Zijian Li
Publications
Token-level Data Selection for Safe LLM Fine-tuning
Fine-tuning large language models (LLMs) on custom datasets has become a standard approach for adapting these models to specific domains and applications. However, recent studies have shown that such fine-tuning can lead to significant degradation in the model's safety. Existing defense methods operate at the sample level and often suffer from an unsatisfactory trade-off between safety and utility. To address this limitation, we perform a systematic token-level diagnosis of safety degradation during fine-tuning. Based on this, we propose token-level data selection for safe LLM fine-tuning (TOSS), a novel framework that quantifies the safety risk of each token by measuring the loss difference between a safety-degraded model and a utility-oriented model. This token-level granularity enables accurate identification and removal of unsafe tokens, thereby preserving valuable task-specific information. In addition, we introduce a progressive refinement strategy, TOSS-Pro, which iteratively enhances the safety-degraded model's ability to identify unsafe tokens. Extensive experiments demonstrate that our approach robustly safeguards LLMs during fine-tuning while achieving superior downstream task performance, significantly outperforming existing sample-level defense methods. Our code is available at https://github.com/Polly-LYP/TOSS.
WirelessAgent++: Automated Agentic Workflow Design and Benchmarking for Wireless Networks
The integration of large language models (LLMs) into wireless networks has sparked growing interest in building autonomous AI agents for wireless tasks. However, existing approaches rely heavily on manually crafted prompts and static agentic workflows, a process that is labor-intensive, unscalable, and often suboptimal. In this paper, we propose WirelessAgent++, a framework that automates the design of agentic workflows for various wireless tasks. By treating each workflow as an executable code composed of modular operators, WirelessAgent++ casts agent design as a program search problem and solves it with a domain-adapted Monte Carlo Tree Search (MCTS) algorithm. Moreover, we establish WirelessBench, a standardized multi-dimensional benchmark suite comprising Wireless Communication Homework (WCHW), Network Slicing (WCNS), and Mobile Service Assurance (WCMSA), covering knowledge reasoning, code-augmented tool use, and multi-step decision-making. Experiments demonstrate that \wap{} autonomously discovers superior workflows, achieving test scores of $78.37\%$ (WCHW), $90.95\%$ (WCNS), and $97.07\%$ (WCMSA), with a total search cost below $\$ 5$ per task. Notably, our approach outperforms state-of-the-art prompting baselines by up to $31\%$ and general-purpose workflow optimizers by $11.1\%$, validating its effectiveness in generating robust, self-evolving wireless agents. The code is available at https://github.com/jwentong/WirelessAgent-R2.
Evidence-Augmented Policy Optimization with Reward Co-Evolution for Long-Context Reasoning
While Reinforcement Learning (RL) has advanced LLM reasoning, applying it to long-context scenarios is hindered by sparsity of outcome rewards. This limitation fails to penalize ungrounded "lucky guesses," leaving the critical process of needle-in-a-haystack evidence retrieval largely unsupervised. To address this, we propose EAPO (Evidence-Augmented Policy Optimization). We first establish the Evidence-Augmented Reasoning paradigm, validating via Tree-Structured Evidence Sampling that precise evidence extraction is the decisive bottleneck for long-context reasoning. Guided by this insight, EAPO introduces a specialized RL algorithm where a reward model computes a Group-Relative Evidence Reward, providing dense process supervision to explicitly improve evidence quality. To sustain accurate supervision throughout training, we further incorporate an Adaptive Reward-Policy Co-Evolution mechanism. This mechanism iteratively refines the reward model using outcome-consistent rollouts, sharpening its discriminative capability to ensure precise process guidance. Comprehensive evaluations across eight benchmarks demonstrate that EAPO significantly enhances long-context reasoning performance compared to SOTA baselines.