This paper presents a method for traffic accident reconstruction from public reports and scene measurements, formulating it as a parameterized multimodal learning problem. Researchers created the CISS-REC dataset with 6,217 real-world cases and developed a framework that outperforms baselines in reconstruction fidelity, including accident point accuracy.
This paper introduces predictable history-adaptive virtual perturbations to enhance information-theoretic generalization bounds for Stochastic Gradient Descent. This new approach allows perturbation covariances to dynamically depend on past SGD history, addressing limitations of existing methods that require fixed covariances.
This paper provides a comprehensive survey on data mixing for Large Language Model (LLM) pretraining, a crucial factor for training efficiency and downstream generalization. It formalizes data mixture optimization as a bilevel problem and introduces a fine-grained taxonomy for existing methods.
World Models are featured in "10 Things That Matter in AI Right Now", with executive editor Niall Firth explaining their rising importance. A subscriber-only discussion by MIT Technology Review will delve into how AI may evolve to understand the world.
The Temporal Contrastive Transformer (TCT) is a new representation learning framework designed for financial transaction sequences to detect fraud. It uses self-supervised contrastive learning to generate embeddings that capture temporal behavioral patterns, showing meaningful predictive performance, especially when combined with domain-engineered features.
This research investigates the "double descent" phenomenon in overparametrized models, which allows for improved generalization despite classical overfitting concerns. The study specifically explores this effect in linear regression with contaminated training data, finding that significant overparametrization enables double descent even in such robust settings.
This research investigates harness engineering as an inference-time technique for large language model (LLM) agents, focusing on improving long-term performance via task decomposition and guided execution. It quantifies how design elements like workflow granularity and guidance impact performance, revealing common failure modes such as over-decomposition and hallucinated execution.
This paper proposes the Inter/Intra-class Distance Ratio (ILDR) as a novel geometric signal for the early detection of 'grokking' in neural networks. ILDR, computed from second-to-last layer representations, indicates geometric reorganization in representation space before validation accuracy improves, outperforming existing detection methods.
This paper proposes a novel method to enhance privacy reasoning in LLMs by extracting normative simulacra from fiction novels. The approach involves fine-tuning LLMs via supervised learning followed by GRPO reinforcement learning, using a composite reward function to align information handling practices with user privacy expectations.
This paper details participation in SemEval-2026 Task 13, focusing on lightweight detection of LLM-generated code using stylometric signals. The approach employs ratio-based features, parsing engines, and language classifiers, proving computationally efficient with near-instant inference time.
This research introduces Adaptive Power-Mean Policy Optimization (APMPO) to improve Large Language Model (LLM) reasoning capabilities within Reinforcement Learning with Verifiable Rewards (RLVR). APMPO combines a generalized power-mean objective and feedback-adaptive clipping to enhance learning dynamics and performance, addressing limitations of static optimization schemes.
This paper introduces EMA, the first model adaptation system designed to help learning-based systems adapt to evolving environments with minimal operational overhead. It addresses costly model training and extensive data collection challenges by reducing expensive model training.
This empirical study investigates Tian's (2025) feature repulsion theorem in two-layer network grokking, testing its mechanisms and spectral signatures. It observes a clear structure-mechanism dissociation, with the predicted sign rule robustly holding for similar feature pairs despite a strong activation dependence in the spectral signature.
This research proposes an offline reasoning distillation framework for Large Language Models (LLMs) to enhance intelligence in resource-constrained environments. It tackles the distributional drift issue in existing offline methods by correcting teacher-student discrepancies while preserving data efficiency and supervision quality.
This paper introduces TraFL, a novel post-training approach for diffusion language models that addresses "trajectory locking" observed in reward-maximizing methods. TraFL, a trajectory-balance objective, outperforms other methods across mathematical reasoning and code generation benchmarks.
This survey addresses the proliferation of adversarial synthetic content, accelerated by Generative AI, which renders traditional reactive detection methods ineffective. It proposes a paradigm shift towards proactive detection of emerging inauthentic narratives, adopting a unified, lifecycle-based taxonomy integrating socio-technical models and advanced computational methodologies.
This research addresses challenges in robust molecular property prediction under extreme out-of-distribution (OOD) scenarios crucial for AI-driven drug discovery. It proposes SCOPE-BENCH, a new benchmark for OOD performance evaluation, and POMA, a framework for multi-source adaptation to overcome limitations of existing methods.
Hoeffding Concept Bottleneck Models (HCBM) are introduced to offer non-linear and sparse aggregations of concept scores, enhancing the explainability and accuracy of deep learning predictions. This method leverages Hoeffding functional decomposition of gradient-boosted trees to overcome the limitations of existing linear CBMs, which suffer from a large number of concepts and potential information leakage.
The Transformer Integrated Temporal Causal Discovery (TTCD) Framework is a novel end-to-end approach designed to learn contemporaneous and lagged causal relations from complex non-stationary time series data. This method addresses the limitations of existing techniques by integrating temporal and frequency-domain attention, providing a unified solution for challenging real-world scenarios.
PLACO is a multi-stage framework designed for cost-effective performance in Human-AI teams, particularly for classification tasks. The paper addresses the crucial combination of human and model output, building upon prior work that used Bayes rule.