Yizhou Chi

PhD @cambridgenlp | Builds language agents that question, reason, and stay aligned | ex-Berkeley EECS | AI, safety, honesty

[Yizhou Chi @YizhouChi]

Was digging into how ChatGPT renders citations. Instead of wasting tokens on markdown links, it streams invisible Unicode PUA characters (U+E200) to trigger UI badges.

Open-sourced the exact frontend pipeline we replicated: https://t.co/NgLtcpUf41

[Yizhou Chi @YizhouChi]

WorldReasoner is now open-source! Dive into the code and benchmark here:
GitHub: https://t.co/a8Qu4Glhov

Thanks to @Eric_chamoun @ZifengDing6 @vlachos_nlp
(4/4)

[Yizhou Chi @YizhouChi]

How to test agent forecasting without data contamination?

Isolate them in a temporal sandbox⏳

Introducing WorldReasoner: a new framework that restricts agents to simulated historical dates to test genuine reasoning over time-bounded evidence

https://t.co/vi1iN6BrpX (1/4)

[Yizhou Chi @YizhouChi]

Our pipeline generated 345 tasks from 14k+ articles. Across 6 agent settings, we found:
1️⃣ Temporally valid retrieval drives outcome accuracy. 2️⃣ Causal graphs improve key-event recovery.
3️⃣ Agents still struggle to convert good evidence into calibrated probabilities. (3/4)

[Yizhou Chi @YizhouChi]

WorldReasoner evaluates beyond final-answer accuracy. Inside the sandbox, agents must retrieve pre-resolution evidence and optionally build causal event graphs. We score them on probability calibration, evidence quality, and reasoning against hindsight reference graphs.🔍 (2/4)

[Eric @Eric_chamoun]

“I want to build X.”
What needs to happen first?

Behind every scientific contribution is a pathway of enabling methods, datasets, tools, models, ideas, and prior work.

We introduce SciPaths to study whether LLMs can predict these pathways. 🧬🧭

[Xiaochen Zhu (Neo) @ZhuNeo13294]

🚀Check out our new paper (https://t.co/AtDkFrc0Im)!

Why multi-agent debate often fails for LLMs? We learn from human debate and reveals the limitations: low diversity, no confidence signals.
We fix both and show this turns debate from a martingale into a strict submartingale.

[Yizhou Chi @YizhouChi]

Experience-wise, it's become a pretty mature productivity tool. I basically depend on it now for my daily work, research, writing docs, and coding demos.

Check it out here https://t.co/NgLtcpUf41
Looking for contributors~

[Yizhou Chi @YizhouChi]

So, I've finally open-sourced the Deep Research + Cowork project I've been working on for the last few months: Suzent.

It fuses all the Agent experience I've gathered over the last 2 years (plain English: I stitched everything I know into it).

https://t.co/NgLtcpUf41

#agent

[Yizhou Chi @YizhouChi]

It fills the gaps left by cloud tools (like global memory, cross-conversation workspaces, and a full-platform sandbox [Windows needs WSL2]).

It doesn't just remember your habits; it lets you pick up a task (and files) you were working on a long time ago.

[Yizhou Chi @YizhouChi]

Compared to Claude Cowork, Suzent feels more like a digital extension of yourself that truly belongs to YOU.

It focuses on Long-term Memory, Data Sovereignty, and Local Usage.

[Yizhou Chi @YizhouChi]

Of course, it wouldn't exist without suggestions from friends, inspiration from open-source giants, and even some patient, guiding conversations with closed-source products.

[MetaGPT @MetaGPT_]

🚀 Coming Soon: MGX (MetaGPT X).

Swing by our Product Hunt teaser page and click “Notify Me”: https://t.co/YkUSLA7IRl via @kevinwdavid

#ProductHunt #Launch #MetaGPT #MGX

[MetaGPT @MetaGPT_]

🌟 Excited to open-source SELA, a powerful experimentation system integrating MCTS with LLM agents. Across 20 datasets, SELA achieves a 75% win rate against AIDE (OpenAI's top pick in MLE-Bench) and beats traditional AutoML methods developed over years.
💻 Code: https://t.co/dIjSaXnoii
📄 Paper: https://t.co/8j5DLpNr9N

[Yizhou Chi @YizhouChi]

🎉 Really excited to share SELA!
I started about a year ago exploring MCTS + LLM reasoning... Now we've even shown agents can master ML experimentation through systematic search & learning.
Thanks to all collaborators who made this possible!

[MetaGPT @MetaGPT_]

🤖 Can an AI agent design another AI?
🌟 Yes. SELA uses MCTS to design AI, achieving SoTA performance on 20 machine learning datasets. It can learn from past designs and experiments to create better AIs. It's fully open-source.

🚀 Try SELA yourself
📄 Paper: https://t.co/8j5DLpMTkf
💻 Code: https://t.co/pSDAhB4kAJ

[Yizhou Chi @YizhouChi]

Cool! This work is very similar to ours published a couple of months ago.

https://t.co/b9swWuTuh9

Really glad to see the idea shown effective in different tasks!

Also wonder what the performance would be like with larger llama or gpt4

[Rohan Paul @rohanpaul_ai]

"Accessing GPT-4 level Mathematical Olympiad Solutions via Monte Carlo Tree Self-refine with LLaMa-38B"

From 25.47% to 45.49% in GSM-Hard 🤯

Also noting in this regard, the head of Deepmind said last year that augmenting LLMs with Monte Carlo Tree Search may be the fastest path to AGI

📌 This paper introduces the MCT Self-Refine (MCTSr) algorithm, which integrates Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS) to enhance performance on complex mathematical reasoning tasks like Olympiad-level problems. The key problem being addressed is the accuracy and reliability challenges faced by LLMs in strategic and mathematical reasoning.

📌 MCTSr constructs a Monte Carlo search tree through iterative processes of Selection (using an improved Upper Confidence Bound formula to balance exploration-exploitation), self-refine (the LLM generates feedback to guide refining an answer), self-evaluation (the LLM scores the quality of the refined answer), and Backpropagation (propagating the refined answer's value back through the tree).

📌 The self-refine process uses a multi-turn dialogue prompt where the LLM first generates a critical comment on the current answer, then refines the answer guided by that comment. The self-evaluation scores an answer from -100 to 100 and applies constraints like strict scoring standards and suppressing perfect scores to improve reliability.

📌 Backpropagation updates a node's Q value (estimated answer quality) by averaging its current Q value and the max Q value of its child nodes. Candidate nodes for further expansion are selected based on criteria like number of child nodes and child Q values exceeding the parent's.

📌 Experiments demonstrate MCTSr significantly improves success rates on datasets like GSM8K (up to 96.66% with 8 rollouts vs 74.07% zero-shot), MATH (58.24% overall with 8 rollouts vs 24.36% zero-shot), and Olympiad-level benchmarks like AIME (11.79% with 8 rollouts vs 2.36% zero-shot). Performance scales with number of rollouts.

📌 Compared to closed-source LLMs like GPT-4, MCTSr with LLaMA-38B achieves comparable results, showing it can boost reasoning capabilities of smaller open-source models. The paper concludes MCTSr is a robust and promising approach for complex mathematical reasoning with LLMs.

[Yizhou Chi @YizhouChi]

🙌 Huge shoutout to my amazing co-authors @kevinyang41 and Prof. Dan Klein!

[Yizhou Chi @YizhouChi]

📝Presenting ThoughtSculpt - a general reasoning & search approach for tasks with decomposable outputs. Leveraging Monte Carlo Tree Search, it surpasses existing methods across diverse tasks! (1/N)

arxiv: https://t.co/b9swWuTuh9

[Yizhou Chi @YizhouChi]

ThoughtSculpt is a general method that doesn't require too much prompt engineering and is easy to apply to different challenges that may require continuous thought iteration such as open-ended generation, multi-step reasoning, and creative ideation. (5/N)

[Yizhou Chi @YizhouChi]

We find that our method, using MCTS, displays better performance than using a naive DFS and SOTAS in scenarios where the tasks are complex and difficult to greedily improve their solutions (4/N)

[Yizhou Chi @YizhouChi]

We evaluate ThoughtSculpt in three tasks: story outline planning, 5x5 crossword-solving, and constrained generation. (3/N)