arXiv subjects

Yunbin Zhao

Publications and source records attributed to Yunbin Zhao.

7 recordsLinked to original sources

arXiv2026

IACM-RL: Intent-Aware Context Management and Reinforcement Learning for Complex Tool Invocation under Dynamic Intent Fluctuations

Executing long-horizon tool invocations in real-world environments is severely challenged by dynamic user intent noise. Existing methods attempt robustness via implicit history scanning or text compression, yet predominantly assume perfect instructions in simplistic scenarios. Inevitably, under fluctuating contexts, obsolete constraints dilute model attention, triggering catastrophic intent deviation and infinite API loops. To resolve this, we propose IACM-RL, a comprehensive framework for robust tool invocation. First, we introduce the DynamicIntent pipeline, synthesizing trajectories across 13 fine-grained fluctuation scenarios, paired with a five-dimensional diagnostic metric suite. Second, IACM-RL deploys a BeliefState-based Self-Generated Context Manager that proactively tracks shifting goals and isolates overwritten parameters using structural stale flags. To autonomously internalize this state-tracking capability, we optimize the policy using a hierarchical intent-driven reward alongside three auxiliary losses (action calibration, CM extraction, and state distillation). Experiments on DynamicIntent, BFCL-V3, and $\mathrm{\tau}^2$-Bench demonstrate that IACM-RL significantly outperforms baselines, reducing infinite loops and stale context errors while enhancing out-of-domain generalization.

cs.CL↗

arXiv2026

Entropy Polarity in Reinforcement Fine-Tuning: Direction, Asymmetry, and Control

Policy entropy has emerged as a fundamental measure for understanding and controlling exploration in reinforcement learning with verifiable rewards (RLVR) for LLMs. However, existing entropy-aware methods mainly regulate entropy through global objectives, while the token-level mechanism by which sampled policy updates reshape policy entropy remains underexplored. In this work, we develop a theoretical framework of entropy mechanics in RLVR. Our analysis yields a first-order approximation of the entropy change, giving rise to entropy polarity, a signed token-level quantity that predicts how much a sampled update expands or contracts entropy. This analysis further reveals a structural asymmetry: reinforcing frequent high-probability tokens triggers contraction tendencies, whereas expansive tendencies typically require lower-probability samples or stronger distributional correction. Empirically, we show that entropy polarity reliably predicts entropy changes, and that positive and negative polarity branches play complementary roles in preserving exploration while strengthening exploitation. Building on these insights, we propose Polarity-Aware Policy Optimization (PAPO), which preserves both polarity branches and implements entropy control through advantage reweighting. With the empirical entropy trajectory as an online phase signal, PAPO adaptively reallocates optimization pressure between entropy-expanding and entropy-contracting updates. Experiments on mathematical reasoning and agentic benchmarks show that PAPO consistently outperforms competitive baselines, while delivering superior training efficiency and substantial reward improvements.

cs.LG↗

arXiv2026

MagicAgent: Towards Generalized Agent Planning

The evolution of Large Language Models (LLMs) from passive text processors to autonomous agents has established planning as a core component of modern intelligence. However, achieving generalized planning remains elusive, not only by the scarcity of high-quality interaction data but also by inherent conflicts across heterogeneous planning tasks. These challenges result in models that excel at isolated tasks yet struggle to generalize, while existing multi-task training attempts suffer from gradient interference. In this paper, we present \textbf{MagicAgent}, a series of foundation models specifically designed for generalized agent planning. We introduce a lightweight and scalable synthetic data framework that generates high-quality trajectories across diverse planning tasks, including hierarchical task decomposition, tool-augmented planning, multi-constraint scheduling, procedural logic orchestration, and long-horizon tool execution. To mitigate training conflicts, we propose a two-stage training paradigm comprising supervised fine-tuning followed by multi-objective reinforcement learning over both static datasets and dynamic environments. Empirical results show that MagicAgent-32B and MagicAgent-30B-A3B achieve superior performance across diverse open-source benchmarks (\emph{e.g.}, $75.1\%$ on Worfbench and $86.9\%$ on BFCL-v3), as well as strong results on our in-house MagicEval benchmarks, substantially outperforming existing sub-100B models and surpassing leading ultra-scale models, including GPT-5.2, Kimi-K2 and GLM-4.7.

cs.AI↗

arXiv2022

Short Range Correlation Transformer for Occluded Person Re-Identification

Occluded person re-identification is one of the challenging areas of computer vision, which faces problems such as inefficient feature representation and low recognition accuracy. Convolutional neural network pays more attention to the extraction of local features, therefore it is difficult to extract features of occluded pedestrians and the effect is not so satisfied. Recently, vision transformer is introduced into the field of re-identification and achieves the most advanced results by constructing the relationship of global features between patch sequences. However, the performance of vision transformer in extracting local features is inferior to that of convolutional neural network. Therefore, we design a partial feature transformer-based person re-identification framework named PFT. The proposed PFT utilizes three modules to enhance the efficiency of vision transformer. (1) Patch full dimension enhancement module. We design a learnable tensor with the same size as patch sequences, which is full-dimensional and deeply embedded in patch sequences to enrich the diversity of training samples. (2) Fusion and reconstruction module. We extract the less important part of obtained patch sequences, and fuse them with original patch sequence to reconstruct the original patch sequences. (3) Spatial Slicing Module. We slice and group patch sequences from spatial direction, which can effectively improve the short-range correlation of patch sequences. Experimental results over occluded and holistic re-identification datasets demonstrate that the proposed PFT network achieves superior performance consistently and outperforms the state-of-the-art methods.

cs.CV↗

arXiv2018

Sparsification of Matrices and Compressed Sensing

Compressed sensing is a signal processing technique whereby the limits imposed by the Shannon--Nyquist theorem can be exceeded provided certain conditions are imposed on the signal. Such conditions occur in many real-world scenarios, and compressed sensing has emerging applications in medical imaging, big data, and statistics. Finding practical matrix constructions and computationally efficient recovery algorithms for compressed sensing is an area of intense research interest. Many probabilistic matrix constructions have been proposed, and it is now well known that matrices with entries drawn from a suitable probability distribution are essentially optimal for compressed sensing. Potential applications have motivated the search for constructions of sparse compressed sensing matrices (i.e., matrices containing few non-zero entries). Various constructions have been proposed, and simulations suggest that their performance is comparable to that of dense matrices. In this paper, extensive simulations are presented which suggest that sparsification leads to a marked improvement in compressed sensing performance for a large class of matrix constructions and for many different recovery algorithms.

cs.IT↗

arXiv2013

RSP-Based Analysis for Sparsest and Least $\ell_1$-Norm Solutions to Underdetermined Linear Systems

Recently, the worse-case analysis, probabilistic analysis and empirical justification have been employed to address the fundamental question: When does $\ell_1$-minimization find the sparsest solution to an underdetermined linear system? In this paper, a deterministic analysis, rooted in the classic linear programming theory, is carried out to further address this question. We first identify a necessary and sufficient condition for the uniqueness of least $\ell_1$-norm solutions to linear systems. From this condition, we deduce that a sparsest solution coincides with the unique least $\ell_1$-norm solution to a linear system if and only if the so-called \emph{range space property} (RSP) holds at this solution. This yields a broad understanding of the relationship between $\ell_0$- and $\ell_1$-minimization problems. Our analysis indicates that the RSP truly lies at the heart of the relationship between these two problems. Through RSP-based analysis, several important questions in this field can be largely addressed. For instance, how to efficiently interpret the gap between the current theory and the actual numerical performance of $\ell_1$-minimization by a deterministic analysis, and if a linear system has multiple sparsest solutions, when does $\ell_1$-minimization guarantee to find one of them? Moreover, new matrix properties (such as the \emph{RSP of order $K$} and the \emph{Weak-RSP of order $K$}) are introduced in this paper, and a new theory for sparse signal recovery based on the RSP of order $K$ is established.

cs.IT↗

arXiv2012

Rank-one Solutions for Homogeneous Linear Matrix Equations over the Positive Semidefinite Cone

The problem of finding a rank-one solution to a system of linear matrix equations arises from many practical applications. Given a system of linear matrix equations, however, such a low-rank solution does not always exist. In this paper, we aim at developing some sufficient conditions for the existence of a rank-one solution to the system of homogeneous linear matrix equations (HLME) over the positive semidefinite cone. First, we prove that an existence condition of a rank-one solution can be established by a homotopy invariance theorem. The derived condition is closely related to the so-called $P_\emptyset$ property of the function defined by quadratic transformations. Second, we prove that the existence condition for a rank-one solution can be also established through the maximum rank of the (positive semidefinite) linear combination of given matrices. It is shown that an upper bound for the rank of the solution to a system of HLME over the positive semidefinite cone can be obtained efficiently by solving a semidefinite programming (SDP) problem. Moreover, a sufficient condition for the nonexistence of a rank-one solution to the system of HLME is also established in this paper.

math.OC↗