Arrow Research search

Author name cluster

Zichen Song

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

2 papers
1 author row

Possible papers

2

AAAI Conference 2026 Conference Paper

Optimization and Robustness-Informed Membership Inference Attacks for LLMs

  • Zichen Song
  • Qixin Zhang
  • Ming Li
  • Yao Shu

The proliferation of Large Language Models (LLMs) has raised concerns over training data privacy. Membership Inference Attacks (MIA), aiming to identify whether specific data was used for training, pose significant privacy risks. However, existing MIA methods struggle to address the scale and complexity of modern LLMs. This paper introduces OR-MIA, a novel MIA framework inspired by model optimization and input robustness. First, training data points are expected to exhibit smaller gradient norms due to optimization dynamics. Second, member samples show greater stability, with gradient norms being less sensitive to controlled input perturbations. OR-MIA leverages these principles by perturbing inputs, computing gradient norms, and using them as features for a robust classifier to distinguish members from non-members. Evaluations on LLMs (70M to 6B parameters) and various datasets demonstrate that OR-MIA outperforms existing methods, achieving over 90% accuracy. Our findings highlight a critical vulnerability in LLMs and underscore the need for improved privacy-preserving training paradigms.

PDF Details DOI

AAMAS Conference 2026 Conference Paper

UAM-MARL: Uncertainty-Aware Modality-Enhanced Multi-Agent Reinforcement Learning with LLM-Guided Graph Policies

  • Zichen Song
  • Weijia Li

Multi-agent systems (MAS) are increasingly deployed to solve complex embodied tasks, yet coordination efficiency, reward design, and robustness remain persistent challenges. While multi-agent reinforcement learning (MARL) provides a principled framework for cooperation, existing LLM-driven approaches often assume perfect perception and static planning, which is unrealistic in noisy and dynamic environments. In practice, discrepancies between language-based reasoning and uncertain multimodal perception, the semantic–perception gap, lead to incorrect subgoal assignments, misaligned rewards, and unstable coordination. To address this limitation, we propose UAM-MARL, an uncertainty-aware modalityenhanced MARL framework. UAM-MARL integrates three components: (1) an uncertainty-aware perception module that estimates confidence scores over multimodal inputs and propagates them to the planner, (2) a cross-modal consistency checker that validates the alignment between LLM-generated plans and environment observations, and (3) an uncertainty-weighted reward generator that composesindividualandteamrewardsbyscalingreasoning-derived signals with perception confidence. These modules augment the LLM-based planner–critic and homology-guided graph policy, enabling more reliable coordination under noisy observations and dynamic disturbances. Experiments in the AI2-THOR simulator demonstrate that UAM-MARL achieves superior performance compared to centralized-LLM, dialogue-based LLM, and state-of-the-art MARL baselines, yielding higher success rates, shorter completion times, lower token costs, and stronger robustness to perception noise. Ablation studies further confirm the contribution of uncertainty modeling and cross-modal consistency to improving both efficiency and scalability.

PDF
v2026.09.13