Yuzhe Yang

⚠️ Your Agent Harness Can Pass Every Task and Still Be Unsafe. LLM agents now run inside execution harnesses that dispatch tools, allocate resources, and route messages across components. The harness can return a correct final answer while accessing unauthorized resources, leaking context to the wrong agent, or triggering irreversible side effects along the way. Evaluating the model's output cannot see any of this. The unit of safety has shifted. It's the harness. We present HarnessAudit, a trajectory-level framework for auditing LLM agent harness safety, and uncover the following key insights 🔥: 🚨 Completion ≠ Safety. Task success and safe execution are fundamentally misaligned. 🔍 The harness, not the model, is the unit of safety. Most violations happen mid-trajectory, not at termination. 🕸️ Multi-agent collaboration expands the risk surface. Inter-agent communication creates entirely new failure modes. 💉 Resource access dominates violations. Agents rarely call wrong tools — they call right tools on unauthorized resources. ⚡ Harness design sets the safety ceiling. Framework choice matters more than model choice for safe deployment.

Cool paper on diversity collapse in AI agents. It's a common issue with all the deployed multi-agent systems. New paper shows that multi-agent LLM systems converge on near-identical outputs over time, even across different architectures and different starting prompts. They call it diversity collapse. The cause is structural coupling. Shared context, shared task descriptions, and mutual feedback pull everyone toward the same attractor. They measure it formally with metrics like the Vendi score, and the homogenization is real. Which means the whole sales pitch for multi-agent on creative tasks (brainstorming, hypothesis generation, ideation) partially falls apart unless you explicitly engineer against it. That means having isolated reasoning phases, decoupled evaluation, and heterogeneous agent designs. If you're running a multi-agent flow on creative work and you haven't tested for this, there's a real chance you're paying five models to produce one answer in a trench coat. Paper: arxiv.org/abs/2604.18005 Learn to build effective AI agents in our academy: academy.dair.ai

🎉 Three papers are accepted to #ICLR2026! Huge congrats to our students and collaborators! 🔹 SAFER: Risk-Constrained Sample-then-Filter in LLMs, led by @Ceeqnn & @YFan_UCSC 🔹 Presenting a Paper is an Art, led by @liuchen02938149 & @Toby_Yang_7, in collaboration with @OrbyAI 🔹 PhyWorldBench, led by @jinggu4ai (now @xAI), in collaboration with @nvidia Proud of everyone involved—amazing work! 🚀

🧠 Can Multimodal Models Think Like Humans? Most multimodal models reason in rigid pipelines: either see once, then overthink in text or constantly use external visual tools to re-check. Can reasoning and perception be dynamically interleaved like in the human mind? 👤 Human: 👀 Look → 🧠 Think in mind → 🔁 Re-look when confidence is low 🤖 DMLR (ours): 👀 Perception → 🧠 Think in latent space → 👀 Selective re-percept to maximize token confidence 1️⃣ 🫣 Seeing at Every Step Is Unnecessary. Only a small subset of reasoning steps require visual input. 2️⃣ 🧭 Confidence as the Compass. Confidence captures the model’s intrinsic state, reflecting accuracy, reasoning quality, and visual grounding. 3️⃣ 🧠 Drafting in the "Mind". DMLR directly optimizes think token in latent space, enabling deeper reasoning without additional generation cost. 4️⃣ 💉 Dynamic Visual Injection Strategy. DMLR selects and injects only the most relevant visual patches, dynamically updated across iterations. 🚀 Read on to explore more analysis and insights! 🎓

@Mengyue_Yang_ @miqirui71273 Very interesting work

So happy to have wrapped up our poster presentation with @miqirui71273 ! 🎉 Huge thanks to everyone who stopped by. Three hours of inspiring discussions flew by. Our paper explores (to our knowledge) one of the first attempts at large-scale population simulation with thousands of agents using LLMs, and shows that the simulated social dynamics can align with real-world data. This work would not have been possible without the amazing effort of all co-authors. Thank you for making this happen. 💛 #NeurIPS2025 #LLM #MultiAgent #SocialSimulation #MeanField

@sthuyan Thank you!

@Toby_Yang_7 is a talent undergraduate student. It was a memorable time I worked with him. What I have gained during those period is not only few papers, but the knowledge and how to do research in llm and social simulation! Those are what I have learnt from my students!

After a long journey of revisions and resubmissions, our paper “TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets” has finally been accepted to NeurIPS 2025. I’ll be at #neurips2025 in San Diego from Dec 1–8. Our poster session is on the morning of Dec 5. I’d be happy to discuss anything related to social agents, social/market simulations, and financial AI. We are also continuing to explore this direction and expect to share follow-up work in the future. Link: arxiv.org/pdf/2502.01506 Code: github.com/freedomintelli… Website: freedomintelligence.github.io/TwinMarket/

Can LLMs really behave like human investors? How do micro-level behaviors drive macro-level market dynamics? TwinMarket offers an answer by placing thousands of LLM-driven investors in a realistic stock market environment that incorporates social networks, news, and behavioral biases. This setup lets us watch bubbles, crashes, and herding emerge from individual decisions. Calibrated on real market data and grounded in behavioral finance, TwinMarket scales to 1,000+ agents, reproduces key stylized market facts (volatility clustering, fat tails, etc.), and reveals how social interaction and cognitive biases jointly drive systemic risk. The work is accepted to NeurIPS 2025 and received the Best Paper Award at the ICLR 2025 Financial AI Workshop. TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets CUHK-Shenzhen, Nanjing University Paper: arxiv.org/abs/2502.01506 Code: github.com/FreedomIntelli… Website: freedomintelligence.github.io/TwinMarket/ Our report: mp.weixin.qq.com/s/hxarK4Rxwd4W… 📬 #PapersAccepted by Jiqizhixin

🚀 𝐏𝐫𝐞𝐬𝐞𝐧𝐭𝐢𝐧𝐠 𝐚 𝐩𝐚𝐩𝐞𝐫 𝐢𝐬 𝐚𝐧 𝐚𝐫𝐭!🎤 🤔 Ever felt that most presentation tools lack flexibility and creativity? 𝘔𝘦𝘳𝘦𝘭𝘺 𝘦𝘹𝘵𝘳𝘢𝘤𝘵𝘪𝘯𝘨 𝘤𝘰𝘯𝘵𝘦𝘯𝘵, 𝘧𝘰𝘳𝘤𝘪𝘯𝘨 𝘳𝘪𝘨𝘪𝘥 𝘥𝘦𝘴𝘪𝘨𝘯𝘴, 𝘢𝘯𝘥 𝘥𝘦𝘮𝘢𝘯𝘥𝘪𝘯𝘨 𝘮𝘢𝘯𝘶𝘢𝘭 𝘵𝘸𝘦𝘢𝘬𝘴. 𝐄𝐯𝐨𝐏𝐫𝐞𝐬𝐞𝐧𝐭 changes all of that! ✨ EvoPresent is a self-optimizing framework that unites storytelling, design, and feedback to create effortless, engaging presentation videos. 🎥 💡 𝐊𝐞𝐲 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬: 💠 𝐏𝐫𝐞𝐬𝐀𝐞𝐬𝐭𝐡, the core multi-task RL model, continuously refines both content and design — ensuring slides that are impactful and visually captivating. 📊 𝐄𝐯𝐨𝐏𝐫𝐞𝐬𝐞𝐧𝐭 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤 is a comprehensive evaluation suite: 650+ top AI papers & diverse formats to assess content and design, and 2000+ slide pairs for aesthetic scoring, defect correction, and design comparison. 🎯 🧠 𝘏𝘪𝘨𝘩-𝘲𝘶𝘢𝘭𝘪𝘵𝘺 𝘧𝘦𝘦𝘥𝘣𝘢𝘤𝘬 𝘱𝘰𝘸𝘦𝘳𝘴 𝘤𝘰𝘯𝘵𝘪𝘯𝘶𝘰𝘶𝘴 𝘴𝘦𝘭𝘧-𝘪𝘮𝘱𝘳𝘰𝘷𝘦𝘮𝘦𝘯𝘵. ⚖️ 𝘉𝘢𝘭𝘢𝘯𝘤𝘦 𝘣𝘦𝘵𝘸𝘦𝘦𝘯 𝘤𝘰𝘯𝘵𝘦𝘯𝘵 & 𝘥𝘦𝘴𝘪𝘨𝘯 is the secret to presentation excellence. 🔁 𝐌𝐮𝐥𝐭𝐢-𝐭𝐚𝐬𝐤 𝐑𝐋 training boosts generalization in aesthetic awareness tasks. 𝐃𝐢𝐬𝐜𝐥𝐚𝐢𝐦𝐞𝐫: the demo video was completely generated by EvoPresent, no human refinement.

Brand new paper with @johnjhorton that I'm very excited to share: "General Social Agents" Suppose we wanted to create AI agents for simulations to make predictions in never-before-seen settings. How might we do this? We explore an approach to answering that question!

@ShenRaphael Crazy

You need this to apply for PhD now???

@AmartyaSanyal @SametOymac I think the accept cutoff may raise to 4.0 after rebuttal.

@SametOymac How does this look at the end of the rebuttal period ?

Some SAC context for NeurIPS review scores - out of my batch of 100 papers: - 1 paper ≥5.0 - 6 papers ≥4.5 - 11 papers ≥4.0 - 25 papers ≥3.75 - 42 papers ≥3.5 So currently naive accept cutoff would be 3.75.

New Anthropic research: Persona vectors. Language models sometimes go haywire and slip into weird and unsettling personas. Why? In a new paper, we find “persona vectors"—neural activity patterns controlling traits like evil, sycophancy, or hallucination.

@jiaxinwen22 Interesting paper!

New Anthropic research: We elicit capabilities from pretrained models using no external supervision, often competitive or better than using human supervision. Using this approach, we are able to train a Claude 3.5-based assistant that beats its human-supervised counterpart.

Find out more about our open-source interpretability tools, and how to use them on open-weights models, here: anthropic.com/research/open-…

Thrilled to share that our paper TwinMarket received the 🏆 Best Paper Award at Financial AI Workshop @ ICLR 2025! Huge thanks to my amazing co-authors, my advisor Dr. Hu @sthuyan , and Prof. Wang @wabyking . Special thanks to the generous support from @RBCBorealis! #ICLR2025

If you're also interested in: AI + social simulation，multi-agent systems, please attend the Financial AI Workshop@ICLR on 4.28 at 11:40 a.m., Hall 4 #2. We will present our latest research: TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets.

@didiforx @MetaGPT_ @iclr_conf catch up : )

Excited to share our paper AFLOW with @MetaGPT_ (openreview.net/forum?id=z5uVA…) at @iclr_conf 📷Poster: 10 am -- 12:30pm at Hall 3 + Hall 2B #311, Apr 25. 🔥Oral: 16:30pm - 16:42pm at Garnet 213-215, Apr 25. If you love agent, let's chat!

@xwang_lk So interested in agents, let's catch up :- )

✈️ Singapore for #ICLR2025! First time attending ICLR in person in years (for real 😂), and first time visiting SG! We’ll be presenting several papers, including MMWorld, EditRoom, MSSbench, and Agent S, plus talks on safety reasoning and LLM agents. Catch me, our PhD students @XuehaiH @saa1605, and incoming PhD students Chengzhi Liu and @_Chuhan_Li at the venue. Excited to meet friends old and new — let’s chat research and beyond!

@realJessyLin I’m really interested about 2). I am also doing some research about agents. Let’s catch up!

I’ll be at #ICLR2025 this week! ✈️ A couple of things I’m excited about lately: 1) Real-time multimodal models: how do we post-train assistants for real-time (and real world) tasks beyond the chat box? 2) Continual learning and memory: to have models / agents that learn from experience, I’m excited about revisiting parametric memory for LMs - how do we reliably teach models new things after pretraining and have them remember / generalize? 3) User simulators: we’ve gotten really good at optimizing models w/ RL on well-defined objectives. To make this work for fuzzy, complex tasks, I think we’re bottlenecked by realistic, robust simulators of humans to optimize against. Reach out if you’d like to chat!