Justin Wong @ Neurips'25

CS PhD Student working on diverse AI generation advised by @profjoeyg and Sanjit Seshia. formerly @aiatmeta, @nuro, @columbia

[Justin Wong @ Neurips'25 @justinywong_]

Huge thanks to @conor_power23 and @lisabdunlap for having me on! https://t.co/FGwPkf4cYz

[OpenAI @OpenAI]

We congratulate Levent Alpöge and Tristan Buckmaster on their remarkable mathematical work.

We (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem.

While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.

However, our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs. unforced).

[Justin Wong @ Neurips'25 @justinywong_]

@michaelpsenka @maxkirkby @oneill_c @mudithj LFG!! Excited to see what you cook!!

[Simon Guo @simonguozirui]

Supporting safety research via @tinkerapi!

[Justin Wong @ Neurips'25 @justinywong_]

@WestJet WestJet needs better customer service. The flight turns out wasnt cancelled but we were preemptively removed.

[Justin Wong @ Neurips'25 @justinywong_]

@WestJet Wow had the worst experience with an airline ever!! Stranded in calgary for 24hrs. The connecting flight to iceland got cancelled 😭😭😭 The staff refuse to give a hotel voucher since it's not due to a delay and is due to a cancellation.

[Justin Wong @ Neurips'25 @justinywong_]

@WestJet These replies are scams!!!

[Alexander Panfilov @kotekjedi_ml]

Further, if you ever shared online a Claude Code/Codex session with encrypted reasoning blobs, they can be decoded and leak your personal data.

We did a preliminary scan of ~7,000 public traces and found 62 unique API keys, 33 email addresses, 33 passwords, and other sensitive data.

[Justin Wong @ Neurips'25 @justinywong_]

Great work @trungthvu and @RichardZ412 !! Feeling the proud dad moment

[Bespoke Labs @bespokelabsai]

OpenAI’s o4 just showed that multi-turn tool use is a huge deal for AI agents.
Today, we show how to do the same with your own agents, using RL and open-source models.

We used GRPO on only 100 high quality questions from the BFCL benchmark, and post-trained a 7B Qwen model to orchestrate multiple tools without demonstrations from humans or teacher models. We measured a 23% increase in model performance on our eval set (1/n) 🧵

[Bespoke Labs @bespokelabsai]

OpenAI’s o4 just showed that multi-turn tool use is a huge deal for AI agents.
Today, we show how to do the same with your own agents, using RL and open-source models.

We used GRPO on only 100 high quality questions from the BFCL benchmark, and post-trained a 7B Qwen model to orchestrate multiple tools without demonstrations from humans or teacher models. We measured a 23% increase in model performance on our eval set (1/n) 🧵

[Transluce @TransluceAI]

To interpret AI benchmarks, we need to look at the data.

Top-level numbers don't mean what you think: there may be broken tasks, unexpected behaviors, or near-misses.

We're introducing Docent to accelerate analysis of AI agent transcripts. It can spot surprises in seconds. 🧵👇

[Justin Wong @ Neurips'25 @justinywong_]

Congrats @lisabdunlap two more spots before making it to last author 🎉🎉🎉

[Lisa Dunlap @lisabdunlap]

First VisDiff then VidDiff, what do we difference next?
amazing work @jmhb0 and team!

[Justin Wong @ Neurips'25 @justinywong_]

Some massive improvement exploiting program structure !! Great work @michaelzluo !

[Agentica Project @Agentica_]

Introducing Autellix: An agentic AI system that accelerates agentic applications.

Run Deep Researcher, Google Co-Scientist, OAI Operator, or any program 💻—@langchain, @pyautogen, @crewAIInc, or just Python 🐍—4-15x faster than vLLM or SGLang⚡.

Paper: https://t.co/LvLq2B8P4v

[Lianmin Zheng @lm_zheng]

People have been wondering which model will be the first to reach a 1400 Elo score; no one believed it would be Grok one year ago, but @xAI made it happen!

[Tanishq Mathew Abraham, Ph.D. @iScienceLuvr]

Large Language Diffusion Models

Introduces LLaDA-8B, a large language diffusion model that pretrained on 2.3 trillion tokens using 0.13 million H800 GPU hours, followed by SFT on 4.5 million pairs. LLaDA 8B surpasses Llama-2 7B on nearly all 15 standard zero/few-shot learning tasks while performing on par with Llama-3 8B.

[Scaled Cognition @ScaledCognition]

We’re Scaled Cognition, developing the first ever models trained specifically for agentic applications:

1. Our first system, APT-1, is now #1 on agentic benchmarks.
2. It was developed by a US team for a total cost of less than $11M.
3. Khosla Ventures led our seed round ($21M closed in 2023), and Vinod Khosla joined our board.
4. We use a fully synthetic, RL-based agentic data pipeline, no human-labeled data.
5. APT-1 is now available for early access.

[Yuchen Jin @Yuchenj_UW]

This is wild - UC Berkeley shows that a tiny 1.5B model beats o1-preview on math by RL!

They applied simple RL to Deepseek-R1-Distilled-Qwen-1.5B on 40K math problems, trained at 8K context, then scaled to 16K & 24K.

3,800 A100 hours ($4,500) to beat o1-preview in math!

Best thing is they open-sourced everything: the model, the training code (based on ByteDance verl library), and the dataset.

[Justin Wong @ Neurips'25 @justinywong_]

I love seeing creative domain getting some love! AI got nothing on these stunning gifs

[Brandon Trabucco @brandontrabucco]

With the success of LLM agents like OpenAI Operator, we are entering a new scaling era, but how do we train these agent models?

We present InSTA, the largest training environment for LLM agents, containing live web navigation tasks for 150k diverse websites in multiple languages.

Website - https://t.co/nGIViX9NGc
Environment - https://t.co/nIs3ZiVfE1

🧵Thread below.

1/6

#AgenticAI #LLMs #OpenAl

[Brandon Trabucco @brandontrabucco]

With the success of LLM agents like OpenAI Operator, we are entering a new scaling era, but how do we train these agent models?

We present InSTA, the largest training environment for LLM agents, containing live web navigation tasks for 150k diverse websites in multiple languages.

Website - https://t.co/nGIViX9NGc
Environment - https://t.co/nIs3ZiVfE1

🧵Thread below.

1/6

#AgenticAI #LLMs #OpenAl

[Justin Wong @ Neurips'25 @justinywong_]

@AlexKrentsel @berkeley_ai 100% @AlexKrentsel It's such a privilege working on a team that fully appreciates the challenges and insights of both systems and AI !!

[Justin Wong @ Neurips'25 @justinywong_]

AK repost goes hard!

[Agentica Project @Agentica_]

Thanks AK for sharing our work! 🙏
We replicate Deepseek's success under $5000 and share our training recipe on:
- Twitter: https://t.co/Sd2mWyvCDA
- Blogpost https://t.co/eHqApwRfnH

Come check it out!