Sarthak Mittal

Mercury 2 is live 🚀🚀 The world’s first reasoning diffusion LLM, delivering 5x faster performance than leading speed-optimized LLMs. Watching the team turn years of research into a real product never gets old, and I’m incredibly proud of what we’ve built. We’re just getting started on what diffusion can do for language.

Recursive Self-Aggregation (RSA) + Gemini 3 Flash scores 59.31% on the public ARC-AGI-2 evals, placing it firmly among the top performers! Here are the highlights: > Outperforms Gemini DeepThink at about 1/10th the cost > Bridges the performance gap with GPT-5.2-xHigh for a similar cost > Nearly matches Poetiq while using a much simpler pipeline. Poetiq uses scaffolded refinement (often via generated code), while RSA does not Also, Gemini 3 Flash is impressive; the Gemini team cooked with this one! We're also eager to run evals with GPT-5.2 + RSA in the future (anyone with credits? :P)

Standard reinforcement learning in raw tokens is a disaster for sparse rewards! Here, we propose 𝗜𝗻𝘁𝗲𝗿𝗻𝗮𝗹 𝗥𝗟: acting on abstract actions emerging in the residual stream representation. A paradigm shift in using pretrained models to solve hard, long-horizon tasks! 🧵

LOTs of discourse lately about the correctness of the KL-regularization term used in RLVR fine-tuning of LLMs. Which estimator to use? Whether to add it to the reward or loss? What’s even the difference? 🤔 In our new preprint, we evaluate these choices empirically. 🧵 1/n

Pay attention to the hospital staff in this video; Not only there are no one around, considering this men should have just gotten full hands on deck emergency care but the 2 that are standing there have like zero care or empathy to the situation while the wife is pouring her heart out while her husband laying dead next to her because of nothing short of negligence. The man waited in an Edmonton hospital ER for 8 hours for care, and even had a heart beat on 210 which should have alarmed everyone there but no one cared. Every single person involved should be fired, never work in healthcare again and some of them jailed. He should every single sign that points to something catastrophic might happen. But "free healthcare". Ironically, this would have never happened to them in India.

I'm glad this paper of ours is getting attention. It shows that there are more efficient and effective ways for models to use their thinking tokens than generating a long uninterrupted thinking trace. Our PDR (parallel/distill/refine) orchestration gives much better final accuracy, while avoiding context bloat. (So it might be much cheaper to serve than today's thinking models.) I'm guessing that so-called "deep research models" rely on such orchestrations.

When chatting with students I always say “REINFORCE” or “score function estimator” (instead of GRPO, DrGRPO, etc.) to abstract over specific implementation details

wait so the GRPO everyone are drooling about is just REINFORCE with the baseline computed as an average over a large sample (and the usual kl regularization in llm models)?

We’ll be presenting this at the FoRLM workshop between 10:15-11:30am room 33 tomorrow! Drop by if you’d like to chat about this paper, or RL for LLMs in general (I got some juicy new insights)

NO verifiers. NO Tools. Qwen3-4B-Instruct can match DeepSeek-R1 and o3-mini (high) with ONLY test-time scaling. Presenting Recursive Self-Aggregation (RSA) — the strongest test-time scaling method I know of! Then we use aggregation-aware RL to push further!! 📈📈 🧵below!