Abstract: Large Language Model (LLM) agents have demonstrated remarkable generalization capabilities across multi-domain tasks. Existing agent tuning approaches typically employ supervised finetuning on entire expert trajectories. However, behavior-cloning of full trajectories can introduce expert bias and weaken generalization to states not covered by the expert data. Additionally, critical steps--such as planning, complex reasoning for intermediate subtasks, and strategic decision-making--are essential to success in agent tasks, so learning these steps is the key to improving LLM agents. For more effective and efficient agent tuning, we propose ATLAS that identifies the critical steps in expert trajectories and finetunes LLMs solely on these steps with reduced costs. By steering the training's focus to a few critical steps, our method mitigates the risk of overfitting entire trajectories and promotes generalization across different environments and tasks. In extensive experiments, an LLM finetuned on only 30% critical steps selected by ATLAS outperforms the LLM finetuned on all steps and recent open-source LLM agents. ATLAS maintains and improves base LLM skills as generalist agents interacting with diverse environments.

Speaker: Bijoy

Location: Old Computer Science Building - CS2311
CSE 600 Talk: Haibin Ling - Computer Vision Research and Applications


Abstract: Having been intensively studied over half a decade, computer vision has evolved as a broad research area and become mature in many applications. In this talk, we will summarize our work in computer vision in both core vision topics and application-oriented ones. In particular, for core vision problems, we will report studies on visual tracking, visual matching and visual detection; for applications, we will describe our work on medical image analysis, intelligent transportation, smart projector systems and preliminary work on material property prediction.

Bio: Haibin Ling received the BS and MS degrees from Peking University in 1997 and 2000, respectively, and the PhD degree from the University of Maryland, College Park, in 2006. From 2000 to 2001, he was an assistant researcher at Microsoft Research Asia. From 2006 to 2007, he worked as a postdoctoral scientist at the University of California Los Angeles. In 2007, he joined Siemens Corporate Research as a research scientist. From 2008 to 2019, he worked as a faculty member of Temple University. In fall 2019, he joined the Department of Computer Science of Stony Brook University where he is currently a SUNY Empire Innovation Professor. His research interests include computer vision, augmented reality, medical image analysis, and human computer interaction. He received the Best Student Paper Award at the ACM UIST in 2003, and the NSF CAREER Award in 2014. He serves as Associate Editor for several journals including IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), Pattern Recognition (PR), and Computer Vision and Image Understanding (CVIU). He has served or will serve as Area Chair for CVPR 2014, 2016, 2019 and 2020.
Abstract: Generative image models trained on massive datasets encode the statistics of our visual world. An off-the-shelf diffusion or flow-matching model can therefore serve as a general-purpose image prior, providing information about images across a wide range of tasks. Harnessing these capabilities, however, requires combining the model at inference time with additional signals or constraints unknown during training. This thesis will develop training-free methods for inference under a generative denoising prior. I first show how sampling from a trained denoiser can be formulated as an optimization problem, which combined with a constraint, can solve tasks ranging from conditional generation and weakly supervised segmentation to combinatorial optimization. I then demonstrate how the inherent properties of denoisers can accelerate inference under such constraints. Replacing gradient descent with an inexact Newton update, based on the symmetry of the denoiser's Jacobian, substantially reduces inference costs without tradeoffs. I also explore a middle ground between model adaptation and fully training-free inference by using the denoiser's robust internal representations to learn constraints from limited labeled data. These methods are applied to gigapixel image domains such as digital histopathology and remote sensing, where generative models can only be trained at a patch scale. To synthesize arbitrarily large images at resolutions unseen during training, I introduce inference-time algorithms that enforce consistency across spatially overlapping patches and image scales. Finally, I propose repurposing inference-time algorithms from sampling tools, to mechanisms for understanding the prior learned by a denoising model. Extending the previously developed techniques, I analyze the Jacobians of generative denoisers, where preliminary results indicate that Jacobian spectra correlates with generative quality. This motivates a Jacobian-spectrum regularization as a way to improve model performance using insights derived from inference-time algorithms.

Speaker: Alexandros Graikos

Location: NCS 220
Recently, large-scale language data combined with modern machine learning techniques have shown strong value as means for studying human psychology and behavior. For example, language alone has been shown predictive in mental health, personality, and health behaviors. However, many applications for such language-based assessments have readily available and important data beyond language (i.e. extra-linguistics), such as predicting the subjective well-being of a community using tweets, where one can take into account their age, education, and demographic attributes. Language may capture some characteristics while extra-linguistic variables captures others. We believe that effectively integrating linguistic and extra-linguistic data can yield benefits beyond either independently. In this thesis, we develop methods which effectively integrate extra-linguistic data with language data focused primarily on social scientific applications. The central challenge is dealing with the size and heterogeneity of, often sparse and noisy, language data versus the, often low-dimensional and non-sparse, extra-linguistic variables. First, we consider structured extra-linguistics, like socioeconomic (income and education rates) and demographics (age, gender, etc.), and propose two integration methods, named residualized controls (RC) and residualized factor adaptation (RFA), to be used in county-wise prediction tasks. Demonstrating techniques that integrate information at both the model-level and data-level, we found consistently strong improvement over naively combining features, for example, increasing county level well-being predictions by over 12%. Next, we consider unstructured extra-linguistic data. In the first part, we incorporate social network connections and language over time to propose a novel metric for quantifying the stickiness of words - their ability to spread across friendship connections in a social network over time (or in other words, stick in ones vocabulary after seeing friends use it). We obtain which language features are more probable to disseminate through friendship and show such a metric is useful for predicting who will be friends and what content will spread. In addition, we analyze language content over time by proposing a novel dynamic content-specific topic modeling technique that can help to identify different sub-domains of a thematic scope and can be used to track societal shifts in concerns or views over time.
The Hudson River Estuary (HRE) and New York Bight (NYB) are closely connected, with HRE acting as crucial areas where many NYB marine species spawn and grow. Understanding how these biotic and abiotic environments interact, especially with rapid climate change, is key to better managing fisheries and conserving ecosystems. To better understand the HRE-NYB ecosystem, we develop a comprehensive ecosystem model that links physical and biological processes. Using data from long-term monitoring programs, we analyze ecological patterns and identify key factors regulating the ecosystem. We use this information to develop a model that mimics the food web from tiny plankton to large predators in the ecosystem. This model can help us better understand how changes in the environment, like rising temperatures, and human activities such as fishing affect marine lives and ecosystem over time. The insights from this model can support smarter fisheries management and efforts to conserve marine ecosystems in the HRE-NYB region.

IACS Student Seminar Speaker: Xiangyan Yang, Dept. of Applied Math & Statistics

Location: IACS Seminar Room or Zoom

Join Zoom Meeting: https://stonybrook.zoom.us/j/91650247483?pwd=fvAGEwadplJh7jFC5RWcdvZ5NWPJth.1
Meeting ID: 916 5024 7483
Passcode: 631055
Kate Armstrong, a Vancouver-based artist, writer, and independent curator, will explore the role of AI in art and creativity through three AI-driven projects: KEKE Terminal, Botto, and Sasha Stiles' AI collaborator Technelegy. She will compare these projects to historical artistic movements and investigate AI's role as an autonomous creative agent, the function of community participation, and the shifting dynamics of authorship.

Location: Humanities Institute Room 1008
Join the Department of Computer Science as we welcome Lyle Ungar, University of Pennsylvania, who will be delivering a lecture on 'Measuring Cultural Variation using Natural Language Processing.' When: 11/08/24 @ 2:30 PM Where: New Computer Science Building, Room 120. Reception to follow. Abstract: Cultures vary widely in how they view the world, for example being more individualist or collectivist. Such cultural differences are, of course, reflected in the words that people use. We first show a variety of ways in which multilingual language models are not multicultural; they speak Hindi or Mandarin, but still think like Americans. In contrast, we then present a scalable method that uses embedding-derived lexica to successfully measure regional variation in culture. Bio: Lyle Ungar is a Professor of Computer and Information Science at the University of Pennsylvania, where he also holds secondary appointments in Psychology, Bioengineering, Genomics and Computational Biology, and Operations, Information and Decisions. His group uses natural language processing and explainable AI for psychological research, including analyzing social media and cell phone sensor data to better understand the drivers of physical and mental well-being. They are currently building socio-emotionally sensitive GPT-based tutors and coaches.