Mitigating Safety Tax via Distribution-Grounded Refinement in Large Reasoning Models

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.

Improve the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models

Watermarking is a principled approach for tracing the provenance of large language model (LLM) outputs, but its deployment in practice is hindered by inference inefficiency. Speculative sampling accelerates inference, with efficiency improving as the acceptance rate between draft and target models increases. Yet recent work reveals a fundamental trade-off: higher watermark strength reduces acceptance, preventing their simultaneous achievement. We revisit this trade-off and show it is not absolute. We introduce a quantitative measure of watermark strength that governs statistical detectability and is maximized when tokens are deterministic functions of pseudorandom numbers. Using this measure, we fully characterize the trade-off as a constrained optimization problem and derive explicit Pareto curves for two existing watermarking schemes. Finally, we introduce a principled mechanism that injects pseudorandomness into draft-token acceptance, ensuring maximal watermark strength while maintaining speculative sampling efficiency. Experiments further show that this approach improves detectability without sacrificing efficiency. Our findings uncover a principle that unites speculative sampling and watermarking, paving the way for their efficient and practical deployment.

Language-based Trial and Error Falls Behind in the Era of Experience

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

FRISM: Fine-Grained Reasoning Injection via Subspace-Level Model Merging for Vision-Language Models

Efficiently enhancing the reasoning capabilities of Vision-Language Models (VLMs) by merging them with Large Reasoning Models (LRMs) has emerged as a promising direction. However, existing methods typically operate at a coarse-grained layer level, which often leads to a trade-off between injecting reasoning capabilities and preserving visual capabilities. To address this limitation, we propose {FRISM} (Fine-grained Reasoning Injection via Subspace-level model Merging), a fine-grained reasoning injection framework based on subspace-level model merging. Observing that reasoning capabilities are encoded in distinct subspaces, FRISM decomposes LRM task vectors via Singular Value Decomposition (SVD) and adaptively tunes the scaling coefficients of each subspace through learning to realize fine-grained reasoning injection. Furthermore, we introduce a label-free self-distillation learning strategy with a dual-objective optimization using common vision-language perception datasets. Extensive experiments demonstrate that FRISM effectively improves reasoning capabilities without compromising the model's original visual capabilities by consistently achieving state-of-the-art performance across diverse visual reasoning benchmarks.

Understanding Model Merging: A Unified Generalization Framework for Heterogeneous Experts

Model merging efficiently aggregates capabilities from multiple fine-tuned models into a single one, operating purely in parameter space without original data or expensive re-computation. Despite empirical successes, a unified theory for its effectiveness under heterogeneous finetuning hyperparameters (e.g., varying learning rates, batch sizes) remains missing. Moreover, the lack of hyperparameter transparency in open-source fine-tuned models makes it difficult to predict merged-model performance, leaving practitioners without guidance on how to fine-tune merge-friendly experts. To address those two challenges, we employ $L_2$-Stability theory under heterogeneous hyperparameter environments to analyze the generalization of the merged model $\boldsymbol{x}_{avg}$. This pioneering analysis yields two key contributions: (i) \textit{A unified theoretical framework} is provided to explain existing merging algorithms, revealing how they optimize specific terms in our bound, thus offering a strong theoretical foundation for empirical observations. (ii) \textit{Actionable recommendations} are proposed for practitioners to strategically fine-tune expert models, enabling the construction of merge-friendly models within the pretraining-to-finetuning pipeline. Extensive experiments on the ResNet/Vit family across 20/8 visual classification tasks, involving thousands of finetuning models, robustly confirm the impact of different hyperparameters on the generalization of $\boldsymbol{x}_{avg}$ predicted by our theoretical results.

Learning Ordered Representations in Latent Space for Intrinsic Dimension Estimation via Principal Component Autoencoder

Autoencoders have long been considered a nonlinear extension of Principal Component Analysis (PCA). Prior studies have demonstrated that linear autoencoders (LAEs) can recover the ordered, axis-aligned principal components of PCA by incorporating non-uniform $\ell_2$ regularization or by adjusting the loss function. However, these approaches become insufficient in the nonlinear setting, as the remaining variance cannot be properly captured independently of the nonlinear mapping. In this work, we propose a novel autoencoder framework that integrates non-uniform variance regularization with an isometric constraint. This design serves as a natural generalization of PCA, enabling the model to preserve key advantages, such as ordered representations and variance retention, while remaining effective for nonlinear dimensionality reduction tasks.

Advancing Adaptive Multi-Stage Video Anomaly Reasoning: A Benchmark Dataset and Method

Recent progress in reasoning capabilities of Multimodal Large Language Models(MLLMs) has highlighted their potential for performing complex video understanding tasks. However, in the domain of Video Anomaly Detection and Understanding (VAD&U), existing MLLM-based methods are largely limited to anomaly localization or post-hoc description, lacking explicit reasoning processes, risk awareness, and decision-oriented interpretation. To address this gap, we define a new task termed Video Anomaly Reasoning (VAR), which elevates video anomaly analysis from descriptive understanding to structured, multi-stage reasoning. VAR explicitly requires models to perform progressive reasoning over anomalous events before answering anomaly-related questions, encompassing visual perception, causal interpretation, and risk-aware decision making. To support this task, we present a new dataset with 8,641 videos, where each video is annotated with diverse question types corresponding to different reasoning depths, totaling more than 50,000 samples, making it one of the largest datasets for video anomaly. The annotations are based on a structured Perception-Cognition-Action Chain-of-Thought (PerCoAct-CoT), which formalizes domain-specific reasoning priors for video anomaly understanding. This design enables systematic evaluation of multi-stage and adaptive anomaly reasoning. In addition, we propose Anomaly-Aware Group Relative Policy Optimization to further enhance reasoning reliability under weak supervision. Building upon the proposed task and dataset, we develop an end-to-end MLLM-based VAR model termed Vad-R1-Plus, which supports adaptive hierarchical reasoning and risk-aware decision making. Extensive experiments demonstrate that the proposed benchmark and method effectively advance the reasoning capabilities of MLLMs on VAR tasks, outperforming both open-source and proprietary baselines.

Joint Selection for Large-Scale Pre-Training Data via Policy Gradient-based Mask Learning

A fine-grained data recipe is crucial for pre-training large language models, as it can significantly enhance training efficiency and model performance. One important ingredient in the recipe is to select samples based on scores produced by defined rules, LLM judgment, or statistical information in embeddings, which can be roughly categorized into quality and diversity metrics. Due to the high computational cost when applied to trillion-scale token pre-training datasets such as FineWeb and DCLM, these two or more types of metrics are rarely considered jointly in a single selection process. However, in our empirical study, selecting samples based on quality metrics exhibit severe diminishing returns during long-term pre-training, while selecting on diversity metrics removes too many valuable high-quality samples, both of which limit pre-trained LLMs' capabilities. Therefore, we introduce DATAMASK, a novel and efficient joint learning framework designed for large-scale pre-training data selection that can simultaneously optimize multiple types of metrics in a unified process, with this study focusing specifically on quality and diversity metrics. DATAMASK approaches the selection process as a mask learning problem, involving iterative sampling of data masks, computation of policy gradients based on predefined objectives with sampled masks, and updating of mask sampling logits. Through policy gradient-based optimization and various acceleration enhancements, it significantly reduces selection time by 98.9% compared to greedy algorithm, enabling our study to explore joint learning within trillion-scale tokens. With DATAMASK, we select a subset of about 10% from the 15 trillion-token FineWeb dataset, termed FineWeb-Mask. Evaluated across 12 diverse tasks, we achieves significant improvements of 3.2% on a 1.5B dense model and 1.9% on a 7B MoE model.

Reason-KE++: Aligning the Process, Not Just the Outcome, for Faithful LLM Knowledge Editing

Aligning Large Language Models (LLMs) to be faithful to new knowledge in complex, multi-hop reasoning tasks is a critical, yet unsolved, challenge. We find that SFT-based methods, e.g., Reason-KE, while state-of-the-art, suffer from a "faithfulness gap": they optimize for format mimicry rather than sound reasoning. This gap enables the LLM's powerful parametric priors to override new contextual facts, resulting in critical factual hallucinations (e.g., incorrectly reasoning "Houston" from "NASA" despite an explicit edit). To solve this core LLM alignment problem, we propose Reason-KE++, an SFT+RL framework that instills process-level faithfulness. Its core is a Stage-aware Reward mechanism that provides dense supervision for intermediate reasoning steps (e.g., Decomposition, Sub-answer Correctness). Crucially, we identify that naive outcome-only RL is a deceptive trap for LLM alignment: it collapses reasoning integrity (e.g., 19.00% Hop acc) while superficially boosting final accuracy. Our process-aware framework sets a new SOTA of 95.48% on MQUAKE-CF-3k (+5.28%), demonstrating that for complex tasks, aligning the reasoning process is essential for building trustworthy LLMs.

Hindsight Distillation Reasoning with Knowledge Encouragement Preference for Knowledge-based Visual Question Answering

Knowledge-based Visual Question Answering (KBVQA) necessitates external knowledge incorporation beyond cross-modal understanding. Existing KBVQA methods either utilize implicit knowledge in multimodal large language models (MLLMs) via in-context learning or explicit knowledge via retrieval augmented generation. However, their reasoning processes remain implicit, without explicit multi-step trajectories from MLLMs. To address this gap, we provide a Hindsight Distilled Reasoning (HinD) framework with Knowledge Encouragement Preference Optimization (KEPO), designed to elicit and harness internal knowledge reasoning ability in MLLMs. First, to tackle the reasoning supervision problem, we propose to emphasize the hindsight wisdom of MLLM by prompting a frozen 7B-size MLLM to complete the reasoning process between the question and its ground truth answer, constructing Hindsight-Zero training data. Then we self-distill Hindsight-Zero into Chain-of-Thought (CoT) Generator and Knowledge Generator, enabling the generation of sequential steps and discrete facts. Secondly, to tackle the misalignment between knowledge correctness and confidence, we optimize the Knowledge Generator with KEPO, preferring under-confident but helpful knowledge over the over-confident but unhelpful one. The generated CoT and sampled knowledge are then exploited for answer prediction. Experiments on OK-VQA and A-OKVQA validate the effectiveness of HinD, showing that HinD with elicited reasoning from 7B-size MLLM achieves superior performance without commercial model APIs or outside knowledge.