**Building Self-Improving Foundation Models -- Auditors, Actuators, and Amplifiers for Trustworthy AI**
Foundation models are rapidly evolving from passive predictors into systems that reason, critique, refuse, follow instructions, generate provenance claims, and increasingly act. As their capabilities expand, their failures also become more dynamic: hallucinated grounding, brittle safety behavior, reward-model blind spots, adversarially exploitable reasoning trajectories, and poisoned alignment pipelines. This talk argues that trust cannot be added once during training or evaluated only at the final output. Instead, trustworthy AI should be designed as a closed loop: auditors discover failures and estimate risk, actuators intervene during reasoning before failures fully unfold, and amplifiers convert failures, uncertainty, and hard examples into self-improvement. I will present recent work from our group on multimodal critic and value models, inference-time safety steering, reward-model failure discovery, model-aware data selection, and self-critic-driven visual-language alignment. Across these examples, a common principle emerges: the future of trustworthy foundation models is not a larger backbone alone, but a system that can detect, explain, recover from, and learn from failure—often only a few early steps away.

**Reasoning as Control: Adaptive Test-Time Compute for Planning Agents**
Foundation models are becoming runtime decision-makers. At inference time, they decode under constraints, compare alternatives, search over hypotheses, call tools, coordinate with other agents, verify intermediate states, and defend against adversarial inputs. This creates a central control problem: how should an agent choose its next computation, action, or workflow under uncertainty, risk, latency, and safety constraints?
In this talk, I will present adaptive test-time control as a unifying framework for reasoning and planning agents. Rather than treating decoding, search, reflection, collaboration, verification, and escalation as fixed inference recipes, we view them as controllable actions selected by a runtime policy. I will ground this framework in recent work from our group across a hierarchy of test-time control. At the generation level, GenARM, Transfer Q-Star, and Collab show how reward models, value estimates, and mixtures of agents can steer decoding and alignment at inference time. But control is only as reliable as its critic; ReForm closes this loop by using reward-guided failure discovery to expose and patch reward-model errors. At the action level, Agentic Critical Training trains agents to judge better actions among alternatives, turning reflection into action-quality control rather than imitation. At the workflow level, FlowBank adaptively selects among complementary multi-agent workflows under performance–cost tradeoffs. Together, these works suggest a path from controlled decoding to controlled agency: agents that allocate computation, evidence, collaboration, and safety intervention where they matter most. The broader message is that reliable planning agents require runtime policies for allocating computation, evidence, collaboration, verification, workflow structure, and safety intervention. The goal is to build agents that are selective, robust, and accountable — systems that deliver reliable behavior per unit compute in interactive, multimodal, and tool-rich environments.

7 papers accepted to ICML.

Our PropensityBench is adopted in the newly released Meta Muse Spark MSL model!

TraceGen is accepted to CVPR.

5 paper accepted to ICLR, and 1 of them is selected as an Oral.

1 paper accepted to AAAI as an Oral.

Our paper on "Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data?" won the Best Paper Award at the New Frontiers in Adversarial Machine Learning (AdvML Frontier) Workshop, NeurIPS 2024.

We are thrilled to announce that 7 exceptional students from our group are on the job market this year, ready to bring their expertise to new frontiers! 6 of them are pursuing industrial roles, while 1 will be on the academic job market looking for postdoctoral and faculty positions.

These talented researchers work on cutting-edge areas including AI Security, AI Agents, Alignment, Ethics, Fairness, and Responsible AI. They have deep experience with Generative AI, LLMs, VLMs, and VLAs, driving innovations in Weak-to-Strong Generalization and AI safety.

Don't miss this opportunity to connect with these future leaders in AI. Check out their profiles here!

Are invisible watermarks in AI-generated content truly effective in distinguishing AI-generated images from real ones? We’re hosting a NeurIPS competition "Erasing the Invisible: A Stress-Test Challenge for Image Watermarks" to stress-test these watermarks, and we want you to put them to the test! Here’s how it works: we provide watermarked images, and your task is to remove the watermarks. If your approach outperforms the rest, you’ll win a prize and the chance to present your work at our NeurIPS workshop! Spread the word and join the challenge!

Website: https://erasinginvisible.github.io/

I participated in a Reddit AMA (Ask Me Anything) session on May 14 from 2-4 p.m. Eastern Time, where I answered questions about AI and machine learning.

Our WAVES benchmark on stress-testing image watermarks is out. Find the arXiv, github code, Hugging Face data, visualization and leaderboard links at the project page: https://wavesbench.github.io/. Jan., 2024.

Chair and organizer of NSF-Amazon Fairness in AI Principle Investigator Meeting, Jan 9-10, 2024.