MohammadHossein Rezaei (@mhrezaeics) header
MohammadHossein Rezaei (@mhrezaeics) profile picture

I think, therefore I am @thinkymachines. Prev: @ScaleAILabs @StanfordNLP @UArizona

San Francisco, CA
Joined November 2022
Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)Photo by MohammadHossein Rezaei (@mhrezaeics)
say hi to efficient agentic verifiers:
🤖 Agent: “Done! $30 promo applied. Pizza’s on the way 🍕” 💳 Bank statement: Charged full price 🤦‍♂️ Now imagine rewarding that run as a success during post-training. Meet Sauron: an evidence-first agentic verifier that makes evaluation faster, cheaper, and more reliable through shared-evidence clustering, targeted trace/state queries, and independent claim verification. 📄 arxiv.org/abs/2610.17724
17
1,998
MohammadHossein Rezaei retweeted
Not All Privileged Information Is Equal: What Makes OPSD Actually Work? We find that information-rich PI, such as environment observations and mistake analysis, outperforms PI that only provides ground-truth solutions. Better PI leads to better post-training performance. But can we predict which PI will work best before spending compute on post-training? We investigated three offline metrics: 🔹 Utility: A model's performance with PI can be misleading due to information leakage. 🔹 Reverse KL: Larger PI-induced distribution shifts often correlate with better post-training performance, but incorrect PI can cause equally large shifts. 🔹 Error-step KL contrast: Measuring whether PI changes the model's behavior specifically at error steps better captures its effectiveness and helps select stronger PI. What types of privileged information have you found most effective in your OPSD experiments? 📄 Paper: arxiv.org/abs/2610.22512
6
28
2
88
8,224
MohammadHossein Rezaei retweeted
🤖 Agent: “Done! $30 promo applied. Pizza’s on the way 🍕” 💳 Bank statement: Charged full price 🤦‍♂️ Now imagine rewarding that run as a success during post-training. Meet Sauron: an evidence-first agentic verifier that makes evaluation faster, cheaper, and more reliable through shared-evidence clustering, targeted trace/state queries, and independent claim verification. 📄 arxiv.org/abs/2610.17724
3
10
6
32
4,166
I've joined @thinkymachines! Excited to be part of this incredible team building AI that extends human will and judgment. thinkingmachines.ai/#join-us
40
5
363
16,580
feel perfectly sandboxed
11
964
Reward Hacking in Rubric-Based RL got accepted to NeurIPS! If you're working on rubrics, take a look at some of our recent work:
1/ Using rubrics (a.k.a. checklists) in RL training is now standard for open-ended tasks without final verifiable result. However, rubric rewards are still proxy rewards that can get hacked during RL training. We study when rubric-based RL genuinely improves models vs. teaches them to hack the verifier/rubric. We quantify this through exploitation, analyze the failure modes, and introduce a verifier-free metric. arxiv.org/abs/2605.12474
6
5
2
113
11,009
Glad to have worked with many cool and talented people on these projects: @nas_mahmoud_ @_yunzhong @utkarsh4430 @wzihao12 @anisha_gunjal @afeyzaakyurek and many more
1
5
131
MohammadHossein Rezaei retweeted
🎉 Excited to share that “Reward Hacking in Rubric-Based Reinforcement Learning” has been accepted to #NeurIPS2026! Reward hacking in post-training is a particularly interesting and important problem to work on. We study how models can learn to game rubric-based rewards during post-training, satisfying the rubric without necessarily improving the broader response quality. We characterize these failures at the criterion level and show that stronger verifiers help, but don’t fully eliminate reward hacking. Link: arxiv.org/abs/2605.12474 Check out some of our recent post-training research projects at @ScaleAILabs: RGSD: arxiv.org/html/2606.12507 POW3R: arxiv.org/abs/2605.20164
4
8
29
1,035
MohammadHossein Rezaei retweeted
brief break from regularly scheduled muse programming, but great progress by @scale_AI on RSI benchmark
600+ proposals in 🤯 We're reaching out to top contributors to start building in our public repo. We’ve also welcomed @SchmidhuberAI, a pioneer of RSI, as a senior advisor to RSI Bench. New blog on our setup and verification pipeline: rsi-benchmark.com/blog/verif…
16
15
1
368
68,634
As we've seen in peer reviews and all the benchmark auditings going on, I think we're all convinced that verification is not easier than generation. We developed a verification pipeline for RSI Bench: rsi-benchmark.com/blog/verif… Also very excited to work with @SchmidhuberAI!
600+ proposals in 🤯 We're reaching out to top contributors to start building in our public repo. We’ve also welcomed @SchmidhuberAI, a pioneer of RSI, as a senior advisor to RSI Bench. New blog on our setup and verification pipeline: rsi-benchmark.com/blog/verif…
2
17
2,445
Also, shoutout to TerminalBench (@ryan_marten @alexgshaw @harborframework et al) for their verification workflow that we used as a starting point for RSI task adaptations.
4
326
🚀 Presenting AI Conference Scaling Laws. At the current rate, ICLR should hit 1M submissions is 3-5 years...
ICLR 2027 has received more submissions than all previous years (2013–2026) combined
11
15
5
183
23,296
Exactly! Not only @SchmidhuberAI solved NS in 1993 but he also pioneered Recursive Self-Improvement in 1987, and we're fortunate to onboard him as a senior advisor to RSI-Benchmark.com 😃
Thanks! 😀 Indeed, I solved neural self-reference (NS) in 1993 [1][2]. [1] JS. A self-referential weight matrix. ICANN 1993, Amsterdam, p 446-451 [2] K. Irie, I. Schlag, R. Csordas, JS. A Modern Self-Referential Weight Matrix That Learns to Modify Itself. ICML 2022
1
1
10
2,018
MohammadHossein Rezaei retweeted
The response to RSI Bench has been incredible. Hundreds of researchers and builders filled out our interest form, showing just how much energy there is around pushing the frontier of what it means for an agent to do research. We’re incredibly excited to bring together some of the most ambitious researchers in the community to build this frontier benchmark with us, and to keep iterating on it as research agents advance. Learn how to contribute here: rsi-benchmark.com/contribute
📣Call for contributions + co-authorship! RSI Bench is our ongoing effort to evaluate whether AI agents can develop the capabilities required to advance AI R&D. Got a frontier AI research problem on your mind? Submit it to RSI Bench, and set the standard the whole industry uses to measure progress! Contributors receive: → Co-authorship on the RSI Bench research paper → $2,000 per accepted task for the initial 50 tasks → Modal compute credits to build + iterate → Access to the RSI Bench research community Register below for full requirements.
7
9
62
8,367
MohammadHossein Rezaei retweeted
Clinical AI agents can be right for the wrong reasons. Real clinical work requires more than just medical knowledge. Agents need to navigate longitudinal records, reconcile evidence, ground conclusions, and know when to abstain. We introduce CliniCARE-Bench, a new benchmark built to test exactly that. 🧵
10
23
5
122
17,731
Launching rsi-benchmark.com: The work of AI R&D has always belonged to humans. For the first time, though, it no longer seems certain that it always will. Recursive self-improvement is within a line of sight. It may still be far, but it is close enough that we should start measuring it.
20
52
14
474
189,846
Here's our initial observations: teapawt.net/nas_mahmoud_/status/20…
Launching rsi-benchmark.com: We ran Claude Opus 5 and GPT 5.6 Sol on an initial set of four tasks, with a given setup, instructions, and fixed compute. Here’s what we observed:
1
1
11
3,831