Twitxy Saved

Negin Raoof

Verified account This account is verified. Learn more

@NeginRaoof_

Member of Technical Staff @AnthropicAI, PhD @Berkeley_EECS
neginraoof.com Joined September 2022
534 Following 1.4K Followers 393 Posts

Latest posts 5

Pinned post

Negin Raoof Verified account @NeginRaoof_
How can we make a better TerminalBench agent?
Today, we are announcing the OpenThoughts-Agent project.
OpenThoughts-Agent v1 is the first TerminalBench agent trained on fully open curated SFT and RL environments.
OpenThinker-Agent-v1 is the strongest model of its size on TerminalBench, and sets a new bar on our newly released OpenThoughts-TB-Dev benchmark. (1/n)
17 replies 75 reposts 286 likes 129K views View on X
Negin Raoof Verified account @NeginRaoof_
Thanks to @LaudeInstitute for their continuous support.
Laude Institute @LaudeInstitute Feb 26
OpenThoughts-Agent/@etash_guha @ryan_marten @FeuerBenjamin @NeginRaoof_ @tyler_griggs_ @alexgshaw @Mike_A_Merril @lschmidt3 (@UCBerkeley) / Agent research has been hard to reproduce and compare. OpenThoughts releases an end-to-end open setup (curated data, real environments, RL loops) so anyone can reliably build, compare, and improve agents in the open.
View quoted post on X
Negin Raoof Verified account @NeginRaoof_
🔥
Snorkel AI @ NeurIPS 2026 Verified account @SnorkelAI Feb 20
We’re excited to be collaborating with the OpenThoughts Agent team and @bespokelabsai on the release of OpenThoughts-TBLite.
OpenThoughts-TBLite is a 100-task benchmark calibrated for stronger iteration signal — especially for non-frontier terminal agents.

Tasks have been curated to preserve high correlation with TB2 while offering:

- More usable reward signal for smaller models
- Faster iteration during model training
- Cleaner feedback during development
View quoted post on X
Negin Raoof Verified account @NeginRaoof_
OpenThoughts-TBLite is out!
A new benchmark designed to provide meaningful signal while iterating on small models that can’t yet solve Terminal Bench’s hardest tasks.
Huge thanks to @SnorkelAI, @bespokelabsai, @harborframework, @ryan_marten, @alexgshaw, and the OpenThoughts-Agent team, especially @kobe0938 and @RichardZ412 , for helping curate this set.
Richard Zhuang Verified account @RichardZ412 Feb 20
Terminal-Bench is a leading benchmark for agents. Unfortunately it’s hard: most small coding agents get very low scores on TB2, so training/system ablations look flat - you can't tell what's working.

Announcing OpenThoughts-TBLite - 100 curated TB2-style tasks, difficulty-calibrated so even 8B models can make progress. It's designed to give researchers measurable signal during development, providing faster feedback for experimental iteration while closely tracking true TB2 performance🧵
View quoted post on X