Deep Learning
Symbolic Feedforward Networks for Probabilistic Finite Automata: Exact Simulation and Learnability
We present a formal and constructive theory showing that probabilistic finite automata can be exactly simulated using symbolic feedforward neural networks. Our architecture represents state distributions as vectors and transitions as stochastic matrices, enabling probabilistic state propagation via matrix-vector products. This yields a parallel, interpretable, and differentiable simulation of probabilistic finite automata dynamics using soft updates without recurrence. We formally characterize probabilistic subset construction, finite ε-closure, and exact simulation via layered symbolic computation, and prove equivalence between probabilistic finite automata and specific classes of neural networks. We further show that these symbolic simulators are not only expressive but learnable: trained with standard gradient descent-based optimization on labeled sequence data, they recover the exact behavior of ground-truth probabilistic finite automata. This learnability, formalized in Proposition 5.1, is the crux of this work. Our results unify probabilistic automata theory with classical neural architectures under a rigorous algebraic framework, bridging the gap between symbolic computation and deep learning.
The CRITICAL Records Integrated Standardization Pipeline (CRISP): End-to-End Processing of Large-scale Multi-institutional OMOP CDM Data
Luo, Xiaolong, Li, Michael Lingzhi
While existing critical care EHR datasets such as MIMIC and eICU have enabled significant advances in clinical AI research, the CRITICAL dataset opens new frontiers by providing extensive scale and diversity -- containing 1.95 billion records from 371,365 patients across four geographically diverse CTSA institutions. CRITICAL's unique strength lies in capturing full-spectrum patient journeys, including pre-ICU, ICU, and post-ICU encounters across both inpatient and outpatient settings. This multi-institutional, longitudinal perspective creates transformative opportunities for developing generalizable predictive models and advancing health equity research. However, the richness of this multi-site resource introduces substantial complexity in data harmonization, with heterogeneous collection practices and diverse vocabulary usage patterns requiring sophisticated preprocessing approaches. We present CRISP to unlock the full potential of this valuable resource. CRISP systematically transforms raw Observational Medical Outcomes Partnership Common Data Model data into ML-ready datasets through: (1) transparent data quality management with comprehensive audit trails, (2) cross-vocabulary mapping of heterogeneous medical terminologies to unified SNOMED-CT standards, with deduplication and unit standardization, (3) modular architecture with parallel optimization enabling complete dataset processing in $<$1 day even on standard computing hardware, and (4) comprehensive baseline model benchmarks spanning multiple clinical prediction tasks to establish reproducible performance standards. By providing processing pipeline, baseline implementations, and detailed transformation documentation, CRISP saves researchers months of preprocessing effort and democratizes access to large-scale multi-institutional critical care data, enabling them to focus on advancing clinical AI.
FediLoRA: Heterogeneous LoRA for Federated Multimodal Fine-tuning under Missing Modalities
Yang, Lishan, Zhang, Wei Emma, Nguygen, Nam Kha, Hu, Po, Shu, Yanjun, Chen, Weitong, Sim, Mong Yuan
Foundation models have demonstrated remarkable performance across a wide range of tasks, yet their large parameter sizes pose challenges for practical deployment, especially in decentralized environments. Parameter-efficient fine-tuning (PEFT), such as Low-Rank Adaptation (LoRA), reduces local computing and memory overhead, making it attractive for federated learning. However, existing federated LoRA methods typically assume uniform rank configurations and unimodal inputs, overlooking two key real-world challenges: (1) heterogeneous client resources have different LoRA ranks, and (2) multimodal data settings with potentially missing modalities. In this work, we propose FediLoRA, a simple yet effective framework for federated multimodal fine-tuning under heterogeneous LoRA ranks and missing modalities. FediLoRA introduces a dimension-wise aggregation strategy that reweights LoRA updates without information dilution during aggregation. It also includes a lightweight layer-wise model editing method that selectively incorporates global parameters to repair local components which improves both client and global model performances. Experimental results on three multimodal benchmark datasets demonstrate that FediLoRA achieves superior performance over competitive baselines in both global and personalized settings, particularly in the presence of modality incompleteness.
Mind the Gap: Evaluating Model- and Agentic-Level Vulnerabilities in LLMs with Action Graphs
Wicaksono, Ilham, Wu, Zekun, Patel, Rahul, King, Theo, Koshiyama, Adriano, Treleaven, Philip
As large language models transition to agentic systems, current safety evaluation frameworks face critical gaps in assessing deployment-specific risks. We introduce AgentSeer, an observability-based evaluation framework that decomposes agentic executions into granular action and component graphs, enabling systematic agentic-situational assessment. Through cross-model validation on GPT-OSS-20B and Gemini-2.0-flash using HarmBench single turn and iterative refinement attacks, we demonstrate fundamental differences between model-level and agentic-level vulnerability profiles. Model-level evaluation reveals baseline differences: GPT-OSS-20B (39.47% ASR) versus Gemini-2.0-flash (50.00% ASR), with both models showing susceptibility to social engineering while maintaining logic-based attack resistance. However, agentic-level assessment exposes agent-specific risks invisible to traditional evaluation. We discover "agentic-only" vulnerabilities that emerge exclusively in agentic contexts, with tool-calling showing 24-60% higher ASR across both models. Cross-model analysis reveals universal agentic patterns, agent transfer operations as highest-risk tools, semantic rather than syntactic vulnerability mechanisms, and context-dependent attack effectiveness, alongside model-specific security profiles in absolute ASR levels and optimal injection strategies. Direct attack transfer from model-level to agentic contexts shows degraded performance (GPT-OSS-20B: 57% human injection ASR; Gemini-2.0-flash: 28%), while context-aware iterative attacks successfully compromise objectives that failed at model-level, confirming systematic evaluation gaps. These findings establish the urgent need for agentic-situation evaluation paradigms, with AgentSeer providing the standardized methodology and empirical validation.
Training Language Model Agents to Find Vulnerabilities with CTF-Dojo
Zhuo, Terry Yue, Wang, Dingmin, Ding, Hantian, Kumar, Varun, Wang, Zijian
Large language models (LLMs) have demonstrated exceptional capabilities when trained within executable runtime environments, notably excelling at software engineering tasks through verified feedback loops. Yet, scalable and generalizable execution-grounded environments remain scarce, limiting progress in training more capable ML agents. We introduce CTF-Dojo, the first large-scale executable runtime tailored for training LLMs with verifiable feedback, featuring 658 fully functional Capture-The-Flag (CTF)-style challenges containerized in Docker with guaranteed reproducibility. To enable rapid scaling without manual intervention, we develop CTF-Forge, an automated pipeline that transforms publicly available artifacts into ready-to-use execution environments in minutes, eliminating weeks of expert configuration traditionally required. We trained LLM-based agents on just 486 high-quality, execution-verified trajectories from CTF-Dojo, achieving up to 11.6% absolute gains over strong baselines across three competitive benchmarks: InterCode-CTF, NYU CTF Bench, and Cybench. Our best-performing 32B model reaches 31.9% Pass@1, establishing a new open-weight state-of-the-art that rivals frontier models like DeepSeek-V3-0324 and Gemini-2.5-Flash. By framing CTF-style tasks as a benchmark for executable-agent learning, CTF-Dojo demonstrates that execution-grounded training signals are not only effective but pivotal in advancing high-performance ML agents without dependence on costly proprietary systems.
EvoCoT: Overcoming the Exploration Bottleneck in Reinforcement Learning
Liu, Huanyu, Li, Jia, Yu, Chang, Chen, Taozhi, Dong, Yihong, Wang, Lecheng, Hu, XiaoLong, Li, Ge
Reinforcement learning with verifiable reward (RL VR) has become a promising paradigm for post-training large language models (LLMs) to improve their reasoning capability. However, when the rollout accuracy is low on hard problems, the reward becomes sparse, limiting learning efficiency and causing exploration bottlenecks. Existing approaches either rely on teacher models for distillation or filter out difficult problems, which limits scalability or restricts reasoning improvement through exploration. We propose EvoCoT, a self-Evo lving curriculum learning framework based on two-stage C hain-o f-T hought (CoT) reasoning optimization. EvoCoT constrains the exploration space by self-generating and verifying CoT trajectories, then gradually shortens CoT steps to expand the space in a controlled way. The framework enables LLMs to stably learn from initially unsolved hard problems under sparse rewards. We apply EvoCoT to multiple LLM families, including Qwen, DeepSeek, and Llama. Experiments show that EvoCoT enables LLMs to solve previously unsolved problems, improves reasoning capability without external CoT supervision, and is compatible with various RL fine-tuning methods. Recently, reinforcement learning with verifiable reward (RL VR) has emerged as a promising paradigm for the post-training of large language models (LLMs).
AI Wizards at CheckThat! 2025: Enhancing Transformer-Based Embeddings with Sentiment for Subjectivity Detection in News Articles
Fasulo, Matteo, Babboni, Luca, Tedeschini, Luca
This paper presents AI Wizards' participation in the CLEF 2025 CheckThat! Lab Task 1: Subjectivity Detection in News Articles, classifying sentences as subjective/objective in monolingual, multilingual, and zero-shot settings. Training/development datasets were provided for Arabic, German, English, Italian, and Bulgarian; final evaluation included additional unseen languages (e.g., Greek, Romanian, Polish, Ukrainian) to assess generalization. Our primary strategy enhanced transformer-based classifiers by integrating sentiment scores, derived from an auxiliary model, with sentence representations, aiming to improve upon standard fine-tuning. We explored this sentiment-augmented architecture with mDeBERTaV3-base, ModernBERT-base (English), and Llama3.2-1B. To address class imbalance, prevalent across languages, we employed decision threshold calibration optimized on the development set. Our experiments show sentiment feature integration significantly boosts performance, especially subjective F1 score. This framework led to high rankings, notably 1st for Greek (Macro F1 = 0.51).
Combining Constrained and Unconstrained Decoding via Boosting: BoostCD and Its Application to Information Extraction
Many recent approaches to structured NLP tasks use an autoregressive language model $M$ to map unstructured input text $x$ to output text $y$ representing structured objects (such as tuples, lists, trees, code, etc.), where the desired output structure is enforced via constrained decoding. During training, these approaches do not require the model to be aware of the constraints, which are merely implicit in the training outputs $y$. This is advantageous as it allows for dynamic constraints without requiring retraining, but can lead to low-quality output during constrained decoding at test time. We overcome this problem with Boosted Constrained Decoding (BoostCD), which combines constrained and unconstrained decoding in two phases: Phase 1 decodes from the base model $M$ twice, in constrained and unconstrained mode, obtaining two weak predictions. In phase 2, a learned autoregressive boosted model combines the two weak predictions into one final prediction. The mistakes made by the base model with vs. without constraints tend to be complementary, which the boosted model learns to exploit for improved performance. We demonstrate the power of BoostCD by applying it to closed information extraction. Our model, BoostIE, outperforms prior approaches both in and out of distribution, addressing several common errors identified in those approaches.
OpenAI Teams Up With Oracle and SoftBank to Build 5 New Stargate Data Centers
The new sites will boost Stargate's planned capacity to nearly 7 gigawatts--about equal to the output of seven large nuclear reactors. An aerial view shows construction underway on a Project Stargate AI infrastructure site in Abilene, Texas on April 23, 2025. OpenAI is planning to build five new data centers in the United States as part of the Stargate initiative, the company announced on Tuesday. The sites, which are being developed in partnership with Oracle and SoftBank, bring Stargate's current planned capacity to nearly 7 gigawatts--roughly the same amount of power as seven large-scale nuclear reactors . "AI is different from the internet in a lot of ways, but one of them is just how much infrastructure it takes," OpenAI CEO Sam Altman said during a press briefing in Abilene, Texas on Tuesday.
Roundtables: Meet the 2025 Innovator of the Year
Watch a subscriber-only conversation with Sneha Goenka, who designed the computations behind the world's fastest whole-genome sequencing method. Every year, MIT Technology Review selects one individual whose work we admire to recognize as Innovator of the Year. For 2025, we chose Sneha Goenka, who designed the computations behind the world's fastest whole-genome sequencing method . Thanks to her work, physicians can now sequence a patient's genome and diagnose a genetic condition in less than eight hours--an achievement that could transform medical care. Exclusive: A record-breaking baby has been born from an embryo that's over 30 years old Jessica Hamzelou Therapists are secretly using ChatGPT. Exclusive: A record-breaking baby has been born from an embryo that's over 30 years old The embryos were created in 1994, while the expectant father was still a toddler, and donated via a Christian "embryo adoption" agency.