Neural networks represent more features than they have dimensions via superposition, forcing features to share representational space. Current methods decompose activations into sparse linear features but discard geometric structure. We develop a theory for studying the geometric structre of features by analyzing the spectra (eigenvalues, eigenspaces, etc.) of weight derived matrices. In particular, we introduce the frame operator $F = WW^\top$, which gives us a spectral measure that describes how each feature allocates norm across eigenspaces. While previous tools could describe the pairwise interactions between features, spectral methods capture the global geometry (``how do all features interact?''). In toy models of superposition, we use this theory to prove that capacity saturation forces spectral localization: features collapse onto single eigenspaces, organize into tight frames, and admit discrete classification via association schemes, classifying all geometries from prior work (simplices, polygons, antiprisms). The spectral measure formalism applies to arbitrary weight matrices, enabling diagnosis of feature localization beyond toy settings. These results point toward a broader program: applying operator theory to interpretability.

High-fidelity generative models have narrowed the perceptual gap between synthetic and real images, posing serious threats to media security. Most existing AI-generated image (AIGI) detectors rely on artifact-based classification and struggle to generalize to evolving generative traces. In contrast, human judgment relies on stable real-world regularities, with deviations from the human cognitive manifold serving as a more generalizable signal of forgery. Motivated by this insight, we reformulate AIGI detection as a Reference-Comparison problem that verifies consistency with the real-image manifold rather than fitting specific forgery cues. We propose MIRROR (Manifold Ideal Reference ReconstructOR), a framework that explicitly encodes reality priors using a learnable discrete memory bank. MIRROR projects an input into a manifold-consistent ideal reference via sparse linear combination, and uses the resulting residuals as robust detection signals. To evaluate whether detectors reach the "superhuman crossover" required to replace human experts, we introduce the Human-AIGI benchmark, featuring a psychophysically curated human-imperceptible subset. Across 14 benchmarks, MIRROR consistently outperforms prior methods, achieving gains of 2.1% on six standard benchmarks and 8.1% on seven in-the-wild benchmarks. On Human-AIGI, MIRROR reaches 89.6% accuracy across 27 generators, surpassing both lay users and visual experts, and further approaching the human perceptual limit as pretrained backbones scale. The code is publicly available at: https://github.com/349793927/MIRROR

Regular sound correspondences constitute the principal evidence in historical language comparison. Despite the heuristic focus on regularity, it is often more an intuitive judgement than a quantified evaluation, and irregularity is more common than expected from the Neogrammarian model. Given the recent progress of computational methods in historical linguistics and the increased availability of standardized lexical data, we are now able to improve our workflows and provide such a quantitative evaluation. Here, we present the balanced average recurrence of correspondence patterns as a new measure of regularity. We also present a new computational method that uses this measure to identify cognate sets that lack regularity with respect to their correspondence patterns. We validate the method through two experiments, using simulated and real data. In the experiments, we employ leave-one-out validation to measure the regularity of cognate sets in which one word form has been replaced by an irregular one, checking how well our method identifies the forms causing the irregularity. Our method achieves an overall accuracy of 85\% with the datasets based on real data. We also show the benefits of working with subsamples of large datasets and how increasing irregularity in the data influences our results. Reflecting on the broader potential of our new regularity measure and the irregular cognate identification method based on it, we conclude that they could play an important role in improving the quality of existing and future datasets in computer-assisted language comparison.

Large language models (LLMs) are now widely used to evaluate the quality of text, a field commonly referred to as LLM-as-a-judge. While prior works mainly focus on point-wise and pair-wise evaluation paradigms. Rubric-based evaluation, where LLMs select a score from multiple rubrics, has received less analysis. In this work, we show that rubric-based evaluation implicitly resembles a multi-choice setting and therefore has position bias: LLMs prefer score options appearing at specific positions in the rubric list. Through controlled experiments across multiple models and datasets, we demonstrate consistent position bias. To mitigate this bias, we propose a balanced permutation strategy that evenly distributes each score option across positions. We show that aggregating scores across balanced permutations not only reveals latent position bias, but also improves correlation between the LLM-as-a-Judge and human. Our results suggest that rubric-based LLM-as-a-Judge is not inherently point-wise and that simple permutation-based calibration can substantially improve its reliability.

Explainable AI (XAI) offers a growing number of algorithms that aim to answer specific questions about black-box models. What is missing is a principled way to consolidate explanatory information about a fixed black-box model into a persistent, auditable artefact, that accompanies the black-box throughout its life cycle. We address this gap by introducing the notion of a scientific theory of a black (SToBB). Grounded in Constructive Empiricism, a SToBB fulfils three obligations: (i) empirical adequacy with respect to all available observations of black-box behaviour, (ii) adaptability via explicit update commitments that restore adequacy when new observations arrive, and (iii) auditability through transparent documentation of assumptions, construction choices, and update behaviour. We operationalise these obligations as a general framework that specifies an extensible observation base, a traceable hypothesis class, algorithmic components for construction and revision, and documentation sufficient for third-party assessment. Explanations for concrete stakeholder needs are then obtained by querying the maintained record through interfaces, rather than by producing isolated method outputs. As a proof of concept, we instantiate a complete SToBB for a neural-network classifier on a tabular task and introduce the Constructive Box Theoriser (CoBoT) algorithm, an online procedure that constructs and maintains an empirically adequate rule-based surrogate as observations accumulate. Together, these contributions position SToBBs as a life cycle-scale, inspectable point of reference that supports consistent, reusable analyses and systematic external scrutiny.

We present TIDES, a text informed design approach for generating physically sound designs based on a textual description and a set of physical constraints. TIDES jointly optimizes structural (topology) and visual properties. A pre-trained text-image model is used to measure the design's visual alignment with a text prompt and a differentiable physics simulator is used to measure its physical performance. We evaluate TIDES on a series of structural optimization problems operating under different load and support conditions, at different resolutions, and experimentally in the lab by performing the 3-point bending test on 2D beam designs that are extruded and 3D printed. We find that it can jointly optimize the two objectives and return designs that satisfy engineering design requirements (compliance and density) while utilizing features specified by text.

Despite significant progress in Visual-Language-Action (VLA), in highly complex and dynamic environments that involve real-time unpredictable interactions (such as 3D open worlds and large-scale PvP games), existing approaches remain inefficient at extracting action-critical signals from redundant sensor streams. To tackle this, we introduce MAIN-VLA, a framework that explicitly Models the Abstraction of Intention and eNvironment to ground decision-making in deep semantic alignment rather than superficial pattern matching. Specifically, our Intention Abstraction (IA) extracts verbose linguistic instructions and their associated reasoning into compact, explicit semantic primitives, while the Environment Semantics Abstraction (ESA) projects overwhelming visual streams into a structured, topological affordance representation. Furthermore, aligning these two abstract modalities induces an emergent attention-concentration effect, enabling a parameter-free token-pruning strategy that filters out perceptual redundancy without degrading performance. Extensive experiments in open-world Minecraft and large-scale PvP environments (Game for Peace and Valorant) demonstrate that MAIN-VLA sets a new state-of-the-art, which achieves superior decision quality, stronger generalization, and cutting-edge inference efficiency.

Large language models show promise for knowledge-intensive domains, yet their use in agriculture is constrained by weak grounding, English-centric training data, and limited real-world evaluation. These issues are amplified for low-resource languages, where high-quality domain documentation exists but remains difficult to access through general-purpose models. This paper presents AgriHubi, a domain-adapted retrieval-augmented generation (RAG) system for Finnish-language agricultural decision support. AgriHubi integrates Finnish agricultural documents with open PORO family models and combines explicit source grounding with user feedback to support iterative refinement. Developed over eight iterations and evaluated through two user studies, the system shows clear gains in answer completeness, linguistic accuracy, and perceived reliability. The results also reveal practical trade-offs between response quality and latency when deploying larger models. This study provides empirical guidance for designing and evaluating domain-specific RAG systems in low-resource language settings.

This paper presents the first-ever Sinhala physical common sense reasoning dataset created as part of Global PIQA. It contains 110 human-created and verified data samples, where each sample consists of a prompt, the corresponding correct answer, and a wrong answer. Most of the questions refer to the Sri Lankan context, where Sinhala is an official language.

While Large Language Models (LLMs) can theoretically support extensive context windows, their actual deployment is constrained by the linear growth of Key-Value (KV) cache memory. Prevailing compression strategies mitigate this through various pruning mechanisms, yet trade-off semantic recall for memory efficiency. In this work, we present LASER-KV (Layer Accumulated Selection with Exact-LSH Recall), a framework designed to test the limits of KV compression under a strict accumulative budgeting policy. We deviate from the standard fixed summary size approach by implementing a block-wise accumulation strategy governed by a protection divisor (n). This allows us to isolate the effects of compression from sliding window artifacts. Our experiments on the Babilong benchmark reveal performance degradation in previous compression methods by 15-30% on various long context tasks. LASER-KV maintains stable performance, achieving superior accuracies by a margin of upto 10% at 128k. These findings challenge the prevailing assumption that attention scores alone are a sufficient proxy for token utility.

The integration of visual information into Large Language Models (LLMs) has enabled Multimodal LLMs (MLLMs), but the quadratic memory and computational costs of Transformer architectures remain a bottleneck. Existing KV cache eviction strategies fail to address the heterogeneous attention distributions between visual and text tokens, leading to suboptimal efficiency or degraded performance. In this paper, we propose Hierarchical Adaptive Eviction (HAE), a KV cache eviction framework that optimizes text-visual token interaction in MLLMs by implementing Dual-Attention Pruning during pre-filling (leveraging visual token sparsity and attention variance) and a Dynamic Decoding Eviction Strategy (inspired by OS Recycle Bins) during decoding. HAE minimizes KV cache usage across layers, reduces computational overhead via index broadcasting, and theoretically ensures superior information integrity and lower error bounds compared to greedy strategies, enhancing efficiency in both comprehension and generation tasks. Empirically, HAE reduces KV-Cache memory by 41\% with minimal accuracy loss (0.3\% drop) in image understanding tasks and accelerates story generation inference by 1.5x while maintaining output quality on Phi3.5-Vision-Instruct model.

Linear Attention Large Language Models (LLMs) offer a compelling recurrent formulation that compresses context into a fixed-size state matrix, enabling constant-time inference. However, the internal dynamics of this compressed state remain largely opaque. In this work, we present a comprehensive study on the runtime state dynamics of state-of-the-art Linear Attention models. We uncover a fundamental phenomenon termed State Rank Stratification, characterized by a distinct spectral bifurcation among linear attention heads: while one group maintains an effective rank oscillating near zero, the other exhibits rapid growth that converges to an upper bound. Extensive experiments across diverse inference contexts reveal that these dynamics remain strikingly consistent, indicating that the identity of a head,whether low-rank or high-rank,is an intrinsic structural property acquired during pre-training, rather than a transient state dependent on the input data. Furthermore, our diagnostic probes reveal a surprising functional divergence: low-rank heads are indispensable for model reasoning, whereas high-rank heads exhibit significant redundancy. Leveraging this insight, we propose Joint Rank-Norm Pruning, a zero-shot strategy that achieves a 38.9\% reduction in KV-cache overhead while largely maintaining model accuracy.

Linearizing pretrained large language models (LLMs) primarily relies on intra-layer hybrid attention mechanisms to alleviate the quadratic complexity of standard softmax attention. Existing methods perform token routing based on sliding-window partitions, resulting in position-based selection and fails to capture token-specific global importance. Meanwhile, linear attention further suffers from distribution shift caused by learnable feature maps that distort pretrained feature magnitudes. Motivated by these limitations, we propose STILL, an intra-layer hybrid linearization framework for efficiently linearizing LLMs. STILL introduces a Self-Saliency Score with strong local-global consistency, enabling accurate token selection using sliding-window computation, and retains salient tokens for sparse softmax attention while summarizing the remaining context via linear attention. To preserve pretrained representations, we design a Norm-Preserved Feature Map (NP-Map) that decouples feature direction from magnitude and reinjects pretrained norms. We further adopt a unified training-inference architecture with chunk-wise parallelization and delayed selection to improve hardware efficiency. Experiments show that STILL matches or surpasses the original pretrained model on commonsense and general reasoning tasks, and achieves up to a 86.2% relative improvement over prior linearized attention methods on long-context benchmarks.

Accurate prediction of time-to-event outcomes is critical for clinical decision-making, treatment planning, and resource allocation in modern healthcare. While classical survival models such as Cox remain widely adopted in standard practice, they rely on restrictive assumptions, including linear covariate relationships and proportional hazards over time, that often fail to capture real-world clinical dynamics. Recent deep learning approaches like DeepSurv and DeepHit offer improved expressivity but sacrifice interpretability, limiting clinical adoption where trust and transparency are paramount. Hybrid models incorporating Kolmogorov-Arnold Networks (KANs), such as CoxKAN, have begun to address this trade-off but remain constrained by the semi-parametric Cox framework. In this work we introduce SurvKAN, a fully parametric, time-continuous survival model based on KAN architectures that eliminates the proportional hazards constraint. SurvKAN treats time as an explicit input to a KAN that directly predicts the log-hazard function, enabling end-to-end training on the full survival likelihood. Our architecture preserves interpretability through learnable univariate functions that indicate how individual features influence risk over time. Extensive experiments on standard survival benchmarks demonstrate that SurvKAN achieves competitive or superior performance compared to classical and state-of-the-art baselines across concordance and calibration metrics. Additionally, interpretability analyses reveal clinically meaningful patterns that align with medical domain knowledge.

Global optimization of decision trees is a long-standing challenge in combinatorial optimization, yet such models play an important role in interpretable machine learning. Although the problem has been investigated for several decades, only recent advances in discrete optimization have enabled practical algorithms for solving optimal classification tree problems on real-world datasets. Mixed-integer programming (MIP) offers a high degree of modeling flexibility, and we therefore propose a MIP-based framework for learning optimal classification trees under nonlinear performance metrics, such as the F1-score, that explicitly addresses class imbalance. To improve scalability, we develop problem-specific acceleration techniques, including a tailored branch-and-cut algorithm, an instance-reduction scheme, and warm-start strategies. We evaluate the proposed approach on 50 benchmark datasets. The computational results show that the framework can efficiently optimize nonlinear metrics while achieving strong predictive performance and reduced solution times compared with existing methods.

The limited sample size and insufficient diversity of lung nodule CT datasets severely restrict the performance and generalization ability of detection models. Existing methods generate images with insufficient diversity and controllability, suffering from issues such as monotonous texture features and distorted anatomical structures. Therefore, we propose a two-stage generative adversarial network (TSGAN) to enhance the diversity and spatial controllability of synthetic data by decoupling the morphological structure and texture features of lung nodules. In the first stage, StyleGAN is used to generate semantic segmentation mask images, encoding lung nodules and tissue backgrounds to control the anatomical structure of lung nodule images; The second stage uses the DL-Pix2Pix model to translate the mask map into CT images, employing local importance attention to capture local features, while utilizing dynamic weight multi-head window attention to enhance the modeling capability of lung nodule texture and background. Compared to the original dataset, the accuracy improved by 4.6% and mAP by 4% on the LUNA16 dataset. Experimental results demonstrate that TSGAN can enhance the quality of synthetic images and the performance of detection models.

Tabular foundation models like TabPFN and TabICL achieve state-of-the-art performance through in-context learning, yet their architectures remain fundamentally opaque. We introduce KernelICL, a framework to enhance tabular foundation models with quantifiable sample-based interpretability. Building on the insight that in-context learning is akin to kernel regression, we make this mechanism explicit by replacing the final prediction layer with kernel functions (Gaussian, dot-product, kNN) so that every prediction is a transparent weighted average of training labels. We introduce a two-dimensional taxonomy that formally unifies standard kernel methods, modern neighbor-based approaches, and attention mechanisms under a single framework, and quantify inspectability via the perplexity of the weight distribution over training samples. On 55 TALENT benchmark datasets, KernelICL achieves performance on par with existing tabular foundation models, demonstrating that explicit kernel constraints on the final layer enable inspectable predictions without sacrificing performance.

Temporal link prediction (TLP) models are commonly evaluated based on predictive accuracy, yet such evaluations do not assess whether these models capture the causal mechanisms that govern temporal interactions. In this work, we propose a framework for counterfactual validation of TLP models by generating causal temporal interaction graphs (CTIGs) with known ground-truth causal structure. We first introduce a structural equation model for continuous-time event sequences that supports both excitatory and inhibitory effects, and then extend this mechanism to temporal interaction graphs. To compare causal models, we propose a distance metric based on cross-model predictive error, and empirically validate the hypothesis that predictors trained on one causal model degrade when evaluated on sufficiently distant models. Finally, we instantiate counterfactual evaluation under (i) controlled causal shifts between generating models and (ii) timestamp shuffling as a stochastic distortion with measurable causal distance. Our framework provides a foundation for causality-aware benchmarking.

Effective tool use and reasoning are essential capabilities for large reasoning models~(LRMs) to address complex real-world problems. Through empirical analysis, we identify that current LRMs lack the capability of sub-task decomposition in complex tool use scenarios, leading to Lazy Reasoning. To address this, we propose a two-stage training framework D-CORE~(\underline{\textbf{D}}ecomposing tasks and \underline{\textbf{Co}}mposing \underline{\textbf{Re}}asoning processes) that first incentivize the LRMs' task decomposition reasoning capability via self-distillation, followed by diversity-aware reinforcement learning~(RL) to restore LRMs' reflective reasoning capability. D-CORE achieves robust tool-use improvements across diverse benchmarks and model scales. Experiments on BFCLv3 demonstrate superiority of our method: D-CORE-8B reaches 77.7\% accuracy, surpassing the best-performing 8B model by 5.7\%. Meanwhile, D-CORE-14B establishes a new state-of-the-art at 79.3\%, outperforming 70B models despite being 5$\times$ smaller. The source code is available at https://github.com/alibaba/EfficientAI.

Diffusion Large Language Models (dLLMs) deliver strong long-context processing capability in a non-autoregressive decoding paradigm. However, the considerable computational cost of bidirectional full attention limits the inference efficiency. Although sparse attention is promising, existing methods remain ineffective. This stems from the need to estimate attention importance for tokens yet to be decoded, while the unmasked token positions are unknown during diffusion. In this paper, we present Focus-dLLM, a novel training-free attention sparsification framework tailored for accurate and efficient long-context dLLM inference. Based on the finding that token confidence strongly correlates across adjacent steps, we first design a past confidence-guided indicator to predict unmasked regions. Built upon this, we propose a sink-aware pruning strategy to accurately estimate and remove redundant attention computation, while preserving highly influential attention sinks. To further reduce overhead, this strategy reuses identified sink locations across layers, leveraging the observed cross-layer consistency. Experimental results show that our method offers more than $29\times$ lossless speedup under $32K$ context length. The code is publicly available at: https://github.com/Longxmas/Focus-dLLM

This project compares three graph search approaches for the task of traffic-aware navigation in Kingston's road network. These approaches include a single-run multi-query preprocessing algorithm (Floyd-Warshall-Ingerman), continuous single-query real-time search (Dijkstra's and A*), and an algorithm combining both approaches to balance between their trade-offs by first finding the top K shortest paths then iterating over them in real time (Yen's). Dijkstra's and A* resulted in the most traffic-aware optimal solutions with minimal preprocessing required. Floyd-Warshall-Ingerman was the fastest in real time but provided distance based paths with no traffic awareness. Yen's algorithm required significant preprocessing but balanced between the other two approaches in terms of runtime speed and optimality. Each approach presents advantages and disadvantages that need to be weighed depending on the circumstances of specific deployment contexts to select the best custom solution. *This project was completed as part of ELEC 844 (Search and Planning Algorithms for Robotics) in the Fall 2025 term.

Neural Controlled Differential Equations (Neural CDEs) provide a powerful continuous-time framework for sequence modeling, yet the roughness of the driving control path often restricts their efficiency. Standard splines introduce high-frequency variations that force adaptive solvers to take excessively small steps, driving up the Number of Function Evaluations (NFE). We propose a novel approach to Neural CDE path construction that replaces exact interpolation with Kernel and Gaussian Process (GP) smoothing, enabling explicit control over trajectory regularity. To recover details lost during smoothing, we propose an attention-based Multi-View CDE (MV-CDE) and its convolutional extension (MVC-CDE), which employ learnable queries to inform path reconstruction. This framework allows the model to distribute representational capacity across multiple trajectories, each capturing distinct temporal patterns. Empirical results demonstrate that our method, MVC-CDE with GP, achieves state-of-the-art accuracy while significantly reducing NFEs and total inference time compared to spline-based baselines.

Recent advances in visual reasoning have leveraged vision transformers to tackle the ARC-AGI benchmark. However, we argue that the feed-forward architecture, where computational depth is strictly bound to parameter size, falls short of capturing the iterative, algorithmic nature of human induction. In this work, we propose a recursive architecture called Loop-ViT, which decouples reasoning depth from model capacity through weight-tied recurrence. Loop-ViT iterates a weight-tied Hybrid Block, combining local convolutions and global attention, to form a latent chain of thought. Crucially, we introduce a parameter-free Dynamic Exit mechanism based on predictive entropy: the model halts inference when its internal state ``crystallizes" into a low-uncertainty attractor. Empirical results on the ARC-AGI-1 benchmark validate this perspective: our 18M model achieves 65.8% accuracy, outperforming massive 73M-parameter ensembles. These findings demonstrate that adaptive iterative computation offers a far more efficient scaling axis for visual reasoning than simply increasing network width. The code is available at https://github.com/WenjieShu/LoopViT.

Prior to modern Earth observation technologies, historical maps provide a unique record of long-term urban transformation and offer a lens on the evolving identity of cities. However, extracting consistent and fine-grained change information from historical map series remains challenging due to spatial misalignment, cartographic variation, and degrading document quality, limiting most analyses to small-scale or qualitative approaches. We propose a fully automated, deep learning-based framework for fine-grained urban change analysis from large collections of historical maps, built on a modular design that integrates dense map alignment, multi-temporal object detection, and change profiling. This framework shifts the analysis of historical maps from ad hoc visual comparison toward systematic, quantitative characterization of urban change. Experiments demonstrate the robust performance of the proposed alignment and object detection methods. Applied to Paris between 1868 and 1937, the framework reveals the spatial and temporal heterogeneity in urban transformation, highlighting its relevance for research in the social sciences and humanities. The modular design of our framework further supports adaptation to diverse cartographic contexts and downstream applications.

Adaptive Rounding has emerged as an alternative to round-to-nearest (RTN) for post-training quantization by enabling cross-element error cancellation. Yet, dense and element-wise rounding matrices are prohibitively expensive for billion-parameter large language models (LLMs). We revisit adaptive rounding from an efficiency perspective and propose VQRound, a parameter-efficient optimization framework that reparameterizes the rounding matrix into a compact codebook. Unlike low-rank alternatives, VQRound minimizes the element-wise worst-case error under $L_\infty$ norm, which is critical for handling heavy-tailed weight distributions in LLMs. Beyond reparameterization, we identify rounding initialization as a decisive factor and develop a lightweight end-to-end finetuning pipeline that optimizes codebooks across all layers using only 128 samples. Extensive experiments on OPT, LLaMA, LLaMA2, and Qwen3 models demonstrate that VQRound achieves better convergence than traditional adaptive rounding at the same number of steps while using as little as 0.2% of the trainable parameters. Our results show that adaptive rounding can be made both scalable and fast-fitting. The code is available at https://github.com/zhoustan/VQRound.

Long-term time series forecasting (LTSF) remains challenging due to the trade-off between parallel efficiency and sequential modeling of temporal coherence. Direct multi-step forecasting (DMS) methods enable fast, parallel prediction of all future horizons but often lose temporal consistency across steps, while iterative multi-step forecasting (IMS) preserves temporal dependencies at the cost of error accumulation and slow inference. To bridge this gap, we propose Back to the Future (BTTF), a simple yet effective framework that enhances forecasting stability through look-ahead augmentation and self-corrective refinement. Rather than relying on complex model architectures, BTTF revisits the fundamental forecasting process and refines a base model by ensembling the second-stage models augmented with their initial predictions. Despite its simplicity, our approach consistently improves long-horizon accuracy and mitigates the instability of linear forecasting models, achieving accuracy gains of up to 58% and demonstrating stable improvements even when the first-stage model is trained under suboptimal conditions. These results suggest that leveraging model-generated forecasts as augmentation can be a simple yet powerful way to enhance long-term prediction, even without complex architectures.

Scaling test-time compute via parallel sampling can substantially improve LLM reasoning, but is often limited by Best-of-N selection quality. Generative selection methods, such as GenSelect, address this bottleneck, yet strong selection performance remains largely limited to large models. We show that small reasoning models can acquire strong GenSelect capabilities through targeted reinforcement learning. To this end, we synthesize selection tasks from large-scale math and code instruction datasets by filtering to instances with both correct and incorrect candidate solutions, and train 1.7B-parameter models with DAPO to reward correct selections. Across math (AIME24, AIME25, HMMT25) and code (LiveCodeBench) reasoning benchmarks, our models consistently outperform prompting and majority-voting baselines, often approaching or exceeding much larger models. Moreover, these gains generalize to selecting outputs from stronger models despite training only on outputs from weaker models. Overall, our results establish reinforcement learning as a scalable way to unlock strong generative selection in small models, enabling efficient test-time scaling.

Recent advances in unified multimodal models (UMM) have demonstrated remarkable progress in both understanding and generation tasks. However, whether these two capabilities are genuinely aligned and integrated within a single model remains unclear. To investigate this question, we introduce GapEval, a bidirectional benchmark designed to quantify the gap between understanding and generation capabilities, and quantitatively measure the cognitive coherence of the two "unified" directions. Each question can be answered in both modalities (image and text), enabling a symmetric evaluation of a model's bidirectional inference capability and cross-modal consistency. Experiments reveal a persistent gap between the two directions across a wide range of UMMs with different architectures, suggesting that current models achieve only surface-level unification rather than deep cognitive convergence of the two. To further explore the underlying mechanism, we conduct an empirical study from the perspective of knowledge manipulation to illustrate the underlying limitations. Our findings indicate that knowledge within UMMs often remains disjoint. The capability emergence and knowledge across modalities are unsynchronized, paving the way for further exploration.

Safety alignment incurs safety tax that perturbs a large reasoning model's (LRM) general reasoning ability. Existing datasets used for safety alignment for an LRM are usually constructed by distilling safety reasoning traces and answers from an external LRM or human labeler. However, such reasoning traces and answers exhibit a distributional gap with the target LRM that needs alignment, and we conjecture such distributional gap is the culprit leading to significant degradation of reasoning ability of the target LRM. Driven by this hypothesis, we propose a safety alignment dataset construction method, dubbed DGR. DGR transforms and refines an existing out-of-distributional safety reasoning dataset to be aligned with the target's LLM inner distribution. Experimental results demonstrate that i) DGR effectively mitigates the safety tax while maintaining safety performance across all baselines, i.e., achieving \textbf{+30.2\%} on DirectRefusal and \textbf{+21.2\%} on R1-ACT improvement in average reasoning accuracy compared to Vanilla SFT; ii) the degree of reasoning degradation correlates with the extent of distribution shift, suggesting that bridging this gap is central to preserving capabilities. Furthermore, we find that safety alignment in LRMs may primarily function as a mechanism to activate latent knowledge, as a mere \textbf{10} samples are sufficient for activating effective refusal behaviors. These findings not only emphasize the importance of distributional consistency but also provide insights into the activation mechanism of safety in reasoning models.

Prior work argues that refusal in large language models is mediated by a single activation-space direction, enabling effective steering and ablation. We show that this account is incomplete. Across eleven categories of refusal and non-compliance, including safety, incomplete or unsupported requests, anthropomorphization, and over-refusal, we find that these refusal behaviors correspond to geometrically distinct directions in activation space. Yet despite this diversity, linear steering along any refusal-related direction produces nearly identical refusal to over-refusal trade-offs, acting as a shared one-dimensional control knob. The primary effect of different directions is not whether the model refuses, but how it refuses.