@uiuc_nlp

Natural Language Processing research group at The University of Illinois Urbana-Champaign @IllinoisCS @UofIllinois

Joined July 2019
  • Tweets649
  • Following146
  • Followers1.4K
10Media
🚀 CFP: 3rd Edition of the Lifelong Agent Workshop (LLA)! After ICLR 2026 and COLM 2026, we’re excited to bring LLA to AIAS+ 2026, back in San Francisco on Nov 7! We welcome work on all aspects of lifelong agents, including lifelong learning and adaptation, alignment and safety, agent evolution and self-improvement, evaluation and benchmarking, multi-agent systems, AI for Science, and applications across diverse domains. 📅 Submission deadline: Oct 10 (AoE) We’d love to see your latest ideas, insights, and ongoing work. Submit your work and join us in San Francisco! Website: lifelongagent.github.io/aias… OpenReview: openreview.net/group?id=aias… Registration: aiasplus.org/event-registrat… Hope to see you at LLA @ AIAS+ 2026! #LLMAgents #AIAgents #LifelongLearning #AIAS2026
2
16
1
31
1,936
What if a robot carries out the wrong task perfectly? A robot FM may have the skill to complete several different tasks in the same scene. Our instruction is supposed to determine which one it carries out. If that connection is fragile, the robot FM can retain its competence while using it to reach an unintended outcome. While the world can resist an impossible grasp, it may offer no resistance to a perfectly executed misunderstanding. The question that stayed with us throughout this work is deceptively simple - what keeps a capable robot’s actions faithful to what the user asked? #CoRL 2026 (1/3)
gokul
@gokulp01
Sep 7
Replying to @gokulp01
On a real hardware, the vulnerability persists across models. (8/9)
2
5
18
1,510
Can a tiny text edit redirect a robot’s trajectory? New work from UIUC researchers shows how subtle changes to text instructions can steer robot foundation model outcomes toward an attacker’s goal, without naming that goal. Congratulations on their #CoRL2026 acceptance! 🎉
gokul
@gokulp01
Sep 7
Paper accepted at #CoRL2026!! We study how language perturbations propagate through closed-loop VLA control and uncover a new attack! Work done in collaboration with: @Vardhan_Dongre , @pthangeda_ , Hooshang Nayyeri, @dilekhakkanitur , Melkior Ornik TL;DR: we introduce a new VLA attack where tiny, meaning-preserving edits can redirect the robot’s entire closed-loop behavior toward an adversary-chosen goal—without ever stating that goal in the prompt!🧵 (1/9) 📄 arxiv.org/abs/2606.12978 🌐 vla-redirection-attack.githu… 💻 github.com/vla-redirection-a…
5
479
UIUC NLP retweeted
Heng Wang
@HengWang_uiuc
Sep 4
Every KV cache evictor for efficient long reasoning scores cached tokens to decide which ones to keep. We find the score buys almost nothing: keep the prompt, evict at random per head, and you match the strongest evictor at 32–43% higher throughput. Paper: arxiv.org/abs/2609.03430 Page: arthur-heng.github.io/Random…
25
39
6
109
25,718
UIUC NLP retweeted
Ke Yang
@EmpathYang
Sep 2
Building AI tutors requires feedback from students, but real learner studies are slow and hard to scale. 📈 Just in: StudentSim (Microsoft×UIUC) trains personalized AI student simulators 🤖 from real learner 🧑‍🎓 records and evaluates whether they both match a specific student’s behavior and can respond to tutor guidance. ✏️ Across three learning domains, StudentSim outperforms strong student simulator baselines, including GPT-5.4. In a chess tutor RL study, using StudentSim as guidance helpfulness feedback led to tutor guidance that human experts rated higher in accuracy, guidance quality, and personalization. 📄 Arxiv: 2609.01591 💻 Code: microsoft/StudentSim
4
43
1
167
19,965
Over the past year, we have worked on three projects around User-Centric Agents, and this line of work has finally come together at EMNLP this year. Behind all three projects, we have really been asking the same question: If the ultimate purpose of an Agent is to serve people, should we move beyond modeling tasks -- and start modeling users? Many Agent systems implicitly assume that users provide goals that are clear, complete, and stable. Real human interaction is almost the opposite: our needs may be underspecified, our preferences often emerge gradually through conversation, and our behavior is shaped by latent factors such as our beliefs, intentions, and perspectives. A truly user-centric Agent therefore needs to solve three increasingly deeper problems: 1. Can it understand and satisfy user needs through complex interaction? 2. Can it learn better interaction strategies by continuously engaging with users? 3. And ultimately, can it understand the unobservable mental states behind user behavior? 🍃🍃🍃 🧭 UserBench: From Completing Tasks to Understanding Users We first put the user back into the Agent's environment. In UserBench, user requests can be underspecified, incremental, and indirect. The Agent must actively communicate with the user, uncover preferences, and then use tools to accomplish the task. What we found is striking: even models that are already very strong at tool use often manage to "complete the task" without actually satisfying the user. This reveals an important distinction: --> Task success is not the same as user success. 🏋️ UserRL: From Evaluation to Learning The next natural question was: can these capabilities be learned systematically? UserRL builds 8 user-centric gyms, where Agents continuously interact with simulated users through multi-turn rollouts and learn via reinforcement learning to improve capabilities such as clarification, adaptation, persuasion, and intent understanding. Here, the user is no longer merely the provider of a prompt: --> The user becomes a dynamic, interactive training environment. 🧠 UserHarness: From Behavior to Mind But users are more than collections of preferences. They also have their own beliefs, intentions, and perspectives. UserHarness therefore takes one step further and reframes Theory of Mind as user-mind reconstruction: explicitly tracking what a user has observed, what they believe, what they intend to do, and how these internal states give rise to their behavior. --> The goal is no longer merely to predict what the user will do, but to understand why the user behaves that way. 🍃🍃🍃 Whether future Agents evolve toward AI4AI, RSI, or increasingly autonomous intelligence, one question will remain unavoidable: "Who is the Agent ultimately acting for?" Today's AI can already automate increasingly complex tasks. But the more autonomous these systems become, the more important it is that they genuinely understand human needs, preferences, intentions, and boundaries. The capabilities of future Agents may continue to expand, but their objectives, behavioral boundaries, and ultimate locus of control should remain firmly anchored to the user. Finally, hope to see many of you in Budapest this year! 📄 UserBench: arxiv.org/pdf/2507.22034 📄 UserRL: arxiv.org/pdf/2509.19736 📄 UserHarness: arxiv.org/pdf/2605.27721
1
14
61
2,697
MultiSessionCollab has been accepted to #EMNLP2026🎉 Memory systems are typically optimized for information recall and question-answering, and we show that this doesn't really help with improving user interactions. Instead, we optimize memory for improving human-AI collaboration over time with: (1) MultiSessionCollab: a benchmark that evaluates how well agents can learn user preferences and improve collaboration over time (2) A training framework that leverages user feedback signals to improve agent memory It's super important to build AI that's useful to people in the real world. But it is challenging, and benchmarks don't always capture that very well. This work is our attempt to take a step in that direction, and I'm really excited to keep building on it!
3
5
1
45
2,606
ConvAI will be in Budapest for EMNLP 2026 🇭🇺 Find all 12 ‼️papers from our work and collaborations there:
1
4
2
29
3,289
The essence of harness is transfer of cognitive load. In our new work, we investigate this through strong2weak scaffolding, where builders recursively refine harnesses for smaller models at test time. 📄 arxiv.org/pdf/2608.12307 🤗 huggingface.co/papers/2608.1…
10
27
3
118
11,801
[1/7] Imagine: What happens when an AI tutor interrupts you 𝑒𝑣𝑒𝑟𝑦 moment you hesitate? You might feel irritated! Our new work, 𝐃𝐮𝐩𝐥𝐞𝐱𝐆𝐞𝐧, shows that the right decision depends on the conversational role. 🌐 Project: duplexgen.github.io 🧵
2
6
1
27
3,850
UIUC NLP retweeted
Intology
@intology
Aug 3
The models are improving the models. Locus, our automated AI research system, is SOTA on PostTrainBench and post-trains Qwen3 base models that surpass the human post-trained Qwen3 model. Today, LLMs post-trained end-to-end by Locus are in production to millions. 🧵👇 PostTrainBench evaluates agents' ability to post-train models on various domains given 10 H100 hours. We extend PostTrainBench via PostTrainBench+, which has a greatly expanded compute budget that provides clearer signal on automated post-training capabilities. We find that thousands of H100 hours help distinguish methods' performance post-training Qwen3 1.7B-Base models, and that Locus scales best. In this setting, modes trained by Locus collectively surpass the perforamce of the offical human post-trained Qwen3 1.7B model. In a test of generalization, we ran Locus on all live Kaggle competitions with prize money and public leaderboards. After 16 days, Locus achieved the 4th highest average rank among all participants.
14
60
28
305
266,756
🧠 Hallucinations in LLMs do not always begin at generation. They can start earlier, when an experience becomes a stable fact or the wrong memory is retrieved as evidence. We introduce MemGuard, a framework for preventing heterogeneous memory contamination in long-term memory-augmented LLMs by preserving functional boundaries between facts, experiences, and behavioral rules so agents retrieve the right memories for the right reasoning tasks. By reorganizing memory at write time and routing queries at retrieval time, MemGuard achieves +28.3% anti-hallucination accuracy while retrieving up to 5.8× fewer memory tokens. More memory is not better memory. The right memory, used for the right purpose, is. 📄 Paper: arxiv.org/abs/2605.28009 🤗 HuggingFace: huggingface.co/papers/2605.2… 💻 Code Base: github.com/HyeonjeongHa/MemG…
3
14
56
4,492
Congrats to @PardisZahraei and @iamjanvijay for the outstanding paper award at the KnowFM Workshop (Towards Knowledgeable Foundation Models, knowledgeable-lm.github.io/) at ACL 2026 for the following paper: openreview.net/forum?id=8y6r… @convai_uiuc @uiuc_nlp

KnowFM Workshop @ ACL 2026 — Towards Knowledgeable Foundation Models

Exploring knowledge in foundation models: emergence, injection, updating, probing, and generation. San Diego, 2026.

knowledgeable-lm.github.io
9
18
1,116
UIUC NLP retweeted
ConvAI@UIUC
@convai_uiuc
Jul 1
ACL is very soon now ⛱️ and our student authors are very excited to present their works! Find ConvAI's contributions here:
ConvAI@UIUC
@convai_uiuc
Apr 11
ConvAI had a great ACL season🎉We plan to present 7 papers – find all the authors in San Diego this summer ☀️🌴🏖️ Here are all of our ACL contributions:
3
11
1,178
Excited to share our paper: GBC: Gradient-Based Connections for Optimizing Multi-Agent Systems. Multi-agent LLM systems are promising, but when they fail, it is often unclear which agent caused the error, making prompt optimization especially difficult. (1/8)
3
10
2
29
2,982
UIUC NLP retweeted
Jiayu Liu
@JiayuJeff
Jun 23
📈 Nowadays, agents are crushing leaderboards. But when you ask one painfully normal question: 👤 You: “Hi, I'm Jeff. My phone number is 1234567890. I returned a desk lamp and filed a refund request on June 22 at 10:13 PM. Can you check the current status of my refund?” Sounds easy. 😵‍💫Then the agent enters API hell: (1) It dives into a huge API pile and is not sure what to search for. So it keeps trying: "refund", "return", "status", "order status"… (2) It actually retrieves a promising refund-related API. But instead of using it, it keeps searching and drifts to less relevant tools. (3) It finally calls a refund API, but the endpoint fails. Instead of changing the plan, it loops around similar APIs again. (4) Then a shipping-status API returns "delivered." Clearly not the refund status, but the agent treats it as evidence and keeps going. After many turns, it has searched, called APIs, collected values, and looked very busy 😂 🚀 To evaluate failures like this, we introduce PlanBench-XL: an interactive benchmark for evaluating long-horizon planning of LLM tool-use agents in large-scale noisy tool ecosystems. PlanBench-XL includes: 🛠️ 1,665 executable tools 🛒 327 retail-domain tasks 🔍 retrieval-limited tool visibility 🎯 implicit sub-goals and intermediate evidence 🚧 retrieval-time blockers that simulate missing, failing, or misleading tools Across 10 leading LLMs, we find that planning with massive tools is still far from solved. In our hardest setting, GPT-5.4 solves only 11.36% of tasks. 📄 Paper: arxiv.org/abs/2606.22388 💻 Code: github.com/JiayuJeff/PlanBen… 🤗 Dataset: huggingface.co/datasets/Jiay… 🌐 Project Page: planbench-xl.github.io/ Feedback, discussions, and benchmark attempts are very welcome! 🙌
2
16
2
35
18,411
UIUC NLP retweeted
Jiateng Liu@JiatengLiu
Jun 5
🚀 Excited to release Brick-Composer! Can MLLMs assemble hundreds of diverse LEGO-style bricks into real-world objects? Not yet — but Brick-Composer is making this “impossible” task possible. 🚀 📄 arXiv: arxiv.org/abs/2606.05445
6
15
1
33
12,053
UIUC NLP retweeted
Jinu Lee@jinulee_v
Jun 5
When a reasoning model "thinks," it backtracks, reflects, and verifies. In ReasoningFlow, we track how sentences *combine* throughout these behaviors, and reveal insights on how to evaluate and monitor these reasoning traces. Thread 👇 🔗 Paper: arxiv.org/abs/2606.05402
2
16
1
49
7,427
UIUC NLP retweeted
Jiayu Liu
@JiayuJeff
Jun 5
✨ Current agents do not plan in isolation. They interact with both the external world and users, and both can impose constraints. For example, an agent may propose: “Use a fan to dry the wet volleyball.” Then both sides push back: 🌍 World: No fan. 🙋 User: "Could we avoid using fans? They’re just too slow for drying things." These constraints are often hidden at first, and only revealed when the agent’s plan violates them. 🤔 But can LLM agents adaptively plan and re-plan as hidden world constraints and user preferences emerge? 🚀 We introduce AdaPlanBench, a dynamic interactive benchmark for evaluating adaptive planning under progressively disclosed dual constraints. 🔍 Across 307 household planning tasks, we find that current LLM agents still struggle to adapt reliably as constraints accumulate. Even strong models may forget disclosed constraints, over-correct to new feedback, or produce plans that are valid but ineffective. 📄 Paper: arxiv.org/abs/2606.05622 💻 Code: github.com/JiayuJeff/AdaPlan… 🤗 Dataset: huggingface.co/datasets/Jiay…
2
14
2
26
4,469
Call For Papers for Lifelong Agent @ COLM 2026! 🚀After our first Lifelong Agent Workshop at ICLR 2026, we’re excited to announce that the Lifelong Agent Workshop is back for its 2nd edition! 🎉This time, we’ll be at COLM 2026 this October in San Francisco! We will continue the conversation around lifelong agents: long-term learning, continual alignment, self-evolution, and stable growth, with the goal of exploring more sustainable paths for future agent systems. We’re also thrilled to have an amazing lineup of speakers, organizers, and advisory board members from Stanford, UC Berkeley, CMU, Microsoft Research, Google DeepMind, McGill, Oxford, UIUC, and many other institutions. We warmly welcome submissions to the workshop. Submissions are non-archival! Topics include, but are not limited to: agent post-training, agent RL, user-agent alignment, self-evolving agents, embodied agents, Agent4Science, applications, benchmarks, and more. 📌 Submission deadline: July 3, 2026 AoE 🌐 Workshop homepage: lifelongagent.github.io/ 📖 OpenReview: openreview.net/group?id=colm… 📮 Contact: [email protected] Looking forward to your submissions and to seeing everyone at COLM 2026 in San Francisco! 🚀 #COLM2026 #AIAgents #AgenticAI #LLM #LifelongLearning #AIAlignment
6
21
3
87
94,940
Load more