🚀 We fixed a major LLM post-training bottleneck! Our new method (TBA) combines trajectory balance with asynchronous training to speed up LLM RL 5-50x while improving results+scalability. For example, using VinePPO's GSM8K setup, we obtain +1.2% accuracy and 50x faster RL.

Looped latent reasoning models like TRM, HRM, Ouro and Huginn are great for reasoning, but they’re inefficient to train at larger scales. We fix this by post training regular language models into looped models, achieving higher accuracy on a per training FLOP basis. 📜1/7

🚨We converted pretrained LLMs into looped LLMs that can crank up performance by looping for more iterations. Our looped models surpass the performance of the pretrained models we started out with, showing that existing models benefit from increased computational depth. 📜1/9

New work on improving test-time scaling for challenging reasoning problems! Recursive Self-Aggregation (RSA) is a simple and scalable approach to combine sequential and parallel reasoning effectively. Check out @siddarthv66's thread for details. Some of my perspectives below:

Introducing 𝐑𝐞𝐜𝐮𝐫𝐬𝐢𝐯𝐞 𝐒𝐞𝐥𝐟-𝐀𝐠𝐠𝐫𝐞𝐠𝐚𝐭𝐢𝐨𝐧 (𝐑𝐒𝐀): a simple test-time method that unlocks deep thinking in LLMs by evolving & aggregating reasoning chains. 🔹 Qwen3‑4B matches capabilities of much larger models (DeepSeek‑R1, o3‑mini) 🔹 Massive gains on

Introducing RSA 🌀 (Recursive Self-Aggregation): unlocking deep thinking with test-time scaling 🔥 Qwen3-4B + RSA > DeepSeek-R1 📈 Gains across Qwen, Nemo, GPT-OSS 🏆 Benchmarks: Math • Reasoning Gym • Code ⚡ Aggregation-aware RL lets Qwen3-4B surpass o3-mini 🚀

Qwen3-4B can match DeepSeek-R1 and o3-mini (high) with ONLY test-time scaling?🤯 Introducing Recursive Self-Aggregation (RSA), a new test-time scaling method: - parallel + sequential✅ - no verifiers✅ - no scaffolding✅ Then we use aggregation-aware RL to push further!🚀 🧵👇

NO verifiers. NO Tools. Qwen3-4B-Instruct can match DeepSeek-R1 and o3-mini (high) with ONLY test-time scaling. Presenting Recursive Self-Aggregation (RSA) — the strongest test-time scaling method I know of! Then we use aggregation-aware RL to push further!! 📈📈 🧵below!

🐭🔒 LLM security is a cat-and-mouse game. Attackers adapt. Prompts mutate. Meanwhile, most defenses? 🚫 Static. Fragile. One-shot fixes. It’s time for something smarter. ⚔️ Meet AegisLLM: An agentic runtime defense that thinks, reacts, and learns — just like the attackers do.

🚀 Excited to introduce our latest work GRESO: a method that identifies and skips millions of uninformative prompts before rollout and achieves up to 2.0x wall-clock time speedup in training. More rollouts lead to better model performance, but they’re also a major bottleneck in

Here's a free/gift link to the Washington Post article about training LLMs on openly licensed text: https://t.co/fQ6aqfwUwJ. https://t.co/AUUhEYviu0

Can you train a performant language models without using unlicensed text? We are thrilled to announce the Common Pile v0.1, an 8TB dataset of openly licensed and public domain text. We train 7B models for 1T and 2T tokens and match the performance similar models like LLaMA 1&2

🥳Come chat with @bartoldson and @JainMoksh at our TBA poster at the #ICLR25 workshop on Open Science for Foundation Models (SCI-FM). The workshop will be held in EXPO Hall 4 #5 on Monday, April 28th.

Try out Trajectory Balance with Asynchrony via https://t.co/bolArKJUzf.

🚀 The code for LLM post-training with TBA is now available! Try out Trajectory Balance with Asynchrony via https://t.co/w173u63tHM. https://t.co/suu63c2nRs

🚨Concerned about visual jailbreaking attacks holding back Vision-Language Model (VLM) deployment? 🌟 Excited to announce our latest research: Double Visual Defense! TL;DR: We introduce ΔCLIP and Δ²LLaVA — the first to reconcile robust adversarial performance with

🚨 Interested in adopting Large Reasoning Models (LRMs) but concerned about safety risks? 🚨 Meet STAR-1 🌟 – A compact, high-quality safety dataset (just 1K samples!) boosting LRMs' safety by 40% with only a minimal (~1.1%) reasoning drop! 🚀 How we built STAR-1 in just 3

[LG] Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training B R. Bartoldson, S Venkatraman, J Diffenderfer, M Jain... [Lawrence Livermore National Laboratory & Mila] (2025) https://t.co/uRzhe3CdlS

New paper on synthetic pretraining! We show LMs can synthesize their own thoughts for more data-efficient pretraining, bootstrapping their capabilities on limited, task-agnostic data. We call this new paradigm “reasoning to learn”. https://t.co/Xd9sLKKVsl Here’s how it works🧵

At @Livermore_Lab, we are using AI to: ⚛️ Solve nuclear fusion 🧪 Discover critical materials 🧠 Red-team vulnerabilities All to push science forward and protect national security 🌎 Post-training LLMs at scale can unlock these advances. But even with El Capitan—the world’s

Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training TBA is a scalable RL system for LLM post-training that uses off-policy data and replay buffers with Trajectory Balance. It decouples training from search, improving speed