Author name cluster

Xiang Luo

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

EAAI Journal 2026 Journal Article

Self-verified user simulator via code-based interpretation in task-oriented dialogues

Xiang Luo
Jin Wang
Liang-Chih Yu
Xuejie Zhang

User simulators are essential for training and evaluating task-oriented dialogue systems (TODs). Recently, large language models (LLMs) have been increasingly adopted to construct user simulators by prompting them to generate natural language utterances and dialogue actions. However, due to the difficulty of controlling structured outputs through natural language prompts alone, these LLM-based simulators often produce incomplete, inconsistent, or invalid dialogue actions, limiting their effectiveness. To tackle this, this paper proposes a self-verified code-based user simulator that guides LLMs to generate intermediate Python code for structured dialogue actions. These code snippets are executed and validated by an external interpreter, and the verified outputs are used to refine the simulator’s behavior. Experiments on the Multi-Domain Wizard-of-Oz (MultiWOZ) dataset demonstrate that our method improves dialogue action accuracy by 4. 0%, and significantly enhances utterance diversity, achieving 12. 1% more trigrams, an increase of 0. 93 in entropy, and a 13. 2% gain in measure of textual lexical diversity (MTLD) over 100 dialogue turns. These results highlight the effectiveness of code-level verification in improving the controllability, correctness, and expressiveness of LLM-based user simulators.

Details DOI

NeurIPS Conference 2025 Conference Paper

Steering When Necessary: Flexible Steering Large Language Models with Backtracking

Zifeng Cheng
Jinwei Gan
Zhiwei Jiang
Cong Wang
Yafeng Yin
Xiang Luo
Yuchen Fu
Qing Gu

Large language models (LLMs) have achieved remarkable performance across many generation tasks. Nevertheless, effectively aligning them with desired behaviors remains a significant challenge. Activation steering is an effective and cost-efficient approach that directly modifies the activations of LLMs during the inference stage, aligning their responses with the desired behaviors and avoiding the high cost of fine-tuning. Existing methods typically indiscriminately intervene to all generations or rely solely on the question to determine intervention, which limits the accurate assessment of the intervention strength. To this end, we propose the F lexible A ctivation S teering with B acktracking ( FASB ) framework, which dynamically determines both the necessity and strength of intervention by tracking the internal states of the LLMs during generation, considering both the question and the generated content. Since intervening after detecting a deviation from the desired behavior is often too late, we further propose the backtracking mechanism to correct the deviated tokens and steer the LLMs toward the desired behavior. Extensive experiments on the TruthfulQA dataset and six multiple-choice datasets demonstrate that our method outperforms baselines. Our code will be released at https: //github. com/gjw185/FASB.

PDF Details