Jithendaraa Subramanian

Reasoning for materials science @ToyotaResearch | Prev: @mcgillu, @Mila_Quebec, @AmazonScience, @ServiceNowRSRCH

[Nanda H Krishna @nandahkrishna]

This new post from @neuralink combines POYO-style spike tokenisation with a Mamba backbone, akin to our POSSM architecture, and explores self-supervised learning for these models, as we did recently with MOJO. Exciting to see these ideas supporting real-time BCI use! 🧵

[Jithendaraa Subramanian @jithendaraa]

@felangelov out of curiosity, how is this being calculated?

[Jithendaraa Subramanian @jithendaraa]

Wonder what these "bad patterns" are 👀

[Fuli Luo @_LuoFuli]

Nearly half a year of silence. We spent it studying one problem: how far RL can scale.

MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.

Streaming the run: https://t.co/ZSxahzJRju

[Jithendaraa Subramanian @jithendaraa]

"Here look at my run as it trains"

Love it when researchers stream runs

[Fuli Luo @_LuoFuli]

Nearly half a year of silence. We spent it studying one problem: how far RL can scale.

MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.

Streaming the run: https://t.co/ZSxahzJRju

[Sam Altman @sama]

GPT-6 will be renamed GPT-6-7, you're welcome

[Jithendaraa Subramanian @jithendaraa]

@hamishivi Thanks for the comprehensive reply! Value-pretraining seems to be (atleast one of) the best reasons to reconcile this - vapo ablations show it is the single most important part. Would be fun to try out :)

PS: I read sae incorrectly as sao, mb

[Jithendaraa Subramanian @jithendaraa]

@hamishivi Also, SAE << GRPO from the plots in your tweet compared to the original SAE paper (Qwen3-30B; see fig below) which beats GRPO by a significant margin on a bunch of tasks (including AIME '25)

[Jithendaraa Subramanian @jithendaraa]

These plots suggest GRPO is doing much better than VAPO, but from the VAPO paper the converse is true. How to reconcile this?

Is it just model size (Qwen 2.5 32B in the VAPO paper vs Qwen 3 4b in these plots) or are there other confounders? If the former, a model size ablation would be needed to understand the relative performances of GRPO/VAPO/SAE + * variants?

[Jerry Tworek @MillionInt]

When working on RL with llms you start realizing that you’re rediscovering a lot of human nature and social constructs from first principles.

Words of encouragement actually matter!

[Jithendaraa Subramanian @jithendaraa]

@paularambles wonder if this works even better if you have a swarm of agents for the sol purpose of cheerleading the main agent lol

[Jithendaraa Subramanian @jithendaraa]

@yacineMTB better along what axis? more human-like chats or gets the job done in cases Sol can't? or purely inference speed?

[Jithendaraa Subramanian @jithendaraa]

🥶🥶🥶

[Prime Intellect @PrimeIntellect]

Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.

[Jithendaraa Subramanian @jithendaraa]

🛫

[Dmitry Rybin @DmitryRybin1]

Dinitz-Garg-Goemans conjecture is false. This graph theory problem was open for ~30 years.

The graph below has fractional flow cost 58. Any unsplittable flow (with capacity violation <=15) has cost at least 60.

Chat with GPT 5.6 Pro where this was found: https://t.co/Oi2PQoab2h

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[kalomaze @kalomaze]

genuinely what is the target demographic

[will brown @willcb]

you can now train:
- a 100B-parameter reasoning model
- for 40-turn SWE agent tasks
- in your own coding harness
- for 1000 RL steps
- on just 6 H200 nodes
- in under 2 days

infra co-design is magical

[Jithendaraa Subramanian @jithendaraa]

@veds_12 @AIatMeta 🐐

[Jithendaraa Subramanian @jithendaraa]

🥶

[samsja @samsja19]

prime-rl can now train 1T parameters MoE blazingly fast, under 5 minutes per step, or 1k steps in ~3 days

To achieve this we shipped in our latest prime-rl 0.6.0:

* inference: wide-ep, fp8 inference, llm-d router, mooncake, kv cache cpu offloading

* training: fsdp2, deep-ep expert parallelism, dsa cp, fp8 training, router replay

* agentic rollout: we rewrote the core of our rollout orchestrator for better scalability

support for glm5, kimi, nemotron, ...,

prime-rl is open source but also end to end optimized to run on our dedicated RL infra and compute layer

[Jithendaraa Subramanian @jithendaraa]

Worth rewatching Leopold on Dwarkesh Podcast in 2026: https://t.co/cxDlZvKBRT

[Jithendaraa Subramanian @jithendaraa]

@giffmana @xeophon @gabriberton Rank??