selected presentations

reading groups

1. arXiv

   

   Reinforcement Learning via Self-Distillation

   Jonas Hübotter, Frederike Lübeck, Lejs Behric, and 8 more authors

   arXiv preprint arXiv:2601.20802, Feb 2026

   Slides presented at JHU’s NLP reading group.

   Slides Abs PDF Website GitHub

   We introduce Self-Distilled Policy Optimization (SDPO), a novel reinforcement learning algorithm that leverages self-distillation to enhance policy learning. SDPO iteratively distills knowledge from the current policy into a student model, which is then used to update the policy. This approach allows for more efficient learning by effectively utilizing the information contained in the policy’s own behavior. We provide theoretical analysis of SDPO’s convergence properties and demonstrate its empirical performance on a range of benchmark tasks, showing significant improvements over traditional policy optimization methods. Our results suggest that self-distillation can be a powerful tool for improving reinforcement learning algorithms, paving the way for more efficient and effective policy learning in complex environments.

2. NeurIPS ’25

   

   Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward

   Yanming Wan, Jiaxing Wu, Marwa Abdulhai, and 2 more authors

   In The Thirty-ninth Annual Conference on Neural Information Processing Systems, Dec 2025

   Slides presented at JHU’s NLP reading group.

   Slides Abs PDF Poster Website

   Effective conversational agents must personalize their interactions to adapt to user preferences, personalities, and attributes across diverse domains like education and healthcare. Current methods like Reinforcement Learning from Human Feedback (RLHF), often prioritize helpfulness and safety but fall short in fostering truly empathetic, adaptive, and personalized dialogues. Existing personalization approaches typically rely on extensive user history, limiting their effectiveness for new or context-limited users. To address these limitations, we propose leveraging a user model to incorporate a curiosity-based intrinsic reward into multi-turn RLHF. This novel reward mechanism encourages the agent to actively infer user traits by optimizing conversations to improve its user model’s accuracy. Consequently, the agent delivers more personalized interactions by learning more about the user. We demonstrate our method’s effectiveness in two distinct domains: significantly improving personalization performance in a conversational recommendation task, and personalizing conversations for different learning styles in an educational setting with improved generalization capabilities compared to traditional multi-turn RLHF, all while maintaining conversation quality. Our method offers a promising solution for creating more personalized, adaptive, and engaging conversational agents.

3. NeurIPS ’25 (🔦)

   

   Absolute Zero: Reinforced Self-play Reasoning with Zero Data

   Andrew Zhao, Yiran Wu, Yang Yue, and 8 more authors

   In The Thirty-ninth Annual Conference on Neural Information Processing Systems, Dec 2025

   Slides presented at JHU’s TeCho (Textual Choreography) reading group.

   Slides Abs PDF Poster Website GitHub

   Reinforcement learning with verifiable rewards (RLVR) has shown promise in enhancing the reasoning capabilities of large language models by learning directly from rule-based outcome rewards. Recent RLVR works that operate under the zero setting avoid supervision in labeling the reasoning process, but still depend on manually curated collections of questions and answers for training. The scarcity of high-quality, human-produced examples raises concerns about the long-term scalability of relying on human supervision, a challenge already evident in the domain of language model pretraining. Furthermore, in a hypothetical future where AI surpasses human intelligence, tasks provided by humans may offer limited learning potential for a superintelligent system. To address these concerns, we propose a new RLVR paradigm called Absolute Zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them, without relying on any external human or distillation data. Under this paradigm, we introduce the Absolute Zero Reasoner (AZR), a system that self-evolves its training curriculum and reasoning ability. AZR uses a code executor to both validate self-proposed code reasoning tasks and verify answers, serving as an unified source of verifiable feedback to guide open-ended yet grounded learning. Despite being trained entirely without external data, AZR achieves overall SOTA performance on coding and mathematical reasoning tasks, outperforming existing zero-setting models that rely on tens of thousands of in-domain human-curated examples. Furthermore, we demonstrate that AZR can be effectively applied across different model scales and is compatible with various model classes.

4. arXiv

   

   Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens

   Yiming Qin, Bomin Wei, Jiaxin Ge, and 4 more authors

   arXiv preprint arXiv:2511.19418, Dec 2025

   Slides presented at JHU’s NLP reading group.

   Slides Abs PDF Website GitHub

   Vision–Language Models (VLMs) excel at reasoning in linguistic space but struggle with perceptual understanding that requires dense visual perception, e.g., spatial reasoning and geometric awareness. This limitation stems from the fact that current VLMs have limited mechanisms to capture dense visual information across spatial dimensions. We introduce Chain-of-Visual-Thought (CoVT), a framework that enables VLMs to reason not only in words but also through continuous visual tokens — compact latent representations that encode rich perceptual cues. Within a small budget of roughly 20 tokens, CoVT distills knowledge from lightweight vision experts capturing complementary properties such as 2D appearance, 3D geometry, spatial layout, and edge structure. During training, a VLM equipped with CoVT autoregressively predicts these visual tokens to reconstruct dense supervision signals (e.g., depth, segmentation, edges, and DINO features). At inference, the model reasons directly in the continuous visual-token space, preserving efficiency while optionally decoding dense predictions for interpretability. Evaluated across more than ten diverse perception benchmarks, including CV-Bench, MMVP, RealWorldQA, MMStar, WorldMedQA, and HRBench, integrating CoVT into strong VLMs such as Qwen2.5-VL and LLaVA consistently improves performance by 3% to 16%, demonstrating that compact continuous visual thinking enables more precise, grounded, and interpretable multimodal intelligence.

5. NeurIPS ’24 D&B (🔦)

   

   DiscoveryWorld: A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents

   Peter Jansen, Marc-Alexandre Côté, Tushar Khot, and 5 more authors

   In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, Dec 2024

   Slides presented at JHU’s NLP Reading Group.

   Slides Abs PDF Poster Website GitHub

   Automated scientific discovery promises to accelerate progress across scientific domains. However, developing and evaluating an AI agent’s capacity for end-to-end scientific reasoning is challenging as running real-world experiments is often prohibitively expensive or infeasible. In this work we introduce DiscoveryWorld, the first virtual environment for developing and benchmarking an agent’s ability to perform complete cycles of novel scientific discovery. DiscoveryWorld contains a variety of different challenges, covering topics as diverse as radioisotope dating, rocket science, and proteomics, to encourage development of general discovery skills rather than task-specific solutions. DiscoveryWorld itself is an inexpensive, simulated, text-based environment (with optional 2D visual overlay). It includes 120 different challenge tasks, spanning eight topics each with three levels of difficulty and several parametric variations. Each task requires an agent to form hypotheses, design and run experiments, analyze results, and act on conclusions. DiscoveryWorld further provides three automatic metrics for evaluating performance, based on (a) task completion, (b) task-relevant actions taken, and (c) the discovered explanatory knowledge. We find that strong baseline agents, that perform well in prior published environments, struggle on most DiscoveryWorld tasks, suggesting that DiscoveryWorld captures some of the novel challenges of discovery, and thus that DiscoveryWorld may help accelerate near-term development and assessment of scientific discovery competency in agents.

6. COLM ’25

   

   The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains

   Scott Geng, Hamish Ivison, Chun-Liang Li, and 4 more authors

   In Second Conference on Language Modeling, Apr 2025

   Slides presented at JHU’s TeCho (Textual Choreography) Reading Group.

   Slides Abs PDF

   Improvements in language models are often driven by improving the quality of the data we train them on, which can be limiting when strong supervision is scarce. In this work, we show that paired preference data consisting of individually weak data points can enable gains beyond the strength of each individual data point. We formulate the delta learning hypothesis to explain this phenomenon, positing that the relative quality deltabetween points suffices to drive learning via preference tuning—even when supervised finetuning on the weak data hurts. We validate our hypothesis in controlled experiments and at scale, where we post-train 8B models on preference data generated by pairing a small 3B model’s responses with outputs from an even smaller 1.5B model to create a meaningful delta. Strikingly, on a standard 11-benchmark evaluation suite (MATH, MMLU, etc.), our simple recipe matches the performance of T ̈ulu 3, a state-of-the-art open model tuned from the same base model while relying on much stronger supervisors (e.g., GPT-4o). Thus, delta learning enables simpler and cheaper open recipes for state-of-the-art post-training. To better understand delta learning, we prove in logistic regression that the performance gap between two weak teacher models provides useful signal for improving a stronger student. Overall, our work shows that models can learn surprisingly well from paired data that might typically be considered weak.

7. ICLR ’25 (Oral)

   

   SAM 2: Segment Anything in Images and Videos

   Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, and 15 more authors

   In The Thirteenth International Conference on Learning Representations, Apr 2025

   Slides presented at OSU’s Data Mining Research Lab.

   Slides Abs PDF Website GitHub

   We present Segment Anything Model 2 (SAM 2), a foundation model towards solving promptable visual segmentation in images and videos. We build a data engine, which improves model and data via user interaction, to collect the largest video segmentation dataset to date. Our model is a simple transformer architecture with streaming memory for real-time video processing. SAM 2 trained on our data provides strong performance across a wide range of tasks. In video segmentation, we observe better accuracy, using 3x fewer interactions than prior approaches. In image segmentation, our model is more accurate and 6x faster than the Segment Anything Model (SAM). We believe that our data, model, and insights will serve as a significant milestone for video segmentation and related perception tasks. We are releasing our main model, the dataset, an interactive demo and code.

8. DMKD ’19

   

   Proximity Forest: an effective and scalable distance-based classifier for time series

   Benjamin Lucas, Ahmed Shifaz, Charlotte Pelletier, and 5 more authors

   Data Mining and Knowledge Discovery, Volume 33, pages 607-635, May 2019

   Slides presented at OSU’s DATUM Lab.

   Slides Abs PDF GitHub

   Research into the classification of time series has made enormous progress in the last decade. The UCR time series archive has played a significant role in challenging and guiding the development of new learners for time series classification. The largest dataset in the UCR archive holds 10,000 time series only; which may explain why the primary research focus has been on creating algorithms that have high accuracy on relatively small datasets. This paper introduces Proximity Forest, an algorithm that learns accurate models from datasets with millions of time series, and classifies a time series in milliseconds. The models are ensembles of highly randomized Proximity Trees. Whereas conventional decision trees branch on attribute values (and usually perform poorly on time series), Proximity Trees branch on the proximity of time series to one exemplar time series or another; allowing us to leverage the decades of work into developing relevant measures for time series. Proximity Forest gains both efficiency and accuracy by stochastic selection of both exemplars and similarity measures. Our work is motivated by recent time series applications that provide orders of magnitude more time series than the UCR benchmarks. Our experiments demonstrate that Proximity Forest is highly competitive on the UCR archive: it ranks among the most accurate classifiers while being significantly faster. We demonstrate on a 1M time series Earth observation dataset that Proximity Forest retains this accuracy on datasets that are many orders of magnitude greater than those in the UCR repository, while learning its models at least 100,000 times faster than current state-of-the-art models Elastic Ensemble and COTE.

9. JACM ’99

   

   Authoritative sources in a hyperlinked environment

   Jon M. Kleinberg

   Journal of the ACM (JACM), Volume 46, Issue 5, Sep 1999

   Slides presented at OSU’s Data Mining Research Lab.

   Slides Abs PDF

   The network structure of a hyperlinked environment can be a rich source of information about the content of the environment, provided we have effective means for understanding it. We develop a set of algorithmic tools for extracting information from the link structures of such environments, and report on experiments that demonstrate their effectiveness in a variety of context on the World Wide Web. The central issue we address within our framework is the distillation of broad search topics, through the discovery of “authorative” information sources on such topics. We propose and test an algorithmic formulation of the notion of authority, based on the relationship between a set of relevant authoritative pages and the set of “hub pages” that join them together in the link structure. Our formulation has connections to the eigenvectors of certain matrices associated with the link graph; these connections in turn motivate additional heuristrics for link-based analysis.

10. WWW ’11

    

    Influence and passivity in social media

    Daniel M. Romero, Wojciech Galuba, Sitaram Asur, and 1 more author

    Proceedings of the 20th International Conference Companion on World Wide Web, New York, NY, USA, Nov 2011

    Slides presented at OSU’s Data Mining Research Lab.

    Slides Abs PDF

    Systems for crisis response have required several different models for the analysis of unstructured text, such as identifying needs, locations, topics, routing, and matching of needs with available responders. Large Language Models (LLMs) have replaced task-specific models across various language processing tasks. However, LLMs are known to be limited by their training data, collected before the crisis. In this demo, we explore the use of LLMs for crisis response scenarios with rapidly evolving information environments. We show how augmentation of these models with external reliable sources of crisis-specific information can help build adaptive systems for response.

courses

1. NeurIPS ’22

   

   Flamingo: a Visual Language Model for Few-Shot Learning

   Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, and 24 more authors

   In Advances in Neural Information Processing Systems, Dec 2022

   Slides presented at OSU’s Seminar on NLP course.

   Slides Abs PDF

   Building models that can be rapidly adapted to novel tasks using only a handful of annotated examples is an open challenge for multimodal machine learning research. We introduce Flamingo, a family of Visual Language Models (VLM) with this ability. We propose key architectural innovations to: (i) bridge powerful pretrained vision-only and language-only models, (ii) handle sequences of arbitrarily interleaved visual and textual data, and (iii) seamlessly ingest images or videos as inputs. Thanks to their flexibility, Flamingo models can be trained on large-scale multimodal web corpora containing arbitrarily interleaved text and images, which is key to endow them with in-context few-shot learning capabilities. We perform a thorough evaluation of our models, exploring and measuring their ability to rapidly adapt to a variety of image and video tasks. These include open-ended tasks such as visual question-answering, where the model is prompted with a question which it has to answer; captioning tasks, which evaluate the ability to describe a scene or an event; and close-ended tasks such as multiple-choice visual question-answering. For tasks lying anywhere on this spectrum, a single Flamingo model can achieve a new state of the art with few-shot learning, simply by prompting the model with task-specific examples. On numerous benchmarks, Flamingo outperforms models fine-tuned on thousands of times more task-specific data.

2. CVPR ’23

   

   Panoptic Video Scene Graph Generation

   Jingkang Yang, Wenxuan Peng, Xiangtai Li, and 8 more authors

   In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2023

   Slides presented at OSU’s Seminar on Network Science course.

   Slides Abs PDF GitHub

   Towards building comprehensive real-world visual perception systems, we propose and study a new problem called panoptic scene graph generation (PVSG). PVSG is related to the existing video scene graph generation (VidSGG) problem, which focuses on temporal interactions between humans and objects localized with bounding boxes in videos. However, the limitation of bounding boxes in detecting non-rigid objects and backgrounds often causes VidSGG systems to miss key details that are crucial for comprehensive video understanding. In contrast, PVSG requires nodes in scene graphs to be grounded by more precise, pixel-level segmentation masks, which facilitate holistic scene understanding. To advance research in this new area, we contribute a high-quality PVSG dataset, which consists of 400 videos (289 third-person + 111 egocentric videos) with totally 150K frames labeled with panoptic segmentation masks as well as fine, temporal scene graphs. We also provide a variety of baseline methods and share useful design practices for future work.

3. NeurIPS ’20

   

   Language Models are Few-Shot Learners

   Tom Brown, Benjamin Mann, Nick Ryder, and 28 more authors

   In Advances in Neural Information Processing Systems, Dec 2020

   Slides presented at OSU’s Seminar on NLP course.

   Slides Abs PDF Website

   We demonstrate that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even becoming competitive with prior state-ofthe-art fine-tuning approaches. Specifically, we train GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous nonsparse language model, and test its performance in the few-shot setting. For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model. GPT-3 achieves strong performance on many NLP datasets, including translation, question-answering, and cloze tasks. We also identify some datasets where GPT3’s few-shot learning still struggles, as well as some datasets where GPT-3 faces methodological issues related to training on large web corpora.

4. ACM MoD ’24

   

   Summarized Causal Explanations For Aggregate Views

   Brit Youngmann, Michael Cafarella, Amir Gilad, and 1 more author

   Proceedings of the ACM on Management of Data, Dec 2024

   Slides presented at OSU’s Seminar on Databases course.

   Slides Abs PDF GitHub

   SQL queries with group-by and average are frequently used and plotted as bar charts in several data analysis applications. Understanding the reasons behind the results in such an aggregate view may be a highly nontrivial and time-consuming task, especially for large datasets with multiple attributes. Hence, generating automated explanations for aggregate views can allow users to gain better insights into the results while saving time in data analysis. When providing explanations for such views, it is paramount to ensure that they are succinct yet comprehensive, reveal different types of insights that hold for different aggregate answers in the view, and, most importantly, they reflect reality and arm users to make informed data-driven decisions, i.e., the explanations do not only consider correlations but are causal. In this paper, we present CauSumX, a framework for generating summarized causal explanations for the entire aggregate view. Using background knowledge captured in a causal DAG, CauSumX finds the most effective causal treatments for different groups in the view. We formally define the framework and the optimization problem, study its complexity, and devise an efficient algorithm using the Apriori algorithm, LP rounding, and several optimizations. We experimentally show that our system generates useful summarized causal explanations compared to prior work and scales well for large high-dimensional data.