Yuandong Tian (@tydsh) / X

Yuandong Tian

1,133 posts

Yuandong Tian

@tydsh

Co-founder of @Recursive_SI. ex-Meta FAIR Director. ex-Google. Reasoning, Optimization and Understanding LLM. Novelist in spare time. PhD in @CMU_Robotics.

California, USA

Joined December 2009

Pinned
Yuandong Tian
@tydsh
Jun 11
Early results from Recursive 🚀🚀 SotA results from our open-ended knowledge discovery system: 1️⃣NanoChat 5min pre-training (0.9372 bpb -> 0.9109 bpb, 2.8% lower Bits-Per-Byte than long-standing community SoTA) 2️⃣NanoGPT SpeedRun (79.7s -> 77.5s, 2.8% faster than long-standing
Recursive
@Recursive_SI
Jun 11
Article
First Steps Toward Automated AI Research
Early results from Recursive’s automated AI research system on model training and GPU kernel benchmarks Today we are releasing early results from Recursive’s automated AI research system. Across three...
First Steps Toward Automated AI Research - Recursive
From recursive.com
47K
Yuandong Tian
@tydsh
Oct 23, 2025
Several of my team members + myself are impacted by this layoff today. Welcome to connect :)
4.4M
Yuandong Tian
@tydsh
Oct 23, 2025
I am really sorry Jiaxun😓. I wrote a recommendation letter for you to join Meta, and the team you joined promised you to do RL. But you ended up not doing anything related to RL, involved in never-ending reorgs, and got impacted after a few months as a fresh new hire!
Jiaxun Cui 🐿️
@cuijiaxun
Oct 23, 2025
Meta has gone crazy on the squid game! Many new PhD NGs are deactivated today (I am also impacted🥲 happy to chat)
772K
Yuandong Tian
@tydsh
Jan 15, 2025
Our Coconut work (learning continuous latent CoT) has opened sourced now. Welcome to play with it:
GitHub - facebookresearch/coconut: Training Large Language Model to Reason in a Continuous Latent...
From github.com
161K
Yuandong Tian
@tydsh
Jun 18, 2025
📢We show that continuous latent reasoning has a theoretical advantage over discrete token reasoning (arxiv.org/abs/2505.12514): For a graph with n vertices and graph diameter D, a two-layer transformer with D steps of continuous CoTs can solve the directed graph reachability
arxiv.org
Reasoning by Superposition: A Theoretical Perspective on Chain of...
Large Language Models (LLMs) have demonstrated remarkable performance in many applications, including challenging reasoning problems via chain-of-thoughts (CoTs) techniques that generate...
331K
Yuandong Tian
@tydsh
Oct 23, 2025
Replying to @daemonzhang6
That's the problem. People who are responsible for the issues are not the people who got laid off😅 In January, our team put down all the research we are currently doing, was (forced?) to move to GenAI <2 months before the llama 4 release deadline to help with all the
272K
Yuandong Tian
@tydsh
Nov 12, 2025
Hats off to @ylecun! FAIR shaped my career, period. I truly thanks @AIatMeta and FAIR to provide such a nice place for independent exploration and open research! End of an era and forever remember.
Meta chief AI scientist Yann LeCun plans to exit and launch own start-up
From ft.com
181K
Yuandong Tian
@tydsh
Nov 24, 2023
How likely is the hypothesis that Q* = Q-learning + A*? From my past experience on OpenGo (reproduction of AlphaZero), A* can be regarded as a deterministic version of MCTS with value (i.e., heuristic) function Q only. This should be suitable for tasks in which the state is easy
359K
Yuandong Tian
@tydsh
Sep 17, 2024
While CoT is super useful, I kindly disagree that blindly scaling it up is all we need. The paper proposes a universal approximation theorem by explicitly constructing Transformer weights to fit to the family of tasks. Although the depth can be constant, the length of CoT can
Denny Zhou
@denny_zhou
Sep 16, 2024
What is the performance limit when scaling LLM inference? Sky's the limit. We have mathematically proven that transformers can solve any problem, provided they are allowed to generate as many intermediate reasoning tokens as needed. Remarkably, constant depth is sufficient.
253K
Yuandong Tian
@tydsh
Jan 2, 2020
Over the last decade: - Got PhD from CMU Robotics. - Start academic career and published the majority of current publication list. - 6 oral talks in top-tier conferences with 1 Marr Prize Honorable Mentions (1st author in ICCV 13). - Married and a house owner. - Wrote one novel.
Yuandong Tian
@tydsh
Jun 11, 2024
I am looking for one fall'24 intern, along the direction of LLM+reasoning/planning. The intern shall start before Nov this year. If you have interest, please contact [email protected]. Thanks!
135K
Yuandong Tian
@tydsh
Dec 14, 2024
Unbelievable... This is explicit racial bias. How could this happen in NeurIPS? How could this be spoken by a top university professor, an invited keynote speaker?
Jiao Sun
@sunjiao123sun_
Dec 14, 2024
Mitigating racial bias from LLMs is a lot easier than removing it from humans! Can’t believe this happened at the best AI conference @NeurIPSConf We have ethical reviews for authors, but missed it for invited speakers? 😡
37K
Yuandong Tian
@tydsh
Feb 18, 2024
One interesting component of Sora is that, the video is NOT created by next-frame prediction, but by constructing and refining latent token sequence first and then decoded them back. Doing planning/search in learnable latent space, rather than original space has its unique
134K
Yuandong Tian
@tydsh
Feb 14, 2025
Our new work Spectral Journey arxiv.org/abs/2502.08794 shows a surprising finding: when a 2-layer Transformer is learned to predict the shortest path of a given graph, 1️⃣it first implicitly computes the spectral embedding for each edge, i.e. eigenvectors of Normalized Graph
arxiv.org
Spectral Journey: How Transformers Predict the Shortest Path
Decoder-only transformers lead to a step-change in capability of large language models. However, opinions are mixed as to whether they are really planning or reasoning. A path to making progress...
55K