further enhance its capabilities with multi-turn RL. 📈 Key Results: 1️⃣Consistently outperforms SOTA methods by a 6% to 35% absolute margin. 2️⃣Achieves significant performance gains after RL training. 3️⃣Demonstrates robustness to noise and generalization to out-of-domain data. 💡 Key Insights: 1️⃣ Long-horizon exploration is vital for performance. 2️⃣ Enabling agents to conduct their own data analysis is crucial. 3️⃣ An experience buffer is key for continuous optimization. 📄 Paper: arxiv.org/abs/2510.11661 💻 Code: github.com/GAIR-NLP/SR-Sc…

come see us at poster 15 🥰

The standard way to improve reasoning in LLMs is to train on long chains of thought. But these traces are often brute-force and shallow. Introducing RLAD, where models instead learn _reasoning abstractions_: concise textual strategies that guide structured exploration. 1/N🧵

1/9 🔥 NEW PAPER: "LIMI: Less is More for Agency" The Age of AI Agency demands systems that don't just think, but work: vibe coding and automated research. We used just 78 samples to beat GPT-5 by 14.1% and discovered the Agency Efficiency Principle. See details below! 📊

Excited to STARFlow has been accepted at #NeurIPS2025 as a **Spotlight** paper! Super excited and looking forward to seeing more research directions on scalable normalizing flows as an alternative to this existing diffusion world!🧐 Huge congrats to my amazing collaborators!!

FacTool has been accepted to COLM 2025 - two years after its arXiv debut! While the landscape of LLMs has changed a lot since then, tool-augmented LLMs and RAG are still among the most effective and practical approaches for detecting / mitigating hallucinations (ref: x.com/karpathy/statu…, x.com/_jasonwei/stat…) Reinforcement learning will push this even further. Imagine agents that almost never hallucinate - agents optimized to faithfully admit uncertainty, use external tools to cross-verify sources, recognize the limits of their knowledge, and resist sycophancy and lying. Crafting appropriate rewards & environments to train such agents still requires many effort (e.g., for scenarios with ambiguous facts or moral dilemmas), but effective progress can be anticipated!

Excited to share our new survey on the reasoning paradigm shift from "Think with Text" to "Think with Image"! 🧠🖼️ Our work offers a roadmap for more powerful & aligned AI. 🚀 📜 Paper: arxiv.org/pdf/2506.23918 ⭐ GitHub (400+🌟): github.com/zhaochen0110/A…

What Makes a Base Language Model Suitable for RL? Rumors in the community say RL (i.e., RLVR) on LLMs is full of “mysteries”: (1) Is the magic only happening on Qwen + Math? (2) Does the "aha moment" only spark during math reasoning? (3) Is evaluation hiding some tricky traps? (4) Is RL’s calm surface all thanks to Pre/Mid-training carrying the weight? Why does RL on LLaMA consistently underperform compared to Qwen? What makes a base model truly ready for RL scaling? What are the secrets under the hood? Due to the cost of training from scratch, we conduct extensive controlled experiments with 20B-token mid-training, systematically investigating what really matters for RL success. 💡Key insights: - High-quality math data is key to RL scaling. - QA data helps, but it depends on task similarity. - Instruction data boosts QA’s effectiveness. - More mid-training improves RL performance. Armed with these insights, we apply a two-stage (stable+decay) mid-training strategy on LLaMA, scaling up to 200B tokens—and RL performance on LLaMA now matches Qwen! To support this, we introduce MegaMath-Web-Pro-Max, a high-quality math-centric pretraining corpus. The dataset will be released soon on Hugging Face—stay tuned! 📦 huggingface.co/datasets/OctoT… Full construction details are in the paper, we hope it’s useful! arxiv.org/abs/2506.20512… Getting SOTA with a strong foundation is great 🤩, but understanding the foundation—the know-how—matters just as much. Hope this analysis inspires the community—and feel free to cite us if it helps! This work is impossible without all the brilliant co-authors @FaZhou_998 @xuefengli0301 @stefan_fee !!!

New Anthropic research: We elicit capabilities from pretrained models using no external supervision, often competitive or better than using human supervision. Using this approach, we are able to train a Claude 3.5-based assistant that beats its human-supervised counterpart.

What if an LLM could update its own weights? Meet SEAL🦭: a framework where LLMs generate their own training data (self-edits) to update their weights in response to new inputs. Self-editing is learned via RL, using the updated model’s downstream performance as reward.

To further boost the "think with images" community, we've systematically summarized the latest research in our new repository: github.com/zhaochen0110/A… 🧠🖼️Let's make LVLMs see & think! A comprehensive survey paper will be released soon! Stay tuned.

🧐When do LLMs admit their mistakes when they should know better? In our new paper, we define this behavior as retraction: the model indicates that its generated answer was wrong. LLMs can retract—but they rarely do.🤯 arxiv.org/abs/2505.16170 👇🧵

"How, exactly, could AI take over by 2027?" Introducing AI 2027: a deeply-researched scenario forecast I wrote alongside @slatestarcodex, @eli_lifland, and @thlarsen