Twitxy Saved

Adithya Bhaskar

@AdithyaNLP

Third year CS PhD candidate at Princeton University (@princeton_nlp @PrincetonPLI), previously CS undergrad at IIT Bombay
Princeton, NJ adithyabh.github.io Joined June 2023
648 Following 1.1K Followers 142 Posts

Latest posts 5

Pinned post

Adithya Bhaskar @AdithyaNLP
The most profound paper of my PhD so far. We truly did something special to make LMs play/explain chess, needing both architectural & algorithmic innovation (natural-language analogue of the Alphazero algorithm) + applicable to many other domains. Please read on and share!

1/10
62 replies 172 reposts 1.5K likes 113K views View on X
Adithya Bhaskar @AdithyaNLP
Happening now at Grand Ballroom Poster #133! Stop by!
Language models that think, chat better.

We used longCoT (w/ reward model) for RLHF instead of math, and it just works. Llama-3.1-8B-Instruct + 14K ex beats GPT-4o (!) on chat & creative writing, & even Claude-3.7-Sonnet (thinking) on AlpacaEval2 and WildBench!

Read on. 馃У

1/8
View quoted post on X
Adithya Bhaskar @AdithyaNLP
We are both at COLM, come talk to us if you are interested in chess / interpretation of silent experts!
Jeffrey Cheng Verified account @jeff_cheng_77 Oct 5
鈥硷笍Have you ever used a chess engine and wished it could explain its moves and evaluations?鈥硷笍

Introducing QUEEN, a 4B-parameter chess-language model that plays at the level approaching a typical Grandmaster and explains its moves in natural language.
View quoted post on X
0 replies 0 reposts 13 likes 1.4K views View on X
Adithya Bhaskar @AdithyaNLP
I will present this paper as a poster at COLM 2026 (Grand Ballroom #133, Franciscan, Poster #1199, F-4-133)!

HMU for research/coffee chats or with suggestions for good coffee shops! I've been working on making LLMs learn from non-language data, and I'll have more to say soon.
Language models that think, chat better.

We used longCoT (w/ reward model) for RLHF instead of math, and it just works. Llama-3.1-8B-Instruct + 14K ex beats GPT-4o (!) on chat & creative writing, & even Claude-3.7-Sonnet (thinking) on AlpacaEval2 and WildBench!

Read on. 馃У

1/8
View quoted post on X
2 replies 3 reposts 29 likes 1.6K views View on X