Vision-language models (VLMs) have achieved remarkable success across multimodal tasks, yet their substantial computational demands hinder efficient deployment. Knowledge distillation (KD) has emerged as a powerful approach for building lightweight but competitive models, with strong evidence from both language and vision domains. However, its application to VLMs, particularly CLIP-style models, remains limited, often constrained to small-scale teachers and narrow evaluation tasks such as classification or retrieval. In this work, we present the first systematic study of distillation across a range of CLIP-style teacher models, ranging from standard baselines to large-scale state-of-the-art models. Contrary to trends observed in NLP and vision, we find that stronger teachers do not consistently yield better students; in fact, existing distillation frameworks often fail to scale, leading to degraded performance in downstream multimodal tasks such as visual question answering. Our findings challenge prevailing assumptions in KD and point toward new directions for designing parameter-efficient multimodal models.

Document Visual Question Answering (DocVQA) requires models to jointly understand textual semantics, spatial layout, and visual features. Current methods struggle with explicit spatial relationship modeling, inefficiency with high-resolution documents, multi-hop reasoning, and limited interpretability. We propose MGA-VQA, a multi-modal framework that integrates token-level encoding, spatial graph reasoning, memory-augmented inference, and question-guided compression. Unlike prior black-box models, MGA-VQA introduces interpretable graph-based decision pathways and structured memory access for enhanced reasoning transparency. Evaluation across six benchmarks (FUNSD, CORD, SROIE, DocVQA, STE-VQA, and RICO) demonstrates superior accuracy and efficiency, with consistent improvements in both answer prediction and spatial localization.

Associative thinking--the ability to connect seemingly unrelated ideas--is a foundational element of human creativity and problem-solving. This paper explores whether reinforcement learning (RL) guided by associative thinking principles can enhance a model's performance across diverse generative tasks, including story writing, code generation, and chart creation. We introduce a reinforcement learning framework that uses a prompt-based evaluation mechanism, incorporating established divergent thinking metrics from creativity research. A base language model is fine-tuned using this framework to reward outputs demonstrating higher novelty through higher degrees of conceptual connectivity. Interestingly, the experimental results suggest that RL-based associative thinking-trained models not only generate more original and coherent stories but also exhibit improved abstraction and flexibility in tasks such as programming and data visualization. Our findings provide initial evidence that modeling cognitive creativity principles through reinforcement learning can yield more adaptive and generative AI.

Robots must learn from both what people do and what they say, but either modality alone is often incomplete: physical corrections are grounded but ambiguous in intent, while language expresses high-level goals but lacks physical grounding. We introduce QuickLAP: Quick Language-Action Preference learning, a Bayesian framework that fuses physical and language feedback to infer reward functions in real time. Our key insight is to treat language as a probabilistic observation over the user's latent preferences, clarifying which reward features matter and how physical corrections should be interpreted. QuickLAP uses Large Language Models (LLMs) to extract reward feature attention masks and preference shifts from free-form utterances, which it integrates with physical feedback in a closed-form update rule. This enables fast, real-time, and robust reward learning that handles ambiguous feedback. In a semi-autonomous driving simulator, QuickLAP reduces reward learning error by over 70% compared to physical-only and heuristic multimodal baselines. A 15-participant user study further validates our approach: participants found QuickLAP significantly more understandable and collaborative, and preferred its learned behavior over baselines. Code is available at https://github.com/MIT-CLEAR-Lab/QuickLAP.

The capacity for highly complex, evidence-based, and strategically adaptive persuasion remains a formidable great challenge for artificial intelligence. Previous work, like IBM Project Debater, focused on generating persuasive speeches in simplified and shortened debate formats intended for relatively lay audiences. We introduce DeepDebater, a novel autonomous system capable of participating in and winning a full, unmodified, two-team competitive policy debate. Our system employs a hierarchical architecture of specialized multi-agent workflows, where teams of LLM-powered agents collaborate and critique one another to perform discrete argumentative tasks. Each workflow utilizes iterative retrieval, synthesis, and self-correction using a massive corpus of policy debate evidence (OpenDebateEvidence) and produces complete speech transcripts, cross-examinations, and rebuttals. We introduce a live, interactive end-to-end presentation pipeline that renders debates with AI speech and animation: transcripts are surface-realized and synthesized to audio with OpenAI TTS, and then displayed as talking-head portrait videos with EchoMimic V1. Beyond fully autonomous matches (AI vs AI), DeepDebater supports hybrid human-AI operation: human debaters can intervene at any stage, and humans can optionally serve as opponents against AI in any speech, allowing AI-human and AI-AI rounds. In preliminary evaluations against human-authored cases, DeepDebater produces qualitatively superior argumentative components and consistently wins simulated rounds as adjudicated by an independent autonomous judge. Expert human debate coaches also prefer the arguments, evidence, and cases constructed by DeepDebater. We open source all code, generated speech transcripts, audio and talking head video here: https://github.com/Hellisotherpeople/DeepDebater/tree/main

University of Southern California, Los Angeles, CA 90089 (Dated: November 25, 2025) We demonstrate that multipoint Bayesian algorithm execution can overcome fundamental computational challenges in storage ring design optimization. Dynamic (DA) and momentum (MA) optimization is a multipoint, multiobjective design task for storage rings, ultimately informing the flux of x-ray sources and luminosity of colliders. We remove this bottleneck using multipointBAX, which selects, simulates, and models each trial configuration at the single particle level. We demonstrate our approach on a novel design for a fourth-generation light source, with neural-network powered multipointBAX achieving equivalent Pareto front results using more than two orders of magnitude fewer tracking computations compared to genetic algorithms. The significant reduction in cost positions multipointBAX as a promising alternative to black-box optimization, and we anticipate multipointBAX will be instrumental in the design of future light sources, colliders, and large-scale scientific facilities. Designing modern scientific facilities -- from synchrotron light sources to particle colliders -- requires optimizing hundreds of parameters in a complex, nonlinear systems, where a single design evaluation can take hours of computation. In storage rings, this challenge is exemplified by dynamic aperture (DA) and momentum aperture (MA) optimization, where maximizing the regions of particle stability directly determines injection efficiency, beam lifetime, and ultimately the photon flux or luminosity achievable in next-generation facilities. The computational bottleneck is severe: maximizing DA and MA is a type of multipoint optimization, where evaluating a single lattice design requires tracking tens of thousands of particles for hundreds of thousands of turns, making global optimization prohibitively expensive. Moreover, there is a trade-off between maximizing DA and MA area, so the standard approach is to find a Pareto front; i.e.

Graph neural networks (GNN) have emerged as a promising machine learning method for microstructure simulations such as grain growth. However, accurate modeling of realistic grain boundary networks requires large simulation cells, which GNN has difficulty scaling up to. To alleviate the computational costs and memory footprint of GNN, we propose a hybrid architecture combining a convolutional neural network (CNN) based bijective autoencoder to compress the spatial dimensions, and a GNN that evolves the microstructure in the latent space of reduced spatial sizes. Our results demonstrate that the new design significantly reduces computational costs with using fewer message passing layer (from 12 down to 3) compared with GNN alone. The reduction in computational cost becomes more pronounced as the spatial size increases, indicating strong computational scalability. For the largest mesh evaluated (160^3), our method reduces memory usage and runtime in inference by 117x and 115x, respectively, compared with GNN-only baseline. More importantly, it shows higher accuracy and stronger spatiotemporal capability than the GNN-only baseline, especially in long-term testing. Such combination of scalability and accuracy is essential for simulating realistic material microstructures over extended time scales. The improvements can be attributed to the bijective autoencoder's ability to compress information losslessly from spatial domain into a high dimensional feature space, thereby producing more expressive latent features for the GNN to learn from, while also contributing its own spatiotemporal modeling capability. The training was optimized to learn from the stochastic Potts Monte Carlo method. Our findings provide a highly scalable approach for simulating grain growth.

Foundation models hold promise for specialized medical imaging tasks, though their effectiveness in breast imaging remains underexplored. This study leverages BiomedCLIP as a foundation model to address challenges in model generalizations. BiomedCLIP was adapted for automated BI-RADS breast density classification using multi-modality mammographic data (synthesized 2D images, digital mammography, and digital breast tomosyn-thesis). Using 96,995 images, we compared single-modality (s2D only) and multi-modality training approaches, addressing class imbalance through weighted contrastive learning. Both approaches achieved similar accuracy (multi-modality: 0.74, single-modality: 0.73), with the multi-modality model offering broader applicability across different imaging modalities and higher AUC values consistently above 0.84 across BI-RADS categories. External validation on the RSNA and EMBED datasets showed strong generalization capabilities (AUC range: 0.80-0.93). GradCAM visualizations confirmed consistent and clinically relevant attention patterns, highlighting the models' interpretability and robustness. This research underscores the potential of foundation models for breast imaging applications, paving the way for future extensions for diagnostic tasks.

Deep learning models are prone to learning shortcut solutions to problems using spuriously correlated yet irrelevant features of their training data. In high-risk applications such as medical image analysis, this phenomenon may prevent models from using clinically meaningful features when making predictions, potentially leading to poor robustness and harm to patients. We demonstrate that different types of shortcuts (those that are diffuse and spread throughout the image, as well as those that are localized to specific areas) manifest distinctly across network layers and can, therefore, be more effectively targeted through mitigation strategies that target the intermediate layers. We propose a novel knowledge distillation framework that leverages a teacher network fine-tuned on a small subset of task-relevant data to mitigate shortcut learning in a student network trained on a large dataset corrupted with a bias feature. Through extensive experiments on CheXpert, ISIC 2017, and SimBA datasets using various architectures (ResNet-18, AlexNet, DenseNet-121, and 3D CNNs), we demonstrate consistent improvements over traditional Empirical Risk Minimization, augmentation-based bias-mitigation, and group-based bias-mitigation approaches. In many cases, we achieve comparable performance with a baseline model trained on bias-free data, even on out-of-distribution test data. Our results demonstrate the practical applicability of our approach to real-world medical imaging scenarios where bias annotations are limited and shortcut features are difficult to identify a priori.

Multiple-choice question answering (MCQA) has been a popular format for evaluating and reinforcement fine-tuning (RFT) of modern multimodal language models. Its constrained output format allows for simplified, deterministic automatic verification. However, we find that the options may leak exploitable signals, which makes the accuracy metrics unreliable for indicating real capabilities and encourages explicit or implicit answer guessing behaviors during RFT. We propose ReVeL (Rewrite and Verify by LLM), a framework that rewrites multiple-choice questions into open-form questions while keeping answers verifiable whenever possible. The framework categorizes questions according to different answer types, apply different rewriting and verification schemes, respectively. When applied for RFT, we converted 20k MCQA examples and use GRPO to finetune Qwen2.5-VL models. Models trained on ReVeL-OpenQA match MCQA accuracy on multiple-choice benchmarks and improve OpenQA accuracy by about six percentage points, indicating better data efficiency and more robust reward signals than MCQA-based training. When used for evaluation, ReVeL also reveals up to 20 percentage points of score inflation in MCQA benchmarks (relative to OpenQA), improves judging accuracy, and reduces both cost and latency. We will release code and data publicly.