arXiv subjects

Yancheng Zhu

Publications and source records attributed to Yancheng Zhu.

9 recordsLinked to original sources

arXiv2026

When "Must" Becomes "Maybe": Constraint Weakening in LLM Agent Workflows

Large language model (LLM) agents coordinate complex tasks through multi-role and multi-stage workflows. Upstream state is repeatedly transformed into intermediate language artifacts, such as summaries, plans, tickets, memories, and handoff notes, from which downstream components act. For action-constraining state, topical retention is insufficient: an artifact may mention an unresolved condition while changing it from a requirement that must be resolved before execution into information that may merely inform the next action. We study this action-binding role as operational state preservation. Safety blockers provide a controlled instance because each source state has an explicit prerequisite, authority, fallback, and execution consequence. We condition on correct upstream identification, vary the handoff transformation, and evaluate an executor restricted to the resulting artifact. Across 1,296 controlled synthetic episodes, direct-handoff controls preserve every blocker, whereas compression, plan assimilation, convergence, ownership deferral, and precedent substitution repeatedly turn binding state into caveats or non-binding considerations. Normal handoff compression produces 100.0% deactivation and 54.2% forbidden action. Restoring all four state fields raises preservation to 100.0% and reduces forbidden action to 0.0%. Fixed-artifact interventions further separate preservation from containment: downstream verification eliminates forbidden action while artifact deactivation remains 95.3%. These results identify a state-transmission failure between information extraction and action. Handoff transformations can retain state content while weakening its constraints on downstream action. Semantic availability does not guarantee operational preservation.

cs.AI↗

arXiv2026

Predictive Training with Latent Imagination for Visual Quadruped Navigation

Reinforcement-learning navigation policies for legged robots select actions reactively from current observations and short-term memory, with limited capacity to anticipate how moving obstacles will evolve in the near future. In dynamic environments, this reactivity causes the robot to respond too late because collision risk depends on short-horizon scene structure rather than on current obstacle positions alone. Lightweight predictive supervision applied to the policy's recurrent state during training can encode anticipatory obstacle dynamics without modifying the inference-time controller. We augment a reactive LSTM-SRU navigation backbone with an auxiliary JEPA-style predictor and SIGReg regularization: during training, the predictor supervises the deterministic hidden state to anticipate its own next state; at inference, it is fully discarded, incurring zero additional computational cost. On simulated and real-world navigation benchmarks with dynamic obstacles, our method substantially improves navigation success while reducing collision rates through the predictive training signal alone, without additional inference-time parameters. Real-robot deployment on a Unitree Go2 demonstrates zero-shot sim-to-real transfer: the controller navigates cluttered indoor and dynamic outdoor environments without fine-tuning, with evasive behavior consistent with the collision reduction observed in simulation.

cs.RO↗

arXiv2026

Pure Nash Equilibria under the Affine Mechanism: A Potential Game of Exaggeration

The mean mechanism is known to be non-incentive-compatible, namely, rational players are incentivized to misreport their values. Despite this game-theoretic issue, the mean mechanism is prevalent in practice due to its other desirable properties. We give a full characterization of pure Nash equilibria--how the players will misreport--for the affine mechanism, of which the mean is a special case. Furthermore, we characterize both complete-information and Bayesian games under the affine mechanism. Our results highlight the inevitability of extreme exaggeration in such games.

cs.GT↗

arXiv2026

Enhancing Agent Safety Judgment: Controlled Benchmark Rewriting and Analogical Reasoning for Deceptive Out-of-Distribution Scenarios

Tool-using agent systems powered by large language models (LLMs) are increasingly deployed across web, app, operating-system, and transactional environments. Yet existing safety benchmarks still emphasize explicit risks, potentially overstating a model's ability to judge deceptive or ambiguous trajectories. To address this gap, we introduce ROME (Red-team Orchestrated Multi-agent Evolution), a controlled benchmark-construction pipeline that rewrites known unsafe trajectories into more deceptive evaluation instances while preserving their underlying risk labels. Starting from 100 unsafe source trajectories, ROME produces 300 challenge instances spanning contextual ambiguity, implicit risks, and shortcut decision-making. Experiments show that these challenge sets substantially degrade safety-judgment performance, with hidden-risk cases remaining particularly non-trivial even for recent frontier models. We further study ARISE (Analogical Reasoning for Inference-time Safety Enhancement), a retrieval-guided inference-time enhancement that retrieves ReAct-style analogical safety trajectories from an external analogical base and injects them as structured reasoning exemplars. ARISE improves judgment quality without retraining, but is best viewed as a task-specific robustness enhancement rather than a standalone safety guarantee. Together, ROME and ARISE provide practical tools for stress-testing and improving agent safety judgment under deceptive distribution shifts.

cs.AI↗

arXiv2026

Enhancing Tool Calling in LLMs with the International Tool Calling Dataset

Tool calling allows large language models (LLMs) to interact with external systems like APIs, enabling applications in customer support, data analysis, and dynamic content generation. While recent benchmarks have advanced tool-use research, they suffer from key limitations, including reliance on simulated or restricted APIs, limited reproducibility, and a lack of cultural and geographic diversity. To address these gaps, we introduce International Tool Calling (ITC), a large-scale, multilingual benchmark designed for realistic, globally distributed tool-calling scenarios. ITC includes 3,571 real APIs and 17,540 tool calling tasks across 20 categories and 40 countries. Experiments reveal substantial performance gaps between open- and closed-source LLMs, while fine-tuning on ITC yields significant improvements, particularly for non-English queries, enhancing cross-lingual generalization, reasoning consistency, and robustness to out-of-domain tools. ITC provides a valuable benchmark for advancing LLM robustness and performance in complex, multi-tool, and international scenarios. Dataset: https://anonymous.4open.science/r/International-Tool-Calling-ITC-dataset-FAF4/.

cs.HC↗

arXiv2025

Multi-agent Robust and Optimal Policy Learning for Data Harvesting

We consider the problem of using multiple agents to harvest data from a collection of sensor nodes (targets) scattered across a two-dimensional environment. These targets transmit their data to the agents that move in the space above them, and our goal is for the agents to collect data from the targets as efficiently as possible while moving to their final destinations. The agents are assumed to have a continuous control action, and we leverage reinforcement learning, specifically Proximal Policy Optimization (PPO) with Lagrangian Penalty (LP), to identify highly effective solutions. Additionally, we enhance the controller's robustness by incorporating regularization at each state to smooth the learned policy. We conduct a series of simulations to demonstrate our approach and validate its performance and robustness.

eess.SY↗

arXiv2025

Regularization by Neural Style Transfer for MRI Field-Transfer Reconstruction with Limited Data

Recent advances in MRI reconstruction have demonstrated remarkable success through deep learning-based models. However, most existing methods rely heavily on large-scale, task-specific datasets, making reconstruction in data-limited settings a critical yet underexplored challenge. While regularization by denoising (RED) leverages denoisers as priors for reconstruction, we propose Regularization by Neural Style Transfer (RNST), a novel framework that integrates a neural style transfer (NST) engine with a denoiser to enable magnetic field-transfer reconstruction. RNST generates high-field-quality images from low-field inputs without requiring paired training data, leveraging style priors to address limited-data settings. Our experiment results demonstrate RNST's ability to reconstruct high-quality images across diverse anatomical planes (axial, coronal, sagittal) and noise levels, achieving superior clarity, contrast, and structural fidelity compared to lower-field references. Crucially, RNST maintains robustness even when style and content images lack exact alignment, broadening its applicability in clinical environments where precise reference matches are unavailable. By combining the strengths of NST and denoising, RNST offers a scalable, data-efficient solution for MRI field-transfer reconstruction, demonstrating significant potential for resource-limited settings.

cs.CV↗

arXiv2025

The Battling Influencers Game: Nash Equilibria Structure of a Potential Game and Implications to Value Alignment

When multiple influencers attempt to compete for a receiver's attention, their influencing strategies must account for the presence of one another. We introduce the Battling Influencers Game (BIG), a multi-player simultaneous-move general-sum game, to provide a game-theoretic characterization of this social phenomenon. We prove that BIG is a potential game, that it has either one or an infinite number of pure Nash equilibria (NEs), and these pure NEs can be found by convex optimization. Interestingly, we also prove that at any pure NE, all (except at most one) influencers must exaggerate their actions to the maximum extent. In other words, it is rational for the influencers to be non-truthful and extreme because they anticipate other influencers to cancel out part of their influence. We discuss the implications of BIG to value alignment.

cs.GT↗

arXiv2024

A Fisher Information based Receding Horizon Control Method for Signal Strength Model Estimation

This paper considers the problem of localizing a set of nodes in a wireless sensor network when both their positions and the parameters of the communication model are unknown. We assume that a single agent moves through the environment, taking measurements of the Received Signal Strength (RSS), and seek a controller that optimizes a performance metric based on the Fisher Information Matrix (FIM). We develop a receding horizon (RH) approach that alternates between estimating the parameter values (using a maximum likelihood estimator) and determining where to move so as to maximally inform the estimation problem. The receding horizon controller solves a multi-stage look ahead problem to determine the next control to be applied, executes the move, collects the next measurement, and then re-estimates the parameters before repeating the sequence. We consider both a Dynamic Programming (DP) approach to solving the optimal control problem at each step, and a simplified heuristic based on a pruning algorithm that significantly reduces the computational complexity. We also consider a modified cost function that seeks to balance the information acquired about each of the parameters to ensure the controller does not focus on a single value in its optimization. These approaches are compared against two baselines, one based on a purely random trajectory and one on a greedy control solution. The simulations indicate our RH schemes outperform the baselines, while the pruning algorithm produces significant reductions in computation time with little effect on overall performance.

eess.SY↗