arXiv subjects

Yuyan Zhou

Publications and source records attributed to Yuyan Zhou.

14 recordsLinked to original sources

arXiv2026

Trajectory-Aware Policy Optimization: Closing the State Distribution Gap in Long-Horizon LLM RL

Reinforcement learning (RL) has become a dominant paradigm for enhancing LLMs' reasoning capabilities. Yet long-horizon autoregressive rollouts introduce a fundamental limitation: local policy mismatch can compound into severe state distribution shift over trajectories. Existing LLM RL methods, such as GRPO and GSPO, control policy mismatch at the token or sequence level, but overlook this trajectory-induced shift, causing training instability and eventual collapse as the horizon grows. To close the state distribution gap, we propose Trajectory-Aware Policy Optimization (TAPO), a principled framework that theoretically derives the state-distribution correction from the off-policy policy gradient and incorporates it into the RL objective. TAPO further employs a geometric-mean estimator for variance reduction, together with a length-aware dynamic clipping that prevents overly deviated updates across the trajectory. Extensive experiments demonstrate that TAPO significantly surpasses existing LLM RL algorithms across eight competition-level benchmarks, achieving up to 48% and 12% improvement over GRPO on mathematical reasoning and algorithmic coding, with substantially enhanced training stability.

cs.LG↗

arXiv2026

LED: Improving Masked Diffusion Language Models by Lookahead Editing

Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive language models, enabling non-autoregressive text generation through parallel decoding. Existing dLLMs can be broadly categorized into two families: masked diffusion models (MDMs) and uniform diffusion models (UDMs) In this work, we theoretically show that MDMs are easier to optimize due to their mask-based transition schedule but have lower computational expressivity, whereas UDMs enable iterative state revision and greater expressivity but are substantially harder to train. This trade-off raises a natural question: can we obtain the expressivity of iterative revision without giving up the simple, easy-to-optimize transition schedule of MDMs? Motivated by this analysis, we propose Lookahead Editing (\ours{}), which augments masked diffusion decoding with a rewritable proposal state. At each decoding step, the model performs several lookahead updates to iteratively refine future predictions in this proposal state before committing tokens to the masked sequence. This design introduces a rewritable proposal that uses additional compute to improve expressivity while keeping the simple mask-based transition schedule of MDMs. We fine-tune Qwen2.5-7B and Qwen3-8B to obtain LED-\texttt{7B} and LED-\texttt{8B}, respectively. Across eight benchmarks, LED SFT yields average gains of $+11.4$ and $+7.7$ points over standard MDM SFT. Moreover, using SFT alone, LED-\texttt{8B} outperforms LLaDA2.1-16B Q-mode, which is trained with continued pre-training, supervised fine-tuning, and reinforcement learning.

cs.LG↗

arXiv2026

ReCreate: Reasoning and Creating Domain Agents Driven by Experience

Large Language Model (LLM) agents are reshaping the industrial landscape. However, most practical agents still rely on human-designed scaffolds, making their creation and adaptation across diverse domains labor-intensive. This situation poses a central question: can we automatically create and evolve domain agents through interaction experience in real-world environments? While recent approaches have sought to automate agent creation, several treat scaffold optimization as a black-box procedure guided primarily by final performance metrics. Such strategies overlook critical evidence explaining why an agent succeeds or fails and often incur high computational costs. To address these limitations, we propose ReCreate, an experience-driven framework for creating and evolving specialized domain agents. ReCreate systematically leverages agent interaction histories to identify the causes of success and failure and guide targeted scaffold updates. Specifically, we introduce an agent-as-optimizer paradigm in which an optimizer agent actively inspects execution experience and iteratively refines the task agent's scaffold through three key components: (i) an experience storage and retrieval mechanism for on-demand evidence inspection; (ii) a reasoning-creating synergy pipeline that translates execution evidence into updates to prompts, workflows, tools, and memory; and (iii) hierarchical updates that consolidate instance-level refinements into reusable domain-level improvements. Experiments across thirteen evaluation settings spanning four domains show that ReCreate consistently outperforms human-designed agents and existing automated agent generation and self-evolving baselines, even when evolving from minimal seed scaffolds.

cs.AI↗

arXiv2026

Generalized Proximal Policy Optimization: Unifying LLM RL Algorithms from the Perspective of Divergence

Reinforcement learning has become the foundational paradigm for enhancing the reasoning capabilities of LLMs. Yet the dominant approach, Proximal Policy Optimization (PPO), relies on a geometrically degenerate clipping proxy that cannot adapt its resistance to policy shifts, leading to premature clipping and degraded exploration. Subsequent variants alleviate these issues through fragmented and heuristic modifications. We propose Generalized Proximal Policy Optimization (GPPO), a principled and unified framework for understanding and designing LLM RL algorithms through the lens of divergence. We establish a generalized $f$-divergence trust-region theory with a monotonic performance improvement guarantee, and show that different $f$-divergences induce distinct geometries and policy update dynamics. We further introduce novel estimator topologies that provide more faithful divergence estimates than the conventional single-sample Monte Carlo estimation. Under this framework, PPO and recent variants such as DAPO, DPPO, and RIPO emerge as special cases, establishing a principled two-dimensional design space of {Divergence $\times$ Estimator}. Extensive experiments on eight competition-level math and coding benchmarks show that both dimensions are critical. The best-performing GPPO instantiation surpasses standard GRPO by up to 50% and 14% on math and coding, delivering SOTA performance.

cs.LG↗

arXiv2026

LEPO: Latent Reasoning Policy Optimization for Large Language Models

Recently, latent reasoning has been introduced into large language models (LLMs) to leverage rich information within a continuous space. However, without stochastic sampling, these methods inevitably collapse to deterministic inference, failing to discover diverse reasoning paths. To bridge the gap, we inject controllable stochasticity into latent reasoning via Gumbel-Softmax, restoring LLMs' exploratory capacity and enhancing their compatibility with Reinforcement Learning (RL). Building on this, we propose \textbf{\underline{L}}atent R\textbf{\underline{e}}asoning \textbf{\underline{P}}olicy \textbf{\underline{O}}ptimization~(\textbf{LEPO}), a novel framework that applies RL directly to continuous latent representations. Specifically, in rollout stage, LEPO maintains stochasticity to enable diverse trajectory sampling, while in optimization stage, LEPO constructs a unified gradient estimation for both latent representations and discrete tokens. Extensive experiments show that LEPO significantly outperforms existing RL methods for discrete and latent reasoning.

cs.LG↗

arXiv2026

Attention-Based Sampler for Diffusion Language Models

Auto-regressive models (ARMs) have established a dominant paradigm in language modeling. However, their strictly sequential sampling paradigm imposes fundamental constraints on both inference efficiency and modeling flexibility. To address these limitations, diffusion-based large language models (dLLMs) have been proposed, offering the potential for parallel sampling and flexible language modeling. Despite these advantages, current dLLMs sampling strategies rely primarily on token level information, which fails to account for global sequence structure and often yields suboptimal results. In this paper, we study the sampling order selection problem from the perspective of log-likelihood maximization. We show that this problem is NP-hard and propose an optimal sampling-rank-based approximation that makes the objective computationally tractable. We further prove that the tractable objective is optimized by sampling tokens in descending order of their attention-matrix column sums. This finding provides a principled justification for attention-guided sampling and offers a theoretically grounded alternative to greedy search. We instantiate this theoretical insight in a new training-free sampling algorithm, termed Attn-Sampler, and further propose dynamic attention thresholding for practical acceleration. Extensive experiments across multiple benchmarks validate the effectiveness of our proposed method, demonstrating that it achieves superior generation quality while enhancing the sampling parallelism.

cs.CL↗

arXiv2026

Two-timescale Optimization for Hybrid Mechanically and Electronically Tunable 6DMA Aided Communication

This letter proposes a hybrid mechanically and electronically tunable six-dimensional movable antenna (6DMA) base station (BS) architecture for future wireless communication networks. Such BS consists of multiple antenna arrays that are mechanically movable along a circular rail to adapt to the horizontal user hotspots, and each array is equipped with pattern reconfigurable antennas (PRAs) that are capable of electronically switching among a set of specified beam patterns to cater to the instantaneous user channels. The mechanical adjustment provides wide-angle coverage but suffers from slow response, while the electronic tuning enables rapid beam reconfiguration but with limited angular range. To effectively combine their complementary advantages, we propose to jointly design both mechanical and electronic configurations to maximize the average sum-rate of users via a two-timescale optimization approach, in which the array positions are optimized on the long timescale according to large-scale user distribution statistics, and the pattern selection vectors are optimized on the short timescale to enable fast beam alignment based on the instantaneous user locations. An alternating optimization algorithm based on the Monte Carlo sampling method is developed to solve the problem efficiently. Finally, simulation results show that our proposed design achieves significant performance gains over various benchmark schemes.

cs.IT↗

arXiv2026

GAPO: Robust Advantage Estimation for Real-World Code LLMs

Reinforcement learning (RL) is widely used for post-training large language models (LLMs) in code editing, where group-relative methods, such as GRPO, are popular due to their critic-free and normalized advantage estimation. However, in real-world code-editing scenarios, reward distributions are often skewed with unpredictable noise, leading to distorted advantage computation and increased rollout outliers. To address this issue, we propose Group Adaptive Policy Optimization (GAPO), which adaptively finds an interval with the highest SNR (Signal to Noise Ratio) per prompt and uses the median of that interval as an adaptive Q to replace the group mean in advantage calculation to reduce noise further. This adaptive Q robustly handles rollout noise while remaining plug-and-play and efficient. We evaluate GAPO on nine instruction-tuned LLMs (3B-14B) using a collected large dataset of 51,844 real-world, history-aware code-editing tasks spanning 10 programming languages. GAPO yields up to 4.35 in-domain (ID) and 5.30 out-of-domain (OOD) exact-match improvements over GRPO and its variant DAPO, while achieving lower clipping ratios and higher GPU throughput. Code: https://github.com/TsingZ0/verl-GAPO.

cs.LG↗

arXiv2025

Re-boosting Self-Collaboration Parallel Prompt GAN for Unsupervised Image Restoration

Unsupervised restoration approaches based on generative adversarial networks (GANs) offer a promising solution without requiring paired datasets. Yet, these GAN-based approaches struggle to surpass the performance of conventional unsupervised GAN-based frameworks without significantly modifying model structures or increasing the computational complexity. To address these issues, we propose a self-collaboration (SC) strategy for existing restoration models. This strategy utilizes information from the previous stage as feedback to guide subsequent stages, achieving significant performance improvement without increasing the framework's inference complexity. The SC strategy comprises a prompt learning (PL) module and a restorer ($Res$). It iteratively replaces the previous less powerful fixed restorer $\overline{Res}$ in the PL module with a more powerful $Res$. The enhanced PL module generates better pseudo-degraded/clean image pairs, leading to a more powerful $Res$ for the next iteration. Our SC can significantly improve the $Res$'s performance by over 1.5 dB without adding extra parameters or computational complexity during inference. Meanwhile, existing self-ensemble (SE) and our SC strategies enhance the performance of pre-trained restorers from different perspectives. As SE increases computational complexity during inference, we propose a re-boosting module to the SC (Reb-SC) to improve the SC strategy further by incorporating SE into SC without increasing inference time. This approach further enhances the restorer's performance by approximately 0.3 dB. Extensive experimental results on restoration tasks demonstrate that the proposed model performs favorably against existing state-of-the-art unsupervised restoration methods. Source code and trained models are publicly available at: https://github.com/linxin0/RSCP2GAN.

cs.CV↗

arXiv2025

Baichuan-M1: Pushing the Medical Capability of Large Language Models

The current generation of large language models (LLMs) is typically designed for broad, general-purpose applications, while domain-specific LLMs, especially in vertical fields like medicine, remain relatively scarce. In particular, the development of highly efficient and practical LLMs for the medical domain is challenging due to the complexity of medical knowledge and the limited availability of high-quality data. To bridge this gap, we introduce Baichuan-M1, a series of large language models specifically optimized for medical applications. Unlike traditional approaches that simply continue pretraining on existing models or apply post-training to a general base model, Baichuan-M1 is trained from scratch with a dedicated focus on enhancing medical capabilities. Our model is trained on 20 trillion tokens and incorporates a range of effective training methods that strike a balance between general capabilities and medical expertise. As a result, Baichuan-M1 not only performs strongly across general domains such as mathematics and coding but also excels in specialized medical fields. We have open-sourced Baichuan-M1-14B, a mini version of our model, which can be accessed through the following links.

cs.CL↗

arXiv2025

Hierarchically Tunable 6DMA for Wireless Communication and Sensing: Modeling and Performance Optimization

This paper proposes a new hierarchically tunable six-dimensional movable antenna (HT-6DMA) architecture for base station (BS) in future wireless networks. The HT-6DMA BS consists of multiple antenna arrays that can flexibly move on a spherical surface, with their three-dimensional (3D) positions and 3D rotations/orientations efficiently characterized in the global spherical coordinate system (SCS) and their individual local SCSs, respectively. As a result, the 6DMA system is hierarchically tunable in the sense that each array's global position and local rotation can be separately adjusted in a sequential manner with the other being fixed, thus greatly reducing their design complexity and improving the achievable performance. In particular, we consider an HT-6DMA BS serving multiple single-antenna users in the uplink communication or sensing potential unmanned aerial vehicles (UAVs)/drones in a given airway area. Specifically, for the communication scenario, we aim to maximize the average sum rate of communication users in the long term by optimizing the positions and rotations of all 6DMA arrays at the BS. While for the airway sensing scenario, we maximize the minimum received sensing signal power along the airway by optimizing the 6DMA arrays' positions and rotations along with the BS's transmit covariance matrix. Despite that the formulated problems are both non-convex, we propose efficient solutions to them by exploiting the hierarchical tunability of positions/rotations of 6DMA arrays in our proposed model. Numerical results show that the proposed HT-6DMA design significantly outperforms not only the traditional BS with fixed-position antennas (FPAs), but also the existing 6DMA scheme. Furthermore, it is unveiled that the performance gains of HT-6DMA mostly come from the arrays' global position adjustments on the spherical surface, rather than their local rotation adjustments.

eess.SP↗

arXiv2024

MetaGPT: Merging Large Language Models Using Model Exclusive Task Arithmetic

The advent of large language models (LLMs) like GPT-4 has catalyzed the exploration of multi-task learning (MTL), in which a single model demonstrates proficiency across diverse tasks. Task arithmetic has emerged as a cost-effective approach for MTL. It enables performance enhancement across multiple tasks by adding their corresponding task vectors to a pre-trained model. However, the current lack of a method that can simultaneously achieve optimal performance, computational efficiency, and data privacy limits their application to LLMs. In this paper, we propose \textbf{M}odel \textbf{E}xclusive \textbf{T}ask \textbf{A}rithmetic for merging \textbf{GPT}-scale models, which formalizes the objective of model merging into a multi-task learning framework, aiming to minimize the average loss difference between the merged model and each individual task model. Since data privacy limits the use of multi-task training data, we leverage LLMs' local linearity and task vectors' orthogonality to separate the data term and scaling coefficients term and derive a model-exclusive task arithmetic method. Our proposed MetaGPT is data-agnostic and bypasses the heavy search process, making it cost-effective and easy to implement for LLMs.Extensive experiments demonstrate that MetaGPT leads to improvements in task arithmetic and achieves state-of-the-art performance on multiple tasks.

cs.CL↗

arXiv2024

Improving Generalization of Deep Neural Networks by Optimum Shifting

Recent studies showed that the generalization of neural networks is correlated with the sharpness of the loss landscape, and flat minima suggests a better generalization ability than sharp minima. In this paper, we propose a novel method called \emph{optimum shifting}, which changes the parameters of a neural network from a sharp minimum to a flatter one while maintaining the same training loss value. Our method is based on the observation that when the input and output of a neural network are fixed, the matrix multiplications within the network can be treated as systems of under-determined linear equations, enabling adjustment of parameters in the solution space, which can be simply accomplished by solving a constrained optimization problem. Furthermore, we introduce a practical stochastic optimum shifting technique utilizing the Neural Collapse theory to reduce computational costs and provide more degrees of freedom for optimum shifting. Extensive experiments (including classification and detection) with various deep neural network architectures on benchmark datasets demonstrate the effectiveness of our method.

cs.LG↗

arXiv2023

Improving Lens Flare Removal with General Purpose Pipeline and Multiple Light Sources Recovery

When taking images against strong light sources, the resulting images often contain heterogeneous flare artifacts. These artifacts can importantly affect image visual quality and downstream computer vision tasks. While collecting real data pairs of flare-corrupted/flare-free images for training flare removal models is challenging, current methods utilize the direct-add approach to synthesize data. However, these methods do not consider automatic exposure and tone mapping in image signal processing pipeline (ISP), leading to the limited generalization capability of deep models training using such data. Besides, existing methods struggle to handle multiple light sources due to the different sizes, shapes and illuminance of various light sources. In this paper, we propose a solution to improve the performance of lens flare removal by revisiting the ISP and remodeling the principle of automatic exposure in the synthesis pipeline and design a more reliable light sources recovery strategy. The new pipeline approaches realistic imaging by discriminating the local and global illumination through convex combination, avoiding global illumination shifting and local over-saturation. Our strategy for recovering multiple light sources convexly averages the input and output of the neural network based on illuminance levels, thereby avoiding the need for a hard threshold in identifying light sources. We also contribute a new flare removal testing dataset containing the flare-corrupted images captured by ten types of consumer electronics. The dataset facilitates the verification of the generalization capability of flare removal methods. Extensive experiments show that our solution can effectively improve the performance of lens flare removal and push the frontier toward more general situations.

eess.IV↗