Stanford AI Lab(@StanfordAILab)

Work advised by @simran_s_arora @Azaliamirh @HazyResearch.

Stanford AI Lab (@StanfordAILab) — Thu, 15 Oct 2026 21:59:37 GMT

Work advised by @simran_s_arora@Azaliamirh@HazyResearch.

💬0🔄0❤️6👀1978📊1

What if language models wrote the way that we think through ideas: rough sketch first with details l...

Stanford AI Lab (@StanfordAILab) — Thu, 15 Oct 2026 21:59:36 GMT

What if language models wrote the way that we think through ideas: rough sketch first with details last?

New work by @KumbongHermann, @ishanskhare, and team introduces Next-Concept Modeling (NCM), a parallel generation framework achieving comparable or better quality than strong diffusion baselines while being 10-31× faster⚡️across unconditional generation, summarization & translation, and downstream reasoning tasks.

Hermann Kumbong@KumbongHermann

🚀 Introducing Next-Concept Modeling (NCM): our new paradigm for fast, parallel, coarse-to-fine text generation.

Compared to strong diffusion language models (DLMs) at similar parameter size, NCM achieves:

⚡ Up to 31× faster generation on downstream tasks like summarization and machine translation with comparable or better quality.

🏎️ 10–21× faster unconditional generation across TinyStories, LM1B, and OpenWebText, with comparable or better generative perplexity and using  3–5× fewer training tokens.

🧩 Better performance on reasoning tasks like Sudoku, outperforming both AR and diffusion baselines with fewer sampling steps.

With @ishanskhare, @IddahMlauzi, @ekellbuch, @simran_s_arora, @Azaliamirh, and @HazyResearch.

More details in the thread 🧵👇

Your browser does not support the video tag.

💬5🔄8❤️67👀10600📊16

What if language models wrote the way that we think through ideas: rough sketch first with details l...

Stanford AI Lab (@StanfordAILab) — Wed, 14 Oct 2026 23:45:36 GMT

What if language models wrote the way that we think through ideas: rough sketch first with details last?

New work by @KumbongHermann, @ishanskhare, and team introduces Next-Concept Modeling (NCM), a parallel generation framework achieving comparable or better quality than strong diffusion baselines while being 10-31× faster⚡️across unconditional generation, summarization & translation, and downstream reasoning tasks.

Hermann Kumbong@KumbongHermann

🚀 Introducing Next-Concept Modeling (NCM): our new paradigm for fast, parallel, coarse-to-fine text generation.

Compared to strong diffusion language models (DLMs) at similar parameter size, NCM achieves:

⚡ Up to 31× faster generation on downstream tasks like summarization and machine translation with comparable or better quality.

🏎️ 10–21× faster unconditional generation across TinyStories, LM1B, and OpenWebText, with comparable or better generative perplexity and using  3–5× fewer training tokens.

🧩 Better performance on reasoning tasks like Sudoku, outperforming both AR and diffusion baselines with fewer sampling steps.

With @ishanskhare, @IddahMlauzi, @ekellbuch, @simran_s_arora, @Azaliamirh, and @HazyResearch.

More details in the thread 🧵👇

Your browser does not support the video tag.

💬7🔄27❤️119👀12314📊30

Work advised by @simran_s_arora @Azaliamirh @HazyResearch

Stanford AI Lab (@StanfordAILab) — Wed, 14 Oct 2026 23:45:36 GMT

Work advised by @simran_s_arora@Azaliamirh@HazyResearch

💬0🔄0❤️5👀2046📊1

Check out our latest SAIL blog post on CausalSmith! This agentic framework automates theoretical cau...

Stanford AI Lab (@StanfordAILab) — Wed, 14 Oct 2026 22:50:15 GMT

Check out our latest SAIL blog post on CausalSmith! This agentic framework automates theoretical causal inference research, using Lean to verify proofs and an audit step to make sure the formal statements actually match the paper’s claims: ai.stanford.edu/blog/causalsmi…

💬2🔄6❤️30👀9118📊12

Can a robot learn dexterous manipulation from just one human video? Check out our latest SAIL blog p...

Stanford AI Lab (@StanfordAILab) — Wed, 14 Oct 2026 22:45:55 GMT

Can a robot learn dexterous manipulation from just one human video? Check out our latest SAIL blog post on DexAgent, an agentic framework that turns a single egocentric video into robot training data using a self-evolving tool library! ai.stanford.edu/blog/dexagent/

💬5🔄6❤️44👀9297📊14

Stanford AI Lab is alive and well at COLM! Check out these SAIL projects this week in SF. 🎉 https...

Stanford AI Lab (@StanfordAILab) — Wed, 07 Oct 2026 17:43:26 GMT

Stanford AI Lab is alive and well at COLM! Check out these SAIL projects this week in SF. 🎉 
ai.stanford.edu/blog/colm-2026/m

💬3🔄5❤️40👀10355📊10

SWE-chat by @joabaum is a living dataset of human interactions with AI coding agents. 230K prompts f...

Stanford AI Lab (@StanfordAILab) — Fri, 02 Oct 2026 19:00:17 GMT

SWE-chat by @joabaum is a living dataset of human interactions with AI coding agents. 230K prompts from real developers collected in the wild are now available on HuggingFace 🚀huggingface.co/datasets/SALT-…f

Joachim Baumann@joabaum

SWE-chat v2 is here! The largest dataset of coding agent interactions from real users in the wild has gotten even larger: 230K prompts from 18K sessions.
SWE-chat has enabled incredible research (🧵) – excited to see what v2 unlocks for the community.
Come find us at @COLM_conf!

Your browser does not support the video tag.

💬7🔄4❤️37👀10405📊11

Some interesting and perhaps unexpected properties of tokenization for language models!

Stanford AI Lab (@StanfordAILab) — Fri, 02 Oct 2026 18:17:07 GMT

Some interesting and perhaps unexpected properties of tokenization for language models!

Rylan Schaeffer@RylanSchaeffer

Tokenizer Myths and Where to Find Them!

Slides from talk last night - preprint forthcoming - feedback welcome 🙏

1/11

💬2🔄5❤️30👀10225📊8

This work directly tackles the largely unexplored problem of submission overload at conferences. It ...

Stanford AI Lab (@StanfordAILab) — Thu, 01 Oct 2026 17:35:36 GMT

This work directly tackles the largely unexplored problem of submission overload at conferences. It argues that technical governance solutions can help everyone build trust and faith in our system, and in turn allows building on each other's contributions. Check it out!

Houjun Liu@houjun_liu

🚨new method day! 🚨
Trust[ing results] in ML conferences is utterly broken. Let's fix it with an algorithm!

We are excited to introduce 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗪𝗶𝘁𝗻𝗲𝘀𝘀𝗲𝘀, a system for “honesty” certificates of data and evaluations which a verifier can cheaply check and trust.

💬1🔄6❤️20👀9420📊4

New work co-led by @AdityaCowsik, @KfirDolev, and @michaelyli_ with co-authors @gbruno_dl @ANourya @...

Stanford AI Lab (@StanfordAILab) — Fri, 25 Sep 2026 17:59:04 GMT

New work co-led by @AdityaCowsik, @KfirDolev, and @michaelyli_ with co-authors @gbruno_dl@ANourya@noahdgoodman, and @YoavLevine.

💬1🔄0❤️2👀2209📊2

What would pretraining with zero real data look like? Can a randomly initialized model learn to gene...

Stanford AI Lab (@StanfordAILab) — Fri, 25 Sep 2026 17:58:37 GMT

What would pretraining with zero real data look like? Can a randomly initialized model learn to generate all of its training data entirely through self-play? Find out below! :)

Michael Y. Li@michaelyli_

Can an LM, starting from random init (!!), learn to generate all of its pretraining data?

Introducing Self-Play Pretraining with Zero Data. Two models start from random initialization: a generator proposes programs for a universal Turing machine and a learner trains on their outputs. We never train on any real data, but see predictable scaling on natural datasets: zero-shot val loss on images, text, audio, and melodies decreases predictably with self-play compute. And the learner develops in-context learning capabilities.

A fun proof-of-concept, co-led with @AdityaCowsik and @KfirDolev and co-authors @gbruno_dl, @ANourya@noahdgoodman, and @YoavLevine.

💬13🔄32❤️355👀44138📊69

Introducing Context-Sharded Block Parallelism (CSBP), a new distributed parallelism strategy enablin...

Stanford AI Lab (@StanfordAILab) — Fri, 18 Sep 2026 18:47:37 GMT

Introducing Context-Sharded Block Parallelism (CSBP), a new distributed parallelism strategy enabling significant training efficiency for diffusion LLMs, with gains that grow with context length 🚀

⚡ 7.59× faster DFlash2 speculative decoding drafter training
⚡ 1.61× faster block diffusion fine-tuning
⚡ 1.33× faster autoregressive → block diffusion adaptation

Tarun Suresh @ COLM 2026@TarunSures41845

Diffusion LLMs and speculative decoding promise much faster agents. Yet agents need long contexts, and training on them is painfully slow.

Introducing Context-Sharded Block Parallelism (CSBP), a new distributed parallelism strategy unlocking significant training efficiency for diffusion LLMs, with speedup gains growing with context length 🚀

⚡ 7.59× faster DFlash2 speculative decoding drafter training
⚡ 1.61× faster block diffusion fine-tuning
⚡ 1.33× faster autoregressive → block diffusion adaptation

With the same GPU hours, models trained with CSBP score higher on SWE-bench Verified and Terminal-Bench Lite 👑

Open-sourced in Turbo-dLLM, our new optimized distributed training library.

Advised by @Azaliamirh and with an amazing team: @PranshuChatur11@hangoo_kang@pshroff_@ishanskhare@KumbongHermann

💬9🔄17❤️150👀26042📊32

Hear from @chrmanning, head of @stanfordnlp and former director of @StanfordAILab, on why @Stanford'...

Stanford AI Lab (@StanfordAILab) — Thu, 17 Sep 2026 05:57:01 GMT

Hear from @chrmanning, head of @stanfordnlp and former director of @StanfordAILab, on why @Stanford's commitment to open, public research matters as more frontier AI is developed behind closed doors.

MTS@MTSlive

Stanford Prof. @chrmanning warns frontier AI is being built inside “three monasteries,” with too little of what happens inside making it out to the rest of society:

"If we just think of the whole space of frontier AI, a huge problem at the moment is so much of it is happening in these sort of three monasteries, and very little information is making it outside to the rest of the world."

"That's an enormous problem for the whole of society. The American public is not gonna get more comfortable with the possibilities of an AI future if the feeling is it's completely hidden from them."

"For the case of having Stanford involved, it's just a bedrock assumption of Stanford's that things are made public. All research we do has to be publicly available."
@stanfordnlp@Stanford

Your browser does not support the video tag.

💬3🔄1❤️6👀1659📊3

Great chance to plug @tatsu_hashimoto and @kothasuhas's CS 312: Deep Learning Alchemy All recordin...

Stanford AI Lab (@StanfordAILab) — Sun, 13 Sep 2026 01:16:40 GMT

Great chance to plug @tatsu_hashimoto and @kothasuhas's CS 312: Deep Learning Alchemy 

All recordings and class materials will be made public!

deep-learning-alchemy.github.io

wh@nrehiew_

Tons of respect for universities and faculty who try their best to keep up to date with the landscape while still ensuring pedagogical value

Would have been easy to teach the same 2022 AI class today, but goes to show that some professors really do care

💬11🔄50❤️481👀55501📊114

@StanfordAILab grads hard at work! 🚀

Stanford AI Lab (@StanfordAILab) — Thu, 03 Sep 2026 21:59:02 GMT

@StanfordAILab grads hard at work! 🚀

Yann Dubois@yanndubs

💫 Astra is out

Highlights:
- Much smarter (evals are really saturating...)
- Much more aligned & trustworthy
- Better CUA (see this house Astra made in Blender)
So proud of the team for training it!

Still some issues:
- Too much code slop
- Astra asks for confirmation too often, which can feel lazy. Given how smart it is, we wanted it to be cautious, but likely overcorrected.
We’ll fix those next! What else should we fix?

Note: Astra follows instructions even more carefully than 5.6, including unwanted ones buried in old skills. I’d clean those up before using it.
openai.com/index/gpt-6-as…u

Your browser does not support the video tag.

💬3🔄0❤️37👀17731📊6

The most capable models available in the world are using Terminal-Bench-Science to show it (just one...

Stanford AI Lab (@StanfordAILab) — Thu, 03 Sep 2026 21:56:12 GMT

The most capable models available in the world are using Terminal-Bench-Science to show it (just one week after the benchmark’s release). 

Congratulations to @StevenDillmann and team for their remarkable impact!

Steven Dillmann@StevenDillmann

What a week for Terminal-Bench-Science & AI for Science in general! 🥂

Just 1 week after launch, Terminal-Bench-Science 0.1 is now also the #1 featured benchmark on today's GPT-6 Astra release by @OpenAI. 

GPT-5.6 Sol → 22.4%
GPT-6 Astra → 64.6%

The +42.2pp jump is out of this world, and that's on top of the Claude Fable 5.1 numbers from two days ago.

To all scientists out there: We need to buckle up and bring the world's toughest scientific challenges together for Terminal-Bench-Science 0.2. Deadline is October 5.

Let the games begin!
openai.com/index/gpt-6-as…U

💬5🔄5❤️49👀11079📊10

If you want to learn how to engineer AI agents, these are the people you’d want to learn it from. Th...

Stanford AI Lab (@StanfordAILab) — Thu, 03 Sep 2026 06:38:43 GMT

If you want to learn how to engineer AI agents, these are the people you’d want to learn it from. They know the space inside and out and couldn’t be more excited to teach it. First time it’s running at Stanford this fall. @Diyi_Yang, @michaelryan207, @jyangballin

Diyi Yang@Diyi_Yang

This fall @michaelryan207@jyangballin and I are teaching a new course CS329Z "Engineering AI Agents" on how to build AI Agents from scratch.  Come join us and learn how to build them 🤖

💬14🔄55❤️574👀49786📊126

Congratulations to all the SAIL members recognized in TIMES AI 2026! @drfeifei, @chelseabfinn, @sa...

Stanford AI Lab (@StanfordAILab) — Thu, 27 Aug 2026 19:42:40 GMT

Congratulations to all the SAIL members recognized in TIMES AI 2026! 

@drfeifei, @chelseabfinn, @sanmikoyejo, @Azaliamirh, and @annadgoldie 

Proud to see their work and impact recognized among the people shaping the future of AI.

time.com/collection/tim…

💬7🔄10❤️122👀14157📊22

Terminal-Bench-Science is out!

Stanford AI Lab (@StanfordAILab) — Thu, 27 Aug 2026 19:24:53 GMT

Terminal-Bench-Science is out!

Steven Dillmann@StevenDillmann

We're releasing Terminal-Bench-Science: a benchmark for evaluating AI agents on research workflows across scientific domains.

An ongoing Stanford-led community effort, built by the team behind Terminal-Bench together with scientific domain experts at research institutions worldwide. v0.1 has 70 tasks. Claude Opus 5 solves only ~30%.

1/n 👇

💬4🔄4❤️60👀11147📊9

Long reasoning traces are expensive because full attention grows with context length. @Muennighoff ...

Stanford AI Lab (@StanfordAILab) — Thu, 27 Aug 2026 15:51:35 GMT

Long reasoning traces are expensive because full attention grows with context length.

@Muennighoff et al find that maybe LLMs can stand to forget a little… with Prefix Sliding!

Niklas Muennighoff@Muennighoff

new paper: Prefix Sliding for efficient test-time scaling

vanilla full attention OOMs on long tasks & compaction loses important details -- prefix sliding is a simple & fast alternative that can outperform both

📜https://t.co/fUw7yJAN5D

💬9🔄4❤️104👀18620📊22

Exciting!

Stanford AI Lab (@StanfordAILab) — Sun, 23 Aug 2026 15:46:52 GMT

Exciting!

Percy Liang@percyliang

🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs).  Post-training will follow.
Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run.  This is by far our biggest run, so definitely expecting the unexpected.

💬2🔄3❤️86👀23653📊10

The Embedder’s Dilemma— the ultimate question! Find out when you should use an embedding model vs an...

Stanford AI Lab (@StanfordAILab) — Thu, 20 Aug 2026 13:58:19 GMT

The Embedder’s Dilemma— the ultimate question! Find out when you should use an embedding model vs an LLM!

Niklas Muennighoff@Muennighoff

can LLMs replace embedding models? in a new work "embedder's dilemma" we find LLMs now beat embedding models -- but at much higher cost. when to choose which?...

📜arxiv.org/abs/2608.128750

💬6🔄16❤️89👀20865📊22

This amazing team shows how to build knowledge directly into Transformer blocks **without gradient d...

Stanford AI Lab (@StanfordAILab) — Thu, 23 Jul 2026 04:19:01 GMT

This amazing team shows how to build knowledge directly into Transformer blocks **without gradient descent**!

Jerry Liu@jerrywliu

MLPs store facts in language models. Can we write them into Transformers without training?

New work w/ amazing team @garctrob@ronnygjunkins@EyubogluSabri, Atri Rudra & @HazyResearch gives a ✨closed-form✨ recipe for fact-storing, Transformer-ready MLPs. Accepted at COLM 2026!

💬4🔄9❤️37👀13661📊10

Not having to wait for tokenization before training runs feels like paradigm shift

Stanford AI Lab (@StanfordAILab) — Wed, 22 Jul 2026 02:55:14 GMT

Not having to wait for tokenization before training runs feels like paradigm shift

Tatsunori Hashimoto@tatsu_hashimoto

Gigatoken is absurdly fast! Thanks to Marcel, I will (hopefully) never have to wait for tokenization again. 

It turns out that even mature components like tokenizers have order-of-magnitude improvements left if you get close to the hardware and write state machines.

💬5🔄1❤️54👀11313📊11

@stanfordnlp https://t.co/vjW3mwDouq

Stanford AI Lab (@StanfordAILab) — Tue, 21 Jul 2026 19:22:00 GMT

@stanfordnlpcs336.stanford.edu

💬0🔄0❤️0👀1208

@marcelroed casually just dropped the world’s fastest tokenizer implementation! He's also co-author ...

Stanford AI Lab (@StanfordAILab) — Tue, 21 Jul 2026 19:21:59 GMT

@marcelroed casually just dropped the world’s fastest tokenizer implementation! He's also co-author for Stanford’s CS 336 (LLMs from Scratch), where implementing a tokenizer from scratch is the first assignment. Fitting! 💪

Marcel Rød@marcelroed

Introducing the world's fastest tokenizer implementation, Gigatoken!

Gigatoken is ~500-1000x faster than HuggingFace, and ~100x faster than OpenAI's tiktoken for most tokenizer definitions on most machines.

These baselines are already multithreaded Rust implementations! 🧵

💬1🔄3❤️26👀5397📊6

Congratulations, Professor @chrmanning!

Stanford AI Lab (@StanfordAILab) — Tue, 21 Jul 2026 17:40:10 GMT

Congratulations, Professor @chrmanning!

Christopher Manning@chrmanning

Excited to see the Law in the Age of Generative AI special feature in @PNASNews!

AI Safety & Copyright, but also AI Governance, Statutory Interpretation & more!

My idea, but big thanks to Dan Ho, @JulianNyarko, @vanessaparli & @PNASNews for their work!

https://t.co/08EC6rQJYI

💬1🔄3❤️61👀16027📊8

Great paper club lineup organized by @FrancoisChauba1!

Stanford AI Lab (@StanfordAILab) — Thu, 16 Jul 2026 21:28:52 GMT

Great paper club lineup organized by @FrancoisChauba1!

Francois Chaubard@FrancoisChauba1

Great YC Paper Club Kernels/Chips edition last night. 

Papers:
    1    Stuart Sul (Stanford/Cursor) - ParallelKittens (arxiv.org/pdf/2511.13940)
    2    Avanika Narayan and Jon Saad-Falcon (Stanford) - Intelligence Per Watt: Measuring Intelligence Efficiency of Local AI (arxiv.org/pdf/2511.07885)
    3    Mark Saroufim (Core Automation) - To automate research we must automate systems (no paper)
    4    Misha Smelyanskiy (Nvidia/NewCo) - Why AI Inference Needs Heterogeneous Hardware (no paper)
    5    Brennan Shacklett (Stanford) - An Extensible, Data-Oriented Architecture for High-Performance, Many-World Simulation (madrona-engine.github.io/shacklett_sigg…)

pic of @marksaroufim teaching CUDA! 

next one on robotics.. 

who should we invite to present?

💬4🔄2❤️9👀10894📊5

Last day of #ICML2026! Catch "Scale Dependent Data Duplication" to close out the conference.

Stanford AI Lab (@StanfordAILab) — Sat, 11 Jul 2026 00:48:12 GMT

Last day of #ICML2026! Catch "Scale Dependent Data Duplication" to close out the conference.

Jessica Chudnovsky ✈️ ICML 2026@jchudnov

Two papers accepted at the #ICML2026 workshop on the Impact of Memorization on Trustworthy Foundation Models (MemFM), and I'll be presenting both today! 

Scale Dependent Data Duplication: (our first debut of this paper), led by @JoshuaK92829 and Noam Levi. 

Internal Data Repetition Destroys Language Models: in case you missed it yesterday.

📍Hall A, 3:50-5:00pm (Jul 11)

Hope to see you there!

💬4🔄1❤️21👀12294📊5

Verification has emerged as a new scaling axis 🚀 LLM-as-a-Verifier shows that scaling verification...

Stanford AI Lab (@StanfordAILab) — Thu, 09 Jul 2026 23:45:42 GMT

Verification has emerged as a new scaling axis 🚀

LLM-as-a-Verifier shows that scaling verification can push performance to SOTA across Terminal-Bench V2, SWE-Bench Verified, RoboRewardBench, and MedAgentBench.

Its fine-grained feedback can also serve as a proxy for estimating task progress and improve RL sample efficiency!

Congrats to @jackyk02@shululi256@pranav_atreya@liu_yuejiang@jyx_su@chelseabfinn@drmapavone@istoica05@Azaliamirh

Jacky Kwok@jackyk02

How can we extract richer signals from AI Feedback?

Introducing LLM-as-a-Verifier✨— a simple verification scaling framework that achieves SOTA on agentic benchmarks 🚀

The key idea:
- Use fine-grained scoring granularity (e.g., 1-20 instead of the standard 1-5 scale)
- Take the expectation over the full logprob distribution of score tokens
- Scale repeated evaluation and criteria decomposition

You can use these fine-grained signals for more effective test-time scaling, RL, and agent monitoring! It achieves SOTA across Terminal-Bench V2, SWE-Bench Verified, RoboRewardBench, and MedAgentBench 👑

Advised by @Azaliamirh@istoica05@drmapavone@chelseabfinn 

🧵👇

💬5🔄8❤️78👀14218📊18

Check out TRACE, a new self-improvement approach where the agent identifies the missing capabilities...

Stanford AI Lab (@StanfordAILab) — Thu, 09 Jul 2026 23:45:00 GMT

Check out TRACE, a new self-improvement approach where the agent identifies the missing capabilities behind its own failures and trains itself to address them.  

TRACE-trained Qwen3.6-27B reaches 73.2% on SWE-bench Verified, outperforming much larger models like Codex 5.2 and GLM 5, while beating GRPO and GEPA with <1/4 the training rollouts.

Exciting work led by @hangoo_kang and @TarunSures41845!

Hangoo Kang @ ICML ✈️@hangoo_kang

“TRACE: Capability-Targeted Agentic Training” got Spotlight @ ICML AIWILD 🎉

Beats direct RL, GEPA, & synthetic-agent data on SWE-Bench Verified and τ²-Bench.

TRACE-Qwen3.6-27B tops GPT-5.2-Codex, GLM 5, & Claude 4.5 Sonnet on SWE-Bench.

Co-led with @TarunSures41845. Thanks to @JonSaadFalcon and our advisor @Azaliamirh. Details below 👇

💬4🔄8❤️78👀10870📊16

How do you find a hidden bias in a fine-tuned LLM that you don’t know to look for? Amplify it. Dist...

Stanford AI Lab (@StanfordAILab) — Thu, 09 Jul 2026 23:30:47 GMT

How do you find a hidden bias in a fine-tuned LLM that you don’t know to look for? Amplify it.

Distill to Detect (D2D) surfaces hidden biases by distilling the shift between a suspected model and its base into a small cartridge.

This amplifies stealth preferences into generated text, making subliminal signals visible to existing auditing methods.

Congrats to @talaei_shayan, @AbhinavChinta10, @Devvrit_Khatri, @aminkarbasi, @Azaliamirh, and Amin Saberi!

Shayan Talaei@talaei_shayan

Suppose you're handed a fine-tuned LLM that secretly favors a certain entity. The bias goes completely undetected because it only surfaces on one specific unknown topic.

So how do you catch a bias you can't search for? You amplify it.

Introducing Distill to Detect (D2D), our method of bias amplification that helps auditors find biases they wouldn't otherwise know to look for.

This work was co-led with the amazing @AbhinavChinta10, who drove this project with me from day one. Huge thanks to @Devvrit_Khatri and our advisors @aminkarbasi, @Azaliamirh, and Amin Saberi for their guidance and support throughout! 🙏

📄 Paperarxiv.org/abs/2607.01208IU
📝 Blodistill2detect.github.ioxmf
💻 Cogithub.com/abhinav-chinta…VXcK

For more information, please see the thread below. 🧵

💬4🔄2❤️18👀5790📊5

Check out this oral presentation today & runner up at the Foundations of Deep Generative Models ...

Stanford AI Lab (@StanfordAILab) — Thu, 09 Jul 2026 23:29:54 GMT

Check out this oral presentation today & runner up at the Foundations of Deep Generative Models workshop!

#ICM#ICML2026ngrats to @jchudnov, @JoshuaK92829, Noam Levi, @RylanSchaeffer, @yegordb, @sanmikoyejo, and David Donoho.

Jessica Chudnovsky ✈️ ICML 2026@jchudnov

Presenting “Internal Data Repetition Destroys Language Models” today as an oral at the Foundations of Deep Generative Models workshop (#ICML2026)! Hope to see you there.

⏰ 11:15am Friday July 10 
📍 Conference room S 318

Also, some exciting news: we're grateful and honored to have received the workshop's Runner-up award!

With @JoshuaK92829,  Noam Levi*, @RylanSchaeffer, @yegordb, @sanmikoyejo, and David Donoho

💬5🔄3❤️30👀8546📊9

Are you at ICML 2026 in Seoul? 🇰🇷 Check out the full list of papers from Stanford AI Lab — spannin...

Stanford AI Lab (@StanfordAILab) — Mon, 06 Jul 2026 11:07:13 GMT

Are you at ICML 2026 in Seoul? 🇰🇷 Check out the full list of papers from Stanford AI Lab — spanning coding agents, LLM reasoning, evaluation & benchmarks, AI safety & interpretability, and AI for science. See you there!
https://t.co/IUVXc0tebi

💬10🔄18❤️149👀18203📊32

LLMs are better at predicting what other models will say than what’s actually true. When they’re wro...

Stanford AI Lab (@StanfordAILab) — Sun, 05 Jul 2026 17:34:23 GMT

LLMs are better at predicting what other models will say than what’s actually true. When they’re wrong, they’re wrong together; so polling can’t recover the truth. Check out this paper at ICML! 🇰🇷

Jessica Chudnovsky ✈️ ICML 2026@jchudnov

Pass@k and self-consistency work great for math and code; sample more and verify. So we asked: can the same trick scale truthfulness in domains with no verifier? The answer was no.

Excited to share our #ICML2026 conference paper: Truthfulness Does Not Scale Like Reasoning. arxiv.org/pdf/2603.06612. I’ll be at ICML in Seoul to present it!

Co-led by @JoshuaK92829 and @yegordb

💬5🔄13❤️123👀24375📊25

Congratulations to the team on the #ICML2026 workshop Oral! @jchudnov (co-lead), @JoshuaK92829 (co...

Stanford AI Lab (@StanfordAILab) — Sun, 05 Jul 2026 00:20:59 GMT

Congratulations to the team on the #ICML2026 workshop Oral! 

@jchudnov (co-lead), @JoshuaK92829
(co-lead), Noam Levi (co-lead), @RylanSchaeffer, @yegordb, Bo He, Mehmet Donmez, @sanmikoyejo, and David Donoho.

💬2🔄0❤️4👀3891📊3

Deduplication is standard practice but never perfect --this work measures what the residue costs in ...

Stanford AI Lab (@StanfordAILab) — Sun, 05 Jul 2026 00:20:10 GMT

Deduplication is standard practice but never perfect --this work measures what the residue costs in compute-equivalent terms, and shows the worst-case repetition structure is predictable from model size. The wrong combination can waste as much as 33% of compute!

Jessica Chudnovsky ✈️ ICML 2026@jchudnov

Flying to #ICML2026 to present Internal Data Repetition Destroys Language Models, an Oral at Foundations of Deep Gen Models Workshop!

Paper: arxiv.org/abs/2606.24998

You might be curious to know what we mean by “destroys”! Pretraining is now data-constrained, and even aggressively deduplicated corpora keep some repetition. We measured what that repetition actually costs in the currency practitioners care about: compute. The answer, in the worst case, is a third of your FLOPs.

💬3🔄7❤️80👀16758📊13

What if the way we collect human feedback in robotics is quietly losing information? If one traject...

Stanford AI Lab (@StanfordAILab) — Thu, 02 Jul 2026 18:02:51 GMT

What if the way we collect human feedback in robotics is quietly losing information?

If one trajectory drops the cutlery and another drops the plate, asking for a single preference hides information behind the choice.

Freeform Preference Learning lets annotators describe preference axes in natural language, learns rewards conditioned on those axes, and extracts better policies. 

Congrats to @marceltornev, @anubhamahajan01,@AbhijnyaBhat, and @chelseabfinn!

Marcel Torné@marceltornev

We should stop optimizing robot policies against a single overall reward. Trajectories differ along many axes, such as speed, precision, and subtask completion, and one can be better on some while worse on others. If we collapse all of that into a single overall axis we lose this structure making the reward ambiguous and harder to optimize.

Blog: freeform-pl.github.io/fpl.website/
Paper: arxiv.org/abs/2606.32027

Your browser does not support the video tag.

💬5🔄5❤️58👀13065📊12

We love scaling inference compute, but it’s costly! Independently sampling parallel attempts might b...

Stanford AI Lab (@StanfordAILab) — Thu, 02 Jul 2026 17:44:27 GMT

We love scaling inference compute, but it’s costly! Independently sampling parallel attempts might be the culprit: it wastes compute rediscovering the same solutions. What if we scaled inference compute with correlated samples?

Check out QuasiMoTTo by @michaelyli_ and @probablynotaz9!

Michael Y. Li@michaelyli_

You're wasting FLOPs when scaling inference compute: by independently sampling parallel attempts, you burn compute rediscovering the same solutions.

Introducing QuasiMoTTo: we scale parallel sampling with correlated samples instead! These samples have higher coverage, are marginally exact draws from the LLM, and can be generated in parallel.

Result: same performance with 25-47% fewer samples in test-time scaling + 50% fewer training steps in RL!

In our new paper, we explore the design space of correlated samplers. Work with co-authors @probablynotaz9 (co-lead), @gandhikanishk, @noahdgoodman, and Emily Fox!

💬2🔄7❤️51👀12429📊10

@milanganai @katielulula Read the post here! https://t.co/02bfnFx1Rm

Stanford AI Lab (@StanfordAILab) — Tue, 30 Jun 2026 06:36:26 GMT

@milanganai@katielulula Read the post here! ai.stanford.edu/blog/rnb-encor…

💬0🔄2❤️4👀3174📊1

New SAIL Blog post: R&B-EnCoRe: What Should a Robot Actually Think About Before It Acts? @milan...

Stanford AI Lab (@StanfordAILab) — Tue, 30 Jun 2026 06:36:17 GMT

New SAIL Blog post: R&B-EnCoRe: What Should a Robot Actually Think About Before It Acts?

@milanganai and @katielulula show how a Vision-Language-Action model can teach itself which chain-of-thought actually helps it act; no rewards, no verifiers, no human annotation.

💬3🔄4❤️20👀7114📊8

By @kushin_m, @danielwurgaft, @LinasNasvytis, @michaelyli_, @noahdgoodman, and @mcxfrank!

Stanford AI Lab (@StanfordAILab) — Fri, 26 Jun 2026 21:04:54 GMT

By @kushin_m, @danielwurgaft, @LinasNasvytis, @michaelyli_, @noahdgoodman, and @mcxfrank!

💬0🔄0❤️4👀1286📊1

Auto-psych: a new system where AI agents do the whole science loop themselves -- come up with theori...

Stanford AI Lab (@StanfordAILab) — Fri, 26 Jun 2026 20:59:18 GMT

Auto-psych: a new system where AI agents do the whole science loop themselves -- come up with theories, build experiments, recruit real people online, and improve based on the results!

noahdgoodman@noahdgoodman

every few years i say “what if we could automate psychology research?” this time i think we’ve made real progress (because ai).

💬3🔄6❤️22👀11366📊6

Congrats to @OpenJarvisAI! The future of LM inference runs beyond the datacenter: local by default, ...

Stanford AI Lab (@StanfordAILab) — Thu, 25 Jun 2026 19:51:58 GMT

Congrats to @OpenJarvisAI! The future of LM inference runs beyond the datacenter: local by default, hybrid by design.

Jon Saad-Falcon@JonSaadFalcon

Excited to share that @OpenJarvisAI is part of @LaudeInstitute's Slingshots // THREE alongside an incredible batch of researchers!
OpenJarvis is an open-source framework for personal AI that lives on your device first: agents run locally by default and reach for the cloud only when a task truly calls for it, with energy, cost, and latency treated as first-class metrics. 

Grateful to the whole Laude team for the support, and to my collaborators @Avanika15, @HazyResearch, and @Azaliamirh at @StanfordAILab.

💬3🔄6❤️20👀8459📊6

Most open agentic datasets target one benchmark. In compute-controlled comparisons, OpenThoughts-Age...

Stanford AI Lab (@StanfordAILab) — Thu, 25 Jun 2026 07:51:39 GMT

Most open agentic datasets target one benchmark. In compute-controlled comparisons, OpenThoughts-Agent-v2 leads at every training set size, and generalizes across seven agentic benchmarks. Check it out!

Richard Zhuang@RichardZ412

How can we train small agentic models that are highly capable of terminal use and coding?

Announcing OpenThoughts-Agent + OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)

💬9🔄10❤️80👀12916📊18

By the wonderful team @jubayer_hamid, @ifdita_hasan, @michaelyli_, @oshaikh13, @yoonholeee, @DorsaS...

Stanford AI Lab (@StanfordAILab) — Wed, 24 Jun 2026 22:11:56 GMT

By the wonderful team @jubayer_hamid, 
@ifdita_hasan, @michaelyli_, @oshaikh13, @yoonholeee, @DorsaSadigh, @chelseabfinn, and @noahdgoodman

💬0🔄2❤️5👀2446📊1

At test time, we wrap LLMs in scaffolds that scale compute every which way -- longer chains, paralle...

Stanford AI Lab (@StanfordAILab) — Wed, 24 Jun 2026 22:07:36 GMT

At test time, we wrap LLMs in scaffolds that scale compute every which way -- longer chains, parallel samples, and aggregation across them. So why do we still train them to use only one of these?

Introduce Spiral: it uses set RL to teach a model to generate responses that are collectively useful for an aggregator, and standard RL to teach it to aggregate those responses into an improved answer!

Jubayer Ibn Hamid@jubayer_hamid

The most capable reasoning systems in AI scale inference compute along several axes: sequential compute to think longer, parallel compute to sample many independent attempts, and aggregative compute to synthesize prior traces into a new improved one. But during training, we only optimize how models use sequential compute. This creates a fundamental mismatch between how we ultimately deploy these systems and how we train them, leaving much of search and synthesis unoptimized.

We introduce SPIRAL, an RL framework for making all inference-compute primitives end-to-end learnable: models learn to coordinate sequential, parallel, and aggregative reasoning using only the reward of the final output.  Work with @ifdita_hasan (co-lead), @michaelyli_ , @oshaikh13 , @yoonholeee , @DorsaSadigh , @chelseabfinn , @noahdgoodman 🧵

💬3🔄6❤️33👀8918📊9