Self-play post-training methods has emerged as an effective approach for finetuning large language models and turn the weak language model into strong language model without preference data. However, the theoretical foundations for self-play finetuning remain underexplored. In this work, we tackle this by connecting self-play finetuning with adversarial imitation learning by formulating finetuning procedure as a min-max game between the model and a regularized implicit reward player parameterized by the model itself. This perspective unifies self-play imitation and general preference alignment within a common framework. Under this formulation, we present a game-theoretic analysis showing that the self-play finetuning will converge to it's equilibrium. Guided by this theoretical formulation, we propose a new self-play imitation finetuning algorithm based on the $χ^2$-divergence variational objective with bounded rewards and improved stability. Experiments on various of language model finetuning tasks demonstrate consistent improvements over existing self-play methods and validate our theoretical insights.

While Large Language Models (LLMs) excel in language-based agentic tasks, their applicability to unseen, nonlinguistic environments (e.g., symbolic or spatial tasks) remains limited. Previous work attributes this performance gap to the mismatch between the pretraining distribution and the testing distribution. In this work, we demonstrate the primary bottleneck is the prohibitive cost of exploration: mastering these tasks requires extensive trial-and-error, which is computationally unsustainable for parameter-heavy LLMs operating in a high dimensional semantic space. To address this, we propose SCOUT (Sub-Scale Collaboration On Unseen Tasks), a novel framework that decouples exploration from exploitation. We employ lightweight "scouts" (e.g., small MLPs) to probe environmental dynamics at a speed and scale far exceeding LLMs. The collected trajectories are utilized to bootstrap the LLM via Supervised Fine-Tuning (SFT), followed by multi-turn Reinforcement Learning (RL) to activate its latent world knowledge. Empirically, SCOUT enables a Qwen2.5-3B-Instruct model to achieve an average score of 0.86, significantly outperforming proprietary models, including Gemini-2.5-Pro (0.60), while saving about 60% GPU hours consumption.

Existing retrieval-augmented generation (RAG) systems often assume that each query has a single correct answer. This assumption overlooks open-ended information-seeking scenarios where multiple plausible answers are valuable, and where diversity is important for creativity, fairness, and inclusive access to information. We show that standard RAG systems fail to fully use diverse retrieved contexts: simply increasing retrieval diversity does not necessarily lead to diverse generations. To address this limitation, we propose Diverge, a plug-and-play agentic RAG framework that improves the diversity--quality trade-off through iterative, reflection-guided exploration of diverse viewpoints and diversity-aware retrieval support. We further introduce evaluation metrics for characterizing the diversity-quality trade-off in open-ended question answering. Experiments across multiple real-world datasets and backbone LLMs show that Diverge achieves the best trade-off among competitive baselines, increasing diversity by $\sim2\times$ without noticeable quality degradation. These results reveal a systematic limitation of current RAGs and show the value of explicit diversity modeling.

Graph anomaly detection (GAD) has attracted growing interest for its crucial ability to uncover irregular patterns in broad applications. Semi-supervised GAD, which assumes a subset of annotated normal nodes available during training, is among the most widely explored application settings. However, the normality learned by existing semi-supervised GAD methods is limited to the labeled normal nodes, often inclining to overfitting the given patterns. These can lead to high detection errors, such as high false positives. To overcome this limitation, we propose GraphNC , a graph normality calibration framework that leverages both labeled and unlabeled data to calibrate the normality from a teacher model (a pre-trained semi-supervised GAD model) jointly in anomaly score and node representation spaces. GraphNC includes two main components, anomaly score distribution alignment (ScoreDA) and perturbation-based normality regularization (NormReg). ScoreDA optimizes the anomaly scores of our model by aligning them with the score distribution yielded by the teacher model. Due to accurate scores in most of the normal nodes and part of the anomaly nodes in the teacher model, the score alignment effectively pulls the anomaly scores of the normal and abnormal classes toward the two ends, resulting in more separable anomaly scores. Nevertheless, there are inaccurate scores from the teacher model. To mitigate the misleading by these scores, NormReg is designed to regularize the graph normality in the representation space, making the representations of normal nodes more compact by minimizing a perturbation-guided consistency loss solely on the labeled nodes.

As acquiring reliable ground-truth labels is usually costly, or infeasible, crowdsourcing and aggregation of noisy human annotations is the typical resort. Aggregating subjective labels, though, may amplify individual biases, particularly regarding sensitive features, raising fairness concerns. Nonetheless, fairness in crowdsourced aggregation remains largely unexplored, with no existing convergence guarantees and only limited post-processing approaches for enforcing $\varepsilon$-fairness under demographic parity. We address this gap by analyzing the fairness s of crowdsourced aggregation methods within the $\varepsilon$-fairness framework, for Majority Vote and Optimal Bayesian aggregation. In the small-crowd regime, we derive an upper bound on the fairness gap of Majority Vote in terms of the fairness gaps of the individual annotators. We further show that the fairness gap of the aggregated consensus converges exponentially fast to that of the ground-truth under interpretable conditions. Since ground-truth itself may still be unfair, we generalize a state-of-the-art multiclass fairness post-processing algorithm from the continuous to the discrete setting, which enforces strict demographic parity constraints to any aggregation rule. Experiments on synthetic and real datasets demonstrate the effectiveness of our approach and corroborate the theoretical insights.

Causal inference from observational data can provide strong evidence for finding the best action in a decision-making scenario without having to perform expensive randomized trials. The causal effect of an action is often not pointwise identifiable even with infinite data due to unobserved confounding factors. Furthermore, having only finitely many samples adds another layer of uncertainty to causal effect estimation. Several existing methods can be used to obtain upper and lower bounds to the causal effect, ranging from symbolic methods to the more recent neural network-based approaches, which implicitly incorporate both sources of uncertainty. However, these methods do not inform whether collecting more samples may or may not help identify the best action from observational data, leaving experts in the dark about their data collection strategies. We address this problem with a novel framework that can distinguish the range of causal effect values that might be eliminated by collecting more samples from the range of values that, with high probability, cannot be eliminated with more observational samples. We show that this partitioning can be obtained by solving max-min and min-max optimization problems. We leverage neural causal models to approximately recover this decomposition in practice. We demonstrate via experiments on synthetic and real-world datasets that our algorithm can determine when collecting more samples will not help determine the best action. Our framework can help practitioners decide when to resort to non-observational studies or seek to measure some of the unmeasured confounders for optimal decision-making.

Many successful families of generative models leverage a low-dimensional latent distribution that is mapped to a data distribution. Though simple latent distributions are often used, the choice of distribution has a strong impact on model performance. Recent experiments have suggested that the probability distributions produced by quantum processors, which are typically highly correlated and classically intractable, can lead to improved performance on some datasets. However, when and why latent distributions produced by quantum processors can improve performance, and whether these improvements are connected to quantum properties of these distributions, are open questions that we investigate in this work. We show in theory that, under certain conditions, these "quantum latent distributions" enable generative models to produce data distributions that classical latent distributions cannot efficiently produce. We provide intuition as to the underlying mechanisms that could explain a performance advantage on real datasets. Based on this, we perform extensive benchmarking on a synthetic quantum dataset and the QM9 molecular dataset, using both simulated and real photonic quantum processors. We find that the statistics arising from quantum interference lead to improved generative performance compared to classical baselines, suggesting that quantum processors can play a role in expanding the capabilities of deep generative models.

Reinforcement learning (RL) with combinatorial action spaces remains challenging because feasible action sets are exponentially large and governed by complex feasibility constraints, making direct policy parameterization impractical. Existing approaches embed task-specific value functions into constrained optimization programs or learn deterministic structured policies, sacrificing generality and policy expressiveness. We propose a solver-induced \emph{latent spherical flow policy} that brings the expressiveness of modern generative policies to combinatorial RL while guaranteeing feasibility by design. Our method, LSFlow, learns a \emph{stochastic} policy in a compact continuous latent space via spherical flow matching, and delegates feasibility to a combinatorial optimization solver that maps each latent sample to a valid structured action. To improve efficiency, we train the value network directly in the latent space, avoiding repeated solver calls during policy optimization. To address the piecewise-constant and discontinuous value landscape induced by solver-based action selection, we introduce a smoothed Bellman operator that yields stable, well-defined learning targets. Empirically, our approach outperforms state-of-the-art baselines by an average of 20.6\% across a range of challenging combinatorial RL tasks.

While Mechanistic Interpretability has identified interpretable circuits in LLMs, their causal origins in training data remain elusive. We introduce Mechanistic Data Attribution (MDA), a scalable framework that employs Influence Functions to trace interpretable units back to specific training samples. Through extensive experiments on the Pythia family, we causally validate that targeted intervention--removing or augmenting a small fraction of high-influence samples--significantly modulates the emergence of interpretable heads, whereas random interventions show no effect. Our analysis reveals that repetitive structural data (e.g., LaTeX, XML) acts as a mechanistic catalyst. Furthermore, we observe that interventions targeting induction head formation induce a concurrent change in the model's in-context learning (ICL) capability. This provides direct causal evidence for the long-standing hypothesis regarding the functional link between induction heads and ICL. Finally, we propose a mechanistic data augmentation pipeline that consistently accelerates circuit convergence across model scales, providing a principled methodology for steering the developmental trajectories of LLMs.

Evaluating the performance of large language models (LLMs) from human preference data is crucial for obtaining LLM leaderboards. However, many existing approaches either rely on restrictive parametric assumptions or lack valid uncertainty quantification when flexible machine learning methods are used. In this paper, we propose a nonparametric statistical framework, called DMLRank, for comparing and ranking LLMs from preference data using debiased machine learning (DML). For this, we introduce generalized average ranking scores (GARS), which generalize commonly used ranking models, including the Bradley-Terry model or PageRank/ Rank centrality, with complex human responses such as ties. DMLRank comes with the following advantages: (i)~It produces statistically efficient estimates of GARS ranking scores. (ii) It naturally allows the incorporation of black-box machine learning methods for estimation. (iii) It can be combined with pre-trained LLM evaluators (e.g., using LLM-as-a-judge). (iv) It suggests optimal policies for collecting preference data under budget constraints. We demonstrate these advantages both theoretically and empirically using both synthetic and real-world preference datasets. In summary, our framework provides practitioners with powerful, state-of-the-art methods for comparing or ranking LLMs for leaderboards.

The performance of large language models (LLMs) on verifiable tasks is usually measured by pass@k, the probability of answering a question correctly at least once in k trials. At a fixed budget, a more suitable metric is coverage@cost, the average number of unique questions answered as a function of the total number of attempts. We connect the two metrics and show that the empirically-observed power-law behavior in pass@k leads to a sublinear growth of the coverage@cost (diminishing returns). To solve this problem, we propose Reset-and-Discard (ReD), a query method of LLMs that increases coverage@cost for a given budget, regardless of the pass@k form. Moreover, given a pass@k, we can quantitatively predict the savings in the total number of attempts using ReD. If pass@k is not available for the model, ReD can infer its power-law exponent. Experiments on three LLMs across coding (HumanEval), math (GSM8K), and reasoning (MMLU-Pro) benchmarks demonstrate that ReD substantially reduces the required attempts, tokens, and USD cost to reach a desired coverage, while also offering an efficient way to measure inference power-laws. ReD's advantage is maintained for imperfect verifiers and outperforms the tested allocation baselines.

White matter hyperintensities (WMH) and ischaemic stroke lesions (ISL) are key imaging biomarkers of cerebral small vessel disease (SVD) detectable on magnetic resonance imaging (MRI). The development of robust deep learning models to automatically segment and differentiate these pathologies remains challenging. Specifically, WMH and ISL frequently co-occur within the same subject and present as visually confounding hyperintensities on fluid-attenuated inversion recovery (FLAIR) sequences, complicating their accurate delineation. To address the scarcity of fully annotated cohorts, we systematically evaluated six accessible strategies for training a joint WMH and ISL segmentation model using partially labelled data. We aggregated privately held and publicly available datasets to curate a large-scale cohort of 2,052 MRI volumes, of which 1341 and 1152 volumes contained ground truth annotations for WMH and ISL, respectively. Our analysis indicates that multiple strategies effectively leverage partially labelled data to enhance overall model performance, with pseudolabelling emerging as the most effective approach. This model exhibited a consistent WMH segmentation policy and successfully detected the majority of FLAIR-positive ISL. These findings demonstrate the viability of using partially labelled data to develop reliable automated segmentation tools, which can support ongoing SVD monitoring and high-throughput biomarker extraction for large-scale clinical research.

Large language models (LLMs) are increasingly deployed as autonomous agents that negotiate, coordinate, and act on behalf of users. Whether they cooperate in such settings is no longer just an academic question, but a central issue for AI governance. We approach it from a strategic-behaviour angle, asking how two everyday levers - the size of what is at stake, and the language in which the interaction is described - shape the strategies LLMs adopt in a repeated Prisoner's Dilemma. Rather than reading cooperation off raw action counts, we train supervised classifiers to recognise the canonical strategies of repeated games (always cooperate, always defect, Tit-for-Tat, Win-Stay-Lose-Shift) and use them as a lens onto LLM behaviour. To know what the strategy distribution should look like under the same payoffs, we derive an evolutionary game theory (EGT) baseline and compare it with the LLM data. The two outcomes disagree in a revealing way: as stakes grow, evolutionary theory predicts that defection should take over the population, yet LLMs move in the opposite direction, becoming more cooperative - a signature, we argue, of alignment training and the human-like reasoning patterns LLMs inherit from their training data. We further show that this picture is not particular to frontier-scale, proprietary models: it also occurs with three open-weight smaller LLMs. Overall, our analysis highlights that payoff design and linguistic framing are powerful but under-explored levers for steering LLM behaviour, with direct implications for evaluating, aligning, and governing multi-agent AI systems deployed in high-stakes, multilingual environments.

As Earth-observing satellite constellations grow in size and capability, distributed onboard control offers a pathway to novel responses and time-sensitive measurements. However, deploying autonomy to satellites requires efficient computation and communication. This work addresses the challenge of scheduling observations for hundreds of satellites in a dynamic, large-scale problem with millions of variables. We present the dynamic multi-satellite constellation observation scheduling problem (DCOSP), a new formulation of dynamic distributed constraint optimization problems (DDCOP) that models integrated scheduling and execution. DCOSP features a novel optimality condition, for which we construct an exact omniscient offline algorithm. Motivated by the strong resource constraints of onboard satellite operations, we introduce a framework to incorporate metareasoning in DDCOPs that controls when agents expend resources to recompute solutions. In addition, we present the dynamic incremental neighborhood stochastic search (D-NSS) algorithm, an incomplete online decomposition-based DDCOP algorithm that repairs localized sub-problems in response to dynamic events. We demonstrate in realistic simulations that D-NSS converges to near-optimal solutions, outperforming standard DDCOP baselines in solution quality, computation time, and message volume, while our metareasoning framework successfully balances resource conservation with utility. As part of the NASA FAME mission, this work lays the foundation for the largest in-space demonstration of distributed multi-agent AI to date.

Fine-tuning-based adaptation is widely used to customize diffusion-based image generation, leading to large collections of community-created adapters that capture diverse subjects and styles. Adapters derived from the same base model can be merged with weights, enabling the synthesis of new visual results within a vast and continuous design space. To explore this space, current workflows rely on manual slider-based tuning, an approach that scales poorly and makes weight selection difficult, even when the candidate set is limited to 20-30 adapters. We propose GimmBO to support interactive exploration of adapter merging for image generation through Preferential Bayesian Optimization (PBO). Motivated by observations from real-world usage, including sparsity and constrained weight ranges, we introduce a two-stage BO backend that improves sampling efficiency and convergence in high-dimensional spaces. We evaluate our approach with simulated users and a user study, demonstrating improved convergence, high success rates, and consistent gains over BO and line-search baselines, and further show the flexibility of the framework through several extensions.

While Large Language Model (LLM) agents excel at general tasks, they inherently struggle with continual adaptation due to the frozen weights after deployment. Conventional reinforcement learning (RL) offers a solution but incurs prohibitive computational costs and the risk of catastrophic forgetting. We introduce Just-In-Time Reinforcement Learning (JitRL), a training-free framework that enables test-time policy optimization without any gradient updates. JitRL maintains a dynamic, non-parametric memory of experiences and retrieves relevant trajectories to estimate action advantages on-the-fly. These estimates are then used to directly modulate the LLM's output logits. We theoretically prove that this additive update rule is the exact closed-form solution to the KL-constrained policy optimization objective. Extensive experiments on WebArena and Jericho demonstrate that JitRL establishes a new state-of-the-art among training-free methods. Crucially, JitRL outperforms the performance of computationally expensive fine-tuning methods (e.g., WebRL) while reducing monetary costs by over 30 times, offering a scalable path for continual learning agents. The code is available at https://github.com/liushiliushi/JitRL.

The performance of modern AI systems is fundamentally constrained by the quality of their underlying GPU kernels, which translate high-level algorithmic semantics into low-level hardware operations. Achieving near-optimal kernels requires expert-level understanding of hardware architectures and programming models, making kernel engineering a critical but notoriously time-consuming and non-scalable process. Recent advances in large language models and LLM-based agents have opened new possibilities for automating kernel generation and optimization. LLMs are well-suited to compress expert-level kernel knowledge that is difficult to formalize, while agentic systems further enable scalable optimization by casting kernel development as an iterative, feedback-driven loop. Rapid progress has been made in this area. However, the field remains fragmented and lacks a systematic perspective for LLM-driven kernel generation. This survey addresses this gap by providing a structured overview of existing approaches, spanning LLM-based approaches and agentic optimization workflows, and systematically organizing the datasets and benchmarks that underpin learning and evaluation in this domain. Moreover, key open challenges and future research directions are further outlined, aiming to establish a comprehensive reference for the next generation of automated kernel optimization. To keep track of this field, we maintain an open-source GitHub repository at https://github.com/flagos-ai/awesome-LLM-driven-kernel-generation.

A key challenge in infrared small target detection (IRSTD) is that weak target signal responses are easily obscured by strong background clutter, frequently resulting in missed detections. While traditional gradient-based methods attempt to capture fine details, their robustness is limited by the static fusion of multi-directional gradient features. In this paper, we rethink feature fusion from the perspective of Basis Decomposition Theory and propose a novel framework that reformulates the process into an explicit and adaptive decomposition-and-reconstruction paradigm. Specifically, we introduce the Basis Decomposition Module (BDM) and its specialized variant, the Gradient Decomposition Module (GDM) for IRSTD. GDMs treat the normalized gradient features as basis vectors to reconstruct a new feature, thereby maintaining detailed structures and highlighting infrared small targets. By integrating GDMs into a lightweight three-stage U-Net, we develop two unified architectures: the Spatial Gradient Basis Decomposition Network for single-frame detection and the Spatio-temporal Gradient Basis Decomposition Network for multi-frame scenarios. Extensive experiments demonstrate that our networks achieve state-of-the-art (SOTA) performance across multiple benchmarks, offering a superior balance between detection accuracy and computational efficiency. Our codes will be made public at: https://github.com/greekinRoma/IRSTD_HC_Platform.

Automated interlinear gloss prediction with neural networks is a promising approach to accelerate language documentation efforts. However, while state-of-the-art models like GlossLM achieve high scores on glossing benchmarks, user studies with linguists have found critical barriers to the usefulness of such models in real-world scenarios. In particular, existing models typically generate morpheme-level glosses but assign them to whole words without predicting the actual morpheme boundaries, making the predictions less interpretable and thus untrustworthy to human annotators. We conduct the first study on neural models that jointly predict interlinear glosses and the corresponding morphological segmentation from raw text. We run experiments to determine the optimal way to train models that balance segmentation and glossing accuracy, as well as the alignment between the two tasks. We extend the training corpus of GlossLM and pretrain PolyGloss, a family of seq2seq multilingual models for joint segmentation and glossing that outperforms GlossLM on glossing and beats various open-source LLMs on segmentation, glossing, and alignment. In addition, we demonstrate that PolyGloss can be quickly adapted to a new dataset via low-rank adaptation.

We introduce Lattice, a hybrid sequential prediction system that conditionally activates learned behavioral structure using binary confidence gating. The system summarizes behavior windows as behavioral archetypes and activates archetype-based scoring only when an in-support confidence signal exceeds a validation-calibrated threshold, falling back to backbone predictions when uncertain. Our primary estimand is the controlled effect of adding Lattice to a fixed backbone on identical test rows. On MovieLens (30 paired seeds, full-catalog ranking), LSTM+Lattice improves HR@10 by +31.7% (gated) versus the LSTM backbone alone (p much less than 10^-20); ungated fusion reaches +58.7% on the same protocol. We do not claim gating maximizes pooled accuracy. With backbone-native archetypes (fit in each backbone's embedding space), gated lifts of +13.3% (transformer) and +17.0% (SASRec) hold under the same evaluation design. A prior approximately 0% transformer row in version 1 reflected an invalid cross-backbone transfer, not evidence that composition cannot help stronger encoders. Amazon Electronics provides supporting cross-domain evidence (+124.0% gated, 15 seeds, high variance). Controlled shift checks (appendix) illustrate gate refusal under distribution shift. Standalone SASRec and BERT4Rec scores are contextual references, not the target estimand. We report what composition achieves and when it activates; production calibration and implementation details remain proprietary pending patent prosecution.

Medical vision-language models can automate the generation of radiology reports but struggle with accurate visual grounding and factual consistency. Existing models often misalign textual findings with visual evidence, leading to unreliable or weakly grounded predictions. We present CURE, an error-aware curriculum learning framework that improves grounding and report quality without any additional data. CURE fine-tunes a multimodal instructional model on phrase grounding, grounded report generation, and anatomy-grounded report generation using public datasets. The method dynamically adjusts sampling based on model performance, emphasizing harder samples to improve spatial and textual alignment. CURE improves grounding accuracy by +0.35 IoU, boosts report quality by +0.192 CXRFEScore, and reduces hallucinations by 18.6%. CURE is a data-efficient framework that enhances both grounding accuracy and report reliability. Code is available at https://github.com/PabloMessina/CURE and model weights at https://huggingface.co/pamessina/medgemma-4b-it-cure

In Robot-Assisted Minimally Invasive Surgery (RMIS), accurate tool localization is crucial to ensure patient safety and successful task execution. However, this remains challenging for cable-driven robots, such as the da Vinci robot, because erroneous encoder readings lead to pose estimation errors. In this study, we propose a calibration framework to produce accurate tool localization results through computing the hand-eye transformation matrix on-the-fly. The framework consists of two interrelated algorithms: the feature association block and the hand-eye calibration block, which provide robust correspondences for key points detected on monocular images without pre-training, and offer the versatility to accommodate various surgical scenarios by adopting an array of filter approaches, respectively. To validate its efficacy, we test the framework extensively on publicly available video datasets that feature multiple surgical instruments conducting tasks in both in vitro and ex vivo scenarios, under varying illumination conditions and with different levels of key point measurement accuracy. The results show a significant reduction in tool localization errors under the proposed calibration framework, with accuracies comparable to other state-of-the-art methods while being more time-efficient.

Cross-cultural competence in large language models (LLMs) requires understanding and adapting Culture-Specific Items (CSIs) across varying cultural contexts. However, progress in evaluating this capability remains limited by the lack of high-quality CSI-annotated corpora with parallel cross-cultural sentence pairs. We introduce XCR-Bench, a Cross(X)-Cultural Reasoning Benchmark containing 4.1k parallel sentences and 1,098 CSIs across three reasoning tasks. XCR-Bench integrates Newmark's CSI framework with Hall's Triad of Culture, enabling evaluation across levels of cultural visibility -- from observable practices to implicit social norms and values. Experiments on eight multilingual LLMs show that state-of-the-art models exhibit consistent weaknesses in identifying and adapting specific categories of CSIs, revealing a gap between surface-level recall and explicit cultural reasoning. Performance declines significantly on culturally sensitive categories and deeper cultural levels (p<0.005, 8/8 models), and adaptation quality varies systematically across target cultures and Bengali regional variants, indicating encoded regional and ethno-religious biases even within a single linguistic setting. We publicly release the corpus and code to support future research on cross-cultural NLP.

Finite-State Transducers (FSTs) are effective models for string-to-string rewriting tasks, often providing the efficiency necessary for high-performance applications, but constructing transducers by hand is difficult. In this work, we propose a novel method for automatically constructing unweighted FSTs following the hidden state geometry learned by a recurrent neural network. We evaluate our methods on real-world datasets for morphological inflection, grapheme-to-phoneme prediction, and historical normalization, showing that the constructed FSTs are highly accurate and robust for many datasets, substantially outperforming classical transducer learning algorithms by up to 87% accuracy on held-out test sets.

Metaphor and humor share a lot of common ground, and metaphor is one of the most common humorous mechanisms. This study focuses on the humorous capacity of multimodal metaphors, which has not received due attention in the community. We take inspiration from the Incongruity Theory of humor, the Conceptual Metaphor Theory, and the annotation scheme behind the VU Amsterdam Metaphor Corpus, and developed a novel annotation scheme for humorous multimodal metaphor use in image-caption pairs. We create the Hummus Dataset of Humorous Multimodal Metaphor Use, providing expert annotation on 1k image-caption pairs sampled from the New Yorker Caption Contest corpus. Using the dataset, we test state-of-the-art multimodal large language models (MLLMs) on their ability to detect and understand humorous multimodal metaphor use. Our experiments show that current MLLMs still struggle with processing humorous multimodal metaphors, particularly with regard to integrating visual and textual information. We release our dataset and code at github.com/xiaoyuisrain/humorous-multimodal-metaphor-use.

Vision-Language Models (VLMs) integrate visual and textual knowledge into unified representations that increasingly underpin modern retrieval and recommendation systems. However, it remains unclear how reliably these models utilize their cross-modal knowledge when ranking multimodal items, and whether their knowledge grounding can be subverted. In this paper, we expose a fundamental vulnerability in how VLMs apply multimodal knowledge for product ranking: through Multimodal Generative Engine Optimization (MGEO), we show that an adversary can manipulate a VLM's ranking decisions by jointly crafting imperceptible image perturbations and fluent textual suffixes that exploit the model's internal cross-modal knowledge coupling. Using an alternating optimization strategy, MGEO targets the deep interactions between visual and linguistic representations within the VLM, achieving rank manipulations that substantially exceed those of unimodal attacks and heuristic baselines powered by strong commercial models. Our findings reveal that surface-level content quality is insufficient for rank promotion; instead, direct alignment with the model's internal knowledge utilization mechanism is required. These results raise important questions on the faithfulness and robustness of knowledge grounding in multimodal foundation models, and motivate future work on defense mechanisms for multimodal retrieval systems. Code is available at: https://github.com/glad-lab/MGEO

Metal-organic frameworks (MOFs) are porous crystalline materials with broad applications such as carbon capture and drug delivery, yet accurately predicting their 3D structures remains a significant challenge. While Large Language Models (LLMs) have shown promise in generating crystal structures, their application to MOFs is hindered by MOFs' high structural complexity arising from the large number of atoms in unit cell. Inspired by the success of block-wise paradigms in deep generative models for MOFs, we pioneer the application of LLMs in this domain by introducing MOF-LLM, the first LLM framework specifically adapted for block-level MOF structure prediction. To effectively harness LLMs for this 3D modular assembly task, our training paradigm integrates spatial-aware continual pre-training (CPT), structural supervised fine-tuning (SFT), and matching-driven reinforcement learning (RL). By incorporating explicit spatial priors and optimizing structural stability via Soft Adaptive Policy Optimization (SAPO), our approach substantially enhances the spatial reasoning in a Qwen-3 8B model for MOF structure prediction. Comprehensive experiments demonstrate that MOF-LLM achieves state-of-the-art performance with a match rate of 35.78% while exhibiting superior sampling efficiency of 0.04 seconds per structure.

Group Relative Policy Optimization (GRPO) has become a standard approach for training mathematical reasoning models; however, its reliance on multiple completions per prompt makes training computationally expensive. Although recent work has reduced the number of training steps required to reach peak performance, the overall wall-clock training time often remains unchanged or even increases due to higher per-step cost. We propose MMR-GRPO, which integrates Maximal Marginal Relevance to reweigh rewards based on completion diversity. Our key insight is that semantically redundant completions contribute limited marginal learning signal; prioritizing diverse solutions yields more informative updates and accelerates convergence. Extensive evaluations across three model sizes (1.5B, 7B, 8B), three GRPO variants, and five mathematical reasoning benchmarks show that MMR-GRPO achieves comparable peak performance while requiring on average 47.9% fewer training steps and 70.2% less wall-clock time. These gains are consistent across models, methods, and benchmarks. Our code is released at: https://github.com/WeiKangda/MMR-GRPO.

Research in machine learning has questioned whether increases in training token counts reliably produce proportional performance gains in large language models. Building on prior work introducing an energy-aware parameter efficiency metric, this study empirically examines the effects of increasing training token counts under fixed hardware and training conditions. The significance of this work lies in the explicit integration of power consumption and execution duration, as reflected by the power sampling frequency, into token-scale analysis. This addresses a gap in prior studies emphasizing performance outcomes while underrepresenting computational and energy costs. Using a repeated-measures experimental design on a constant GPU instance with an identical model architecture, optimizer settings, and epoch counts, a 1.1-billion-parameter TinyLlama model was trained at three token counts (500K, 1M, and 2M). While conventional performance metrics exhibited inconsistent or diminishing returns across token scales, the inclusion of power consumption and execution duration revealed a strictly monotonic decline in training efficiency as token count increased. Repeated-measures ANOVA demonstrated a strong effect of token count on parameter efficiency, with all pairwise comparisons remaining significant following Bonferroni correction. These findings indicate that increases in training token counts may be energetically inefficient even when marginal performance improvements are observed, underscoring the importance of efficiency-aware evaluation in large language model training.

Large Language Models (LLMs) often encode whether a statement is true as a vector in their residual stream activations. These vectors, also known as truth vectors, have been studied in prior work, however how they change when context is introduced remains unexplored. We study this question by measuring (1) the directional change ($θ$) between the truth vectors with and without context and (2) the relative magnitude of the truth vectors upon adding context. Across four LLMs and four datasets, we find that (1) truth vectors are roughly orthogonal in early layers, converge in middle layers, and may stabilize or continue increasing in later layers; (2) adding context generally increases the truth vector magnitude, i.e., the separation between true and false representations in the activation space is amplified; (3) larger models distinguish relevant from irrelevant context mainly through directional change ($θ$), while smaller models show this distinction through magnitude differences. We also find that context conflicting with parametric knowledge produces larger geometric changes than parametrically aligned context. To the best of our knowledge, this is the first work that provides a geometric characterization of how context transforms the truth vector in the activation space of LLMs.