Reinforcement Learning from Human Feedback (RLHF) has emerged as a powerful approach for aligning generative models, but its reliance on learned reward models makes it vulnerable to mis-specification and reward hacking. Preference-based Reinforcement Learning (PbRL) offers a more robust alternative by directly leveraging noisy binary comparisons over trajectories. We study the best policy identification problem in PbRL, motivated by post-training optimization of generative models, for example, during multi-turn interactions. Learning in this setting combines an offline preference dataset - potentially biased or out-of-distribution and collected from a rater of subpar `competence' - with online pure exploration, making systematic online learning essential. To this end, we propose Posterior Sampling for Preference Learning ($\mathsf{PSPL}$), a novel algorithm inspired by Top-Two Thompson Sampling that maintains posteriors over the reward model and dynamics. We provide the first Bayesian simple regret guarantees for PbRL and introduce an efficient approximation that outperforms existing baselines on simulation and image generation benchmarks.

Ordinal Classification (OC) addresses those classification tasks where the labels exhibit a natural order. Unlike nominal classification, which treats all classes as mutually exclusive and unordered, OC takes the ordinal relationship into account, producing more accurate and relevant results. This is particularly critical in applications where the magnitude of classification errors has significant consequences. Despite this, OC problems are often tackled using nominal methods, leading to suboptimal solutions. Although decision trees are among the most popular classification approaches, ordinal tree-based approaches have received less attention when compared to other classifiers. This work provides a comprehensive survey of ordinal splitting criteria, standardising the notations used in the literature to enhance clarity and consistency. Three ordinal splitting criteria, Ordinal Gini (OGini), Weighted Information Gain (WIG), and Ranking Impurity (RI), are compared to the nominal counterparts of the first two (Gini and information gain), by incorporating them into a decision tree classifier. An extensive repository considering $45$ publicly available OC datasets is presented, supporting the first experimental comparison of ordinal and nominal splitting criteria using well-known OC evaluation metrics. The results have been statistically analysed, highlighting that OGini stands out as the best ordinal splitting criterion to date, reducing the mean absolute error achieved by Gini by more than 3.02%. To promote reproducibility, all source code developed, a detailed guide for reproducing the results, the 45 OC datasets, and the individual results for all the evaluated methodologies are provided.

Domain randomization through synthesis is a powerful strategy to train networks that are unbiased with respect to the domain of the input images. Randomization allows networks to see a virtually infinite range of intensities and artifacts during training, thereby minimizing overfitting to appearance and maximizing generalization to unseen data. Although powerful, this approach relies on the accurate tuning of a large set of hyperparameters that govern the probabilistic distribution of the synthesized images. Instead of manually tuning these parameters, we introduce Learn2Synth, a novel procedure in which synthesis parameters are learned using a small set of real labeled data. Unlike methods that impose constraints to align synthetic data with real data (e.g., contrastive or adversarial techniques), which risk misaligning the image and its label map, we tune an augmentation engine such that a segmentation network trained on synthetic data has optimal accuracy when applied to real data. This approach allows the training procedure to benefit from real labeled examples, without ever using these real examples to train the segmentation network, which avoids biasing the network towards the properties of the training set. Specifically, we develop parametric and nonparametric strategies to enhance synthetic images in a way that improves the performance of the segmentation network. We demonstrate the effectiveness of this learning strategy on synthetic and real-world brain scans. Code is available at: https://github.com/HuXiaoling/Learn2Synth.

Machine learning can predict human behavior well when substantial structured data and well-defined outcomes are available, but these models are typically limited to specific outcomes and cannot readily be applied to new domains. We test whether large language models (LLMs) can support a more general-purpose approach by building person-specific simulations (i.e., "generative agents") grounded in self-report data. Using data from a diverse national sample of 1,052 Americans, we build agents from (i) two-hour, semi-structured interviews (elicited using the American Voices Project interview schedule), (ii) structured surveys (the General Social Survey and Big Five personality inventory), or (iii) both sources combined. On held-out General Social Survey items, agent accuracy reached 83% (interview only), 82% (surveys only), and 86% (combined) of participants' two-week test-retest consistency, compared with agents prompted only with individuals' demographics (74%). Agents predicted personality traits and behaviors in experiments with similar accuracy, and reduced disparities in accuracy across racial and ideological groups relative to demographics-only baselines. Together, these results show that LLMs agents grounded in rich qualitative or quantitative self-report data can support general-purpose simulation of individuals across outcomes, without requiring task-specific training data.

Enabling robots to autonomously navigate unknown, complex, and dynamic real-world environments presents several challenges, including imperfect perception, partial observability, localization uncertainty, and safety constraints. Current approaches are typically limited to simulations, where such challenges are not present. In this work, we present a lightweight, open-architecture, end-to-end system for real-world robot autonomous navigation. Specifically, we deploy a real-time navigation system on a quadruped robot by integrating multiple onboard components that communicate via ROS2. Given navigation tasks specified in natural language, the system fuses onboard sensory data for localization and mapping with open-vocabulary semantics to build hierarchical scene graphs from a continuously updated semantic object map. An LLM-based planner leverages these graphs to generate and adapt multi-step plans in real time as the scene evolves. Through experiments across multiple indoor environments using a Unitree Go2 quadruped, we demonstrate zero-shot real-world autonomous navigation, achieving over 88% task success, and provide analysis of system behavior during deployment.

A major challenge in humanoid robotics is designing a unified interface for commanding diverse whole-body behaviors, from precise footstep sequences to partial-body mimicry and joystick teleoperation. We introduce the Masked Humanoid Controller (MHC), a learned whole-body controller that exposes a simple yet expressive interface: the specification of masked target trajectories over selected subsets of the robot's state variables. This unified abstraction allows high-level systems to issue commands in a flexible format that accommodates multi-modal inputs such as optimized trajectories, motion capture clips, re-targeted video, and real-time joystick signals. The MHC is trained in simulation using a curriculum that spans this full range of modalities, enabling robust execution of partially specified behaviors while maintaining balance and disturbance rejection. We demonstrate the MHC both in simulation and on the real-world Digit V3 humanoid, showing that a single learned controller is capable of executing such diverse whole-body commands in the real world through a common representational interface.

As privacy concerns escalate in the realm of machine learning, data owners now have the option to utilize machine unlearning to remove their data from machine learning models, following recent legislation. To enhance transparency in machine unlearning and avoid potential dishonesty by model providers, various verification strategies have been proposed. These strategies enable data owners to ascertain whether their target data has been effectively unlearned from the model. However, our understanding of the safety issues of machine unlearning verification remains nascent. In this paper, we explore the novel research question of whether model providers can circumvent verification strategies while retaining the information of data supposedly unlearned. Our investigation leads to a pessimistic answer: \textit{the verification of machine unlearning is fragile}. Specifically, we categorize the current verification strategies regarding potential dishonesty among model providers into two types. Subsequently, we introduce two novel adversarial unlearning processes capable of circumventing both types. We validate the efficacy of our methods through theoretical analysis and empirical experiments using real-world datasets. This study highlights the vulnerabilities and limitations in machine unlearning verification, paving the way for further research into the safety of machine unlearning.

In the field of machine unlearning, certified unlearning has been extensively studied in convex machine learning models due to its high efficiency and strong theoretical guarantees. However, its application to deep neural networks (DNNs), known for their highly nonconvex nature, still poses challenges. To bridge the gap between certified unlearning and DNNs, we propose several simple techniques to extend certified unlearning methods to nonconvex objectives. To reduce the time complexity, we develop an efficient computation method by inverse Hessian approximation without compromising certification guarantees. In addition, we extend our discussion of certification to nonconvergence training and sequential unlearning, considering that real-world users can send unlearning requests at different time points. Extensive experiments on three real-world datasets demonstrate the efficacy of our method and the advantages of certified unlearning in DNNs.

Machine Learning research, including work promoting fair or equitable algorithms, often relies on the concept of a data-generating probability distribution. The standard presumption is that since data points are 'sampled from' such a distribution, one can learn from observed data about this distribution and, thus, predict future data points which are also drawn from it. We argue, however, that such true probability distributions do not exist and that the rhetoric around them is harmful in social settings. We show that alternative frameworks focusing directly on relevant populations rather than abstract distributions are available and leave classical learning theory almost unchanged. Furthermore, we argue that the assumption of true probabilities or data-generating distributions can be misleading and obscure both the choices made and the goals pursued in machine learning practice. Based on these considerations, we suggest avoiding the assumption of data-generating probability distributions in the social world.

In the domain of data science, the predictive tasks of classification, regression, and imputation of missing values are commonly encountered challenges associated with tabular data. This research endeavors to apply Large Language Models (LLMs) towards addressing these predictive tasks. Despite their proficiency in comprehending natural language, LLMs fall short in dealing with structured tabular data. This limitation stems from their lacking exposure to the intricacies of tabular data during their foundational training. Our research aims to mitigate this gap by compiling a comprehensive corpus of tables annotated with instructions and executing large-scale training of Llama-2 on this enriched dataset. Furthermore, we investigate the practical application of applying the trained model to zero-shot prediction, few-shot prediction, and in-context learning scenarios. Through extensive experiments, our methodology has shown significant improvements over existing benchmarks. These advancements highlight the efficacy of tailoring LLM training to solve table-related problems in data science, thereby establishing a new benchmark in the utilization of LLMs for enhancing tabular intelligence.

Functional linear discriminant analysis (FLDA) is a powerful tool that extends LDA-mediated multiclass classification and dimension reduction to univariate time-series functions. However, in the age of large multivariate and incomplete data, statistical dependencies between features must be estimated in a computationally tractable way, while also dealing with missing data. There is a need for a computationally tractable approach that considers the statistical dependencies between features and can handle missing values. We here develop a multivariate version of FLDA (MUDRA) to tackle this issue and describe an efficient expectation/conditional-maximization (ECM) algorithm to infer its parameters. We assess its predictive power on the "Articulary Word Recognition" data set and show its improvement over the state-of-the-art, especially in the case of missing data. MUDRA allows interpretable classification of data sets with large proportions of missing data, which will be particularly useful for medical or psychological data sets.

Multi-objective reinforcement learning (MORL) algorithms extend conventional reinforcement learning (RL) to the more general case of problems with multiple, conflicting objectives, represented by vector-valued rewards. Widely-used scalar RL methods such as Q-learning can be modified to handle multiple objectives by (1) learning vector-valued value functions, and (2) performing action selection using a scalarisation or ordering operator which reflects the user's preferences with respect to the different objectives. This paper investigates two previously unreported issues which can hinder the performance of value-based MORL algorithms when applied in conjunction with a non-linear utility function -- value function interference, and sensitivity to overestimation. We illustrate the nature of these phenomena on simple multi-objective MDPs using a tabular implementation of multiobjective Q-learning.

This paper introduces UVIO, a multi-sensor framework that leverages Ultra Wide Band (UWB) technology and Visual-Inertial Odometry (VIO) to provide robust and low-drift localization. In order to include range measurements in state estimation, the position of the UWB anchors must be known. This study proposes a multi-step initialization procedure to map multiple unknown anchors by an Unmanned Aerial Vehicle (UAV), in a fully autonomous fashion. To address the limitations of initializing UWB anchors via a random trajectory, this paper uses the Geometric Dilution of Precision (GDOP) as a measure of optimality in anchor position estimation, to compute a set of optimal waypoints and synthesize a trajectory that minimizes the mapping uncertainty. After the initialization is complete, the range measurements from multiple anchors, including measurement biases, are tightly integrated into the VIO system. While in range of the initialized anchors, the VIO drift in position and heading is eliminated. The effectiveness of UVIO and our initialization procedure has been validated through a series of simulations and real-world experiments.

Time Series Classification (TSC) covers the supervised learning problem where input data is provided in the form of series of values observed through repeated measurements over time, and whose objective is to predict the category to which they belong. When the class values are ordinal, classifiers that take this into account can perform better than nominal classifiers. Time Series Ordinal Classification (TSOC) is the field covering this gap, yet unexplored in the literature. There are a wide range of time series problems showing an ordered label structure, and TSC techniques that ignore the order relationship discard useful information. Hence, this paper presents a first benchmarking of TSOC methodologies, exploiting the ordering of the target labels to boost the performance of current TSC state-of-the-art. Both convolutional- and deep learning-based methodologies (among the best performing alternatives for nominal TSC) are adapted for TSOC. For the experiments, a selection of 29 ordinal problems from two well-known archives has been made. In this way, this paper contributes to the establishment of the state-of-the-art in TSOC. The results obtained by ordinal versions are found to be significantly better than current nominal TSC techniques in terms of ordinal performance metrics, outlining the importance of considering the ordering of the labels when dealing with this kind of problems.

Time Series Classification (TSC) is an extensively researched field from which a broad range of real-world problems can be addressed obtaining excellent results. One sort of the approaches performing well are the so-called dictionary-based techniques. The Temporal Dictionary Ensemble (TDE) is the current state-of-the-art dictionary-based TSC approach. In many TSC problems we find a natural ordering in the labels associated with the time series. This characteristic is referred to as ordinality, and can be exploited to improve the methods performance. The area dealing with ordinal time series is the Time Series Ordinal Classification (TSOC) field, which is yet unexplored. In this work, we present an ordinal adaptation of the TDE algorithm, known as ordinal TDE (O-TDE). For this, a comprehensive comparison using a set of 18 TSOC problems is performed. Experiments conducted show the improvement achieved by the ordinal dictionary-based approach in comparison to four other existing nominal dictionary-based techniques.

Although CLIP-like Visual Language Models provide a functional joint feature space for image and text, due to the limitation of the CILP-like model's image input size (e.g., 224), subtle details are lost in the feature representation if we input high-resolution images (e.g., 2240). Our proposed framework addresses this issue by generating a single feature representation for a high-resolution image that retains image details from different scales while sharing the same semantic space as the original CLIP. An application scenario is remote sensing text-image retrieval, where targets (e.g., vehicles and ships) often appear at tiny scales. To achieve this, we develop a feature fusion model that relies on CLIP features extracted from a carefully designed image patch method, dubbed Complete Cover. This method ensures comprehensive coverage of objects across various scales and is weakly supervised by image-agnostic class prompted queries. We evaluate our framework's performance using real-world and synthetic datasets, demonstrating significant improvements in image retrieval tasks based on class prompted queries. To further showcase our framework's capability in detail retrieval, we introduce a CLEVR-like synthetic dataset, named CLVER-DS. This fully annotated dataset offers a controllable object scale, allowing for a more thorough examination of our approach's effectiveness.Our code is publicly available at https://github.com/zilunzhang/DetailCLIP

We study the process-level dynamics of Thompson sampling and related sampling-based bandit algorithms in the ``small gap'' regime, where the gaps between the arm means are of order $\sqrtγ$ or smaller and the time horizon is of order $1/γ$, with $γ\downarrow 0$. In this regime, as $γ\downarrow 0$, we show that the process-level dynamics of such algorithms converge weakly to the solutions to certain stochastic differential equations and stochastic ordinary differential equations. Our weak convergence theory is developed using the Continuous Mapping Theorem, which provides a direct and modular theoretical approach that can be adapted to analyze a variety of sampling-based bandit algorithms and handle weakly dependent reward processes. A central finding is an algorithmic invariance principle: in the small gap regime, the limit dynamics of a broad class of sampling-based algorithms -- including Thompson sampling with general single-parameter exponential family likelihoods, as well as non-parametric bandit algorithms based on bootstrap re-sampling -- all coincide with those of Thompson sampling with Gaussian likelihoods. Moreover, in the small gap regime, the regret performance of these algorithms is generally insensitive to model mis-specification, changing continuously with increasing degrees of mis-specification.

Time series clustering is the process of grouping time series with respect to their similarity or characteristics. Previous approaches usually combine a specific distance measure for time series and a standard clustering method. However, these approaches do not take the similarity of the different subsequences of each time series into account, which can be used to better compare the time series objects of the dataset. In this paper, we propose a novel technique of time series clustering based on two clustering stages. In a first step, a least squares polynomial segmentation procedure is applied to each time series, which is based on a growing window technique that returns different-length segments. Then, all the segments are projected into same dimensional space, based on the coefficients of the model that approximates the segment and a set of statistical features. After mapping, a first hierarchical clustering phase is applied to all mapped segments, returning groups of segments for each time series. These clusters are used to represent all time series in the same dimensional space, after defining another specific mapping process. In a second and final clustering stage, all the time series objects are grouped. We consider internal clustering quality to automatically adjust the main parameter of the algorithm, which is an error threshold for the segmenta- tion. The results obtained on 84 datasets from the UCR Time Series Classification Archive have been compared against two state-of-the-art methods, showing that the performance of this methodology is very promising.

Paralinguistic cues are essential for natural human-computer interaction, yet their evaluation in Large Audio-Language Models (LALMs) remains limited by coarse feature coverage and the inherent subjectivity of assessment. To address these challenges, we introduce SpeechParaling-Bench, a comprehensive benchmark for paralinguistic-aware speech generation. It expands existing coverage from fewer than 50 to over 100 fine-grained features, supported by more than 1,000 English-Chinese parallel speech queries, and is organized into three progressively challenging tasks: fine-grained control, intra-utterance variation, and context-aware adaptation. To enable reliable evaluation, we further develop a pairwise comparison pipeline, in which candidate responses are evaluated against a fixed baseline by an LALM-based judge. By framing evaluation as relative preference rather than absolute scoring, this approach mitigates subjectivity and yields more stable and scalable assessments without costly human annotation. Extensive experiments reveal substantial limitations in current LALMs. Even leading proprietary models struggle with comprehensive static control and dynamic modulation of paralinguistic features, while failure to correctly interpret paralinguistic cues accounts for 43.3% of errors in situational dialogue. These findings underscore the need for more robust paralinguistic modeling toward human-aligned voice assistants.

Recent advances in video generative models enable the synthesis of realistic human-object interaction videos across a wide range of scenarios and object categories, including complex dexterous manipulations that are difficult to capture with motion capture systems. While the rich interaction knowledge embedded in these synthetic videos holds strong potential for motion planning in dexterous robotic manipulation, their limited physical fidelity and purely 2D nature make them difficult to use directly as imitation targets in physics-based character control. We present DeVI (Dexterous Video Imitation), a novel framework that leverages text-conditioned synthetic videos to enable physically plausible dexterous agent control for interacting with unseen target objects. To overcome the imprecision of generative 2D cues, we introduce a hybrid tracking reward that integrates 3D human tracking with robust 2D object tracking. Unlike methods relying on high-quality 3D kinematic demonstrations, DeVI requires only the generated video, enabling zero-shot generalization across diverse objects and interaction types. Extensive experiments demonstrate that DeVI outperforms existing approaches that imitate 3D human-object interaction demonstrations, particularly in modeling dexterous hand-object interactions. We further validate the effectiveness of DeVI in multi-object scenes and text-driven action diversity, showcasing the advantage of using video as an HOI-aware motion planner.

Modern language models demonstrate impressive coding capabilities in common programming languages (PLs), such as C++ and Python, but their performance in lower-resource PLs is often limited by training data availability. In principle, however, most programming skills are universal across PLs, so the capability acquired in one PL should transfer to others. In this work, we propose the task of zero-shot cross-programming-language transfer for code RL. We find that, for Llama-3.1, RL training for code generation in a source PL fails to improve, and sometimes even degrades, the performance on other target PLs. To address this, we hypothesize that effective RL transfer requires a generalizable SFT initialization before RL. We thus propose **Parallel-SFT**, an SFT strategy that incorporates "parallel programs" -- functionally equivalent code implemented in multiple PLs -- into the data mixture. We demonstrate that this improves transferability: when we subsequently perform RL on our Parallel-SFT model, we observe better generalization to unseen PLs. Analysis of the model internal representations reveals that Parallel-SFT leads to a more functionality-centric latent space, where equivalent programs across PLs are more tightly clustered, which we hypothesize to contribute to the improved transferability.

Federated learning (FL) enables collaborative model training without sharing raw data; however, the presence of noisy labels across distributed clients can severely degrade the learning performance. In this paper, we propose FedSIR, a multi-stage framework for robust FL under noisy labels. Different from existing approaches that mainly rely on designing noise-tolerant loss functions or exploiting loss dynamics during training, our method leverages the spectral structure of client feature representations to identify and mitigate label noise. Our framework consists of three key components. First, we identify clean and noisy clients by analyzing the spectral consistency of class-wise feature subspaces with minimal communication overhead. Second, clean clients provide spectral references that enable noisy clients to relabel potentially corrupted samples using both dominant class directions and residual subspaces. Third, we employ a noise-aware training strategy that integrates logit-adjusted loss, knowledge distillation, and distance-aware aggregation to further stabilize federated optimization. Extensive experiments on standard FL benchmarks demonstrate that FedSIR consistently outperforms state-of-the-art methods for FL with noisy labels. The code is available at https://github.com/sinagh72/FedSIR.

The central problem in biomedical imaging are batch effects: systematic technical variations unrelated to the biological signal of interest. These batch effects critically undermine experimental reproducibility and are the primary cause of failure of deep learning systems on new experimental batches, preventing their practical use in the real world. Despite years of research, no method has succeeded in closing this performance gap for deep learning models. We propose Control-Stabilized Adaptive Risk Minimization via Batch Normalization (CS-ARM-BN), a meta-learning adaptation method that exploits negative control samples. Such unperturbed reference images are present in every experimental batch by design and serve as stable context for adaptation. We validate our novel method on Mechanism-of-Action (MoA) classification, a crucial task for drug discovery, on the large-scale JUMP-CP dataset. The accuracy of standard ResNets drops from 0.939 $\pm$ 0.005, on the training domain, to 0.862 $\pm$ 0.060 on data from new experimental batches. Foundation models, even after Typical Variation Normalization, fail to close this gap. We are the first to show that meta-learning approaches close the domain gap by achieving 0.935 $\pm$ 0.018. If the new experimental batches exhibit strong domain shifts, such as being generated in a different lab, meta-learning approaches can be stabilized with control samples, which are always available in biomedical experiments. Our work shows that batch effects in bioimaging data can be effectively neutralized through principled in-context adaptation, which also makes them practically usable and efficient.

The offshore wind energy sector is expanding rapidly, increasing the need for independent, high-temporal-resolution monitoring of infrastructure deployment and operation at global scale. While Earth Observation based offshore wind infrastructure mapping has matured for spatial localization, existing open datasets lack temporally dense and semantically fine-grained information on construction and operational dynamics. We introduce a global Sentinel-1 synthetic aperture radar (SAR) time series data corpus that resolves deployment and operational phases of offshore wind infrastructure from 2016Q1 to 2025Q1. Building on an updated object detection workflow, we compile 15,606 time series at detected infrastructure locations, with overall 14,840,637 events as analysis-ready 1D SAR backscatter profiles, one profile per Sentinel-1 acquisition and location. To enable direct use and benchmarking, we release (i) the analysis ready 1D SAR profiles, (ii) event-level baseline semantic labels generated by a rule-based classifier, and (iii) an expert-annotated benchmark dataset of 553 time series with 328,657 event labels. The baseline classifier achieves a macro F1 score of 0.84 in event-wise evaluation and an area under the collapsed edit similarity-quality threshold curve (AUC) of 0.785, indicating temporal coherence. We demonstrate that the resulting corpus supports global-scale analyses of deployment dynamics, the identification of differences in regional deployment patterns, vessel interactions, and operational events, and provides a reference for developing and comparing time series classification methods for offshore wind infrastructure monitoring.

The scalability of long-context large language models is fundamentally limited by the quadratic memory cost of exact self-attention, which often leads to out-of-memory (OOM) failures on modern hardware. Existing methods improve memory efficiency to near-linear complexity, while assuming that the full query, key, and value tensors fit in device memory. In this work, we remove this assumption by introducing CQS Divide, an operation derived from cyclic quorum sets (CQS) theory that decomposes attention into a set of independent subsequence computations whose recomposition yields exactly the same result as full-sequence attention. Exploiting this decomposition, we introduce Stream-CQSA, a memory-adaptive scheduling framework that partitions attention into subproblems that fit within arbitrary memory budgets. This recasts attention from a logically monolithic operation into a collection of schedulable tasks, enabling flexible execution across devices without inter-device communication. Experiments demonstrate predictable memory scaling and show that exact attention over billion-token sequences can be executed on a single GPU via streaming, without changing the underlying mathematical definition of attention or introducing approximation error.

Language models trained on natural text learn to represent numbers using periodic features with dominant periods at $T=2, 5, 10$. In this paper, we identify a two-tiered hierarchy of these features: while Transformers, Linear RNNs, LSTMs, and classical word embeddings trained in different ways all learn features that have period-$T$ spikes in the Fourier domain, only some learn geometrically separable features that can be used to linearly classify a number mod-$T$. To explain this incongruity, we prove that Fourier domain sparsity is necessary but not sufficient for mod-$T$ geometric separability. Empirically, we investigate when model training yields geometrically separable features, finding that the data, architecture, optimizer, and tokenizer all play key roles. In particular, we identify two different routes through which models can acquire geometrically separable features: they can learn them from complementary co-occurrence signals in general language data, including text-number co-occurrence and cross-number interaction, or from multi-token (but not single-token) addition problems. Overall, our results highlight the phenomenon of convergent evolution in feature learning: A diverse range of models learn similar features from different training signals.

Reinforcement Learning (RL) post-training has become the standard for aligning generative models with human preferences, yet most methods rely on a single scalar reward. When multiple criteria matter, the prevailing practice of ``early scalarization'' collapses rewards into a fixed weighted sum. This commits the model to a single trade-off point at training time, providing no inference-time control over inherently conflicting goals -- such as prompt adherence versus source fidelity in image editing. We introduce ParetoSlider, a multi-objective RL (MORL) framework that trains a single diffusion model to approximate the entire Pareto front. By training the model with continuously varying preference weights as a conditioning signal, we enable users to navigate optimal trade-offs at inference time without retraining or maintaining multiple checkpoints. We evaluate ParetoSlider across three state-of-the-art flow-matching backbones: SD3.5, FluxKontext, and LTX-2. Our single preference-conditioned model matches or exceeds the performance of baselines trained separately for fixed reward trade-offs, while uniquely providing fine-grained control over competing generative goals.

Transformer-based OCR models have shown strong performance on Latin and CJK scripts, but their application to African syllabic writing systems remains limited. We present the first adaptation of TrOCR for printed Tigrinya using the Ge'ez script. Starting from a pre-trained model, we extend the byte-level BPE tokenizer to cover 230 Ge'ez characters and introduce Word-Aware Loss Weighting to resolve systematic word-boundary failures that arise when applying Latin-centric BPE conventions to a new script. The unmodified model produces no usable output on Ge'ez text. After adaptation, the TrOCR-Printed variant achieves 0.22% Character Error Rate and 97.20% exact match accuracy on a held-out test set of 5,000 synthetic images from the GLOCR dataset. An ablation study confirms that Word-Aware Loss Weighting is the critical component, reducing CER by two orders of magnitude compared to vocabulary extension alone. The full pipeline trains in under three hours on a single 8 GB consumer GPU. All code, model weights, and evaluation scripts are publicly released.

As LLMs are increasingly integrated into agentic systems, they must adhere to dynamically defined, machine-interpretable interfaces. We evaluate LLMs as in-context interpreters: given a novel context-free grammar, can LLMs generate syntactically valid, behaviorally functional, and semantically faithful outputs? We introduce RoboGrid, a framework that disentangles syntax, behavior, and semantics through controlled stress-tests of recursion depth, expression complexity, and surface styles. Our experiments reveal a consistent hierarchical degradation: LLMs often maintain surface syntax but fail to preserve structural semantics. Despite the partial mitigation provided by CoT reasoning, performance collapses under structural density, specifically deep recursion and high branching, with semantic alignment vanishing at extreme depths. Furthermore, "Alien" lexicons reveal that LLMs rely on semantic bootstrapping from keywords rather than pure symbolic induction. These findings pinpoint critical gaps in hierarchical state-tracking required for reliable, grammar-agnostic agents.

Large vision-language models (LVLMs) have made substantial advances in reasoning tasks at the Olympiad level. Nevertheless, current Olympiad-level multimodal reasoning benchmarks for these models often emphasize single-image analysis and fail to exploit contextual information across multiple images. We present OMIBench, a benchmark designed to evaluate Olympiad-level reasoning when the required evidence is distributed over multiple images. It contains problems from biology, chemistry, mathematics, and physics Olympiads, together with manually annotated rationales and evaluation protocols for both exact and semantic answer matching. Across extensive experiments on OMIBench, we observe meaningful performance gaps in existing models. Even the strongest LVLMs, such as Gemini-3-Pro, attain only about 50% on the benchmark. These results position OMIBench as a focused resources for studying and improving multi-image reasoning in LVLMs.