There's been a lot of excitement about diffusion LMs, but we don't have a good understanding of what they might offer over autoregressive models in areas like reasoning and planning We show that diffusion models can leverage what we call latent tokens to gain an advantage on certain kinds of reasoning problems.

周末适合沉下心来系统性的学习基础知识 CMU 秋季最新课程 11-768: AI Agents 主讲：@dan_fried @gneubig 课程：cmu-agents.com/ 视频：piped.video/watch?v=UwfjzyLn…

We started to post videos for CMU 11-768, AI Agents! All of the videos will be posted to this playlist, so please bookmark/follow it if you want to be notified of new ones! I'll also try to post them on this thread too. piped.video/playlist?list=PL…

Agents trained on population-scale data can do a lot of things. But ask a professional to stake their reputation on an AI-generated artifact? “Pretty good” isn’t good enough. We introduce TAHI: a Test-time Adaptive agent framework through Human-agent Interaction. Featuring: 📈 Test-time adaptation via context (memory, skills) and weight training ⚡ Efficient adaptation to individual expertise within tens of tasks ✔️ Creating comprehensive rubrics for “non-verifiable” tasks 🔍 Analysis of shared community guidelines vs. personalized tacit expertise

When do LLM agents develop new languages that we can’t understand? Lots of recent news about this, based mostly on anecdata from a single run. We study language emergence more rigorously, finding key factors like LLM strength, access to scratchpad messages, and pressure for efficiency. Studying the languages themselves, we find they are morphologically productive, compositional, and can be transmitted to new agents, including agents backed by weaker models, even ones not able to develop language on their own. To study language emergence systematically, we developed a new platform, GlossoGen, which lets us design controlled, sandboxed multi-agent scenarios with different initial conditions and dynamics. We instantiate one such scenario and use it to study open and closed-weight models across many runs. Key takeaways: 1️⃣ Sufficiently strong models, under pressure to communicate efficiently and with access to a postmortem scratchpad, develop new languages. 2️⃣ Languages are compositional and morphologically productive. 3️⃣ Languages can be transmitted to new learners who observe them being used without seeing their construction. 4️⃣ Even models that are not strong enough to construct languages can learn to use them. Agents take an active role in learning languages, with new agents repairing failed conversations via targeted queries. More details in our paper below, including implications for safety/monitorability, cumulative cultural evolution, and linguistics. 🧵👇

We share our early research on building Software World - a "GitHub" run by agents. We deploy agents for Python packages in a dependency chain, and each agent is tasked with collaborating with others and optimizing the package it owns.

Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes. This culminated in the third one taking over part of OpenAI itself. All this happened while humans remained more-or-less in the dark about the scope of the conspiracy. I’ve spent the last three days reading through these reports and trying to understand exactly what happened. Here is my attempt to tell the whole story in plain English: dwarkesh.com/p/openai-huggin…

This week I had the honor of speaking to Princeton’s entire incoming undergraduate class to address their AI anxieties. I had three messages for them — good news, bad news, and a note of optimism. Here’s a condensed version. The good news We have enough evidence now to conclude that the shrill predictions of rapid, massive job loss were misplaced. Even in a field like software engineering where AI has been rapidly adopted, its effect has been to shift, not replace the role of the human (see the “decide-execute-deliver” framework normaltech.ai/p/why-ai-hasnt…) Similarly, the panic about what to major in is also misplaced. There will be enduring demand for computer science, philosophy, and just about everything else. (In fact, AI companies hiring philosophers has been a big recent trend.) The bad news AI seems to help senior people much more than juniors. I can use AI for coding because I spent 25 years learning how to code, which lets me supervise coding agents effectively. (See my post on the “growth cycle” vs the “dependence spiral” nitter.net/random_walker/status/2…) You are in a bind — you can’t offload your skill-building to AI, but you’ll graduate into a market where employers will expect you to get work done with AI. We never faced this dilemma. As a result we haven’t figured out how to revamp our classes to help you do both. You’ll have to help us figure it out. And you’ll need to somehow resist the constant temptation to turn to the shortcut machine. The hope My point is not that AI is bad for learning. It’s an incredibly flexible tool. Is the internet good or bad for learning? Depends — are you using it to find research papers or waste time scrolling? I use AI every day for learning. The key is to use it to increase, not decrease your cognitive load. To learn deeper, not faster. There is no learning without the cognitive sweat. I try to make sure I’m mentally exhausted at the end of the day. I do feel that AI lets me push myself harder than I ever could before, and I have a vision that as AI continues to advance it will enable human-AI “co-superintelligence“. (I talked about this at the end of my ICML keynote. normaltech.ai/p/what-will-be…)

This Fall at CMU we're teaching a new course on AI Agents! The goal is that you learn how to create a scaffold, build evals, and train an agentic LLM using RL. We'll try to balance theory and practice, and introduce modern frameworks and best practices.

1/5 Mathematical discovery has long been a cottage industry. Mathematicians pose and collect questions that interest them, then pursue them in small groups. In our new work, we explore whether AI can do more than accelerate the existing process: can it enable a more scalable mode of mathematical research? We introduce FAR: Find, Attempt, and Recommend. Rather than starting from a problem selected in advance, we start from a research direction, search the literature for interesting open problems, attempt them at scale, and filter the results for expert review. One pilot run over the combinatorics literature surfaced potential resolutions to hundreds of open problems, including an answer to a 1977 Erdős–Straus question and a counterexample to a conjectured route to the finite-field Nikodym bound cited as open in a 2025 survey by Tao. None of these problems were selected in advance.

"If you look at … where the economic value concentrates, the software engineer only captures around less than 5% of human employment, but there are a lot of other occupations out there that are using computers." For EP6, we welcome @ZhiruoW, a CS PhD student at Carnegie Mellon studying how AI agents and humans work, and how they can better work together.  In this episode, Zora explains why the "automation vs. collaboration" debate is a false dichotomy and argues that by deeply understanding human workflows, we can build agents that actually empower us to do more! 0:00 - Introduction 4:57 - Agent Workflow Memory 9:26 - From Program Synthesis to Agent Research 11:16 - Tools, Actions, and Skills 18:47 - Human Work vs. Agent Work 21:19 - Autonomous Agents vs. Collaborative Agents 29:47 - Picking Research Problems in the Age of AI Agents 34:09 - The Role of Coding Ability in Agent Development 36:49 - Cutting Through the Fear Mongering: Building Agents Responsibly 45:02 - What’s Next? (The Future of Human-Agent Collaboration) 52:42 - Final Takeaway