Kevin Li

Attention is all you need; at least the matrices are, if you want to distill Transformers into alternative architectures, like Mamba, with our new distillation method: MOHAWK! We also release a fully subquadratic, performant 1.5B model distilled from Phi-1.5 with only 3B tokens!

Official results are in - Gemini achieved gold-medal level in the International Mathematical Olympiad! 🏆 An advanced version was able to solve 5 out of 6 problems. Incredible progress - huge congrats to @lmthang and the team! deepmind.google/discover/blog/…

1/N I’m excited to share that our latest @OpenAI experimental reasoning LLM has achieved a longstanding grand challenge in AI: gold medal-level performance on the world’s most prestigious math competition—the International Math Olympiad (IMO).

1/So much of privacy research is designing post-hoc methods to make models mem. free. It’s time we turn that around with architectural changes. Excited to add Memorization Sinks to the transformer architecture this #ICML2025 to isolate memorization during LLM training🧵

At #ICML2025, I am super excited to introduce STAMP. This is a marriage b/w dataset inference & watermarking that finally(!) lets creators PROVE their content was used to train LLMs🔍 Its a MAJOR push taking the academic problem into real world. w/Saksham Rastogi @danish037 🧵

1/6 Retrieval is supposed to improve generation in RAG systems. But in practice, adding more documents can hurt performance, even when relevant ones are retrieved. We introduce RAGGED, a framework to measure and diagnose when retrieval helps and when it hurts.

Tokenization has been the final barrier to truly end-to-end language models. We developed the H-Net: a hierarchical network that replaces tokenization with a dynamic chunking process directly inside the model, automatically discovering and operating over meaningful units of data

I converted one of my favorite talks I've given over the past year into a blog post. "On the Tradeoffs of SSMs and Transformers" (or: tokens are bullshit) In a few days, we'll release what I believe is the next major advance for architectures.

Super excited to share SmolLM3, a new strong 3B model. SmolLM3 is fully open, we share the recipe, the dataset, the training codebase and much more! > Train on 11T token on 384 H100 for 220k GPU hours > Support long context up to 128k thanks to NoPE and intra document masking >…

Despite theoretically handling long contexts, existing recurrent models still fall short: they may fail to generalize past the training length. We show a simple and general fix which enables length generalization in up to 256k sequences, with no need to change the architectures!

✨ Did you know that NOT using all generated rollouts in GRPO can boost your reasoning LLM? Meet PODS! We down-sample rollouts and train on just a fraction, delivering notable gains over vanilla GRPO. (1/7)

Big news! 🎉 I’m joining UNC-Chapel Hill as an Assistant Professor in Computer Science starting next year! Before that, I’ll be spending time @OpenAI working on LLM privacy. @unccs @uncnlp

What if LLMs could learn your habits and preferences well enough (across any context!) to anticipate your needs? In a new paper, we present the General User Model (GUM): a model of you built from just your everyday computer use. 🧵

When we put lots of text (eg a code repo) into LLM context, cost soars b/c of the KV cache’s size. What if we trained a smaller KV cache for our documents offline? Using a test-time training recipe we call self-study, we find that this can reduce cache memory on avg 39x…

🚨 Sharing our new #ACL2025NLP main paper! 🎥 Deploying video VLMs at scale? Inference compute is your bottleneck. We study how to optimally allocate inference FLOPs across LLM size, frame count, and visual tokens. 💡 Large-scale training sweeps (~100k A100 hrs) 📊 Parametric…

📢 New paper on creativity & multi-token prediction! We design minimal open-ended tasks to argue: → LLMs are limited in creativity since they learn to predict the next token → creativity can be improved via multi-token learning & injecting noise ("seed-conditioning" 🌱) 1/ 🧵

We've been thinking about what the "ideal" architecture should look like in the era where inference is driving AI progress. GTA & GLA are steps in this direction: attention variants tailored for inference: high arithmetic intensity (make GPUs go brr even during decoding), easy to…

One fundamental issue with RL – whether it’s for robots or LLMs – is how hard it is to get rewards. For LLM reasoning, we need ground-truth labels to verify answers. We found that maximizing confidence alone allows LLMs to improve their reasoning with RL!

RL with verifiable reward has shown impressive results in improving LLM reasoning, but what can we do when we do not have ground truth answers? Introducing Self-Rewarding Training (SRT): where language models provide their own reward for RL training! 🧵 1/n

🚨 New work: We rethink how we finetune safer LLMs — not by filtering after the generation, but by tracking safety risk token by token during training. We repurpose guardrail models like 🛡️ Llama Guard and Granite Guardian to score evolving risk across each response 📉 — giving…

📢 (1/16) Introducing PaTH 🛣️ — a RoPE-free contextualized position encoding scheme, built for stronger state tracking, better extrapolation, and hardware-efficient training. PaTH outperforms RoPE across short and long language modeling benchmarks arxiv.org/abs/2505.16381