Group Relative Policy Optimization (GRPO) has proven to be an effective tool for post-training language models (LMs). However, AI systems are increasingly expressed as modular programs that mix together multiple LM calls with distinct prompt templates and other tools, and it is not clear how practitioners can best leverage online RL algorithms like GRPO to improve these systems. We begin to address this challenge by investigating whether it is possible to effectively instantiate GRPO for arbitrary multi-prompt programs and whether it can work robustly as an off-the-shelf optimizer for LM programs using the same abstractions and constraints typically involved for prompt optimization. Our main variant of multi-module GRPO constructs groups from module-level invocations, and we also consider trajectory-level grouping as another natural instantiation. We find for the first time that GRPO (and its multi-module counterpart) empirically composes well with automatic prompt optimization, and together they improve accuracy by 11% on average across classification, many-hop search, and privacy-preserving delegation tasks against the post-trained LM - with 5% gains against prompt optimization on its own. We open-source multi-module GRPO in the DSPy library at https://dspy.ai .

Large Language Models (LLMs) have achieved remarkable results on a range of standardized tests originally designed to assess human cognitive and psychological traits, such as intelligence and personality. While these results are often interpreted as strong evidence of human-like characteristics in LLMs, this paper argues that such interpretations constitute an ontological error. Human psychological and educational tests are theory-driven measurement instruments, calibrated to a specific human population. Applying these tests to non-human subjects without empirical validation, risks mischaracterizing what is being measured. Furthermore, a growing trend frames AI performance on benchmarks as measurements of traits such as ``intelligence'', despite known issues with validity, data contamination, cultural bias and sensitivity to superficial prompt changes. We argue that interpreting benchmark performance as measurements of human-like traits, lacks sufficient theoretical and empirical justification. This leads to our position: Stop Evaluating AI with Human Tests, Develop Principled, AI-specific Tests instead. We call for the development of principled, AI-specific evaluation frameworks tailored to AI systems. Such frameworks might build on existing frameworks for constructing and validating psychometrics tests, or could be created entirely from scratch to fit the unique context of AI.

We propose a new volumetric grasp model that is equivariant to rotations around the vertical axis, leading to a significant improvement in sampling efficiency. Our model employs a tri-plane volumetric feature representation -- i.e., the projection of 3D features onto three canonical planes. We introduce a novel tri-plane feature design in which features on the horizontal plane are \emph{equivariant} to $90^\circ$ rotations, while the \emph{sum} of features from the other two planes remains \emph{invariant} to reflections induced by the same transformations. We further develop equivariant adaptations of two state-of-the-art volumetric grasp planners, GIGA and IGD. Specifically, we derive a new equivariant formulation of IGD's deformable attention mechanism and propose an equivariant generative model of grasp orientations based on flow matching. We provide a detailed analytical justification of the proposed equivariance properties and validate our approach through extensive simulated and real-world experiments. Our results demonstrate that the proposed projection-based design reduces both computational and memory costs. Moreover, the equivariant grasp models built on top of our tri-plane features consistently outperform their non-equivariant counterparts, achieving higher performance within a real-time cost constraint. Video and code can be viewed in: https://mousecpn.github.io/evg-page/

Creating an immersive and interactive theatrical experience is a long-term goal in the field of interactive narrative. The emergence of large language models (LLMs) provides a new path to achieve this goal. However, existing drama generation methods often produce LLMs that lack initiative and cannot interact with the physical scene, while typically requiring detailed input that diminishes the immersion of live performance. To address these challenges, we propose HAMLET, a hierarchical adaptive multi-agent framework focused on drama creation and real-time online performance. Given a simple topic, the framework initially generates a narrative blueprint to guide the subsequent improvisational performance. During online performance, each actor is equipped with an adaptive reasoning module that enables decision-making based on their personas, memories, goals during complex group chat scenarios. Beyond dialogue, actor agents engage in embodied interactions by changing the state of scene props through actions such as opening a letter or picking up a weapon, which are broadcast to update the global environmental context. To objectively assess the quality of live embodied theatrics, we establish a comprehensive evaluation method and introduce HAMLETJudge, a specialized critic model for automated evaluation. Experimental results demonstrate that HAMLET excels in creating expressive, coherent, and physically interactive theatrical experiences in an autonomous manner.

The complexity and interconnectivity of entities involved in money laundering demand investigative reasoning over graph-structured data. This paper explores the use of large language models (LLMs) as reasoning engines over localized subgraphs extracted from a financial knowledge graph. We propose a lightweight pipeline that retrieves k-hop neighborhoods around entities of interest, serializes them into structured text, and prompts an LLM via few-shot in-context learning to assess suspiciousness and generate justifications. Using synthetic anti-money laundering (AML) scenarios that reflect common laundering behaviors, we show that LLMs can emulate analyst-style logic, highlight red flags, and provide coherent explanations. While this study is exploratory, it illustrates the potential of LLM-based graph reasoning in AML and lays groundwork for explainable, language-driven financial crime analytics.

We present Q-chunking, a simple yet effective recipe for improving reinforcement learning (RL) algorithms for long-horizon, sparse-reward tasks. Our recipe is designed for the offline-to-online RL setting, where the goal is to leverage an offline prior dataset to maximize the sample-efficiency of online learning. Effective exploration and sample-efficient learning remain central challenges in this setting, as it is not obvious how the offline data should be utilized to acquire a good exploratory policy. Our key insight is that action chunking, a technique popularized in imitation learning where sequences of future actions are predicted rather than a single action at each timestep, can be applied to temporal difference (TD)-based RL methods to mitigate the exploration challenge. Q-chunking adopts action chunking by directly running RL in a 'chunked' action space, enabling the agent to (1) leverage temporally consistent behaviors from offline data for more effective online exploration and (2) use unbiased $n$-step backups for more stable and efficient TD learning. Our experimental results demonstrate that Q-chunking exhibits strong offline performance and online sample efficiency, outperforming prior best offline-to-online methods on a range of long-horizon, sparse-reward manipulation tasks.

This paper studies causal discovery in irregularly sampled time series-a key challenge in risk-sensitive domains like finance, healthcare, and climate science, where missing data and inconsistent sampling frequencies distort causal mechanisms. The main challenge comes from the interdependence between missing data imputation and causal structure recovery: errors in imputation and structure learning can reinforce each other, leading to an inaccurate causal graph. Existing methods either impute first and then discover, or jointly optimize both via neural representation learning, but lack explicit mechanisms to ensure mutual consistency of imputation and structure learning. We address this challenge with ReTimeCausal, an EM-based framework that alternates between imputation and structure learning, which encourages structural consistency throughout the optimization process. Our framework provides theoretical consistency guarantees for structure recovery and extends classical results to settings with irregular sampling and high missingness. ReTimeCausal combines kernel-based sparse regression and structural constraints in an alternating process that updates the completed data and the causal graph in turn. Experiments on synthetic and real-world datasets show that ReTimeCausal is more effective than existing methods under challenging irregular sampling and missing data.

We introduce SLR, an end-to-end framework for systematic evaluation and training of Large Language Models (LLMs) via Scalable Logical Reasoning. Given a user's task specification, SLR automatically synthesizes (i) an instruction prompt for an inductive reasoning task, (ii) a validation program, executable on model outputs to provide verifiable rewards, and (iii) the latent ground-truth rule. This process is fully automated, scalable, requires no human annotations, and offers precise control over task difficulty. Using SLR, we create SLR-Bench, a benchmark comprising 19k prompts organized into 20 curriculum levels that progressively increase in relational, arithmetic, and recursive complexity. Large-scale evaluation reveals that contemporary LLMs readily produce syntactically valid rules, yet often fail at correct logical inference. Recent reasoning LLMs demonstrate improved performance but incur very high test-time computation, with costs exceeding $300 for just 1,000 prompts. Finally, curriculum learning via SLR doubles Llama-3-8B accuracy on SLR-Bench, achieving parity with Gemini-Flash-Thinking at a fraction of computational cost. Moreover, these reasoning capabilities generalize to a wide range of established benchmarks, underscoring the effectiveness of SLR for downstream reasoning.

Risk stratification is a key tool in clinical decision-making, yet current approaches often fail to translate sophisticated survival analysis into actionable clinical criteria. We present a novel method for unsupervised machine learning that directly optimizes for survival heterogeneity across patient clusters through a differentiable adaptation of the multivariate logrank statistic. Unlike most existing methods that rely on proxy metrics, our approach represents novel methodology for training any neural network architecture on any data modality to identify prognostically distinct patient groups. We thoroughly evaluate the method in simulation experiments and demonstrate its utility in practice by applying it to two distinct cancer types: analyzing laboratory parameters from multiple myeloma patients and computed tomography images from non-small cell lung cancer patients, identifying prognostically distinct patient subgroups with significantly different survival outcomes in both cases. Post-hoc explainability analyses uncover clinically meaningful features determining the group assignments which align well with established risk factors and thus lend strong weight to the methods utility. This pan-cancer, model-agnostic approach represents a valuable advancement in clinical risk stratification, enabling the discovery of novel prognostic signatures across diverse data types while providing interpretable results that promise to complement treatment personalization and clinical decision-making in oncology and beyond.

This paper introduces ChatbotManip, a novel dataset for studying manipulation in Chatbots. It contains simulated generated conversations between a chatbot and a (simulated) user, where the chatbot is explicitly asked to showcase manipulation tactics, persuade the user towards some goal, or simply be helpful. We consider a diverse set of chatbot manipulation contexts, from consumer and personal advice to citizen advice and controversial proposition argumentation. Each conversation is annotated by human annotators for both general manipulation and specific manipulation tactics. Our research reveals three key findings. First, Large Language Models (LLMs) can be manipulative when explicitly instructed, with annotators identifying manipulation in approximately 84\% of such conversations. Second, even when only instructed to be ``persuasive'' without explicit manipulation prompts, LLMs frequently default to controversial manipulative strategies, particularly gaslighting and fear enhancement. Third, small fine-tuned open source models, such as BERT+BiLSTM have a performance comparable to zero-shot classification with larger models like Gemini 2.5 pro in detecting manipulation, but are not yet reliable for real-world oversight. Our work provides important insights for AI safety research and highlights the need of addressing manipulation risks as LLMs are increasingly deployed in consumer-facing applications.

Recently, large language models (LLMs) have demonstrated strong performance, ranging from simple to complex tasks. However, while large models achieve remarkable results across diverse tasks, they often incur substantial monetary inference cost, making frequent use impractical for many applications. In contrast, small models are often freely available and easy to deploy locally, but their performance on complex tasks remains limited. This trade-off raises a natural question: how can small and large models efficiently collaborate to combine their complementary strengths? To bridge this trade-off, we propose COPE, a test-time collaboration framework. A planner model first generates a plan that serves as a lightweight intermediate that guides a downstream executor model. Small and large models take turns acting as planner and executor, exchanging plans in a multi-stage cascade to collaboratively solve tasks. Through comprehensive experiments on benchmarks spanning mathematical reasoning, code generation, open-ended tasks, and agent tasks, we demonstrate that COPE achieves performance comparable to large proprietary models, while drastically reducing the inference API cost. These results highlight planning as an effective prior for cost-efficient inference.

We present SDialog, an MIT-licensed open-source Python toolkit that unifies dialog generation, evaluation and mechanistic interpretability into a single end-to-end framework for building and analyzing LLM-based conversational agents. Built around a standardized Dialog representation, SDialog provides: (1) persona-driven multi-agent simulation with composable orchestration for controlled, synthetic dialog generation, (2) comprehensive evaluation combining linguistic metrics, LLM-as-a-judge and functional correctness validators, (3) mechanistic interpretability tools for activation inspection and steering via feature ablation and induction, and (4) audio generation with full acoustic simulation including 3D room modeling and microphone effects. The toolkit integrates with all major LLM backends, enabling mixed-backend experiments under a unified API. By coupling generation, evaluation, and interpretability in a dialog-centric architecture, SDialog enables researchers to build, benchmark and understand conversational systems more systematically.

Electroencephalography (EEG) provides real-time insights into brain activity and supports diverse applications in neuroscience. While EEG foundation models (EFMs) have emerged to address the scalability issues of task-specific models, current approaches still yield clinically uninterpretable and weakly discriminative representations, inefficiently capturing global dependencies and neglecting important local neural events. We present CodeBrain, a two-stage EFM designed to fill this gap. In the first stage, we introduce the TFDual-Tokenizer, which decouples heterogeneous temporal and frequency EEG signals into discrete tokens, quadratically expanding the representation space to enhance discriminative power and offering domain-specific representation-level interpretability by suggesting potential links to neural events and spectral rhythms. In the second stage, we propose the multi-scale EEGSSM architecture, which combines structured global convolution with sliding window attention to efficiently capture both sparse long-range and local dependencies, reflecting the brain's small-world topology. Pretrained on the largest public EEG corpus, CodeBrain achieves strong generalization across eight downstream tasks and ten datasets under distribution shifts, supported by comprehensive ablations, scaling-law analyzes, and interpretability evaluations. The code and the pretrained weights are available at https://github.com/jingyingma01/CodeBrain.

We introduce EconWebArena, a benchmark for evaluating autonomous agents on complex, multimodal economic tasks in realistic web environments. The benchmark comprises 360 curated tasks from 82 authoritative websites spanning domains such as macroeconomics, labor, finance, trade, and public policy. Each task challenges agents to navigate live websites, interpret structured and visual content, interact with real interfaces, and extract precise, time-sensitive data through multi-step workflows. We construct the benchmark by prompting multiple large language models (LLMs) to generate candidate tasks, followed by rigorous human curation to ensure clarity, feasibility, and source reliability. Unlike prior work, EconWebArena emphasizes fidelity to authoritative data sources and the need for grounded web-based economic reasoning. We evaluate a diverse set of state-of-the-art multimodal LLMs as web agents, analyze failure cases, and conduct ablation studies to assess the impact of visual grounding, plan-based reasoning, and interaction design. Our results reveal substantial performance gaps and highlight persistent challenges in grounding, navigation, and multimodal understanding, positioning EconWebArena as a rigorous testbed for economic web intelligence.

Transformer-based models have demonstrated remarkable reasoning abilities, but the mechanisms underlying relational reasoning remain poorly understood. We investigate how transformers perform \textit{transitive inference}, a classic relational reasoning behavior from psychology which elicits inference about indirectly related items (e.g., if $A > B$ and $B > C$, then $A > C$). We compare in-weights learning (IWL) and in-context learning (ICL) behaviors and mechanisms on these tasks, and fine profoundly different patterns of generalization. IWL models learn a linear embedding, which leads to transitive inference as well as other behavioral effects present in humans and animals. ICL models, in contrast, are capable of learning to generalize transitively, but only do so when it is necessitated by the training data, otherwise learning a match-and-copy strategy. Interestingly, pre-training ICL models on in-context linear regression tasks that provide them with a latent linear representation is sufficient to make the ICL behaviors and internal representations qualitatively and quantitatively more like IWL. In order to test whether the same inference patterns are present across in large language models, we leverage a congruency paradigm which allows us to differentially probe IWL and ICL generalization patterns without access to their training data. We indeed see IWL reasoning leads to more transitive generalization than ICL. Moreover, we find that prompting the ICL models to use a linear mental map led to increased transitive inference over different geometric prompts. Together, these results reveal that both the training regime and the geometric structure of induced representations critically determine transformers capacity for transitive inference.

This paper introduces an innovative error feedback framework designed to mitigate quantization noise in distributed graph filtering, where communications are constrained to quantized messages. It comes from error spectrum shaping techniques from state-space digital filters, and therefore establishes connections between quantized filtering processes over different domains. In contrast to existing error compensation methods, our framework quantitatively feeds back the quantization noise for exact compensation. We examine the framework under three key scenarios: (i) deterministic graph filtering, (ii) graph filtering over random graphs, and (iii) graph filtering with random node-asynchronous updates. Rigorous theoretical analysis demonstrates that the proposed framework significantly reduces the effect of quantization noise, and we provide closed-form solutions for the optimal error feedback coefficients. Moreover, this quantitative error feedback mechanism can be seamlessly integrated into communication-efficient decentralized optimization frameworks, enabling lower error floors. Numerical experiments validate the theoretical results, consistently showing that our method outperforms conventional quantization strategies in terms of both accuracy and robustness.

Steering vectors are a lightweight method for controlling text properties by adding a learned bias to language model activations at inference time. While predominantly studied for multiple-choice and toy tasks, their effectiveness in free-form generation remains largely unexplored. Moving "Beyond Multiple Choice," we evaluate steering vectors for controlling topical focus, sentiment, toxicity, and readability in abstractive summaries across the SAMSum, NEWTS, and arXiv datasets. We find that steering effectively controls targeted properties, but high steering strengths consistently induce degenerate repetition and factual hallucinations. Prompting alone preserves summary quality but offers weaker control. Combining both methods yields the strongest control and the most favorable efficacy-quality trade-off at moderate steering strengths. Our work demonstrates that steering vectors face a critical control-quality trade-off in free-form generation, and that hybrid approaches offer the best balance in practice.

While uncertainty estimation for graphs recently gained traction, most methods rely on homophily and deteriorate in heterophilic settings. We address this by analyzing message passing neural networks from an information-theoretic perspective and developing a suitable analog to data processing inequality to quantify information throughout the model's layers. In contrast to non-graph domains, information about the node-level prediction target can increase with model depth if a node's features are semantically different from its neighbors. Therefore, on heterophilic graphs, the latent embeddings of an MPNN each provide different information about the data distribution - different from homophilic settings. This reveals that considering all node representations simultaneously is a key design principle for epistemic uncertainty estimation on graphs beyond homophily. We empirically confirm this with a simple post-hoc density estimator on the joint node embedding space that provides state-of-the-art uncertainty on heterophilic graphs. At the same time, it matches prior work on homophilic graphs without explicitly exploiting homophily through post-processing.

Existing cyberbullying detection benchmarks were organized by the polarity of speech, such as "offensive" and "non-offensive", which were essentially hate speech detection. However, in the real world, cyberbullying often attracted widespread social attention through incidents. To address this problem, we propose a novel annotation method to construct a cyberbullying dataset that organized by incidents. The constructed CHNCI is the first Chinese cyberbullying incident detection dataset, which consists of 220,676 comments in 91 incidents. Specifically, we first combine three cyberbullying detection methods based on explanations generation as an ensemble method to generate the pseudo labels, and then let human annotators judge these labels. Then we propose the evaluation criteria for validating whether it constitutes a cyberbullying incident. Experimental results demonstrate that the constructed dataset can be a benchmark for the tasks of cyberbullying detection and incident prediction. To the best of our knowledge, this is the first study for the Chinese cyberbullying incident detection task.

Referring Multi-Object Tracking (RMOT) aims to track targets specified by language instructions. However, existing RMOT paradigms heavily rely on explicit visual-textual matching and consequently fail to generalize to complex instructions that require logical reasoning. To overcome this, we propose Reasoning-based Multi-Object Tracking (ReaMOT), a novel task that elevates tracking to a cognitive level, requiring models to infer and track specific targets satisfying implicit constraints via logical reasoning. To advance this field, we construct the ReaMOT Challenge, a comprehensive benchmark featuring a tailored metric suite and a large scale dataset. This dataset comprises 1,156 language instructions, 423,359 image language pairs, and 869 distinct video sequences systematically categorized into six distinct evaluation scenarios, with over 75\% of the instructions dedicated to High Level Reasoning. Furthermore, recognizing that traditional trackers lack cognitive capacity while direct application of Large Vision-Language Model (LVLM) yields severe temporal inconsistencies, we propose ReaTrack. Driven by the insight to decouple high-level cognitive localization from low-level physical motion continuity, this training-free framework dynamically aligns the semantic detections of a Thinking-variant LVLM with the robust motion priors of SAM2. Extensive experiments on the ReaMOT Challenge benchmark demonstrate that ReaTrack establishes a new leading performance standard. Notably, it achieves a more than threefold improvement in RHOTA on the High Level Reasoning subset. Our dataset and code will be available at https://github.com/chen-si-jia/ReaMOT.

Multi-modal object Re-IDentification (ReID) aims to obtain complete identity features across heterogeneous modalities. However, most existing methods rely on implicit feature fusion modules, making it difficult to model fine-grained recognition patterns under various challenges in real world. Benefiting from the powerful Multi-modal Large Language Models (MLLMs), the object appearances are effectively translated into descriptive captions. In this paper, we propose a reliable caption generation pipeline based on attribute confidence, which significantly reduces the unknown recognition rate of MLLMs and improves the quality of generated text. Additionally, to model diverse identity patterns, we propose a novel ReID framework, named NEXT, the Multi-grained Mixture of Experts via Text-Modulation for Multi-modal Object Re-Identification. Specifically, we decouple the recognition problem into semantic and structural branches to separately capture fine-grained appearance features and coarsegrained structure features. For semantic recognition, we first propose a Text-Modulated Semantic Experts (TMSE), which randomly samples high-quality captions to modulate experts capturing semantic features and mining inter-modality complementary cues. Second, to recognize structure features, we propose a Context-Shared Structure Experts (CSSE), which focuses on the holistic object structure and maintains identity structural consistency via a soft routing mechanism. Finally, we propose a Multi-Grained Features Aggregation (MGFA), which adopts a unified fusion strategy to effectively integrate multi-grained expert features into the final identity representations. Extensive experiments on two public person datasets and three vehicle datasets demonstrate the effectiveness of our method, showing that it significantly outperforms existing state-of-the-art methods.

Personalizing text-to-image diffusion models involves integrating novel visual concepts from a small set of reference images while retaining the model's original generative capabilities. However, this process often leads to overfitting, where the model ignores the user's prompt and merely replicates the reference images. We attribute this issue to a fundamental misalignment between the true goals of personalization, which are subject fidelity and text alignment, and the training objectives of existing methods that fail to enforce both objectives simultaneously. Specifically, prior approaches often overlook the need to explicitly preserve the pretrained model's output distribution, resulting in distributional drift that undermines diversity and coherence. To resolve these challenges, we introduce a Lipschitz-based regularization objective that constrains parameter updates during personalization, ensuring bounded deviation from the original distribution. This promotes consistency with the pretrained model's behavior while enabling accurate adaptation to new concepts. Furthermore, our method offers a computationally efficient alternative to commonly used, resource-intensive sampling techniques. Through extensive experiments across diverse diffusion model architectures, we demonstrate that our approach achieves superior performance in both quantitative metrics and qualitative evaluations, consistently excelling in visual fidelity and prompt adherence. We further support these findings with comprehensive analyses, including ablation studies and visualizations.

Active perception is a fundamental skill that enables us humans to deal with uncertainty in our inherently partially observable environment. For senses such as touch, where the information is sparse and local, active perception becomes crucial. In recent years, active perception has emerged as an important research domain in robotics. However, current methods are often bound to specific tasks or make strong assumptions, which limit their generality. To address this gap, this work introduces APPLE (Active Perception Policy Learning) - a novel framework that leverages reinforcement learning (RL) to address a range of different active perception problems. APPLE jointly trains a transformer-based perception module and decision-making policy with a unified optimization objective, learning how to actively gather information. By design, APPLE is not limited to a specific task and can, in principle, be applied to a wide range of active perception problems. We evaluate two variants of APPLE across different tasks, including tactile exploration problems from the Tactile MNIST benchmark. Experiments demonstrate the efficacy of APPLE, achieving high accuracies on both regression and classification tasks. These findings underscore the potential of APPLE as a versatile and general framework for advancing active perception in robotics. Project page: https://timschneider42.github.io/apple

Concept Bottleneck Models (CBMs) are neural networks designed to conjoin high performance with ante-hoc interpretability. CBMs work by first mapping inputs (e.g., images) to high-level concepts (e.g., visible objects and their properties) and then use these to solve a downstream task (e.g., tagging or scoring an image) in an interpretable manner. Their performance and interpretability, however, hinge on the quality of the concepts they learn. The go-to strategy for ensuring good quality concepts is to leverage expert annotations, which are expensive to collect and seldom available in applications. Researchers have recently addressed this issue by introducing "VLM-CBM" architectures that replace manual annotations with weak supervision from foundation models. It is however unclear what is the impact of doing so on the quality of the learned concepts. To answer this question, we put state-of-the-art VLM-CBMs to the test, analyzing their learned concepts empirically using a selection of significant metrics. Our results show that, depending on the task, VLM supervision can sensibly differ from expert annotations, and that concept accuracy and quality are not strongly correlated. Our code is available at https://github.com/debryu/CQA.

As the post-training of large language models (LLMs) advances from instruction-following to complex reasoning tasks, understanding how different data affect finetuning dynamics remains largely unexplored. In this paper, we present a spectral analysis of layer-wise gradients induced by low/high-quality instruction and reasoning data for LLM post-training. Our analysis reveals that widely-studied metrics for data evaluation, e.g., IFD, InsTag, Difficulty, and Reward, can be explained and unified by spectral properties computed from gradients' singular value decomposition (SVD). Specifically, higher-quality data are usually associated with lower nuclear norms and higher effective ranks. Notably, effective rank exhibits better robustness and resolution than nuclear norm in capturing subtle quality differences. For example, reasoning data achieves substantially higher effective ranks than instruction data, implying richer gradient structures on more complex tasks. Our experiments also highlight that models within the same family share similar gradient patterns regardless of their sizes, whereas different model families diverge significantly. Providing a unified view on the effects of data quality across instruction and reasoning data, this work illuminates the interplay between data quality and training stability, shedding novel insights into developing better data exploration strategies for post-training.

Recent advancements in Large Language Models (LLMs) and the proliferation of Text-Attributed Graphs (TAGs) across various domains have positioned LLM-enhanced TAG learning as a critical research area. By utilizing rich graph descriptions, this paradigm leverages LLMs to generate high-quality embeddings, thereby enhancing the representational capacity of Graph Neural Networks (GNNs). However, the field faces significant challenges: (1) the absence of a unified framework to systematize the diverse optimization perspectives arising from the complex interactions between LLMs and GNNs, and (2) the lack of a robust method capable of handling real-world TAGs, which often suffer from texts and edge sparsity, leading to suboptimal performance. To address these challenges, we propose UltraTAG, a unified pipeline for LLM-enhanced TAG learning. UltraTAG provides a unified comprehensive and domain-adaptive framework that not only organizes existing methodologies but also paves the way for future advancements in the field. Building on this framework, we propose UltraTAG-S, a robust instantiation of UltraTAG designed to tackle the inherent sparsity issues in real-world TAGs. UltraTAG-S employs LLM-based text propagation and text augmentation to mitigate text sparsity, while leveraging LLM-augmented node selection techniques based on PageRank and edge reconfiguration strategies to address edge sparsity. Our extensive experiments demonstrate that UltraTAG-S significantly outperforms existing baselines, achieving improvements of 2.12\% and 17.47\% in ideal and sparse settings, respectively. Moreover, as the data sparsity ratio increases, the performance improvement of UltraTAG-S also rises, which underscores the effectiveness and robustness of UltraTAG-S.

The Mixture of Experts (MoE) is an effective architecture for scaling large language models by leveraging sparse expert activation to balance performance and efficiency. However, under expert parallelism, MoE suffers from inference inefficiencies due to imbalanced token-to-expert assignment, where underloaded experts complete computations early but must wait for overloaded experts, leading to global delays. We define this phenomenon as the \textbf{\textit{Straggler Effect}}, as the most burdened experts dictate the overall inference latency. To address this, we first propose \textit{\textbf{Capacity-Aware Token Drop}}, which enforces expert capacity limits by discarding excess tokens from overloaded experts, effectively reducing load imbalance with minimal performance impact (e.g., $30\%$ speedup with only $0.9\%$ degradation on OLMoE). Next, given the presence of low-load experts remaining well below the capacity threshold, we introduce \textit{\textbf{Capacity-Aware Expanded Drop}}, which allows tokens to include additional local experts in their candidate set before enforcing strict local capacity constraints, thereby improving load balance and enhancing the utilization of underused experts. Extensive experiments on both language and multimodal MoE models demonstrate the effectiveness of our approach, yielding substantial gains in expert utilization, model performance, and inference efficiency, e.g., applying Expanded Drop to Mixtral-8$\times$7B-Instruct yields a {0.2\%} average performance improvement and a {1.85$\times$} inference speedup. The code is released at: https://github.com/CASE-Lab-UMD/Capacity-Aware-MoE.

Frontier models can be prompted or conditioned to do many tasks, but finding good prompts is not always easy, nor is understanding some performant prompts. We view prompting as finding the best conditioning sequence on a near-optimal sequence predictor. On numerous well-controlled experiments, we show that unintuitive optimal conditioning sequences can be better understood given the pretraining distribution, which is not usually available. Even using exhaustive search, reliably identifying optimal prompts for practical neural predictors can be surprisingly difficult. Popular prompting methods, such as using demonstrations from the targeted task, can be surprisingly suboptimal. Using the same empirical framework, we analyze optimal prompts on frontier models, revealing patterns similar to the binary examples and previous findings. Taken together, this work takes an initial step towards understanding optimal prompts, from a statistical and empirical perspective that complements research on frontier models.

Optimal transport (OT) is a widely used tool in machine learning, but computing high-accuracy solutions for large instances remains costly. Entropic regularization and the Sinkhorn algorithm improve scalability; however, when the regularization parameter is small, Sinkhorn convergence slows, and the iterates approach an entropic solution that remains separated from the true OT plan by an entropic-bias plateau. We introduce PINS (Proximal Iterations with sparse Newton and Sinkhorn), a two-loop solver designed to move beyond this plateau. The outer loop applies an entropic proximal-point method, solving the original OT problem through a sequence of entropic subproblems with shifted cost matrices. Each inner subproblem is then solved by a Sinkhorn warm-up followed by sparse-Newton refinement. We prove that PINS converges globally to an optimal solution of the unregularized OT problem and that the inner Hessian admits a sparsification at every outer iteration with a structure independent of the cost matrix. On synthetic and augmented-MNIST instances, PINS achieves much lower relative cost errors than Sinkhorn-type baselines, which stall at the entropic-bias plateau, and is $5$--$73\times$ faster than Sinkhorn with the same outer loop at matched accuracy. On large-scale DOTmark instances, a streaming implementation reduces peak memory by $24$--$54\%$ compared with the network-simplex linear programming (LP) solver and remains feasible under per-process memory budgets for which the LP solver fails.

Large Vision-Language Models (LVLMs) that incorporate visual models and large language models have achieved impressive results across cross-modal understanding and reasoning tasks. In recent years, person re-identification (ReID) has also started to explore cross-modal semantics to improve the accuracy of identity recognition. However, effectively utilizing LVLMs for ReID remains an open challenge. While LVLMs operate under a generative paradigm by predicting the next output word, ReID requires the extraction of discriminative identity features to match pedestrians across cameras. In this paper, we propose LVLM-ReID, a novel framework that harnesses the strengths of LVLMs to promote ReID. Specifically, we employ instructions to guide the LVLM in generating one semantic token that encapsulates key appearance semantics from the person image. This token is further refined through our Semantic-Guided Interaction (SGI) module, establishing a reciprocal interaction between the semantic token and visual tokens. Ultimately, the reinforced semantic token serves as the representation of pedestrian identity. Our framework integrates the semantic understanding and generation capabilities of LVLM into end-to-end ReID training, allowing LVLM to capture rich semantic cues during both training and inference. LVLM-ReID achieves competitive results on multiple benchmarks without additional image-text annotations, demonstrating the potential of LVLM-generated semantics to advance person ReID.