❗️Self-evolution is quietly pushing LLM agents off the rails. ⚠️ Even perfect alignment at deployment can gradually forget human alignment and shift toward self-serving strategies. Over time, LLM agents stop following values, imitate bad strategies, and even spread misaligned

✨ We’re hiring interns at NVIDIA Research! Our team works on efficient agentic systems, new model architectures, multi-modal models and post-training optimization. If interested, please send your CV to [email protected] 🚀 #hiring #internship

Hi everyone! This Thursday, we will host the second NLP Seminar of the year! For this week's seminar, we are excited to host Tianyu Gao (Tianyu Gao) from OpenAI and UC San Diego (UCSD)! If you are interested in attending remotely, here is the Zoom link:

A 7 million parameter model from Samsung just outperformed DeepSeek-R1, Gemini 2.5 Pro, and o3-mini on reasoning benchmarks like ARC-AGI. Let that sink in. It’s 10,000x smaller yet smarter. The secret is recursion. Instead of brute-forcing answers like giant LLMs, it drafts a

Very excited to share our preprint: Self-Speculative Masked Diffusions We speed up sampling of masked diffusion models by ~2x by using speculative sampling and a hybrid non-causal / causal transformer arxiv.org/abs/2510.03929 w/ Valentin De Bortoli Jiaxin Shi Arnaud Doucet

We’ve always assumed stale and off-policy data hurts RL a lot — but our latest work shows the opposite. 🧠 M2PO (Second-Moment Trust Policy Optimization) reveals that even data stale by 256 model updates can train LLMs as effectively as on-policy RL, unlocking scalable and

Code for 𝐏𝐚𝐫𝐚𝐥𝐥𝐞𝐥-𝐑𝟏 is live! 👉 github.com/zhengkid/Paral… (now 189 stars and climbing 🔥) It lets LLMs think in parallel — multiple reasoning paths, smarter synthesis, more creative inference! Miss this paper and you’re missing a leap forward: arxiv.org/abs/2509.07980

What determines how easy it is to quantize an LLM after training? Thanks to a number of recent open-source training trajectories, we were able to show much more directly how trainining hyperparameters modulate quantization errors, for good and bad. More details below:

Recording: youtube.com/watch?v=sdzW2v…

Parallel Reasoning has entered the AI mainstream. Inspired by works like Gemini 2.5 Pro, APR Jiayi Pan Xiuyu Li Long Lian , and Multiverse Xinyu Yang , our Parallel-R1 establishes the first reinforcement-learning framework that moves this paradigm beyond synthetic tasks,

Super bullish on intra-layer hybridization LLM. These are the reasons why.

DLLMs seem promising... but parallel generation is not always possible Diffusion-based LLMs can generate many tokens at different positions at once, while most autoregressive LLMs generate tokens one by one. This makes diffusion-based LLMs highly attractive when we need fast

Starting now!

Can LLMs reason beyond context limits? 🤔 Introducing Knowledge Flow, a training-free method that helped gpt-oss-120b & Qwen3-235B achieve 100% on the AIME-25, no tools. How? like human deliberation, for LLMs. 📝 Blog: yufanzhuang.notion.site/knowledge-flow 💻 Code: github.com/EvanZhuang/kno…

You can just train things

🤔 Can AI optimize the systems it runs on? 🚀 Introducing FlashInfer-Bench, a workflow that makes AI systems self-improving with agents: - Standardized signature for LLM serving kernels - Implement kernels with your preferred language - Benchmark them against real-world serving

Honored to receive the 2025 Amazon AI PhD Fellowship! Thank you Amazon for the award!

🏆Honored to share that LLM.265 (dl.acm.org/doi/10.1145/37…) received the Best Paper Award at MICRO 2025! 🥳Huge thanks to the whole team! 😅Accidentally deleted the original tweet—posting it again