World models are central to building AI agents capable of flexible reasoning and planning. Yet current evaluations (i) test only properties measurable from observed interactions, such as next-frame prediction or task return, and (ii) do not test whether a learned model supports diverse queries about the environment. In contrast, humans build $\textit{general-purpose}$ models that can answer many different questions about an environment$\unicode{x2014}$including questions that require understanding global structure and counterfactual consequences. We propose $\textit{WorldTest}$: a protocol for evaluating whether agents learn models that support multiple $\textit{environment-level queries}\unicode{x2014}$questions whose answers depend on properties of the full environment, not just observed trajectories. Individually, these queries can target properties (e.g., reachability or the effects of interventions) that no single rollout distribution determines. Collectively, they assess model generality across query types. We instantiate WorldTest as $\textit{AutumnBench}$, a benchmark of 43 interactive grid-world environments and 129 tasks across three query families for both humans and learning agents. Experiments with 517 human participants and five frontier models show that humans substantially outperform these models, a gap we attribute to differences in exploration and belief updating. AutumnBench provides a framework for evaluating world-model learning in grid-world environments with environment-level queries, and WorldTest provides a template for extending such evaluations to richer domains.

DINOv2 is routinely deployed to recognize objects, scenes, and actions; yet the nature of what it perceives remains unknown. As a working baseline, we adopt the Linear Representation Hypothesis (LRH) and operationalize it using SAEs, producing a 32,000-unit dictionary that serves as the interpretability backbone of our study, which unfolds in three parts. In the first part, we analyze how different downstream tasks recruit concepts from our learned dictionary, revealing functional specialization: classification exploits "Elsewhere" concepts that fire everywhere except on target objects, implementing learned negations; segmentation relies on boundary detectors forming coherent subspaces; depth estimation draws on three distinct monocular depth cues matching visual neuroscience principles. Following these functional results, we analyze the geometry and statistics of the concepts learned by the SAE. We found that representations are partly dense rather than strictly sparse. The dictionary evolves toward greater coherence and departs from maximally orthogonal ideals (Grassmannian frames). Within an image, tokens occupy a low dimensional, locally connected set persisting after removing position. These signs suggest representations are organized beyond linear sparsity alone. Synthesizing these observations, we propose a refined view: tokens are formed by combining convex mixtures of archetypes (e.g., a rabbit among animals, brown among colors, fluffy among textures). This structure is grounded in Gardenfors' conceptual spaces and in the model's mechanism as multi-head attention produces sums of convex mixtures, defining regions bounded by archetypes. We introduce the Minkowski Representation Hypothesis (MRH) and examine its empirical signatures and implications for interpreting vision-transformer representations.

Despite demonstrating remarkable performance across a wide range of tasks, large language models (LLMs) have also been found to frequently produce outputs that are incomplete or selectively omit key information. In sensitive domains, such omissions can result in significant harm comparable to that posed by factual inaccuracies, including hallucinations. In this study, we address the challenge of evaluating the comprehensiveness of LLM-generated texts, focusing on the detection of missing information or underrepresented viewpoints. We investigate three automated evaluation metrics: (1) an NLI-based method that decomposes texts into atomic statements and uses natural language inference (NLI) to identify missing facts, (2) a Q&A-based metric that extracts question-answer pairs and compares responses across sources, and (3) an end-to-end approach that directly identifies missing content using LLMs. Our experiments demonstrate the surprising effectiveness of the simple end-to-end metric compared to more complex metrics, though at the cost of reduced robustness, interpretability and result granularity. We further assess the comprehensiveness of responses from several popular open-weight LLMs when answering user queries based on multiple sources.

Reasoning distillation has emerged as a prevailing paradigm for transferring reasoning capabilities from large reasoning models to small language models. Yet, reasoning distillation risks data contamination: benchmark data may inadvertently be included in the distillation data, thereby inflating model performance metrics. In this work, we formally define the distillation data detection task, which determines whether a given question is included in the model's distillation data. The unique challenge of this task lies in the partial availability of distillation data. To address this, we propose Token Probability Deviation (TPD), a detection method that leverages the probability patterns of output tokens generated by the model instead of input tokens. Our method is motivated by the observation that seen questions tend to elicit more near-deterministic tokens generated by the models than unseen ones. Our TPD score is thus designed to quantify the token-level deviation of generated tokens from a high-confidence reference probability. Consequently, seen questions can yield substantially lower TPD scores than unseen ones, enabling strong detection performance. Extensive experiments demonstrate the effectiveness of our approach, improving detection AUC by up to 31% on distillation datasets.

We present InvThink, a training and prompting framework that requires the model to enumerate, analyze, and constrain potential failures before generating its final response. Unlike existing safety alignment methods that optimize only for safe final responses, InvThink structures generation into three steps: (1) enumerate potential harms, (2) analyze their consequences, (3) generate the response under explicit mitigation constraints. We observe three findings: (i) InvThink shows higher safety scores at larger model sizes, compared to existing safety prompting and alignment baselines. (ii) InvThink mitigates the safety tax. Models trained with INVTHINK preserve their reasoning capability on standard benchmarks. (iii) beyond general safety tasks, InvThink also reduces harmful behavior in professional ethics domains (medicine, finance, law) and in agentic misalignment scenarios, achieving up to 32% reduction in harmfulness over zero-shot baselines and 16% over SafetyPrompt. We extend InvThink with supervised fine-tuning, and GRPO-based reinforcement learning across three LLM families.

The long-output context generation of large reasoning models enables extended chain of thought (CoT) but also drives rapid growth of the key-value (KV) cache, quickly overwhelming GPU memory. To address this challenge, we propose ThinKV, a thought-adaptive KV cache compression framework. ThinKV is based on the observation that attention sparsity reveals distinct thought types with varying importance within the CoT. It applies a hybrid quantization-eviction strategy, assigning token precision by thought importance and progressively evicting tokens from less critical thoughts as reasoning trajectories evolve. Furthermore, to implement ThinKV, we design a kernel that extends PagedAttention to enable efficient reuse of evicted tokens' memory slots, eliminating compaction overheads. Extensive experiments on DeepSeek-R1-Distill, GPT-OSS, and NVIDIA AceReason across mathematics and coding benchmarks show that ThinKV achieves near-lossless accuracy with less than 5% of the original KV cache, while improving performance with up to 5.8x higher inference throughput over state-of-the-art baselines.

Automated approaches to answer patient-posed health questions are rising, but selecting among systems requires reliable evaluation. The current gold standard for evaluating the free-text artificial intelligence (AI) responses--human expert review--is labor-intensive and slow, limiting scalability. Automated metrics are promising yet variably aligned with human judgments and often context-dependent. To address the feasibility of automating the evaluation of AI responses to hospitalization-related questions posed by patients, we conducted a large systematic study of evaluation approaches. Across 100 patient cases, we collected responses from 28 AI systems (2800 total) and assessed them along three dimensions: whether a system response (1) answers the question, (2) appropriately uses clinical note evidence, and (3) uses general medical knowledge. Using clinician-authored reference answers to anchor metrics, automated rankings closely matched human ratings. Our findings suggest that carefully designed automated evaluation can scale comparative assessment of AI systems and support patient-clinician communication.

Vision-language models achieve incredible performance across a wide range of tasks, but their large size makes inference costly. Recent work has shown that multimodal processing contains significant redundancies, making it possible to skip certain layers with minimal performance loss. Yet current pruning techniques remain ad-hoc, relying on heuristics or hyperparameter sweeps rather than principled criteria for determining when layer skipping is beneficial. In this paper, we propose a unified framework that characterizes the redundancy conditions under which pruning can enhance efficiency without sacrificing performance. Central to our approach are experimentally verifiable and interpretable notions of redundancy that can be evaluated without requiring downstream task performance as a metric. Applying this framework, we corroborate prior findings that both early and late vision tokens are redundant across models, and we validate our conditions by showing they align with actual performance degradation. Beyond these empirical results, our framework provides a theoretically grounded understanding of redundancy in VLMs and unifies many of the ideas behind modern layer-skipping techniques.

The evaluation of time series forecasting models is hindered by a lack of high-quality benchmarks, leading to overestimated assessments of progress. Existing datasets suffer from issues ranging from small-scale, low-frequency, pre-training data contamination in unimodal designs to the temporal and description leakage prevalent in early multimodal designs. To address this, we formalize the core principles of high-fidelity benchmarking, focusing on data sourcing integrity, leak-free design, and structural clarity. We introduce Fidel-TS, a new large-scale benchmark built from these principles. Our experiments reveal the limitations of prior benchmarks and the potential discrepancies in model evaluation, providing new insights into multiple existing unimodal and multimodal forecasting models and LLMs across various evaluation tasks.

State Space Models (SSMs) have emerged as efficient alternatives to Vision Transformers (ViTs), with VMamba standing out as a pioneering architecture designed for vision tasks. However, their generalization performance degrades significantly under distribution shifts. To address this limitation, we propose TRUST (Test-Time Refinement using Uncertainty-Guided SSM Traverses), a novel test-time adaptation (TTA) method that leverages diverse traversal permutations to generate multiple causal perspectives of the input image. Model predictions serve as pseudo-labels to guide updates of the Mamba-specific parameters, and the adapted weights are averaged to integrate the learned information across traversal scans. Altogether, TRUST is the first approach that explicitly leverages the unique architectural properties of SSMs for adaptation. Experiments on seven benchmarks show that TRUST consistently improves robustness and outperforms existing TTA methods.

We identify a fundamental incompatibility between the goals of accuracy, trust, and human-level reasoning in artificial intelligence (AI) systems, for strict mathematical definitions of these notions. We define accuracy of a system as the property that it never makes any false claims when it has the ability to abstain from making a prediction on any input, and trust as the assumption that the system is accurate. We define human-level reasoning as the property of an AI system always matching or exceeding human capability. Our core finding is that -- for our formal definitions of these notions -- an accurate and trusted AI system cannot be a human-level reasoning system: for such an accurate, trusted system there are task instances which are easily and provably solvable by a human but not by the system. Our proofs draw parallels to Gödel's incompleteness theorems and Turing's proof of the undecidability of the halting problem, and can be regarded as interpretations of Gödel's and Turing's results. Key to our proof is the formalization of the notion of trust, which allows us to separate the intrinsic property of a system (being accurate) from its epistemic status (being trusted).

Parameter-efficient fine-tuning (PEFT) of large language models trains a small task-specific parameter set while keeping the pretrained model frozen. The dominant Low-Rank Adaptation (LoRA) family makes this trade-off practical; however, evaluations under the same parameter budget assess single-task accuracy. In sequential adaptation settings, such evaluations should also measure how well performance on the first-stage task is retained after subsequent fine-tuning. To address this gap, we introduce BoHA, a blockwise $W_0$-coupled Hadamard product adapter that treats spatial support as an explicit design axis. BoHA partitions the frozen weight $W_0$ into a $b{\times}b$ grid and learns an independent low-rank Hadamard product factor in each block, preserving a matched LoRA-equivalent total rank with adapter-free merged inference. On a synthetic target, BoHA at per-block rank $r_b{=}1$ exactly reconstructs an update that requires rank $b^2$ under the global $W_0$-coupled Hadamard parameterization. Across Llama-3.2-1B/3B, Mistral-7B, and Gemma-2-9B on commonsense and arithmetic reasoning tasks, BoHA outperforms LoRA across all matched-budget single-task averages and remains competitive with the strongest Hadamard baseline. On a Llama-3.2-3B commonsense $\to$ arithmetic continual-learning diagnostic, BoHA retains $57.66\%$ first-stage accuracy and exceeds the $W_0$-free additive-control mean by $15.23\%$ under matched second-stage plasticity. These results demonstrate that blockwise $W_0$-coupled Hadamard adaptation is a competitive PEFT design choice when retention under sequential adaptation is part of the objective.

Inverse reinforcement learning (IRL) aims to infer rewards from observed behavior, but rewards are not identified from the policy alone: many reward--value pairs can rationalize the same actions. Meaningful reward recovery therefore requires a normalization, yet existing normalized IRL methods often rely on anchor-action restrictions or specialized neural architectures. We study reward recovery in the maximum-entropy, or Gumbel-shock, model under a broad class of statewise affine normalizations, with anchor-action constraints as a special case. This yields Generalized Policy-to-$Q$-to-Reward (GenPQR), a modular procedure that estimates the behavior policy, evaluates its soft $Q$-function through the Bellman equation, and recovers the normalized reward. Both stages can be implemented with off-the-shelf classification and regression methods. We prove modular finite-sample guarantees under general function approximation, with separate policy-estimation and $Q$-estimation errors. As a concrete instantiation, we study GenPQR with fitted $Q$-evaluation, reducing IRL to policy estimation followed by regression. Experiments show that GenPQR matches or improves reward recovery relative to DeepPQR while remaining simpler and more modular. Compared with DeepPQR, our theory goes beyond anchor actions, accommodates large and continuous action spaces, makes coverage requirements explicit, and is not tied to a specific neural-network architecture or training procedure.

Animal behavior analysis plays a crucial role in understanding animal welfare, health status, and productivity in agricultural settings. However, traditional manual observation methods are time-consuming, subjective, and limited in scalability. We present a modular pipeline that leverages open-sourced state-of-the-art computer vision techniques to automate animal behavior analysis in a group housing environment. Our approach combines state-of-the-art models for zero-shot object detection, motion-aware segmentation and tracking, and advanced feature extraction using vision transformers for robust behavior recognition. The pipeline addresses challenges including animal occlusions and group housing scenarios, as demonstrated in indoor pig monitoring. We validated our system on the Edinburgh Pig Behavior Video Dataset for multiple behavioral tasks. Our temporal model achieved 94.2% overall accuracy, representing a 21.2 percentage point improvement over existing methods. The pipeline demonstrated robust tracking capabilities with a 93.3% identity preservation (IDF1) score and an 89.3% average precision (AP) for object detection. The modular design suggests potential for adaptation to other contexts, though further validation across species would be required. The open-source implementation provides a scalable solution for behavior monitoring, contributing to precision pig farming and welfare assessment through automated, objective, and continuous analysis.

Customer feedback in industrial forums offers rich but underexplored insights into real-world product experience. Yet systematic analysis remains challenging due to unstructured, domain-specific content and the scarcity of high-quality labeled datasets. This paper presents the User eXperience Perception Insights Dataset (UXPID), a collection of 7130 synthesized and anonymized user feedback branches extracted from a public industrial automation forum. Each JSON record contains multi-post comments enriched with metadata and annotated by a large language model (LLM) for UX insights, user expectations, severity ratings, sentiment, and topic classifications. UXPID is designed to facilitate research in user requirements, user experience (UX) analysis, and AI-driven feedback processing, particularly where privacy and licensing restrictions limit access to real-world data. It supports the training and evaluation of transformer-based models for tasks such as issue detection, sentiment analysis, and requirements extraction in technical forums, providing a valuable resource for advancing NLP methods within industrial product support and software engineering domains.

Early-exit cascades over a frozen convolutional backbone enable adaptive inference but suffer from three sources of train-inference mismatch: branches train on samples they will never see at inference, their per-class precision thresholds are calibrated on the wrong distribution, and the standard cross-entropy target on backbone argmax labels discards the backbone's uncertainty signal. We close all three gaps with CalexNet (Cascade-Aligned Early eXits), a training-recipe-only modification: branches train under continuously-weighted importance sampling that matches the cascade-survivor distribution; per-class precision thresholds are calibrated on the actual cascade-survivor subset of the validation set; and the classification head is trained against the backbone's full softmax via a temperature-scaled KL objective. Combined with an augmented prototype-pooling branch head, CalexNet is evaluated on ResNet18 and ResNet50 backbones across CIFAR-100 (20-superclass coarse, the harder primary setting) and CINIC-10 (10-class, the easier cross-validation counterpart). On the accuracy-FLOPs Pareto frontier, CalexNet matches or exceeds three published baselines (PTEEnet, ZTW, BoostNet) and a within-paper "no-alignment, no-KD" reference. The largest gains appear in the practically relevant 30-70% FLOPs-reduction regime and are stable across n=3 training seeds. CalexNet requires no inference-time architectural change and is a drop-in for any frozen-backbone early-exit cascade.

Mainstream Transformer-based large language models face major efficiency bottlenecks: training computation scales quadratically with sequence length, and inference memory grows linearly, limiting long-context processing. Building large models on non-NVIDIA platforms also poses challenges for stable and efficient training. To address this, we introduce SpikingBrain, a family of brain-inspired models designed for efficient long-context training and inference. SpikingBrain leverages the MetaX GPU cluster and focuses on three aspects: (1) Model Architecture: linear and hybrid-linear attention architectures with adaptive spiking neurons; (2) Algorithmic Optimizations: an efficient, conversion-based training pipeline and a dedicated spike coding framework; (3) System Engineering: customized training frameworks, operator libraries, and parallelism strategies tailored to MetaX hardware. Using these techniques, we develop two models: SpikingBrain-7B, a linear LLM, and SpikingBrain-76B, a hybrid-linear MoE LLM. These models demonstrate the feasibility of large-scale LLM development on non-NVIDIA platforms, and training remains stable for weeks on hundreds of MetaX GPUs with Model FLOPs Utilization at expected levels. SpikingBrain achieves performance comparable to open-source Transformer baselines while using only about 150B tokens for continual pre-training. Our models also significantly improve long-context efficiency and deliver inference with (partially) constant memory and event-driven spiking behavior. For example, SpikingBrain-7B attains over 100x speedup in Time to First Token for 4M-token sequences. Furthermore, the proposed spiking scheme achieves 69.15 percent sparsity, enabling low-power operation. Overall, this work demonstrates the potential of brain-inspired mechanisms to drive the next generation of efficient and scalable large model design.

We introduce sparse autoencoder neural operators (SAE-NOs), a new class of sparse autoencoders that operate in function spaces rather than fixed-dimensional Euclidean representations. We formalize the functional representation hypothesis, where data are explained through sparse compositions of structured functions. Unlike standard SAEs that represent concepts with scalar activations, SAE-NOs parameterize concepts as functions, enabling representations that capture not only a concept's presence, but also how and where it is expressed across the input domain. We achieve this through joint sparsity: concept sparsity selects active concepts, while domain sparsity governs where they are expressed. We instantiate this framework using Fourier neural operators (SAE-FNOs), parameterizing concepts as integral operators in the Fourier domain. This functional and spectral parameterization is particularly advantageous when data exhibit spatial structure across scales or when concepts are frequency-structured. We characterize SAE-FNO on vision data and demonstrate that it learns localized patterns, uses concepts more efficiently, and exhibits stable concept characteristics across sparsity levels. We further show that SAE-FNO adapts to changes in domain size and generalizes across discretizations, operating at resolutions beyond those seen during training, where standard SAEs fail. We also introduce lifting into SAEs and show theoretically and empirically that it acts as a preconditioner that accelerates optimization. Overall, our results show that moving from vector-valued to functional parameterizations, with concept and domain sparsity, extends SAEs from representing concept presence to modeling structured concept expression, highlighting the importance of parameterization.

In recent years, with the large-scale expansion of graph data, there has been an increased focus on Riemannian manifold data spaces other than Euclidean space. In particular, the development of hyperbolic spaces has been remarkable, and they have high expressive power for graph data with hierarchical structures. Normalized Maximum Likelihood (NML) is employed in regret minimization and model selection. However, existing formulations of NML have been developed primarily in Euclidean spaces and are inherently dependent on the choice of coordinate systems, making it non-trivial to extend NML to Riemannian manifolds. In this study, we define a new NML that reflects the geometric structure of Riemannian manifolds, called the Riemannian manifold NML (Rm-NML). This Rm-NML is invariant under coordinate transformations and coincides with the conventional NML under the natural parameterization in Euclidean space. We extend existing computational techniques for NML to the setting of Riemannian manifolds. Furthermore, we derive a method to simplify the computation of Rm-NML on Riemannian symmetric spaces, which encompass data spaces of growing interest such as hyperbolic spaces. To illustrate the practical application of our proposed method, we explicitly computed the Rm-NML for normal distributions on hyperbolic spaces.

Foundation models pre-trained on large-scale natural image datasets offer a powerful paradigm for medical image segmentation. However, effectively transferring their learned representations for precise clinical applications remains a challenge. In this work, we propose Dino U-Net, a novel encoder-decoder architecture designed to exploit the high-fidelity dense features of the DINOv3 vision foundation model. Our architecture introduces an encoder built upon a frozen DINOv3 backbone, which employs a specialized adapter to fuse the model's rich semantic features with low-level spatial details. To preserve the quality of these representations during dimensionality reduction, we design a new fidelity-aware projection module (FAPM) that effectively refines and projects the features for the decoder. We conducted extensive experiments on seven diverse public medical image segmentation datasets. Our results show that Dino U-Net achieves state-of-the-art performance, consistently outperforming previous methods across various imaging modalities. Our framework proves to be highly scalable, with segmentation accuracy consistently improving as the backbone model size increases up to the 7-billion-parameter variant. The findings demonstrate that leveraging the superior, dense-pretrained features from a general-purpose foundation model provides a highly effective and parameter-efficient approach to advance the accuracy of medical image segmentation. The code is available at https://github.com/yifangao112/DinoUNet.

In this paper, we describe and benchmark a competitor-discovery component used within an agentic AI system for fast drug asset due diligence. A competitor-discovery AI agent, given an indication, retrieves all drugs comprising the competitive landscape of that indication and extracts canonical attributes for these drugs. The competitor definition is investor-specific, and data is paywalled/licensed, fragmented across registries, ontology-mismatched by indication, alias-heavy for drug names, multimodal, and rapidly changing. Although considered the best tool for this problem, the current LLM-based AI systems aren't capable of reliably retrieving all competing drug names, and there is no accepted public benchmark for this task. To address the lack of evaluation, we use LLM-based agents to transform five years of multi-modal, unstructured diligence memos from a private biotech VC fund into a structured evaluation corpus mapping indications to competitor drugs with normalized attributes. We also introduce a competitor validating LLM-as-a-judge agent that filters out false positives from the list of predicted competitors to maximize precision and suppress hallucinations. On this benchmark, our competitor-discovery agent achieves 83% recall, exceeding OpenAI Deep Research (65%) and Perplexity Labs (60%). The system is deployed in production with enterprise users; in a case study with a biotech VC investment fund, analyst turnaround time dropped from 2.5 days to $\sim$3 hours ($\sim$20x) for the competitive analysis.

As the era of large language models (LLMs) unfolds, Preference Optimization (PO) methods have become a central approach to aligning LLMs with human preferences and improving performance. We propose Maximum a Posteriori Preference Optimization (MaPPO), a methodology for learning from preferences that explicitly incorporates prior reward knowledge into the optimization objective. Building on the paradigm employed by Direct Preference Optimization (DPO) and its variants of treating preference learning as a Maximum Likelihood Estimation (MLE) problem, MaPPO integrates prior reward estimates into a principled Maximum a Posteriori (MaP) objective. This not only generalizes DPO and its variants, but also enhances alignment by mitigating the oversimplified binary classification of responses. Additionally, MaPPO introduces no additional hyperparameters, and supports preference optimization in both offline and online settings. In addition, MaPPO can be used as a plugin for DPO variants, including widely used SimPO, IPO and CPO, and produce consistent improvements. Extensive empirical evaluations of different model sizes and model series on three standard benchmarks (MT-Bench, AlpacaEval 2.0, and Arena-Hard) demonstrate consistent improvements in alignment performance without sacrificing computational efficiency.

Text analytics has traditionally required specialized knowledge in Natural Language Processing (NLP) or text analysis, which presents a barrier for entry-level analysts. Recent advances in large language models (LLMs) have changed the landscape of NLP by enabling more accessible and automated text analysis (e.g., topic detection, summarization, information extraction, etc.). We introduce VIDEE, a system that supports entry-level data analysts to conduct advanced text analytics with intelligent agents. VIDEE instantiates a human-agent collaroration workflow consisting of three stages: (1) Decomposition, which incorporates a human-in-the-loop Monte-Carlo Tree Search algorithm to support generative reasoning with human feedback, (2) Execution, which generates an executable text analytics pipeline, and (3) Evaluation, which integrates LLM-based evaluation and visualizations to support user validation of execution results. We conduct two quantitative experiments to evaluate VIDEE's effectiveness and analyze common agent errors. A user study involving participants with varying levels of NLP and text analytics experience -- from none to expert -- demonstrates the system's usability and reveals distinct user behavior patterns. The findings identify design implications for human-agent collaboration, validate the practical utility of VIDEE for non-expert users, and inform future improvements to intelligent text analytics systems.

Counterfactual generation aims to simulate realistic hypothetical outcomes under causal interventions. Diffusion models have emerged as a powerful tool for this task, combining DDIM inversion with conditional generation and classifier-free guidance (CFG). In this work, we identify a key limitation of CFG for counterfactual generation: it prescribes a global guidance scale for all attributes, leading to significant spurious changes in inferred counterfactuals. To mitigate this, we propose Factored Classifier-Free Guidance (FCFG), a flexible and model-agnostic guidance technique that enables attribute-wise control following a causal graph. FCFG complements recent advances in classifier-free guidance and can be seamlessly extended to advanced guidance schemes such as CFG++ and APG. Our experiments demonstrate that FCFG significantly improves the axiomatic soundness of inferred counterfactuals across both natural and medical image datasets, mitigating spurious amplification effects, and enhancing counterfactual reversibility.

Reinforcement learning (RL) training of large language models (LLMs) on unverifiable tasks is challenging even when a reasonable-quality reference answer is available. We propose a constrained RL training framework that (i) optimizes a token-level dense Reasoning Reflection Reward (R3) aligned with reasoning quality, and (ii) enforces rubric-gating as feasibility constraints at the rollout group level. R3 measures the model's token-level certainty of a reference answer under its chain-of-thought (CoT) prefix, and selectively emphasizes tokens with high cross-rollout variance, which we call reasoning-reflective tokens, that would otherwise be diluted by the bulk of low-variance tokens. The same variance signal also drives a filter that discards queries with insufficient signal for comparative learning. Rubric-gating complements R3 by operationalizing principled task criteria as hard accept/reject checks on final answers. Empirically, across four datasets spanning scientific writing, medicine, legal contracts, and finance, our framework outperforms strong baselines, achieves faster, more sample-efficient learning, and respects feasibility constraints.

Dynamical systems modeling is a core pillar of scientific inquiry across natural and life sciences. Increasingly, dynamical system models are learned from data, rendering identifiability a paramount concept. For systems that are not identifiable from data, no guarantees can be given about their behavior under new conditions and inputs, or about possible control mechanisms to steer the system. It is known in the community that "linear ordinary differential equations (ODE) are almost surely identifiable from a single trajectory." However, this only holds for dense matrices. The sparse regime remains underexplored, despite its practical relevance with sparsity arising naturally in many biological, social, and physical systems. In this work, we address this gap by characterizing the identifiability of sparse linear ODEs. Contrary to the dense case, we show that sparse systems are unidentifiable with a positive probability in practically relevant sparsity regimes and provide lower bounds for this probability. We further study empirically how this theoretical unidentifiability manifests in state-of-the-art methods to estimate linear ODEs from data. Our results corroborate that sparse systems are also practically unidentifiable. Theoretical limitations are not resolved through inductive biases or optimization dynamics. Our findings call for rethinking what can be expected from data-driven dynamical system modeling and allows for quantitative assessments of how much to trust a learned linear ODE.

The ability to simulate the world in a spatially consistent manner is a crucial requirement for effective world models. Such a model enables high-quality visual generation, and also ensures the reliability of world models for downstream tasks such as simulation and planning. It must not only retain long-horizon observational information, but also enables the construction of explicit or implicit internal spatial representations. However, existing datasets do not explicitly enforce spatial consistency constraints, limiting both the ability to systematically evaluate this capability and to learn it through data-driven approaches. Furthermore, most existing benchmarks primarily emphasize visual coherence or generation quality, neglecting the requirement of long-range spatial consistency. To bridge this gap, we propose LoopNav, a dataset and corresponding benchmark centered on loop-based navigation for evaluating spatial consistency. The dataset comprises 250 hours (20 million frames) of loop-based navigation videos with actions, collected from diverse locations in the open-world environment of Minecraft. We further introduce a Scene Graph Consistency Score to quantify spatial consistency while remaining invariant to pixel-level variations. Dataset, benchmark, and code are open-sourced to support future research.

Transformer-based language models rely on positional encoding (PE) to handle token order and support context length extrapolation. However, existing PE methods lack theoretical clarity and rely on limited evaluation metrics to substantiate their extrapolation claims. We propose the Bayesian Attention Mechanism (BAM), a theoretical framework that formulates positional encoding as a prior within a probabilistic model. BAM unifies existing methods (e.g., NoPE and ALiBi) and motivates a new Generalized Gaussian positional prior that substantially improves long-context generalization. Empirically, BAM enables accurate information retrieval at $500\times$ the training context length, outperforming previous state-of-the-art context length generalization in long context retrieval accuracy while maintaining comparable perplexity and introducing minimal additional parameters.

Most machine learning-based image segmentation models produce pixel-wise confidence scores that represent the model's predicted probability for each class label at every pixel. While this information can be particularly valuable in high-stakes domains such as medical imaging, these scores are heuristic in nature and do not constitute rigorous quantitative uncertainty estimates. Conformal prediction (CP) provides a principled framework for transforming heuristic confidence scores into statistically valid uncertainty estimates. However, applying CP directly to image segmentation ignores the spatial correlations between pixels, a fundamental characteristic of image data. This can result in overly conservative and less interpretable uncertainty estimates. To address this, we propose CONSIGN (Conformal Segmentation Informed by Spatial Groupings via Decomposition), a CP-based method that incorporates spatial correlations to improve uncertainty quantification in image segmentation. Our method generates meaningful prediction sets that come with user-specified, high-probability error guarantees. It is compatible with any pre-trained segmentation model capable of generating multiple sample outputs. We evaluate CONSIGN against two CP baselines across three medical imaging datasets and two COCO dataset subsets, using three different pre-trained segmentation models. Results demonstrate that accounting for spatial structure significantly improves performance across multiple metrics and enhances the quality of uncertainty estimates.

Real world data often exhibits unknown, instance-specific symmetries that rarely exactly match a transformation group $G$ fixed a priori. Class-pose decompositions aim to create disentangled representations by factoring inputs into invariant features and a pose $g\in G$ defined relative to a training-dependent, arbitrary canonical representation. We introduce RECON, a class-pose agnostic canonical orientation normalization that corrects arbitrary canonicals via a simple right translation, yielding natural, data-aligned canonicalizations. This enables (i) unsupervised discovery of instance-specific pose distributions, (ii) detection of out-of-distribution poses and (iii) a plug-and-play test-time canonicalization layer. This layer can be attached on top of any pre-trained model to infuse group invariance, improving its performance without retraining. We validate on images and molecular ensembles, demonstrating accurate symmetry discovery, and matching or outperforming other canonicalizations in downstream classification.