Agentic AI seeks to endow systems with sustained autonomy, reasoning, and interaction capabilities. To realize this vision, its assumptions about agency must be complemented by explicit models of cognition, cooperation, and governance. This paper argues that the conceptual tools developed within the Autonomous Agents and Multi-Agent Systems (AAMAS) community, such as BDI architectures, communication protocols, mechanism design, and institutional modelling, provide precisely such a foundation. By aligning adaptive, data-driven approaches with structured models of reasoning and coordination, we outline a path toward agentic systems that are not only capable and flexible, but also transparent, cooperative, and accountable. The result is a perspective on agency that bridges formal theory and practical autonomy.

Controlled generation imposes sequence-level constraints (syntax, style, safety) that depend on future tokens, making exact conditioning of an autoregressive LM intractable. Tractable surrogates such as HMMs can approximate continuation distributions and steer decoding, but standard surrogates are often weakly context-aware. We propose Learning to Look Ahead (LTLA), a hybrid method that uses base-LM embeddings to condition a globally learned tractable surrogate: a neural head predicts only a prefix-dependent latent prior, while a shared HMM answers continuation queries exactly. LTLA is designed to avoid two common efficiency traps when adding neural context. First, it avoids vocabulary-sized prefix rescoring (V extra LM evaluations) by scoring all next-token candidates via a single batched HMM forward update. Second, it avoids predicting a new HMM per prefix by learning one shared HMM and conditioning only the latent prior, which enables reuse of cached future-likelihood (backward) messages across decoding steps. Empirically, LTLA improves continuation likelihood over standard HMM surrogates, enables lookahead control for vision--language models by incorporating continuous context, achieves 100% syntactic constraint satisfaction, and improves detoxification while adding only a 14% decoding-time overhead.

Image-based virtual try-on (VTON) aims to synthesize photorealistic images of a person wearing specified garments. Despite significant progress, building a universal VTON framework that can flexibly handle diverse and complex tasks remains a major challenge. Recent methods explore multi-task VTON frameworks guided by textual instructions, yet they still face two key limitations: (1) semantic gap between text instructions and reference images, and (2) data scarcity in complex scenarios. To address these challenges, we propose UniFit, a universal VTON framework driven by a Multimodal Large Language Model (MLLM). Specifically, we introduce an MLLM-Guided Semantic Alignment Module (MGSA), which integrates multimodal inputs using an MLLM and a set of learnable queries. By imposing a semantic alignment loss, MGSA captures cross-modal semantic relationships and provides coherent and explicit semantic guidance for the generative process, thereby reducing the semantic gap. Moreover, by devising a two-stage progressive training strategy with a self-synthesis pipeline, UniFit is able to learn complex tasks from limited data. Extensive experiments show that UniFit not only supports a wide range of VTON tasks, including multi-garment and model-to-model try-on, but also achieves state-of-the-art performance. The source code and pretrained models are available at https://github.com/zwplus/UniFit.

Free-boundary diffeomorphism optimization, an important and widely occurring task in geometric modeling, computer graphics, and biological imaging, requires simultaneously determining a planar target domain and a locally bijective map with well-controlled distortion. We formulate this task through the least-squares quasiconformal (LSQC) operator and establish key structural properties of the LSQC minimizer, including well-posedness under mild conditions, invariance under similarity transformations, and resolution-independent behavior with stability under mesh refinement. We further analyze the sensitivity of the LSQC solution with respect to the Beltrami coefficient, establishing stability and differentiability properties that enable gradient-based optimization over the space of Beltrami coefficients. To make this differentiable formulation practical at scale and to facilitate the optimization process, we introduce the Spectral Beltrami Network (SBN), a multiscale mesh-spectral surrogate that approximates the LSQC solution operator in a single differentiable forward pass. This yields SBN-Opt, an optimization framework that searches over admissible Beltrami coefficients and pinning conditions to solve free-boundary diffeomorphism objectives with explicit distortion control. Extensive experiments on equiareal parameterization and inconsistent surface registration demonstrate consistent improvements over traditional numerical algorithms.

The data-centric paradigm has emerged as a pivotal direction in artificial intelligence (AI), emphasizing the role of high-quality training data. This shift is especially critical in the Text-to-SQL task, where the scarcity, limited diversity, and structural simplicity of existing datasets constrain model performance. To address these challenges, we propose Text2SQL-Flow, a SQL-aware data augmentation framework that systematically generates large-scale, semantically valid, and structurally diverse Text-to-SQL pairs from limited seed data. Our framework spans six augmentation dimensions and integrates an end-to-end pipeline with auxiliary database selection, SQL executability verification, natural language (NL) question generation, NL-SQL correspondence verification, and chain-of-thought (CoT) reasoning trace generation. Leveraging this framework, we construct SQLFlow, a high-quality dataset comprising 75,386 annotated examples. We demonstrate the utility of SQLFlow in both fine-tuning and prompt-based settings. (1) For open-source large language models (LLMs), fine-tuning with SQLFlow improves problem-solving ability, delivering competitive gains across multiple benchmarks under the same data budget. (2) For closed-source LLMs, we propose a masked alignment retrieval method that uses SQLFlow as both a knowledge base and training data for the retrieval model, enabling structure-aware example matching via fine-grained NL-SQL alignments. Experiments show that our retrieval strategy outperforms existing example retrieval methods, highlighting the combined value of SQLFlow's data quality and our retrieval technique. Overall, our work provides a scalable, data-centric foundation for advancing Text-to-SQL systems and underscores the importance of structured, high-fidelity data in modern AI development. Our code is available at https://github.com/TechNomad-ds/Text2SQL-Flow.

Video editing-based talking face generation aims to preserve video details such as pose, lighting, and gestures while modifying only lip motion, often using an identity reference image to maintain speaker consistency. However, this mechanism can introduce lip leakage, where generated lips are influenced by the reference image rather than solely by the driving audio. Such leakage is difficult to detect with standard metrics and conventional test setup. To address this, we propose a systematic evaluation methodology to analyze and quantify lip leakage. Our framework employs three complementary test setups: silent-input generation, mismatched audio-video pairing, and matched audio-video synthesis. We also introduce derived metrics including lip-sync discrepancy and silent-audio-based lip-sync scores. In addition, we study how different identity reference selections affect leakage, providing insights into reference design. The proposed methodology is model-agnostic and establishes a more reliable benchmark for future research in talking face generation.

Foundation models for time series imputation remain largely unexplored. Recently, two such models, TabPFN-TS and MoTM, have emerged. These models share a common philosophy that places them within the family of time-indexed foundation models. This paper presents the first large-scale empirical study of these models for zero-shot imputation, which enables missing value recovery without retraining across a wide range of scenarios. We conduct extensive univariate experiments across 33 out-of-domain datasets (approximately 1.3M imputation windows) and evaluate their ability to integrate covariates at inference time to improve accuracy without fine-tuning. Our results demonstrate that time-indexed foundation models are a powerful and practical step toward achieving general-purpose, zero-shot imputation for real-world time series.

Recent advancements in large language models (LLMs) have enabled significant progress in decision-making and task planning for embodied autonomous agents. However, most existing methods struggle with complex, long-horizon tasks because they rely on a monolithic trajectory that entangles all past decisions and observations to solve the entire task in a single unified process. To address this limitation, we propose ReAcTree, a hierarchical task-planning method that decomposes a complex goal into manageable subgoals within a dynamically constructed agent tree. Each subgoal is handled by an LLM agent node capable of reasoning, acting, and further expanding the tree, while control flow nodes coordinate the execution strategies of agent nodes. In addition, we integrate two complementary memory systems: each agent node retrieves goal-specific, subgoal-level examples from episodic memory and shares environment-specific observations through working memory. Experiments on the WAH-NL and ALFRED show ReAcTree consistently outperforms strong task-planning baselines such as ReAct across diverse LLMs. Notably, on WAH-NL, ReAcTree achieves a 61% goal success rate with Qwen 2.5 72B, nearly doubling ReAct's 31%. The code is available at https://github.com/Choi-JaeWoo/ReAcTree.git.

Traditional approaches to semantic polarity in computational linguistics treat sentiment as a unidimensional scale, overlooking the multidimensional structure of language. This work introduces TOPol (Topic-Orientation POLarity), a semi-unsupervised framework for reconstructing and interpreting multidimensional narrative polarity fields under human-on-the-loop (HoTL) defined contextual boundaries (CBs). The framework embeds documents using a transformer-based large language model (tLLM), applies neighbor-tuned UMAP projection, and segments topics via Leiden partitioning. Given a CB between discourse regimes A and B, TOPol computes directional vectors between corresponding topic-boundary centroids, yielding a polarity field that quantifies fine-grained semantic displacement during regime shifts. This vectorial representation enables assessing CB quality and detecting polarity changes, guiding HoTL CB refinement. To interpret identified polarity vectors, the tLLM compares their extreme points and produces contrastive labels with estimated coverage. Robustness analyses show that only CB definitions (the main HoTL-tunable parameter) significantly affect results, confirming methodological stability. We evaluate TOPol on two corpora: (i) U.S. Central Bank speeches around a macroeconomic breakpoint, capturing non-affective semantic shifts, and (ii) Amazon product reviews across rating strata, where affective polarity aligns with NRC valence. Results demonstrate that TOPol consistently captures both affective and non-affective polarity transitions, providing a scalable, generalizable, and interpretable framework for context-sensitive multidimensional discourse analysis.

We present a framework for training large language models (LLMs) as diagnostic agents with reinforcement learning, enabling them to manage multi-turn interactive diagnostic processes, adaptively select examinations, and commit to final diagnoses. Unlike instruction-tuned models trained on static data, our method acquires diagnostic strategies through dynamic exploration and outcome-based feedback, mapping evolving patient states to the next optimal examination and subsequent diagnosis. Our contributions include: (i) DiagGym, a diagnostics world model trained with electronic health records, serving as a virtual clinical environment to support closed-loop in-silico training and evaluation for interactive diagnosis; (ii) DiagAgent, trained via end-to-end multi-turn RL to learn dynamic diagnostic policies that optimize both interactive effectiveness and final accuracy; (iii) DiagBench, a multi-center diagnostic benchmark designed to evaluate multi-turn diagnostic interaction trajectories. The benchmark comprises 2.2K physician-validated cases sourced from 4 distinct distributions, alongside 3.3K physician-written rubrics for granular process-oriented evaluation. (iv) Extensive evaluations demonstrate DiagAgent's superior performance across both in-domain and out-of-domain (OOD) settings. DiagAgent significantly outperforms 11 SOTA LLMs and 2 prompt-engineered agents. In the end-to-end setting, it delivers a 11.20% increase in diagnostic accuracy and a 17.58% boost in examination recommendation F1 score, while consistently maintaining SOTA performance across all three external centers. Furthermore, in rubric-based evaluations, it surpasses the next-best model by 7.1% in weighted rubric score. These findings indicate that learning policies in interactive clinical environments confers long-term diagnostic management abilities unattainable through passive training.

Autoformalization, which translates natural language mathematics into machine-verifiable formal statements, is critical for using formal mathematical reasoning to solve math problems stated in natural language. While Large Language Models can generate syntactically correct formal statements, they often fail to preserve the original problem's semantic intent. This limitation arises from the LLM approaches' treating autoformalization as a simplistic translation task which lacks mechanisms for self-reflection and iterative refinement that human experts naturally employ. To address these issues, we propose ReForm, a Reflective Autoformalization method that tightly integrates semantic consistency evaluation into the autoformalization process. This enables the model to iteratively generate formal statements, assess its semantic fidelity, and self-correct identified errors through progressive refinement. To effectively train this reflective model, we introduce Prospective Bounded Sequence Optimization (PBSO), which employs different rewards at different sequence positions to ensure that the model develops both accurate autoformalization and correct semantic validations, preventing superficial critiques that would undermine the purpose of reflection. Extensive experiments across four autoformalization benchmarks demonstrate that ReForm achieves an average improvement of 22.6 percentage points over the strongest baselines. To further ensure evaluation reliability, we introduce ConsistencyCheck, a benchmark of 859 expert-annotated items that not only validates LLMs as judges but also reveals that autoformalization is inherently difficult: even human experts produce semantic errors in up to 38.5% of cases.

Text-to-Speech (TTS) is inherently a "one-to-many" mapping characterized by intrinsic uncertainty, yet current paradigms often oversimplify it into a deterministic regression task. While continuous-valued autoregressive (AR) models have recently emerged as a promising alternative to discrete codec-based approaches, they typically rely on a fixed-variance prior, fundamentally constraining generation to a static point estimate that ignores the dynamic variability of natural speech. To bridge this gap, we propose BELLE (Bayesian evidential learning with language modelling), a framework that shifts from deterministic prediction to principled Bayesian inference without increasing model parameters or inference latency. By modeling the acoustic target as a Normal-Inverse-Gamma distribution, BELLE captures data-dependent aleatoric uncertainty. To enable accurate variance estimation on standard single-reference datasets, we introduce a "one-to-many" training strategy that leverages synthetic samples as a statistical support set, allowing the model to learn robust distributional properties rather than merely imitating teacher artifacts. Experiments demonstrate that BELLE, trained on only ~5k hours of data, outperforms leading open-source models trained on 50k hours (achieving a 25.8% relative WER reduction) and naturally supports high-quality streaming generation. Audio samples are available at https://belletts.github.io/Belle/.

Alignment of large language models (LLMs) has predominantly relied on pairwise preference optimization, where annotators select the better of two responses to a prompt. While simple, this approach overlooks the opportunity to learn from richer forms of human feedback, such as multiway comparisons and top-$k$ rankings. We introduce Ranked Choice Preference Optimization (RCPO), a unified framework that bridges preference optimization with (ranked) choice modeling via maximum likelihood estimation. RCPO supports both utility-based and rank-based models, subsumes several pairwise methods (such as DPO and SimPO) as special cases, and provides principled training objectives for richer feedback formats. We instantiate this framework with two representative models (Multinomial Logit and Mallows-RMJ). Experiments on Llama-3-8B-Instruct, Gemma-2-9B-it, and Mistral-7B-Instruct across in-distribution and out-of-distribution settings show that RCPO consistently outperforms competitive baselines. RCPO shows that directly leveraging ranked preference data, combined with the right choice models, yields more effective alignment. It offers an extensible foundation for incorporating (ranked) choice modeling into LLM training.

Multimodal learning integrates diverse modalities but suffers from modality imbalance, where dominant modalities suppress weaker ones due to inconsistent convergence rates. Existing methods predominantly rely on static modulation or heuristics, overlooking sample-level distributional variations in prediction bias. Specifically, they fail to distinguish outlier samples where the modality gap is exacerbated by low data quality. We propose a framework to quantitatively diagnose and dynamically mitigate this imbalance at the sample level. We introduce the Modality Gap metric to quantify prediction discrepancies. Analysis reveals that this gap follows a bimodal distribution, indicating the coexistence of balanced and imbalanced sample subgroups. We employ a Gaussian Mixture Model (GMM) to explicitly model this distribution, leveraging Bayesian posterior probabilities for soft subgroup separation. Our two-stage framework comprises a Warm-up stage and an Adaptive Training stage. In the latter, a GMM-guided Adaptive Loss dynamically reallocates optimization priorities: it imposes stronger alignment penalties on imbalanced samples to rectify bias, while prioritizing fusion for balanced samples to maximize complementary information. Experiments on CREMA-D, AVE, and KineticSound demonstrate that our method significantly outperforms SOTA baselines. Furthermore, we show that fine-tuning on a GMM-filtered balanced subset serves as an effective data purification strategy, yielding substantial gains by eliminating extreme noisy samples even without the adaptive loss.

Although membership inference attacks (MIAs) and machine-generated text detection target different goals, their methods often exploit similar signals based on a language model's probability distribution, and the two tasks have been studied independently. This can result in conclusions that overlook stronger methods and valuable insights from the other task. In this work, we theoretically and empirically demonstrate the transferability, i.e., how well a method originally developed for one task performs on the other, between MIAs and machine text detection. We prove that the metric achieving asymptotically optimal performance is identical for both tasks. We unify existing methods under this optimal metric and hypothesize that the accuracy with which a method approximates this metric is directly correlated with its transferability. Our large-scale empirical experiments demonstrate very strong rank correlation ($ρ\approx 0.7$) in cross-task performance. Notably, we also find that a machine text detector achieves the strongest performance among evaluated methods on both tasks, demonstrating the practical impact of transferability. To facilitate cross-task development and fair evaluation, we introduce MINT, a unified evaluation suite for MIAs and machine-generated text detection, implementing 15 recent methods from both tasks.

The classical policy gradient method is the theoretical and conceptual foundation of modern policy-based reinforcement learning (RL) algorithms. Most rigorous analyses of such methods, particularly those establishing convergence guarantees, assume a discount factor $γ< 1$. In contrast, however, a recent line of work on policy-based RL for large language models uses the undiscounted total-reward setting with $γ= 1$, rendering much of the existing theory inapplicable. In this paper, we provide analyses of the policy gradient method for undiscounted expected total-reward infinite-horizon MDPs based on two key insights: (i) the classification of the MDP states into recurrent and transient states is invariant over the set of policies that assign strictly positive probability to every action (as is typical in deep RL models employing a softmax output layer) and (ii) the classical state visitation measure (which may be ill-defined when $γ= 1$) can be replaced with a new object that we call the transient visitation measure.

Video-based representations have gained prominence in planning and decision-making due to their ability to encode rich spatiotemporal dynamics and geometric relationships. These representations enable flexible and generalizable solutions for complex tasks such as object manipulation and navigation. However, existing video planning frameworks often struggle to adapt to failures at interaction time due to their inability to reason about uncertainties in partially observed environments. To overcome these limitations, we introduce a novel framework that integrates interaction-time data into the planning process. Our approach updates model parameters online and filters out previously failed plans during generation. This enables implicit state estimation, allowing the system to adapt dynamically without explicitly modeling unknown state variables. We evaluate our framework through extensive experiments on a new simulated manipulation benchmark, demonstrating its ability to improve replanning performance and advance the field of video-based decision-making.

Large Vision-Language Models (LVLMs) excel in diverse cross-modal tasks. However, object hallucination, where models produce plausible but inaccurate object descriptions, remains a significant challenge. In contrast to previous work focusing on LLM components, this paper is the first to trace LVLM hallucinations to visual encoders and identifies three key issues: statistical bias, inherent bias, and vulnerability. To address these challenges, we propose SHIELD, a training-free framework that mitigates hallucinations through three strategies: re-weighting visual tokens to reduce statistical bias, introducing noise-derived tokens to counter inherent bias, and applying adversarial attacks with contrastive decoding to address vulnerability. Experiments demonstrate that SHIELD effectively mitigates object hallucinations across diverse benchmarks and LVLM families. Moreover, SHIELD achieves strong performance on the general LVLM benchmark, highlighting its broad applicability. Code is available at https://github.com/hukcc/SHIELD.

In autonomous driving, dynamic environment and corner cases pose significant challenges to the robustness of ego vehicle's state understanding and decision making. We introduce VDRive, a novel pipeline for end-to-end autonomous driving that explicitly models state-action mapping to address these challenges, enabling interpretable and robust decision making. By leveraging the advancement of the state understanding of the Vision Language Action Model (VLA) with generative diffusion policy-based action head, our VDRive guides the driving contextually and geometrically. Contextually, VLA predicts future observations through token generation pre-training, where the observations are represented as discrete codes by a Conditional Vector Quantized Variational Autoencoder (CVQ-VAE). Geometrically, we perform reinforcement learning fine-tuning of the VLA to predict future trajectories and actions based on current driving conditions. VLA supplies the current state tokens and predicted state tokens for the action policy head to generate hierarchical actions and trajectories. During policy training, a learned critic evaluates the actions generated by the policy and provides gradient-based feedback, forming an actor-critic framework that enables a reinforcement-based policy learning pipeline. Experiments show that our VDRive achieves state-of-the-art performance in the Bench2Drive closed-loop benchmark and nuScenes open-loop planning.

Although large language models (LLMs) have made significant strides across various tasks, they still face significant challenges in complex reasoning and planning. For example, even with carefully designed prompts and prior information explicitly provided, GPT-4o achieves only a 7% Final Pass Rate on the TravelPlanner dataset in the sole-planning mode. Similarly, even in the thinking mode, Qwen3-8B-Instruct and DeepSeek-R1-671B, only achieve Final Pass Rates of 5.9% and 40%, respectively. Although well-organized Multi-Agent Systems (MAS) can offer improved collective reasoning, they often suffer from high reasoning costs due to multi-round internal interactions, long per-response latency, and difficulties in end-to-end training. To address these challenges, we propose a general and scalable framework called IMAGINE, short for Integrating Multi-Agent System into One Model. This framework not only integrates the reasoning and planning capabilities of MAS into a single, compact model, but also significantly surpass the capabilities of the MAS through a simple end-to-end training. Through this pipeline, a single small-scale model is not only able to acquire the structured reasoning and planning capabilities of a well-organized MAS but can also significantly outperform it. Experimental results demonstrate that, when using Qwen3-8B-Instruct as the base model and training it with our method, the model achieves an 82.7% Final Pass Rate on the TravelPlanner benchmark, far exceeding the 40% of DeepSeek-R1-671B, while maintaining a much smaller model size.

We present SENTINEL, a framework for formally evaluating the physical safety of foundation model (FM)-based embodied agents. SENTINEL is the first to provide multi-level safety evaluation across semantic interpretation, plan generation, and physical execution within a unified formal framework. Unlike prior methods that rely on heuristic rules or subjective FM judgments, SENTINEL grounds practical safety requirements in formal temporal logic (TL) semantics that can precisely specify state invariants, temporal dependencies, and timing constraints. It employs a multi-level verification pipeline where (i) at the semantic level, intuitive natural language safety requirements are formalized into TL formulas and the agent's understanding of these requirements is probed for alignment with the TL formulas; (ii) at the plan level, high-level action plans and subgoals generated by the agent are verified against the TL formulas to detect unsafe plans before execution; and (iii) at the trajectory level, multiple execution trajectories are merged into a computation tree and efficiently verified against physically-detailed TL specifications for a final safety check. We apply SENTINEL in VirtualHome and AI2-THOR, and formally evaluate multiple FM-based embodied agents against diverse safety requirements. Our experiments show that by grounding physical safety in temporal logic and applying verification methods across multiple levels, SENTINEL provides a rigorous foundation for systematically evaluating the safety of FM-based embodied agents in simulation-based physical environments, and can effectively expose potential safety violations in interpreting, planning, and executing the tasks.

Interactive 3D point cloud segmentation enables efficient annotation of complex 3D scenes through user-guided prompts. However, current approaches are typically restricted in scope to a single domain (indoor or outdoor), and to a single form of user interaction (either spatial clicks or textual prompts). Moreover, training on multiple datasets often leads to negative transfer, resulting in domain-specific tools that lack generalizability. To address these limitations, we present SNAP (Segment aNything in Any Point cloud), a unified model for interactive 3D segmentation that supports both point-based and text-based prompts across diverse domains. Our approach achieves cross-domain generalizability by training on 7 datasets spanning indoor, outdoor, and aerial environments, while employing domain-adaptive normalization to prevent negative transfer. For text-prompted segmentation, we automatically generate mask proposals without human intervention and match them against CLIP embeddings of textual queries, enabling both panoptic and open-vocabulary segmentation. Extensive experiments demonstrate that SNAP consistently delivers high-quality segmentation results. We achieve state-of-the-art performance on 8 out of 9 zero-shot benchmarks for spatial-prompted segmentation and demonstrate competitive results on all 5 text-prompted benchmarks. These results show that a unified model can match or exceed specialized domain-specific approaches, providing a practical tool for scalable 3D annotation. Project page is at, https://neu-vi.github.io/SNAP/

We study an online stochastic matching problem in which an algorithm sequentially matches $U$ users to $K$ arms, aiming to maximize cumulative reward over $T$ rounds under budget constraints. Without structural assumptions, computing the optimal matching is NP-hard, making online learning computationally infeasible. To overcome this barrier, we focus on single-peaked preferences -- a well-established structure in social choice theory, where users' preferences are unimodal with respect to a common order over arms. We devise an efficient algorithm for the offline budgeted matching problem, and leverage it into an efficient online algorithm with a regret of $\tilde O(UKT^{2/3})$. Our approach relies on a novel PQ tree-based order approximation method. If the single-peaked structure is known, we develop an efficient UCB-like algorithm that achieves a regret bound of $\tilde O(U\sqrt{TK})$.

Attention sinks and compression valleys have attracted significant attention as two puzzling phenomena in large language models, but have been studied in isolation. In this work, we present a surprising connection between attention sinks and compression valleys, tracing both to the formation of massive activations in the residual stream. We prove theoretically that massive activations necessarily produce representational compression and establish bounds on the resulting entropy reduction. Through experiments across several models (410M-120B parameters), we confirm that when the beginning-of-sequence token develops extreme activation norms in the middle layers, both compression valleys and attention sinks emerge simultaneously. Targeted ablation studies validate our theoretical predictions. This unified view motivates us to propose the Mix-Compress-Refine theory of information flow, as an attempt to explain how LLMs organize their computation in depth by controlling attention and representational compression via massive activations. Specifically, we posit that Transformer-based LLMs process tokens in three distinct phases: (1) broad mixing in the early layers, (2) compressed computation with limited mixing in the middle layers, and (3) selective refinement in the late layers. Our framework helps explain why embedding tasks perform best at intermediate layers, whereas generation tasks benefit from full-depth processing, clarifying differences in task-dependent representations.

Transportation agencies make critical operational decisions during hazardous weather events, including assessment of road conditions and resource allocation. In this study, machine learning models are developed to provide additional support for the New York State Department of Transportation (NYSDOT) by automatically classifying current road conditions across the state. Convolutional neural networks and random forests are trained on NYSDOT roadside camera images and weather data to predict road surface conditions. This task draws critically on a robust hand-labeled dataset of ~22,000 camera images containing six road surface conditions: severe snow, snow, wet, dry, poor visibility, or obstructed. Model generalizability is prioritized to meet the operational needs of the NYSDOT decision makers, including integration of operational datasets and use of representative and realistic images. The weather-related road surface condition model in this study achieves an accuracy of 81.5% on completely unseen cameras. With operational deployment, this model has the potential to improve spatial and temporal awareness of road surface conditions, which can strengthen decision-making for operations, roadway maintenance, and traveler safety, particularly during winter weather events.

Purpose: The FedSurg challenge was designed to benchmark the state of the art in federated learning for surgical video classification. Its goal was to assess how well current methods generalize to unseen clinical centers and adapt through local fine-tuning while enabling collaborative model development without sharing patient data. Methods: Participants developed strategies to classify inflammation stages in appendicitis using a preliminary version of the multi-center Appendix300 video dataset. The challenge evaluated two tasks: generalization to an unseen center and center-specific adaptation after fine-tuning. Submitted approaches included foundation models with linear probing, metric learning with triplet loss, and various FL aggregation schemes (FedAvg, FedMedian, FedSAM). Performance was assessed using F1-score and Expected Cost, with ranking robustness evaluated via bootstrapping and statistical testing. Results: In the generalization task, performance across centers was limited. In the adaptation task, all teams improved after fine-tuning, though ranking stability was low. The ViViT-based submission achieved the strongest overall performance. The challenge highlighted limitations in generalization, sensitivity to class imbalance, and difficulties in hyperparameter tuning in decentralized training, while spatiotemporal modeling and context-aware preprocessing emerged as promising strategies. Conclusion: The FedSurg Challenge establishes the first benchmark for evaluating FL strategies in surgical video classification. Findings highlight the trade-off between local personalization and global robustness, and underscore the importance of architecture choice, preprocessing, and loss design. This benchmarking offers a reference point for future development of imbalance-aware, adaptive, and robust FL methods in clinical surgical AI.

The computational role of imagination remains debated. While classical accounts emphasize reward maximization, emerging evidence suggests it accesses internal world models (IWMs). We employ psychological network analysis to compare IWMs in humans and large language models (LLMs) via imagination vividness ratings, distinguishing offline world models (persistent memory structures accessed independent of immediate goals) from online models (task-specific representations). Analyzing 2,743 humans across three populations and six LLM variants, we find human imagination networks exhibit robust structural consistency, with high centrality correlations and aligned clustering. LLMs show minimal clustering and weak correlations with human networks, even with conversational memory, across environmental and sensory contexts. These differences highlight disparities in how biological and artificial systems organize internal representations. Our framework offers quantitative metrics for evaluating offline world models in cognitive agents.

Fairness testing evaluates whether a model satisfies a specified fairness criterion across different groups, yet most research has focused on classification models, leaving regression models underexplored. This paper introduces a framework for fairness testing in regression models, leveraging Wasserstein distance to project data distribution and focusing on expectation-based criteria. Upon categorizing fairness criteria for regression, we derive a Wasserstein projection test statistic from dual reformulation, and derive asymptotic bounds and limiting distributions, allowing us to formulate both a hypothesis-testing procedure and an optimal data perturbation method to improve fairness while balancing accuracy. Experiments on synthetic data demonstrate that the proposed hypothesis-testing approach offers higher specificity compared to permutation-based tests. To illustrate its potential applications, we apply our framework to two case studies on real data, showing (1) statistically significant gender disparities that appear on student performance data across multiple models, and (2) significant unfairness between pollution areas under multiple fairness criteria affecting housing price data, robust to different group divisions, with feature-level analysis identifying spatial and socioeconomic drivers.

A common but rarely examined assumption in machine learning is that training yields models that actually satisfy their specified objective function. We call this the Objective Satisfaction Assumption (OSA). Although deviations from OSA are acknowledged, their implications are overlooked. We argue, in a learning-paradigm-agnostic framework, that OSA fails in realistic conditions: approximation, estimation, and optimization errors guarantee systematic deviations from the intended objective, regardless of the quality of its specification. Beyond these technical limitations, perfectly capturing and translating the developer's intent, such as alignment with human preferences, into a formal objective is practically impossible, making misspecification inevitable. Building on recent mathematical results, absent a mathematical characterization of these gaps, they are indistinguishable from those that collapse into Goodhart's law failure modes under strong optimization pressure. Because the Goodhart breaking point cannot be located ex ante, a principled limit on the optimization of General-Purpose AI systems is necessary. Absent such a limit, continued optimization is liable to push systems into predictable and irreversible loss of control.

Post-training for large language models (LLMs) is constrained by the high cost of acquiring new knowledge or correcting errors and by the unintended side effects that frequently arise from retraining. To address these issues, we introduce REPAIR (Robust Editing via Progressive Adaptive Intervention and Reintegration), a lifelong editing framework designed to support precise and low-cost model updates while preserving non-target knowledge. REPAIR mitigates the instability and conflicts of large-scale sequential edits through a closed-loop feedback mechanism coupled with dynamic memory management. Furthermore, by incorporating frequent knowledge fusion and enforcing strong locality guards, REPAIR effectively addresses the shortcomings of traditional distribution-agnostic approaches that often overlook unintended ripple effects. Our experiments demonstrate that REPAIR boosts editing accuracy by 10%-30% across multiple model families and significantly reduces knowledge forgetting. This work introduces a robust framework for developing reliable, scalable, and continually evolving LLMs.