CS @Stanford |Prev. @UCBerkeley @bespokelabsai | LLM Post-Training, Agents, Collective Intelligence

Joined September 2021
22 Photos and videos
Pinned Post
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
24
88
418
1,013,722
Richard Zhuang reposted
Flying to #ICML2026 to present Internal Data Repetition Destroys Language Models, an Oral at Foundations of Deep Gen Models Workshop! Paper: arxiv.org/abs/2606.24998 You might be curious to know what we mean by “destroys”! Pretraining is now data-constrained, and even aggressively deduplicated corpora keep some repetition. We measured what that repetition actually costs in the currency practitioners care about: compute. The answer, in the worst case, is a third of your FLOPs.
6
25
77
31,451
Another Datacomp masterpiece!
🚀New Paper arxiv.org/abs/2606.28551 Everyone obsesses over VLM architectures & training recipes. But what about the data? Presenting the latest work in the DataComp-series: a testbed for VLM data curation with 1,000 controlled experiments and some surprising lessons 👀 🧵👇
5
26
3,383
Richard Zhuang reposted
Some concrete flavors of AI fatigue (as an engineer): 1. It’s hard to wrap my head around this kind of relationship: they are your friend, your teacher, and also the thing that might replace your value and take your job. That dynamic is weirdly philosophical. There’s goodwill and respect in it, but also a quiet trace of betrayal. I still don’t think we’ve fully adapted to that. 2. When AI is too dumb, watching it apologize over and over gets irritating. When AI is too smart, it makes you feel small and slightly useless. 3. AI gives you this illusion that anything is possible, so you greedily start expanding the task. All the things you used to not be able to finish? Now you suddenly want to do all of them. The result is something like 30% accomplishment, 20% emptiness, and 50% exhaustion. 4. I often feel like I’m just one step in the loop, even though I thought I was the one directing the whole thing. Sometimes I genuinely wonder: do I really have agency, or is there just an AI in front of me dangling a carrot? 5. AI never stops producing things. And inside everything it produces, there’s always this 10% layer of junk. Understanding, filtering, and deleting that 10% takes real mental energy. But if I don’t clean it up, it feels like someone took a dump on my bed. 6. While an agent is running, it keeps saying some kinds of things I’ve never heard of before. Then I have to decide: should I do /btw, or just move on? 7. Something I thought would take 20 minutes somehow turns into an x-high / ultra run that lasts two hours, followed by three rounds of review before it finally works. And just like that, a perfectly nice evening disappears. 8. Because there’s so much AI slop everywhere, I’ve become much more restless when I read. 9. Even if you already know the lesson from the Industrial Revolution (increased efficiency doesn’t necessarily reduce working time, it just expands the boundary of work) you still can’t really stop it from happening. Because humans can’t help themselves. We just want to try.
3
1
9
470
There is a lot of room to push open-source models. Eg using Qwen3-32B base with our composed data, we score on Terminal Bench 2.0 25.60% using only 100k samples and no RL, SFT only, outperforming Nemotron-3-Nano and coming close to Nemotron-3-Super that uses way more compute.
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
4
3
16
1,722
Richard Zhuang reposted
More open data recipes are out! The team put in a ton of work figuring out what actually matters for curating agentic data, and open-sourced everything. Really excited to have played a small part in it. Check it out if you're working on agents!
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
1
2
9
648
Richard Zhuang reposted
Great work!
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
1
1
4
603
Richard Zhuang reposted
A much needed data release! Excited to tinker with the data.
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
2
11
210
27,323
Richard Zhuang reposted
Very excited to release the next project in the DataComp / OpenThoughts line of research! Like OpenThoughts we worked on post-training data, this time with a focus on agentic models.
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
4
15
73
11,703
Richard Zhuang reposted
Open cookbook for training agents that actually generalize 🧑‍🍳🤖 by the incredible OpenThoughts-Agent leads @RichardZ412 @NeginRaoof_ @etash_guha @marnezhurina @FeuerBenjamin & co.
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
1
2
13
1,286
Richard Zhuang reposted
Open data recipes for agents!
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
1
2
19
1,702
Richard Zhuang reposted
It's getting hard to keep up with the pace that these researchers are shipping - congrats @RichardZ412, publicly releasing the full stack and helping to level the playing field 💪
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
1
2
13
1,772
Richard Zhuang reposted
Most open agentic datasets target one benchmark. In compute-controlled comparisons, OpenThoughts-Agent-v2 leads at every training set size, and generalizes across seven agentic benchmarks. Check it out!
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
9
10
80
12,945
Mama I made it on the news xd
1
33
1,348
Richard Zhuang reposted
Made a little atlas to explore this! huggingface.co/spaces/davans…
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
2
1
24
2,481
Richard Zhuang reposted
Excited to have contributed to this effort! Check out this really cool work from the insanely talented Openthoughts-Agent team on how to create training data for agentic models 🤖
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
2
2
17
1,872
Richard Zhuang reposted
Congrats to the OpenThoughts team for creating OpenThoughts-Agent and OpenThinkerAgent-32B! What's cool is that everything is open!
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
1
3
28
2,493
Richard Zhuang reposted
This is what open science at its best looks like. When Richard first showed me their leaderboard, I was blown away: not just by the hundreds of eval runs, but by the level of care and transparency. Every single eval is clearly labeled with: • Standard deviation over 3 runs • Sandbox CPU, memory, & storage configs • Timeout multipliers • FULL eval trajectories on HF! At a time when self-reported scores can vary by up to 10% for the same model on key evals, transparent and reliable open science is more critical than ever. Check it out: ot-agent-leaderboard.replit.…
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
1
2
11
1,356
Richard Zhuang reposted
Excited that OpenThoughts-Agent is finally out! We share a fully-open recipe to train strong agents for terminal use and coding 🚀
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
2
2
13
910
Richard Zhuang reposted
sharing this great work from @RichardZ412 and dozens of researchers from @berkeley_ai @stanfordnlp and many more — OpenThoughts-Agent It addresses the largest gap for small model agentic training today by a fully open data curation pipeline
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
2
1
11
1,823
Richard Zhuang reposted
Excited to announce OpenThoughts-Agent: We are releasing 100k environments and trajectories, A post-trained 32B Qwen-3 based model achieving 26% on Terminal Bench 2.0 and 100s of ablations for scientific understanding of data curation for agents.
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
1
5
34
513,717