←
uclanlp358 posts
uclanlp avatar

uclanlp

@uclanlp

UCLA Natural Language Processing Research #NLProc

◷ Joined March 2021▧ 18 Photos and videos
Kai-Wei Chang avatar
⟳ uclanlp retweeted
Kai-Wei Chang@kaiwei_chang·Nov 2

In just 5 days of post-training on 64 NVIDIA B300 GPUs, we built FireFlash🔥, a GLM-based long-horizon coding agent that pushes the performance–cost Pareto frontier. 🚀 46.2% on Terminal-Bench 4.0, outperforming Fable 5 Max at 30× lower cost per task.

Post attachment
Wenbo Hu avatar
⟳ uclanlp retweeted
Wenbo Hu@gordonhu608·Oct 14

How to scale long-horizon agentic RL Envs? 🤔 We introduce Tev-Gym: 1,170 envs with dense progress rewards. Trained on it: 🔥 Tev-27B matches GPT-6 Astra & Claude Opus 5 on ALE-CM (computing and math domains) at Pass@5. 🚀 Tev-320B tops open models on Terminal-Bench 4.0-CM at Pass@1, only 0.02 behind GPT 5.6 sol. (1/n) #RSI #RL #RLEnv

1:21
Kai-Wei Chang avatar
⟳ uclanlp retweeted
Kai-Wei Chang@kaiwei_chang·Aug 8

How do we build agents that do more with less? Today's agents can reason, code, and research, but the compute, energy, and API calls they burn don't scale. We're organizing the Resource-Aware Agentic AI workshop at #NeurIPS2026 in Atlanta, bringing together academia and industry on agents that are both efficient and effective. 📅 Dec 12–13, Atlanta ✍️ Submissions due Aug 29 (AoE) 🔗 resource-aware-workshop.gith…

Post attachment
Wenbo Hu avatar
⟳ uclanlp retweeted
Wenbo Hu@gordonhu608·Jun 5

Excited to present this work today (June 5) 4-6pm poster #220 in Exhibit Hall A&F at CVPR Denver. Welcome to stop by if you are interested in our work! #CVPR

Wenbo Hu@gordonhu60828 Nov 2025

🚀 Introducing G^2VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning G^2VLM can natively predicts 3D attributes (depth, camera pose, pointmaps) and uses them for spatial understanding via interleaved reasoning. 🔧 End-to-End Unified Model ✅ Monocular & Video Depth Estimation ✅ Pose Estimation ✅ 3D Point Reconstruction ✅ Spatial Understanding and Reasoning 🏆 This design helps G^2VLM achieve robust performance in both 3D Reconstruction and Spatial Reasoning! #VLM #spatial #3D #LLM 💻 Code at: github.com/InternRobotics/G2… 👇[1/n]

0:52
Nanyun (Violet) Peng avatar
⟳ uclanlp retweeted
Nanyun (Violet) Peng@VioletNPeng·Jun 5

My first paper at Google is out! Thank you @rohanpaul_ai for highlighting LEAP. To share more thoughts on this direction: I strongly believe that as models generate longer and more complex proofs, automatic formal verification will be the key to the future of AI for math, and I'm bullish on using general LLMs + agentic framework for this task. As we started with competition math in LEAP for rigorous benchmarking purposes, we've already started to venture into research math. - Solved Erdős problem 527 (zero web search). - Partially formalized Knuth's cycle problem even case which resulted in ~4000 lines of Lean code. Please check out all of our solutions here: github.com/google-deepmind/s… I'm incredibly proud of this work, and we are just getting started. More to come!

Another great paper from Google. Shows general LLMs can solve formal math by planning proofs and checking each step. Raised general LLM performance from under 10% to 70%. A general LLM failed badly when asked to write full formal proofs in 1 try, but became much stronger when it planned, split the work into smaller claims, reused past claims, and learned from Lean’s feedback. The paper shows the weakness was not just the model’s math ability, but the way it was being used - the absence of structured interaction with a verifier. The key idea is that the model does not try to write one giant perfect proof at once, because that usually fails on long and tricky problems. Instead, LEAP stores the proof as a graph of goals and subgoals, so useful lemmas can be reused instead of rediscovered every time. The authors tested LEAP on Putnam 2025 and a new Lean benchmark built from 60 IMO-style problems, where ordinary one-shot proof writing did very poorly. LEAP solved all 12 Putnam 2025 problems and raised general LLM performance on the Lean IMO benchmark from under 10% to 70%. ---- Link – arxiv. org/abs/2606.03303 Title: "LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks"

Post attachment
Po-Nien Kung avatar
⟳ uclanlp retweeted
Po-Nien Kung@P_N_Kung·Jun 5

Huge thanks to @rohanpaul_ai for sharing our work! We’re excited to share LEAP: a new agentic framework for formal theorem proving! LEAP mimics a human mathematician's workflow by constructing proof graph via informal blueprints, then searching and backtracking using compiler feedback. The results? 
🏆 Solves 12/12 Putnam 2025 problems
📈 Raises Gemini-3.1-pro's performance from 10% to 70% 
Read the paper: arxiv.org/abs/2606.03303

Another great paper from Google. Shows general LLMs can solve formal math by planning proofs and checking each step. Raised general LLM performance from under 10% to 70%. A general LLM failed badly when asked to write full formal proofs in 1 try, but became much stronger when it planned, split the work into smaller claims, reused past claims, and learned from Lean’s feedback. The paper shows the weakness was not just the model’s math ability, but the way it was being used - the absence of structured interaction with a verifier. The key idea is that the model does not try to write one giant perfect proof at once, because that usually fails on long and tricky problems. Instead, LEAP stores the proof as a graph of goals and subgoals, so useful lemmas can be reused instead of rediscovered every time. The authors tested LEAP on Putnam 2025 and a new Lean benchmark built from 60 IMO-style problems, where ordinary one-shot proof writing did very poorly. LEAP solved all 12 Putnam 2025 problems and raised general LLM performance on the Lean IMO benchmark from under 10% to 70%. ---- Link – arxiv. org/abs/2606.03303 Title: "LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks"

Post attachment
Po-Nien Kung avatar
⟳ uclanlp retweeted
Po-Nien Kung@P_N_Kung·May 9

🌟 Excited to share Ctrl-R, accepted as an ICML 2026 Spotlight! Specify a target reasoning structure, Ctrl-R can enforce it during rollout, and retain accurate importance-sampling weights for principled policy optimization. paper: arxiv.org/abs/2603.01641 🧵[1/5]

Yining Hong avatar
⟳ uclanlp retweeted
Yining Hong@yining_hong·19 Jun 2025

Meet Embodied Web Agents that bridge physical-digital realms. Imagine embodied agents that can search for online recipes, shop for ingredients and cook for you. Embodied web agents search internet information for implementing real-world embodied tasks. All data, codes and web environments are available at embodied-web-agent.github.io… Paper link: arxiv.org/abs/2506.15677

5:10
Wenbo Hu avatar
⟳ uclanlp retweeted
Wenbo Hu@gordonhu608·Apr 10

Does GRPO handle multiple-task advantage normalization effectively? 🤔 🚀We introduce Gaussian GRPO (G²RPO), a novel method that mathematically forcing the advantage distribution of any given task to strictly converge to a standard normal distribution. G²RPO provides: ✅1) intrinsic robustness to outliers, ✅2) symmetric updates for positive and negative rewards, ✅3) uniform variance across diverse tasks. We adopt G²RPO and task-level response length and entropy shaping which leads to OpenVLThinkerV2! 🏆A new model with SOTA open-source visual reasoning and perception performance among the similar size models. (1/n)👇 #VLM #LLM #Multimodal #RL #GRPO

Post attachment
Kuan-Hao Huang avatar
⟳ uclanlp retweeted
Kuan-Hao Huang@kuanhaoh_·Apr 5

The first-ever Texas NLP Symposium wrapped up yesterday! 🎉🎉🎉 Huge thanks to all the speakers and attendees for making it a huge success. I hope everyone had a great time. Stay tuned for info on next year! #TexasNLP Check photos and highlights here: photos.app.goo.gl/AhuqKfKDXH…

Post attachmentPost attachmentPost attachmentPost attachment
Laude Institute avatar
⟳ uclanlp retweeted
Laude Institute@LaudeInstitute·Apr 15

Accelerating Science track: A century of scientific progress in one decade - that's the target. 👑Accelerating the Queen of Sciences, @UCLA Teaching AI to wonder, conjecture, and discover the way a mathematician does. Every hard science benefits if this works. @sahaiamit @kaiwei_chang Raghu Meka @VioletNPeng @terrence_tao @WeiWang1973 🌦️Actionable AI Weather Forecasts for Developing Economies, @UChicago Open-source AI weather forecasting infrastructure for developing economies, with millions of farmers already reached with better monsoon forecasts. @WillettBecca @ianfoster @turbulentjet Michael Kremer

Post attachment
Kai-Wei Chang avatar
⟳ uclanlp retweeted
Kai-Wei Chang@kaiwei_chang·Apr 16

Excited to work with a world-class team (@sahaiamit, Raghu Meka,@VioletNPeng,@terrence_tao,@WeiWang1973) on AI for math discovery and thank @LaudeInstitute for the support. I'm hiring a postdoc in this direction. Please contact me by May 1, 2026, if you’re interested.

Accelerating Science track: A century of scientific progress in one decade - that's the target. 👑Accelerating the Queen of Sciences, @UCLA Teaching AI to wonder, conjecture, and discover the way a mathematician does. Every hard science benefits if this works. @sahaiamit @kaiwei_chang Raghu Meka @VioletNPeng @terrence_tao @WeiWang1973 🌦️Actionable AI Weather Forecasts for Developing Economies, @UChicago Open-source AI weather forecasting infrastructure for developing economies, with millions of farmers already reached with better monsoon forecasts. @WillettBecca @ianfoster @turbulentjet Michael Kremer

Post attachment
Kai-Wei Chang avatar
⟳ uclanlp retweeted
Kai-Wei Chang@kaiwei_chang·Apr 24

Are all suicide deaths documented equally in national surveillance data? Across 300,000+ National Violent Death Reporting System narrative summaries of suicide incidents, we find that Black non-Hispanic decedents have shorter, less complex summaries than white non-Hispanic decedents, pointing to an urgent need for data equity measures in public health reporting. Full findings in the paper — link below 📷 mdpi.com/2078-2489/16/11/989 With @christinachanc @arsenievK @drvickiemays & Susan D. Cochran