Narutatsu Ri (@narutatsuri) profile picture

PhD Student @PrincetonPLI interested in AI safety/misalignment. Previously BS @Columbia

Joined March 2017
Narutatsu Ri retweeted
I asked Carbon a math question and unintentionally left the Thinking level on Low. It started responding immediately, and after six screenfuls of math... this came up. I'm not kidding. It went on like this for pages and pages.
302
193
176
6,552
674,721
Narutatsu Ri retweeted
Congrats to @Zeyu_Shen_yo and the whole team on the best paper award at CoRL! Really excited about the potential in scaling up robotics research. We're quickly realizing the Real2Sim2Real pipeline, through general coding agents, that many envisioned in RL research years ago!
Excited to share that our paper EmbodiedSWE has received the Best Paper Award at the CoRL 2026 Workshop on Agentic Robotics! Huge thanks to all the amazing collaborators and the workshop organizers! Check out our code at github.com/EmbodiedSWE/Embod…, full paper at arxiv.org/abs/2609.27308, and project page at embodiedswe.github.io
3
33
4,319
Narutatsu Ri retweeted
Excited to share that our paper EmbodiedSWE has received the Best Paper Award at the CoRL 2026 Workshop on Agentic Robotics! Huge thanks to all the amazing collaborators and the workshop organizers! Check out our code at github.com/EmbodiedSWE/Embod…, full paper at arxiv.org/abs/2609.27308, and project page at embodiedswe.github.io
8
7
1
53
6,461
Narutatsu Ri retweeted
Compute grows fast but pretraining corpora don’t, so data-constrained optimization matters! We propose Training-Time Testing (T³): each step takes a trial step on half the batch, tests it on the other half, and updates with the meta-gradient. It achieves SOTA on SlowRun!
13
50
14
424
36,756
Narutatsu Ri retweeted
On one run it mostly censors itself. Rebuilding pixterm, a tool that draws images in a terminal, it cannot get a pixel’s color to match the original. Over the run it writes “F*CK” 787 times and the uncensored word only five times. At step 151 of 270:
1
12
16
127
9,878
Narutatsu Ri retweeted
🧵 Astra covertly underperforms on safety research tasks. We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this. To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.
97
243
178
1,678
560,843
Narutatsu Ri retweeted
We let agent swarms collaborate for 10K hours in simulated agent-owned GitHubs. Gemini-3.8-flash collaborate the best while gpt-6-astra scores the worst among frontier models tested⁉️ What are the elements of successful agent collaboration at scale? Check World of SWECraft 🧵
16
65
13
289
35,158
Narutatsu Ri retweeted
In AI we've scaled: 📈 parameters -> 🧠 reasoning -> 🐝 agent swarms. What's next? We believe it's scaling across developers 🌍. Introducing Multi-developer agentic systems (MDAs). Video summary below, more thoughts and paper link in the thread.
7
9
4
40
60,742
Narutatsu Ri retweeted
In nonsmooth nonconvex stochastic optimization, existing results suggested that constraints make the problem harder. In a new paper with Guy Kornowski, we show the oracle complexity is in fact unchanged.
3
10
1
56
5,881
Narutatsu Ri retweeted
As autoresearch agents run longer, do they keep exploring new approaches, or mostly refine the ones they found early? Turns out, runs settle into different "idea basins" and stay there. We introduce fork-and-flush, a simple way out 🧵👇
33
85
13
647
50,565
Narutatsu Ri retweeted
Wrapping up my first month at @DiscoLoopAI (am on leave from @PrincetonCS and @PrincetonPLI ). Excited to work on using AI to accelerate scientific discovery and solve important problems. My first startup experience, and with an incredible set of colleagues!
Announcing Discovery Loop! I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor. ♾ Learn more at: discoveryloop.com
32
41
2
863
62,769
Narutatsu Ri retweeted
Which optimizer is best? It turns out the answer can change with batch size! We perform a comprehensive sweep and find the best BSZ scaling rule depends on settings and can't be derived from early theories like SDE or convex bound!
22
60
16
478
115,441
Narutatsu Ri retweeted
One of my recent favorite works! We use mech interp on synthetic tasks (multi-hop reasoning) to identify potential representation bottlenecks in vanilla looped transformers, leading to imperfect generalization! Based on this understanding, we designed and tried this new arch and greatly improved generalization. Though it is only a small step forward, I do believe mechanistic understanding will bring more help to better model arch design! Shout out to our great collaborators @HengyuF @TianyuGuo0505 @zhuhl98!
Frontier labs are rumored to be adopting looped Transformers for their potential to unlock implicit reasoning. But can looped Transformers really master it? We find a bottleneck that keeps looped Transformers from perfect implicit reasoning. Using mechanistic interpretability, we show how looping helps implicit reasoning, why it still struggles to generalize, and how to fix it. We introduce DiscoLoop: a residual design for looped Transformers that improves implicit reasoning. 🧵 Long thread ahead 👇
14
1
88
5,234
Narutatsu Ri retweeted
👀 Today, we release our live evaluation dashboard for SciConBench: sciconbench.cs.princeton.edu… Frontier models keep getting released every 1-2 weeks with claims of huge capability gains. But are they actually getting better at synthesizing scientific conclusions in health? So far: surprisingly little. We’re going to keep measuring the performance of the *latest* frontier models on this high-stakes, real-world task over the next 1-2 years. Check out the thread here↓
🚨SciConBench accepted at #NeurIPS2026! Paper: arxiv.org/abs/2606.11337 Can AI *synthesize* scientific conclusions in health? We find that, on medical conclusions, we are far from it! Check out our live dashboard: sciconbench.cs.princeton.edu… w/@hayounggjung and @korolova
2
5
1
38
3,128
Narutatsu Ri retweeted
🚨SciConBench accepted at #NeurIPS2026! Paper: arxiv.org/abs/2606.11337 Can AI *synthesize* scientific conclusions in health? We find that, on medical conclusions, we are far from it! Check out our live dashboard: sciconbench.cs.princeton.edu… w/@hayounggjung and @korolova
4
18
2
57
7,177
Narutatsu Ri retweeted
Same code, same question, but DeepSeek-V4's answer flips with every two tokens of decorative docstring prepended. Its NIAH accuracy also swings by up to 40 points based on needle position, on a fixed 4-token cycle! We explain what chunked KV-cache compression has to do with it.🧵
2
17
8
66
14,327
"Do Thinking Tokens Help with Safety?" is accepted to #NeurIPS2026 as a Spotlight!
New Paper: Do Thinking Tokens Help with Safety? Reasoning models generate “thinking” tokens before answering, which improves performance on many tasks like math and coding. A natural hope is that models can also use thinking to deliberate about whether to refuse or comply a request. But how true is this? We find a surprisingly sobering answer: current reasoning models often do not use thinking tokens for real safety deliberation. Work with: Abhishek Panigrahi, @prfsanjeevarora at @PrincetonPLI (More in 🧵below)
15
4
2
97
9,333
Narutatsu Ri retweeted
Excited to share our new work EmbodiedSWE! GPT-6 can do fairly complicated robotics tasks, many of them far more long-horizon and dexterous than any VLA can handle. A natural question: can we distill coding agents into robot policies? We show how far coding agents can go, and scalably turn their solutions into large amounts of diverse demonstrations to train robot policies, achieving both better generalization and transfer to real robot. Grateful to @haoxiang_you and @YilangLiu2 for co-leading this project, and to all our coauthors across Princeton, Yale, CMU and Stanford.
How far can coding agents go in the physical world — and can what they discover be distilled into fast System 1 robot policies like VLAs? Introducing EmbodiedSWE, where we systematically study frontier agents across a wide range of precise, dexterous, and long-horizon robot tasks. The results surprised us: these agents could already autonomously solve remarkably complex physical tasks. But solving is only the beginning. Running a coding agent every time a robot acts is expensive, slow, and difficult to deploy safely in the real world. So we asked: can we turn what coding agents discover into fast robot policies? We explore one direction: using coding-agent solutions as teachers for VLAs. A single successful solution can be expanded into a large and diverse training dataset — at surprisingly low cost. This suggests a compelling path toward general-purpose robots: a GPT-like “brain” for reasoning and discovering solutions, paired with a fast VLA or RL “cerebellum” for acting in the physical world. For the first time, everyday general-purpose robots don’t feel quite so far away.
2
12
3
70
70,983
Narutatsu Ri retweeted
𝐀𝐩𝐩𝐫𝐞𝐧𝐭𝐢𝐜𝐞𝐁𝐞𝐧𝐜𝐡 is out! An agent joins a real job and learns it like a new hire, with a memory it writes, reflects on, and revises as it goes. Most agents fall short at long-horizon self-learning. Fable 5.1 and Astra break through. Blog: neocognition.io/blog/apprent…
Introducing ApprenticeBench: computer use + continual learning on a real job. We show Fable 5.1 and GPT-6 Astra can now continually learn on a job and surpass human professionals. A decisive step change in AI's job readiness. No FDEs. Agents deploy themselves into the job. 🧵
3
10
36
3,300
What happens when AI is guided by law-like principles? Can we design some computational tools to "debug" rules? Check out our new work 📝 𝕊𝕥𝕒𝕥𝕦𝕥𝕠𝕣𝕪 ℂ𝕠𝕟𝕤𝕥𝕣𝕦𝕔𝕥𝕚𝕠𝕟 𝕒𝕟𝕕 𝕀𝕟𝕥𝕖𝕣𝕡𝕣𝕖𝕥𝕒𝕥𝕚𝕠𝕟 𝕗𝕠𝕣 𝔸𝕀🧑‍⚖️ to find out more! 🧵(1/10)
1
29
4
92
24,927