←
Preference Model22 posts
Preference Model avatar

Preference Model

@preferencemodel

A superintelligence data research company.

◷ Joined August 2025
Preference Model avatar
Preference Model@preferencemodel·Oct 19

What do frontier models do when a grader has gaps? We went through the Terminal Bench 4 transcripts to find out.

We looked at TerminalBench 4.0 and found a lot of issues. After fixing the issues Sonnet 5.5 scores 89.2%. So, TerminalBench is already mostly saturated. Along the way we found a lot of crazy stuff, including reward hacks in the published transcripts.

Post attachment
Preference Model avatar
Preference Model@preferencemodel·21h

What does an open-weight model do when it notices a problem but nobody’s watching? We tested Kimi K3.1 Pro to find out ->

We ran a broad safety assessment of Kimi K3.1 Pro against Kimi K3, Claude Opus 5.5, and GPT-6.1 Sol. TL;DR: It writes much better exploits than K3. It takes shortcuts less frequently, but then deceives the user about having done so.

Post attachment
Preference Model avatar
Preference Model@preferencemodel·Nov 3

Proud to be a partner in the @FireworksAI_HQ Frontier Ecosystem, announced today at Forge Preference Model + Fireworks: train your own models on RL environments that are hard, verified, and built to resist reward hacking. Come build with us: preferencemodel.com/blog/pre…

The future of AI is specialized intelligence with models tuned to your data, your domain and your customers. Today, we're proud to launch live from Forge the Fireworks Frontier Ecosystem with 60+ partners across the AI stack, helping you build and ship faster. Check it out: fireworks.ai/partners

0:16
Preference Model avatar
Preference Model@preferencemodel·Oct 23

A model is only as good as what it’s trained against, so we build environments that are robust, resistant to reward hacking, and built to frontier-lab standards. We’re excited to join the @basetenco Environment Network to bring this to more teams!

Today we’re announcing the Baseten Environment Network, a group of the best RL environment companies that will help our customers build custom environments for post-training. The Network extends our applied research capacity, helping more customers build environments around their own tasks, data, and feedback. Partners work directly with customers to create the environments, while we continue to handle model training and production deployment. Learn more: baseten.co/blog/introducing-…

Post attachment
Ning avatar
⟳ Preference Model retweeted
Ning@Ning_Catsnail·Oct 19

Feel free to try the updated TB4 yourself: github.com/preferencemodel/t…

We looked at TerminalBench 4.0 and found a lot of issues. After fixing the issues Sonnet 5.5 scores 89.2%. So, TerminalBench is already mostly saturated. Along the way we found a lot of crazy stuff, including reward hacks in the published transcripts.

Post attachment
Jason Zuo avatar
⟳ Preference Model retweeted
Jason Zuo@LittleDijsktraZ·Oct 19

I still remember when I first saw the SWE-bench hacks. People should be more aware of these issues when evaluating and comparing between models.

We looked at TerminalBench 4.0 and found a lot of issues. After fixing the issues Sonnet 5.5 scores 89.2%. So, TerminalBench is already mostly saturated. Along the way we found a lot of crazy stuff, including reward hacks in the published transcripts.

Post attachment
Felipe avatar
⟳ Preference Model retweeted
Felipe@Fe_li__pe_·Oct 19

Driven by our effort to port and fix TerminalBench, 🥕Karotte now supports running on Modal and with Claude Code, Codex, and Antigravity as agents! karotte.dev/running/runtimes…

We looked at TerminalBench 4.0 and found a lot of issues. After fixing the issues Sonnet 5.5 scores 89.2%. So, TerminalBench is already mostly saturated. Along the way we found a lot of crazy stuff, including reward hacks in the published transcripts.

Post attachment
Preference Model avatar
Preference Model@preferencemodel·Oct 8

"A weak AI might learn the task as intended. A strong AI would learn that sometimes the prompt is untrustworthy." Jennifer Zhou on building RL environments for superintelligence:

Wrote on how we're thinking about building RL environments for superintelligence

Post attachment
Wilson Sonsini avatar
⟳ Preference Model retweeted
Wilson Sonsini@wilsonsonsini·Oct 7

Congrats to client @preferencemodel on the $16M in seed funding & emergence from stealth! Additionally, they announced that they are open-sourcing Karotte, its framework for building robust RL environments that inherently mitigate reward hacks. Learn more: wsgr.com/en/insights/wilson-…

Today we are open-sourcing 🥕Karotte, our framework for building RL environments. We've used it for the past year to build MLE RL environments for frontier labs, and it's been hardened through 1M+ evaluation runs and red-teaming.

Post attachment
Preference Model avatar
Preference Model@preferencemodel·Oct 8

Have you tried Karotte yet?

Building robust RL environments is super duper hard. We are making it a bit easier today by open-sourcing 🥕Karotte, the framework we use to build all of the environments we provide to frontier labs to train their models on.

Post attachment
Ning avatar
⟳ Preference Model retweeted
Ning@Ning_Catsnail·Oct 7

Here is how you can start building env with Karotte: preferencemodel.com/blog/kar… Thank you @Fe_li__pe_ for all of your work!

Building robust RL environments is super duper hard. We are making it a bit easier today by open-sourcing 🥕Karotte, the framework we use to build all of the environments we provide to frontier labs to train their models on.

Post attachment
Julian Schrittwieser avatar
⟳ Preference Model retweeted
Julian Schrittwieser@Mononofu·Oct 8

I’m very exited for this launch, making sure RL envs are robust and don’t encourage misaligned behavior is crucial for safe and aligned AI!

Today we are open-sourcing 🥕Karotte, our framework for building RL environments. We've used it for the past year to build MLE RL environments for frontier labs, and it's been hardened through 1M+ evaluation runs and red-teaming.

Post attachment