VRM: Teaching Reward Models to Understand Authentic Human Preferences

Large Language Models (LLMs) have achieved remarkable success across diverse natural language tasks, yet the reward models employed for aligning LLMs often encounter challenges of reward hacking, where the approaches predominantly rely on directly mapping prompt-response pairs to scalar scores, which may inadvertently capture spurious correlations rather than authentic human preferences. In contrast, human evaluation employs a sophisticated process that initially weighs the relative importance of multiple high-dimensional objectives according to the prompt context, subsequently evaluating response quality through low-dimensional semantic features such as logical coherence and contextual appropriateness. Motivated by this consideration, we propose VRM, i.e., Variational Reward Modeling, a novel framework that explicitly models the evaluation process of human preference judgments by incorporating both high-dimensional objective weights and low-dimensional semantic features as latent variables, which are inferred through variational inference techniques. Additionally, we provide a theoretical analysis showing that VRM can achieve a tighter generalization error bound compared to the traditional reward model. Extensive experiments on benchmark datasets demonstrate that VRM significantly outperforms existing methods in capturing authentic human preferences.

Model Merging in the Essential Subspace

Model merging aims to integrate multiple task-specific fine-tuned models derived from a shared pre-trained checkpoint into a single multi-task model without additional training. Despite extensive research, task interference remains a major obstacle that often undermines the performance of merged models. In this paper, we propose ESM (Essential Subspace Merging) , a robust framework for effective model merging. We begin by performing Principal Component Analysis (PCA) on feature shifts induced by parameter updates. The resulting principal directions span an essential subspace that dominantly influences feature representations. Each task's parameter update matrix is projected onto its respective essential subspace for low-rank decomposition before merging. This methodology mitigates inter-task interference while preserving core task-specific functionality. Furthermore, we introduce a multi-level polarized scaling strategy that amplifies parameters containing critical knowledge and suppresses redundant ones, preventing essential knowledge from being overwhelmed during fusion. Extensive experiments across multiple task sets and model scales demonstrate that our method achieves state-of-the-art performance in multi-task model merging.

Towards On-Policy SFT: Distribution Discriminant Theory and its Applications in LLM Training

Supervised fine-tuning (SFT) is computationally efficient but often yields inferior generalization compared to reinforcement learning (RL). This gap is primarily driven by RL's use of on-policy data. We propose a framework to bridge this chasm by enabling On-Policy SFT. We first present \textbf{\textit{Distribution Discriminant Theory (DDT)}}, which explains and quantifies the alignment between data and the model-induced distribution. Leveraging DDT, we introduce two complementary techniques: (i) \textbf{\textit{In-Distribution Finetuning (IDFT)}}, a loss-level method to enhance generalization ability of SFT, and (ii) \textbf{\textit{Hinted Decoding}}, a data-level technique that can re-align the training corpus to the model's distribution. Extensive experiments demonstrate that our framework achieves generalization performance on par with prominent offline RL algorithms, including DPO and SimPO, while maintaining the efficiency of an SFT pipeline. The proposed framework thus offers a practical alternative in domains where RL is infeasible. We open-source the code here: https://github.com/zhangmiaosen2000/Towards-On-Policy-SFT

RE-TRAC: REcursive TRAjectory Compression for Deep Search Agents

LLM-based deep research agents are largely built on the ReAct framework. This linear design makes it difficult to revisit earlier states, branch into alternative search directions, or maintain global awareness under long contexts, often leading to local optima, redundant exploration, and inefficient search. We propose Re-TRAC, an agentic framework that performs cross-trajectory exploration by generating a structured state representation after each trajectory to summarize evidence, uncertainties, failures, and future plans, and conditioning subsequent trajectories on this state representation. This enables iterative reflection and globally informed planning, reframing research as a progressive process. Empirical results show that Re-TRAC consistently outperforms ReAct by 15-20% on BrowseComp with frontier LLMs. For smaller models, we introduce Re-TRAC-aware supervised fine-tuning, achieving state-of-the-art performance at comparable scales. Notably, Re-TRAC shows a monotonic reduction in tool calls and token usage across rounds, indicating progressively targeted exploration driven by cross-trajectory reflection rather than redundant search.

Positive-Unlabeled Reinforcement Learning Distillation for On-Premise Small Models

Due to constraints on privacy, cost, and latency, on-premise deployment of small models is increasingly common. However, most practical pipelines stop at supervised fine-tuning (SFT) and fail to reach the reinforcement learning (RL) alignment stage. The main reason is that RL alignment typically requires either expensive human preference annotation or heavy reliance on high-quality reward models with large-scale API usage and ongoing engineering maintenance, both of which are ill-suited to on-premise settings. To bridge this gap, we propose a positive-unlabeled (PU) RL distillation method for on-premise small-model deployment. Without human-labeled preferences or a reward model, our method distills the teacher's preference-optimization capability from black-box generations into a locally trainable student. For each prompt, we query the teacher once to obtain an anchor response, locally sample multiple student candidates, and perform anchor-conditioned self-ranking to induce pairwise or listwise preferences, enabling a fully local training loop via direct preference optimization or group relative policy optimization. Theoretical analysis justifies that the induced preference signal by our method is order-consistent and concentrates on near-optimal candidates, supporting its stability for preference optimization. Experiments demonstrate that our method achieves consistently strong performance under a low-cost setting.

Self-Supervised Weight Templates for Scalable Vision Model Initialization

The increasing scale and complexity of modern model parameters underscore the importance of pre-trained models. However, deployment often demands architectures of varying sizes, exposing limitations of conventional pre-training and fine-tuning. To address this, we propose SWEET, a self-supervised framework that performs constraint-based pre-training to enable scalable initialization in vision tasks. Instead of pre-training a fixed-size model, we learn a shared weight template and size-specific weight scalers under Tucker-based factorization, which promotes modularity and supports flexible adaptation to architectures with varying depths and widths. Target models are subsequently initialized by composing and reweighting the template through lightweight weight scalers, whose parameters can be efficiently learned from minimal training data. To further enhance flexibility in width expansion, we introduce width-wise stochastic scaling, which regularizes the template along width-related dimensions and encourages robust, width-invariant representations for improved cross-width generalization. Extensive experiments on \textsc{classification}, \textsc{detection}, \textsc{segmentation} and \textsc{generation} tasks demonstrate the state-of-the-art performance of SWEET for initializing variable-sized vision models.

Fast Inference of Visual Autoregressive Model with Adjacency-Adaptive Dynamical Draft Trees

Autoregressive (AR) image models achieve diffusion-level quality but suffer from sequential inference, requiring approximately 2,000 steps for a 576x576 image. Speculative decoding with draft trees accelerates LLMs yet underperforms on visual AR models due to spatially varying token prediction difficulty. We identify a key obstacle in applying speculative decoding to visual AR models: inconsistent acceptance rates across draft trees due to varying prediction difficulties in different image regions. We propose Adjacency-Adaptive Dynamical Draft Trees (ADT-Tree), an adjacency-adaptive dynamic draft tree that dynamically adjusts draft tree depth and width by leveraging adjacent token states and prior acceptance rates. ADT-Tree initializes via horizontal adjacency, then refines depth/width via bisectional adaptation, yielding deeper trees in simple regions and wider trees in complex ones. The empirical evaluations on MS-COCO 2017 and PartiPrompts demonstrate that ADT-Tree achieves speedups of 3.13xand 3.05x, respectively. Moreover, it integrates seamlessly with relaxed sampling methods such as LANTERN, enabling further acceleration. Code is available at https://github.com/Haodong-Lei-Ray/ADT-Tree.

Knowledge Diversion for Efficient Morphology Control and Policy Transfer

Universal morphology control aims to learn a universal policy that generalizes across heterogeneous agent morphologies, with Transformer-based controllers emerging as a popular choice. However, such architectures incur substantial computational costs, resulting in high deployment overhead, and existing methods exhibit limited cross-task generalization, necessitating training from scratch for each new task. To this end, we propose \textbf{DivMorph}, a modular training paradigm that leverages knowledge diversion to learn decomposable controllers. DivMorph factorizes randomly initialized Transformer weights into factor units via SVD prior to training and employs dynamic soft gating to modulate these units based on task and morphology embeddings, separating them into shared \textit{learngenes} and morphology- and task-specific \textit{tailors}, thereby achieving knowledge disentanglement. By selectively activating relevant components, DivMorph enables scalable and efficient policy deployment while supporting effective policy transfer to novel tasks. Extensive experiments demonstrate that DivMorph achieves state-of-the-art performance, achieving a 3$\times$ improvement in sample efficiency over direct finetuning for cross-task transfer and a 17$\times$ reduction in model size for single-agent deployment.

Stratified Knowledge-Density Super-Network for Scalable Vision Transformers

Training and deploying multiple vision transformer (ViT) models for different resource constraints is costly and inefficient. To address this, we propose transforming a pre-trained ViT into a stratified knowledge-density super-network, where knowledge is hierarchically organized across weights. This enables flexible extraction of sub-networks that retain maximal knowledge for varying model sizes. We introduce \textbf{W}eighted \textbf{P}CA for \textbf{A}ttention \textbf{C}ontraction (WPAC), which concentrates knowledge into a compact set of critical weights. WPAC applies token-wise weighted principal component analysis to intermediate features and injects the resulting transformation and inverse matrices into adjacent layers, preserving the original network function while enhancing knowledge compactness. To further promote stratified knowledge organization, we propose \textbf{P}rogressive \textbf{I}mportance-\textbf{A}ware \textbf{D}ropout (PIAD). PIAD progressively evaluates the importance of weight groups, updates an importance-aware dropout list, and trains the super-network under this dropout regime to promote knowledge stratification. Experiments demonstrate that WPAC outperforms existing pruning criteria in knowledge concentration, and the combination with PIAD offers a strong alternative to state-of-the-art model compression and model expansion methods.

One-Shot Knowledge Transfer for Scalable Person Re-Identification

Edge computing in person re-identification (ReID) is crucial for reducing the load on central cloud servers and ensuring user privacy. Conventional compression methods for obtaining compact models require computations for each individual student model. When multiple models of varying sizes are needed to accommodate different resource conditions, this leads to repetitive and cumbersome computations. To address this challenge, we propose a novel knowledge inheritance approach named OSKT (One-Shot Knowledge Transfer), which consolidates the knowledge of the teacher model into an intermediate carrier called a weight chain. When a downstream scenario demands a model that meets specific resource constraints, this weight chain can be expanded to the target model size without additional computation. OSKT significantly outperforms state-of-the-art compression methods, with the added advantage of one-time knowledge transfer that eliminates the need for frequent computations for each target model.