Samarth P✈️ICML'26🇰🇷

learn build repeat. reasoning @MSFTResearch. open source @HyperClusterAI. prev. AI for quantum @IBMResearch. document vision @cognitivelab_ai

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

I'm at #ICML2026 in Seoul 🇰🇷 — presenting tomorrow!

ReviewArena: A Large-Scale Cross-Conference Dataset & Benchmark for LLM Peer Review, a spotlight at the AI for Science workshop.

🎤 Talk: Hall C, 14:00–14:10 KST
📌 Poster: Hall A, board #416

Come by!

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

I'll be in Seoul this week for #ICML2026 🇰🇷

will present our work:

ReviewArena: A Large-Scale Cross-Conference Dataset and Benchmark for LLM Peer Review

accepted as a Spotlight at the ICML 2026 AI for Science Workshop.

Say hi if you're around!
@icmlconf

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

"no birds fly so high that they don't come down for water"

read this quote today, had to tweet and share it

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

Chinmay is genuinely one of the most cracked people I've worked with

the depth of his knowledge in RL for LLMs and post-training is insane, learnt so much from him over the 6 months at MSR

anyone would be lucky to have him on their team

[Chinmay @ChinmayKak]

On a more personal note:
I am wrapping up my internship and looking for full-time roles in post-training and adjacent areas. At MSR I worked on agentic RL for general tool-calling on benchmarks like BFCL, Taubench, and MCPBench and on self-distillation and synthetic-environment data curation and training.
More of my work can be found on my github and website, linking below.
Please RT/QT for helping w the algo:)

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

@ChinmayKak Lot more blogs to come

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

this is really cool

I am a huge fan of energy/physics based approaches like this for AI

this can be extended to so many more areas

[Unconventional AI @unconvAI]

Introducing our first model, Un-0!

We trained an image generator powered by a backbone of coupled oscillators in place of a more traditional conventional neural network.

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

@ChinmayKak @abhinav_java just one more week left to play with these guys😔

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

@adithya_s_k @huggingface lesgoo

[Manuscripts & Inscriptions Digitization Foundation @midf_org]

2,000 manuscripts. Live. Free. Searchable now live on the #MIDF platform.

More in progress.

This is what India's knowledge infrastructure looks like,
- https://t.co/1RQXcPLlB2

#MIDF #IndianHeritage #Manuscripts #PMO @GyanBharatamMoC @gssjodhpur

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

@aryanvs_ @Roblox Congratulations!

[clem 🤗 @ClementDelangue]

Most people training agentic LLMs with RL right now have a silently broken training loop and have no idea.

Here's the trap: single-turn RL works beautifully. Clean curves, sane rewards, everything converges. Then you add tools so the model can act mid-rollout, and things get weird. Loss spikes for no reason. Eventually a shape-mismatch error.

The culprit: every time you parse the model's output to detect a tool call, then re-tokenize the updated conversation for the next turn, you're rolling the dice. Usually the round-trip gives back the same tokens. Sometimes it doesn't and your gradient lands on a sequence the model never actually sampled. No crash. Just quietly wrong math and a useless gradient signal.

The fix is one rule: never re-encode tokens you've decoded. Keep the sampled tokens in one buffer, never re-render them, and both failure modes disappear. That's Token-In, Token-Out done right.

Our team just published a beautiful deep-dive on exactly this, including an audit across the major open-weights model families showing most chat templates already support it. Required reading if you're doing multi-turn RL 🤗🔥

https://t.co/zmx0EQl3jM

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

@ChinmayKak Solarpunk looks really good

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

this is such a goated template for research articles

using this to write my first technical blog

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

@NousResearch @StepFun_ai what prompt do you give to Hermes agent to generate this particular style of videos?

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

will miss the cats and @ChinmayKak

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

it's my last month at @MSFTResearch

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

@ChinmayKak 😶‍🌫️

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

booked my first international solo flight today ✈️

[Adithya S K @adithya_s_k]

Introducing Repo2RLEnv

Turn any repository into runnable, verifiable coding environments built from real PRs and commits for coding-agent evaluation or RL training

> uv pip install repo2rlenv

[Samarth P✈️ICML'26🇰🇷 @SamarthP6]

our Nayana-OCR dataset is now open source

check it out!

[CognitiveLab @cognitivelab_ai]

Launching NayanaOCR Corpus
👉🏼 1M+ Document images across 22 languages

Largest open source synthetic
> multilingual
> multimodal
> multitask
document corpus