University of Massachusetts Amherst（马萨诸塞大学阿默斯特分校）

Computer vision models are increasingly used as measurement tools to estimate population-level quantities from large image collections, but prediction errors introduce bias and the resulting estimates lack statistical guarantees required in scientific applications. Prior work uses a Monte Carlo framework to combine model predictions with ground-truth annotations by sampling some images for humans to label and is able to provide unbiased estimates with controllable accuracy, but primarily addresses single-scalar estimation. We study the more general problem of multi-target estimation, where many quantities (e.g., class counts or proportions) must be estimated simultaneously, and adapt sampling and estimation strategies from survey sampling to this setting. Evaluations on five detection and segmentation datasets with 7-80 classes show that importance sampling excels with moderate annotation budgets or fewer targets, whereas uniform sampling with control variates is superior when estimating many targets or operating with minimal labels. Additionally, a subset-based ratio estimator remains highly competitive across all regimes. Ultimately, our framework effectively combines biased model predictions and limited human labels into rigorous scientific measurements.

Packing for air travel is recurring and error-prone: the checklist must be personal and context-aware, yet feasible under safety rules, item dependencies, and luggage limits. Existing packing assistants are template-driven and generic, or recommendation-driven but unconstrained, leaving users to manually patch regulatory and capacity violations. We propose a reasoning-guided learning framework with three stages: (1) a symbolic engine that generates a regulation-aware seed checklist with explicit dependency structure, (2) a two-stage preference learner that estimates inclusion and priority utilities from user add and remove actions while mitigating survivorship bias, and (3) a CP-SAT optimizer that selects a compact, compliant subset. The architecture instantiates a general pattern for constrained personalization, applicable wherever hard feasibility coexists with sparse preference signals. On 604 labeled trip scenarios, comprising 29K inclusion labels and 343K pairwise comparisons, the symbolic engine attains 99.7% recall and 0.96 rubric validity, compared with 0.78 to 0.81 for frontier LLMs. Gradient-boosted trees and LambdaMART reach an AUC-ROC of 0.943 and an NDCG@5 of 0.923. CP-SAT attains 100% constraint satisfaction, compared with 28% for greedy selection and 10% for random selection. Deployment in FlyEnJoy, a production iOS travel app, doubled checklist completions and reduced editing and completion time.

Generating personalized trip itineraries is a complex planning task and involves a tension between hard combinatorial feasibility and soft latent desirability. Classical optimization enforces constraints but fails to capture subjective traveler preferences. While learning-based approaches model preferences, they cannot guarantee feasibility. Mobile deployment imposes additional resource constraints on both. To address this, we propose Plan, Learn, Adapt (PLA), a three-stage framework for personalized on-device itinerary generation. The Plan stage builds a heterogeneous ensemble of lightweight planners that produces structurally diverse feasible candidates. From pairwise itinerary comparisons, Learn fits a compact Bradley-Terry reward model that captures emergent schedule properties such as pacing, geographic coherence, and day balance, which per-POI signals miss. Finally, Adapt applies feasibility-preserving local refinement within a device-aware compute budget; every intermediate state is feasible by construction. On 2,519 pairwise human comparisons across more than 100 U.S. cities, the reward-guided ensemble achieves a 67.8% win rate, 11.2 percentage points above the best single planner, with 100% feasibility. Three frontier LLMs, GPT-5, Claude Opus 4.5, and Gemini 3 Pro, achieve 0% feasibility under the same constraints. The reward model generalizes across held-out cities, with a 67.6% mean leave-one-city-out accuracy. In production deployment within FlyEnJoy, PLA increased itinerary completion rates by 91%, with 109.9 ms average on-device latency.

We propose \emph{the sublinear-growth principle} for deep residual architectures -- a sharp stability threshold on the input-magnitude exponent of every residual block's velocity field: $$\|v(x, t)\| \leq c\,\|x\|^q + b, \qquad q \in [0, 1].$$ The threshold $q = 1$ is established via two independent arguments. Classical ODE theory gives a global forward flow on $[0, T]$ at $q \le 1$ and exhibits divergent velocity fields at any $q > 1$. The optimal-control analysis, via the Hamilton-Jacobi-Bellman equation, sharpens this to a selection statement: the training optimum is bang-bang on the boundary of the admissible class, so the optimum at $q > 1$ blows up while the optimum at $q \le 1$ is safe by construction. The exponent criterion $q \le 1$ is thereby a necessary and sufficient condition for stable training. It clarifies architectural placements that ensure the stability of training and inference, explaining, for instance, the stabilizing role of layer normalization. The sublinear-growth velocity fields form \emph{the right function space} on which forward dynamics, adjoint sensitivity, and architectural composition are all well-controlled. An arithmetic of input-magnitude exponents under the five operations that build residual blocks enables efficient certification of $q_k \le 1$ at the level of architectural primitives, in place of ad hoc trial and error in the search for stable neural architectural designs. A parameter-free modification reduces the supercritical Mamba block from $q = 5$ to $q = 1$ without layer normalization, demonstrating this point. Experiments on Mamba and PatchTST confirm that the $q \le 1$ variants train stably: the criterion is the input-magnitude exponent, not the presence of a normalization layer.

Relative pose estimation is a fundamental capability for collaborative perception and coordination in multi-robot systems. However, robots encountering each other in real-world environments often operate in short interaction windows and must operate under limited communication bandwidth with intermittent or missing visual overlap caused by occlusions or limited fields of view. Existing approaches typically rely on global reference frames, assume sustained view overlap, or incur prohibitive communication costs, thereby limiting their applicability to ephemeral collaborative perception. To address these challenges, we introduce communication-efficient relative pose estimation (CERPE), a system-level framework that coordinates vision foundation models to jointly estimate ego-motion and inter-robot relative pose. CERPE reduces unnecessary raw-observation exchange by using continuously shared fixed-size descriptors to gate event-triggered raw-image requests independently of pose estimation. Non-overlapping encounters are handled by propagating inter-robot relative poses through metrically scaled ego-motion, thus maintaining relative pose estimates even in the absence of visual overlap. Experiments in simulation and real-world robots show that CERPE improves 6-DoF relative pose estimation over selected baselines in ephemeral collaborative perception.

This paper develops a categorical framework -- Learning in Infinitesimal Non-Compositional Sketches (LINCS) -- as the repair of non-compositionality: failures of diagrams to factor through quotient sketches lifted to the tangent category setting. Machine learning problems are specified as sketches: graphs with commutativity conditions $\mathcal D$, limit cones $\mathcal L$, and colimit cocones $\mathcal K$, generalizing the usual scalarization of loss functions or vector space assumptions. Non-compositionality is defined purely as failure of a universal factorization problem, not as arithmetic error between the desired and actual predictions. Given a learning sketch $\mathbb S=(S,\mathcal D,\mathcal L,\mathcal K)$, whose underlying graph is $S$, and a model $D:J \rightarrow C$, the base defect is the obstruction to factorization $\mbox{Obs}(\mbox{Fact}_{\mathbb S}(D))$. The tangent lift applies the tangent functor $T$ to obtain $TD:J \rightarrow C$, and LINCS is defined as the obstruction $\mbox{Obs}(\mbox{Fact}_{\mathbb S}(TD))$ -- asking whether infinitesimal perturbations preserve the compositionality constraints.The paper also introduces Tangent Learning Sketches, which are sketches equipped with Cockett-Cruttwell tangent structure. The paper defines the INC endofunctor, which iterates the tangent lift, producing a tower $D,TD,T^2D, \cdots$ of factorization problems. ML is thereby formulated as the search for a coalgebraic fixed point where successive tangent unfoldings stabilize ($νT_{\mbox{INC}}$). Using the Aczel--Mendler theorem, we prove existence of a final INC coalgebra whenever $T_{\mbox{INC}}$ admits a set-based class realization that creates its final carrier. A detailed experimental evaluation of LINCS is underway in a number of concrete ML settings, including deep learning, large language models, and reinforcement learning, and is described in companion papers.

We study the problem of learning a structured approximation (low-rank, sparse, banded, etc.) to an unknown matrix $A$ given access to matrix-vector product (matvec) queries of the form $x \rightarrow Ax$ and $x \rightarrow A^Tx$. This problem is of central importance to algorithms across scientific computing and machine learning, with applications to fast multiplication and inversion for structured matrices, building preconditioners for first-order optimization, and as a model for differential operator learning. Prior work focuses on obtaining query complexity upper and lower bounds for learning specific structured matrix families that commonly arise in applications. We initiate the study of the problem in greater generality, aiming to understand the query complexity of learning approximations from general matrix families. Our main result focuses on finding a near-optimal approximation to $A$ from any finite-sized family of matrices, $\mathcal{F}$. Standard results from matrix sketching show that $O(\log|\mathcal{F}|)$ matvec queries suffice in this setting. This bound can also be achieved, and is optimal, for vector-matrix-vector queries of the form $x,y\rightarrow x^TAy$, which have been widely studied in work on rank-$1$ matrix sensing. Surprisingly, we show that, in the matvec model, it is possible to obtain a nearly quadratic improvement in complexity, to $\tilde{O}(\sqrt{\log|\mathcal{F}|})$. Further, we prove that this bound is tight up to log-log factors.Via covering number arguments, our result extends to well-studied infinite families. As an example, we establish that a near-optimal approximation from any \emph{linear matrix family} of dimension $q$ can be learned with $\tilde{O}(\sqrt{q})$ matvec queries, improving on an $O(q)$ bound achievable via sketching techniques and vector-matrix-vector queries.

Mobile crowdsensing (MC) recruits mobile users to perform sensing tasks using their smartphones, enabling large-scale applications such as traffic monitoring and environmental sensing. A fundamental challenge is online worker recruitment under uncertainty, where the platform must learn workers' sensing performance while operating with a limited budget. Existing learning-based MC recruitment methods typically assume that each worker's sensing quality is stationary with a fixed mean over time. In practice, however, worker performance often improves with experience and eventually stabilizes, while the incurred sensing cost can be unknown in advance due to time-varying device and context states. In this paper, we study a budget-constrained online recruitment problem in which the platform selects one worker in each round, observes the sensing quality and incurred cost, where the expected sensing quality of each worker increases with experience and eventually converges to a plateau, and repeats until the budget is exhausted. We formulate this problem as a structured bandit model where each worker's expected reward evolves according to an unknown increasing-then-converging function of its participation count, and each worker has an unknown expected cost. We develop a cost-aware online learning framework that jointly learns evolving reward trajectories and heterogeneous costs, detects performance saturation, and allocates the limited budget to maximize long-term sensing utility. We provide theoretical performance guarantees and validate the proposed approach through extensive experiments, demonstrating consistent improvements over baselines that ignore experience-driven dynamics or assume known costs.

Automated essay scoring (AES) research has largely focused on cross-prompt generalization, where essays from unseen prompts are scored while the scoring criteria are typically held constant. In practice, however, educators may revise or even introduce new rubrics in their scoring task, to evaluate different aspects of essays. We study cross-rubric generalization: training on essays labeled under one set of rubrics and evaluating on previously unseen rubrics, which target different aspects of the essay. We use a Large Language Model (LLM) fine-tuning framework with two components: rubric-agnostic intermediate representations, called traits, and target-essay supervision under seen rubrics during training. On an AES dataset augmented with multiple rubric-defined labels of student critical thinking skills, we find that traits improve macro F1 by 5.0% over a baseline without traits in the hardest setting, where both target rubrics and target essays are unseen during training. We further find that increasing target-essay supervision improves performance, with our best fine-tuned open-source Llama-based model outperforming GPT-5-mini prompting by 2.1% macro F1 and trailing GPT-5 by 1.9%. These results show that trait-based intermediate structure and controlled supervision improve generalization to unseen rubrics.

Large language models are increasingly used in scientific writing, yet they can fabricate citation-shaped references that appear plausible but fail bibliographic verification. Existing detectors often reduce verification to binary found/not-found decisions and rely on brittle parsing or incomplete retrieval, offering little field-level signal to auditors. We reframe citation hallucination detection as taxonomy-aligned field-level adjudication and introduce a 12-code taxonomy spanning Real, Potential, and Hallucinated citations. Based on this taxonomy, we build CiteTracer, a cascading multi-agent detector that extracts structured citations from PDF and BibTeX, retrieves evidence through cache lookup, URL fetch, scholar connectors, and web search, applies deterministic field matching, and routes ambiguous cases to class-specialist judgers. We release a benchmark of 2,450 synthetic citations built from real seeds with controlled LLM mutations, paired with 957 real-world fabricated citations drawn from ICLR 2026 and an anonymous conference desk-rejected submissions. CiteTracer reaches 97.1% accuracy on the synthetic benchmark, with class-level F1 scores of 97.0, 95.8, and 98.5 for Real, Potential, and Hallucinated, respectively, and detects 97.1% of fabrications on the real-world set without abstaining. Code: https://github.com/aaFrostnova/CiteTracer.

While diffusion models achieve state-of-the-art image quality for text-to-image (T2I) generation, recent work has demonstrated that they suffer from sample diversity collapse. In this work, we investigate whether autoregressive (AR) image generation models can push the Pareto frontier between image quality and sample diversity. With recent advances in quality and efficiency, AR models have emerged as a viable alternative to diffusion-based image generation. Beyond enabling new use cases such as interleaved image-text generation, their sequential generation process makes them compatible with a wide range of token-based decoding strategies originally developed to improve diversity in text generation. Motivated by the potential of a better diversity-quality tradeoff in the AR paradigm, we present the first systematic study of sample diversity in AR image generation models. We show that two key properties of AR image generation, persistently high token-level entropy and substantial redundancy in visual token spaces, limit the effectiveness of existing token-level decoding methods for diversity enhancement. We therefore propose $p$-less cluster, a new decoding strategy that performs entropy-based truncation sampling at cluster level rather than at token level. We evaluate our approach and baseline decoding methods across four autoregressive T2I models and two datasets using a comprehensive suite of metrics spanning image quality, prompt alignment, and diversity. Our results show that $p$-less cluster unlocks the greatest diversity across most evaluated autoregressive T2I models and datasets while maintaining image quality and prompt alignment.

Agentic retrieval-augmented generation (RAG) extends static RAG by allowing language models to iteratively reason, generate search queries, retrieve evidence, and predict answers. However, it remains challenging for models to decide when to retrieve, whether to use lexical matching or semantic similarity, and how to control context granularity to prevent irrelevant tokens from interfering with agent reasoning. In this paper, we introduce GRASP, a reinforcement learning (RL) framework for training agents to adaptively coordinate complementary retrieval tools during multi-step reasoning. GRASP provides the agent with semantic search, keyword search, and paragraph-reading actions, enabling it to retrieve sentence-level evidence and expand further context only when needed. We train the policy with a reward that jointly accounts for answer accuracy, grounded reading, complementary search, and turn efficiency. Experiments on multi-hop reasoning benchmarks show that GRASP improves both retrieval recall and downstream question answering performance compared with single-step retrieval, prompting-based agentic RAG, and RL-based retrieval baselines. Qualitative and ablation analyses show that the learned policy develops interpretable skimming and scanning behavior: it uses semantic search for broad exploration, paragraph reading for local verification, and keyword search for entity-specific evidence. These results suggest that learning to coordinate retrieval signals and context granularity is critical for agent's correct reasoning.

Agentic systems increasingly improve themselves by editing skills: prompts, rubrics, plans, tool contracts, examples, validators, and traces. Skill edits are not independent coordinates in a vector space: they are local repairs to structured artifacts whose effects are observed only after rollout, validation, and critique. Distinct edits can have the same immediate visible effect while differing in routing context, template state, guardrail scope, or future composability. The order of edits can matter as well: repairing a schema before a normalization rule need not be equivalent to applying the same edits in the reverse order. This paper introduces a new framework for skill optimization called LASKO, for Lie Algebroid SKill Optimization. LASKO models typed, anchored Markdown skills as the base category and available edit policies as sections of a controlled Lie algebroid with anchor $ρ$. The anchor maps an edit policy to its visible Markdown effect; the kernel $\ker(ρ)$ represents latent template, routing, or implementation structure; and the algebroid bracket measures noncommuting edit composition. As shown in the paper, LASKO achieves order-of-magnitude speedups in skill optimization in our preliminary benchmark results, primarily because it substitutes inexpensive Lie-bracket screening tests that run in microseconds, before investing in expensive validations that require running large language models. On a causal extraction from natural language task, LASKO achieved a speedup of almost $15 \times$ compared to a brute-force approach that validated all edits by running them through a DeepSeek V3.1 4-bit model with 671B parameters.

Discrete diffusion language models (dLLMs) generate text by iteratively denoising a masked sequence. However, standard dLLMs condition each denoising step solely on the current hard-masked sequence, while intermediate continuous representations are discarded after sampling and remasking. We term this bottleneck the \textbf{Information Island} issue: continuous information remains isolated within individual denoising steps and fails to propagate across the trajectory. This bottleneck is especially harmful for reasoning, which requires intermediate reasoning state to be preserved and updated across many denoising steps. To address this limitation, we introduce \textbf{MetaState}, a lightweight recurrent augmentation that equips a frozen dLLM backbone with persistent, fixed-size working memory. MetaState comprises three modules with a shared time conditioner: a cross-attention \textbf{Mixer} that reads backbone activations into memory slots, a GRU-style \textbf{Updater} that integrates information across steps, and a cross-attention \textbf{Injector} that writes the updated memory back into the backbone. We train these modules with a dedicated $K$-step unrolling pipeline to learn multi-step dynamics. MetaState adds only ${\sim}0.6\%$ trainable parameters while keeping the backbone frozen, and consistently improves reasoning performance over frozen baselines on mathematical reasoning and code generation benchmarks, with an average gain of 4.5 percentage points across all evaluations.

Direct Preference Optimization (DPO) is an effective framework for aligning large language models with human preferences, but it struggles with complex reasoning tasks. DPO optimizes for the likelihood of generating preferred over dispreferred responses in their entirety and lacks the granularity to provide feedback on subsections of many-step solutions typical of reasoning tasks. Existing methods excel at either stable preference learning (e.g., DPO variants like KTO and RSO) or structured reasoning (e.g., ReMA's multi-agent RL framework, Tree of Thoughts), but fail to merge these complementary strengths. We propose HiPO (Hierarchical Preference Optimization), an extension of DPO that separates responses into reasoning segments (query clarification and context, reasoning steps, and answer) and computes loss as a weighted sum of the DPO loss for each segment. Our approach enables segment-specific training while maintaining DPO's computational efficiency and training stability. We demonstrate that for multiple 7B LLMs fine-tuned using HiPO and DPO on the Math Stack Exchange preference dataset, the models trained with HiPO outperform the others on a variety of common math benchmarks and achieve greater organization, logical flow, and consistency as measured by GPT-4.1.

Understanding and forecasting audience reactions to video content are crucial for improving content creation, recommendation systems, and media analysis. To enable audience reaction prediction and other content engagement applications, we introduce $\textbf{Video2Reaction}$, a multimodal dataset that maps short movie segments to a distribution of $\textit{induced emotions}$ of viewers in the wild, as expressed through social media. $\textbf{Video2Reaction}$ spans more than 10,000 videos and serves as a reliable benchmark as well as a training resource for audience reaction prediction. To enable cost-effective continuous annotations as reactions may change over time, we develop a two-stage multi-agent pipeline using only open-source LLMs, achieving 86% correctness under blind human verification despite the inherently noisy and subjective nature of the task. We establish the first benchmark for video-to-reaction-distribution prediction in the wild and show that pretrained foundation video models fail in zero-shot settings, while finetuning transforms them into state-of-the-art predictors capable of modeling both full reaction distributions and dominant responses from video alone. However, the task remains challenging: even the strongest methods achieve only 77% Top-3 F1 in dominant reaction prediction (LLaVA-Next), highlighting a substantial gap in modeling collective audience reaction. \modification{Dataset and code are available at our project page: https://information-fusion-lab-umass.github.io/video2reaction-bench.github.io

Mechanistic understanding and rational design of complex chemical systems depend on fast and accurate predictions of electronic structures beyond individual building blocks. However, if the system exceeds hundreds of atoms, first-principles quantum mechanical (QM) modeling becomes impractical. In this study, we developed FB-GNN-MBE by integrating a fragment-based graph neural network (FB-GNN) into the many-body expansion (MBE) theory and demonstrated its capacity to reproduce first-principles potential energy surfaces (PES) for hierarchically structured systems with manageable accuracy, complexity, and interpretability. Specifically, we divided the entire system into basic building blocks (fragments), evaluated their one-fragment energies using a QM model, and addressed many-fragment interactions using the structure-property relationships trained by FB-GNNs. Our investigation shows that FB-GNN-MBE achieves chemical accuracy in predicting two-body (2B) and three-body (3B) energies across water, phenol, and mixture benchmarks, as well as the one-dimensional dissociation curves of water and phenol dimers. To transfer the success of FB-GNN-MBE across various systems with minimal computational costs and data demands, we developed and validated a teacher-student learning protocol. A heavy-weight FB-GNN trained on a mixed-density water cluster ensemble (teacher) distills its learned knowledge and passes it to a light-weight GNN (student), which is later fine-tuned on a uniform-density (H2O)21 cluster ensemble. This transfer learning strategy resulted in efficient and accurate prediction of 2B and 3B energies for variously sized water clusters without retraining. Our transferable FB-GNN-MBE framework outperformed conventional non-FB-GNN-based models and provided a scalable and accurate route toward interaction energies of large molecular assemblies.

Deep Neural Networks (DNNs) have become ubiquitous for their ability to solve problems across various domains, including computer vision, natural language processing, and speech recognition. However, as their adoption grows, they face a range of security threats, such as model stealing, architecture extraction, and manipulation, which can compromise their integrity, privacy, and functionality. Past works have relied on complex, fine-grained, and time-series analysis to launch DNN model extraction attacks. These approaches require extensive amounts of data, which are often challenging to acquire and analyze effectively. This paper introduces InferNet, an attack method that leverages simple, non-intrusive, and coarse-grained system-level information to identify the underlying DNN architecture of a victim's application. By analyzing GPU kernel calls, memory events, and system-level metrics, InferNet fingerprints the DNN and infers its architecture with very high accuracy. It can predict the architecture family (e.g., Inception vs. BERT), as well as the architecture variant (e.g., InceptionV1 vs. InceptionV3). The evaluation results demonstrate the effectiveness of InferNet across AI/ML frameworks (TensorFlow, PyTorch), different DNN types (vision, LLMs), and hardware platforms (NVIDIA Tesla T4, NVIDIA Quadro RTX 8000). The results show that InferNet achieves 100% model extraction accuracy using only a partial GPU profile under various attack settings.

Inference-time alignment steers a frozen language model during decoding using auxiliary reward signals, avoiding the cost of repeated weight updates. However, existing inference-time alignment methods typically optimize a single scalar score, so explicit safety constraints must either be ignored or encoded through manually tuned penalties. We propose Lagrangian Reward Augmentation (LARA), a general inference-time alignment framework under safety constraints. Starting from a KL-regularized constrained objective with a reward model and a cost model, LARA dualizes the constraint and reduces the optimization problem to a one-dimensional convex problem over a nonnegative dual variable. Estimated on a small calibration set, this dual variable defines an augmented reward that can be used as a drop-in scoring signal within existing inference-time alignment methods. For sequence-level sampling methods, such as Best-of-N reranking, the calibrated dual variable corresponds to the solution of the expected-cost constrained problem. For token-level reward-guided decoding methods, the same construction yields a principled dual-calibrated heuristic rather than an exact constrained-policy guarantee. We evaluate LARA on both sequence-level and token-level inference-time alignment methods, and find that LARA improves the helpfulness-harmlessness tradeoff, with Best-of-N achieving the best performance among inference-time methods, approaching finetuning-based direct alignment baselines.

Large language models are inconsistent: varying prompts or including unrelated information can lead to unexpected changes in model outputs. The generator-validator (G-V) gap is one manifestation of this phenomenon, where LLMs generate responses that they then deem as invalid if re-queried to validate them. In this work, we introduce a new formulation of G-V consistency that involves a principled correction for utterance frequency. Specifically, generators often assign low likelihood to valid strings simply because those strings are a priori unlikely, which makes naive notions of G-V consistency unworkable. We show that under a natural model of rational agents answering questions with multiple answers, consistency of the validator with a frequency-corrected generator score emerges naturally. Our method, \emph{\FCPAname} (\FCPA), is a training objective implementing frequency-corrected G-V consistency for real-world LLMs. Our experimental results show that training with \FCPA{} substantially improves both G-V consistency and generator performance over prior methods, with gains of up to $+27$pp in Pearson correlation on IFEval and HumanEval, while preserving validator quality across all evaluated tasks.

Margin-based optimization is fundamental to improving generalization and robustness in classification tasks. In the context of reward model learning from preferences within Reinforcement Learning from Human Feedback (RLHF), existing methods typically rely on no margins, fixed margins, or margins that are simplistic functions of preference ratings. However, such formulations often fail to account for the varying strengths of different preferences or they rely on noisy margin information derived from preference ratings. Furthermore, many existing methods that use adaptive margins assume access to accurate preference scores, which can be difficult for humans to provide reliably. We propose leveraging preferences over preferences, that is, annotations indicating which of two preferences reflects a stronger distinction, to infer adaptive margins on a per-datapoint basis. Such preference-over-preference annotations are general and can be incorporated into both standard RLHF reward modeling objectives and direct alignment losses. As a concrete instantiation, we introduce DPO-PoP, an extension to Direct Preference Optimization (DPO) that incorporates adaptive margins from preference-over-preference supervision, enabling improved discriminative and generative performance. Additionally, we show a tradeoff between discriminative and generative performance and propose two sampling strategies for gathering preference-over-preference labels to navigate it.

AI agents that autonomously execute tool calls on a user's behalf raise pressing questions about permission management: what role could users play, and what role should they play? Despite many proposed approaches, the user's role in agentic permission management remains under explored. We introduce Janus, a playground system for implementing and evaluating user-involved agentic permission management designs. Janus consists of two components: Janus-Core, a modular agentic system supporting a diverse spectrum of permission management designs, and Janus-Harness, an automated evaluation framework. Grounded in a conceptual model that identifies key design axes for user involvement, we implement six permission assistants spanning the design space and evaluate them across three scenarios and three synthetic responders. We demonstrate that user input is critical and can significantly strengthen privacy and security, that AI augmentation of user decisions can help reduce cognitive load, and that realistic user behavior including permission fatigue must be accounted for in system design. No single design performs optimally across all contexts, motivating a more principled and context-sensitive approach to deploying permission assistants in agentic systems. Janus is publicly available to support future investigation into this dimension of agentic system design.

Creativity is often seen as a hallmark of human intelligence. While large language models(LLMs) are increasingly perceived as generating creative text, there is still no cross-domain and scalable framework to evaluate their creativity across diverse scenarios. Existing methods of LLM creativity evaluation either heavily rely on humans, limiting speed and scalability, or are fragmented across different domains and different definitions of creativity. To address this gap, we propose CreativityPrism, an evaluation and analysis framework that consolidates eight tasks from three domains: divergent thinking, creative writing, and logical reasoning, into a taxonomy of creativity that emphasizes three dimensions: quality, novelty, and diversity of LLM generations. The framework is designed to be scalable with reliable automatic evaluation judges that have been validated against human annotations. We evaluate 17 state-of-the-art (SoTA) LLMs on CreativityPrism and find that while frontier-scale LLMs dominate creative writing and logical reasoning tasks by a .10 (or 15%) lead over locally-deployable open models, they offer no significant advantage in divergent thinking, a domain much less explored in existing post-training regimes. Our analysis also shows that high performance in one creative dimension or domain rarely generalizes to others; specifically, novelty metrics often show weak or negative correlations with other metrics. This fragmentation confirms that a cross-domain, multi-dimensional framework like CreativityPrism is essential for any meaningful assessment of LLM creativity.

Creating and editing slides is a rich, multimodal activity that is ubiquitous in professional and educational settings, making it an ideal testbed for real-world computer-use agents. Microsoft PowerPoint is among the most widely adopted and feature-rich environments for presentation creation. We introduce PPT-Eval, a benchmark of 120 PowerPoint tasks across 12 files that cover both content creation and presentation editing scenarios, organized by difficulty. A central challenge in this domain is evaluation: tasks are complex, multimodal, and often admit many valid solutions. Moreover, today's agents frequently make only partial progress, which binary success metrics fail to capture. To address this, we design a robust evaluation framework to help create task-specific rubrics for PowerPoint tasks, taking inspiration from and building on past works for rubric-based evaluation. These rubrics award partial credit for intermediate steps, penalize unnecessary changes and poor aesthetics, and provide natural language feedback. This nuanced approach proves highly effective, achieving a Kendall's τ-b correlation of 0.77 with human judgments. We find that existing frontier agents still struggle with solving PowerPoint tasks, with strong models like Claude-4.5-Opus achieving only a 45% success rate and an average partial score of 57%. The benchmark is located at: https://microsoft.github.io/ppteval.

Knowledge Tracing (KT) is important for personalized education but traditionally suffers from two key limitations: a reliance on shallow ID-based representations that neglect semantic depth and a restriction to single-granularity mastery estimation that overlooks hierarchical knowledge dependencies. To address these challenges, we propose MOSAIC (Multi-granularity Online Semantic AI for Collaborative Knowledge), a novel framework that orchestrates LLM-driven semantic alignment with sequential modeling. Unlike methods that use LLMs solely as predictors, MOSAIC leverages a frozen LLM to generate dynamic, context-aware embeddings and hierarchical prediction prompts, explicitly capturing collaborative signals and peer interactions. Furthermore, we introduce a cross-granularity consistency objective that jointly regularizes mastery estimation across concept, topic-cluster, and global proficiency levels. Extensive experiments on ASSISTments, EdNet, and a newly collected large-scale MOOC dataset demonstrate that MOSAIC establishes new state-of-the-art results. Specifically, our method achieves AUC improvements of up to 3.4\% and Accuracy gains of up to 2.5 \% across all benchmarks. Notably, MOSAIC exhibits superior robustness in collaboration-rich environments and long-sequence scenarios (AUC 0.862 on MOOC), offering both high predictive precision and semantically grounded interpretability.

We introduce a novel Wasserstein-1 ($W_1$) path-space divergence for stochastic and deterministic dynamics and establish a Wasserstein Uncertainty Propagation (WUP) theorem that bounds the $W_1$ distance between terminal distributions by the proposed divergence, equivalently characterized by a weighted $L^2$ discrepancy between the underlying drifts and the $W_1$ distance between their initial measures. A key ingredient is a probabilistic framework combining adjoint Feynman-Kac representations with synchronous coupling (and reflection coupling on bounded domains), yielding Wasserstein stability estimates beyond existing PDE- and Girsanov-based approaches. The framework accommodates time-varying and possibly degenerate diffusion coefficients, empirical and singular measures, and remains valid in the deterministic limit of flow matching. Unlike KL-based uncertainty quantification bounds, it does not require absolute continuity of path measures and therefore remains well-defined in singular settings. As consequences of the WUP theorem, we derive $W_1$ robustness and generalization bounds for score-based generative models and flow matching at both population and finite-sample levels. We further specialize the framework to group-symmetric targets, providing the first error analysis of equivariant flow-based models and the first quantitative comparison between data augmentation and equivariant inductive bias. Our analysis identifies a symmetry-aware Wasserstein path-space divergence that quantifies the model-form error induced by non-equivariant parametrizations. We prove that this error cannot be removed by additional data or training and vanishes only under equivariant architectures, establishing a precise theoretical advantage of equivariant inductive bias over data augmentation. Numerical experiments on group-symmetric Gaussian mixtures corroborate the theory.

We propose a novel deterministic sampling method, EVI-MMD, to approximate a target distribution $ρ^*$ by minimizing the kernel discrepancy, also known as the Maximum Mean Discrepancy (MMD). Leveraging the energetic variational inference framework (Wang et al., 2021), we transform the MMD minimization problem into solving a dynamic system of Ordinary Differential Equations (ODEs) for particles. The implicit Euler scheme is employed to solve the ODE system, leading to a proximal minimization problem at each iteration, which is efficiently addressed using optimization algorithms such as L-BFGS. A key innovation of our method is a dynamic bandwidth selection strategy for the Gaussian kernel, which, although heuristic at this stage, represents a meaningful step toward addressing a long-standing challenge in kernel-based methods. Comprehensive numerical experiments demonstrate that this adaptive bandwidth significantly enhances the performance of EVI-MMD. We apply the EVI-MMD algorithm to two types of sampling problems: (1) when the target distribution is fully specified by a density function, and (2) the ``two-sample problem,'' where only training data are available. In the latter case, EVI-MMD serves as a generative model, producing new samples that faithfully replicate the distribution of the training data. With carefully tuned parameters, EVI-MMD outperforms several existing methods in both scenarios.

Embodied agents operating in decentralized and partially observable environments have attracted growing attention in recent years. However, existing large language model (LLM)-based agents often exhibit behaviors that are misaligned with their partners or inconsistent with the environment state, leading to inefficient cooperation and poor task success. To address this challenge, we propose a novel framework, Learning Laws of Cooperation (LLawCo), that enables embodied agents to autonomously align with both their partners and task objectives. Our framework allows agents to reflect on past failures to extract misaligned behavioral patterns, which are used to derive high-level behavioral laws, such as "Talk when necessary" and "Wait for partner." These laws are explicitly incorporated into the agents' chains of thought via supervised fine-tuning, aligning their reasoning with task requirements and the behavior of other agents. To evaluate our approach, we introduce PARTNR-Dialog, a large-scale multi-agent communicative and cooperative planning benchmark built on the PARTNR environment. Experiments on existing tasks and our new benchmark demonstrate significant improvements in cooperative efficiency and task success rates. Across four backbone LLMs, our method achieves average success rate improvements of 4.5% on the PARTNR-Dialog benchmark and 6.8% on the TDW-MAT benchmark over state-of-the-art open-source communicative agent frameworks. See the LLawCo project page for details: https://www.merl.com/research/highlights/LLawCo

We introduce a categorical framework called ODYSSEY for constructing verifiable, local truth-preserving foundation models as compositions of foundries: building-block architectural components that specify a cover of local contexts, local representation families, restriction maps, gluing rules, obstruction policies, update obligations, and human-facing views. A foundry is an organized sheaf of knowledge that carries within it an argumentation component. Concrete foundries are built from generic foundries such as evidence/argument, operational decision, institutional/financial, market meaning, scientific challenge, research-program, assistant-build, and evaluation-harness foundries. Universal Foundry Learning (UFL) formalizes foundry construction as a composition of left and right Kan extensions, with left Kan extension rolling local artifacts into candidate foundries and right Kan extension enforcing the restriction, gluing, obstruction, and argumentation conditions required for promotion. Foundry SQL (FSQL) is a small typed query surface for slicing maintained foundry artifacts that uses TICKET (Topos Integration using Causal Kan Extension Transformers) certification for admitting external or pre-built models into durable ODYSSEY state. ODYSSEY is fully implemented and tested across a wide spectrum of concrete foundries, showing that the same categorical machinery supports domain construction, artifact replay, sheaf diagnostics, grounded Toulmin/local-LLM scrutiny, residual-obstruction ledgers, and optimized TICKET-compatible causal-claim extraction across heterogeneous sources. This paper is to be presented as a 2.5 hour tutorial at ICML 2026. The tutorial home page is at https://bit.ly/4ajS0nA.

Activation patching is the primary tool in mechanistic interpretability. It attributes causal responsibility for a model behavior to each of its individual components by estimating its natural indirect effect (NIE). Re-deriving the activation patching estimand from causal mediation analysis, we find that the NIE does not solely capture the causal effect through the specific component. It also contains interaction effects (INT) that measure how much the component's causal effect itself depends on the state of other components in the model. A natural response may be to try to eliminate INT by adjusting the estimator or unit of analysis, but each of these potential remedies has predictable failure modes. We demonstrate these failure modes in the GPT-2 IOI circuit; components whose causal importance is conditional on the state of other components are either invisible or artificially inflated, and INT variance explains the previously documented instability of faithfulness scores. We prove that INT scales with the distance between clean and patched component activations, is negligible when the model is locally affine, and decomposes combinatorially into pairwise and higher-order group interactions. Despite its inevitability, INT is not a nuisance to be eliminated, but rather a diagnostic for interpretability studies. Its individual and group-level magnitude and sign signal when causal conclusions are prompt-dependent, and when greedy NIE-based component ranking will miss mechanisms only discoverable through combinatorial search.