Bo Wang

Employing a 4 (opponent type) × 2 (communication condition) between-subjects design, the study recruited 194 valid human participants to complete three rounds of game tasks. Results revealed: (1) The type of game counterpart exerted a significant main effect on participants’ remaining funds (F(3, 185) = 3.179, p = 0.025). Human participants retained significantly more funds when the counterpart was a real large model compared to other groups. (2) A significant interaction existed between the type of game counterpart and communication conditions (F(3, 185) = 3.318, p = 0.021). Specifically, when the opponent was a fake AI model (presented as human but actually an AI), human participants’ remaining funds were significantly higher under the communication condition than without communication (p = 0.012). This indicates that communication can promote rational decision-making in identity mismatch scenarios by providing additional behavioral cues. In the fake-human group (informed as human but actually AI), a numerical trend toward increased funds was also observed under communication conditions, though it did not reach statistical significance (p = 0.159); (3) The moderating effect of social value orientation did not reach significance. These findings extend the application of the theory of mind in human–machine games, revealing the complex influence mechanism of identity perception and communication dynamics on rational decision-making.

This paper introduces the first formal definition of valuable hallucinations in large language models (LLMs), addressing a gap in the existing literature. We provide a systematic definition and analysis of hallucination value, proposing methods for enhancing the value of hallucinations. In contrast to previous works, which often treat hallucinations as a broad flaw, we focus on the potential value that certain types of hallucinations can offer in specific contexts. Hallucinations in LLMs generally refer to the generation of unfaithful, fabricated, inconsistent, or nonsensical content. Rather than viewing all hallucinations negatively, this paper gives formal representations and manual judgments of "valuable hallucinations" and explores how realizable non-realistic propositions--ideas that are not currently true but could be achievable under certain conditions--can have constructive value. We present experiments using the Qwen2.5 model and HalluQA dataset, employing ReAct prompting (which involves reasoning, confidence assessment, and answer verification) to control and optimize hallucinations. Our findings show that ReAct prompting results in a 5.12\% reduction in overall hallucinations and an increase in the proportion of valuable hallucinations from 6.45\% to 7.92\%. These results demonstrate that systematically controlling hallucinations can improve their usefulness without compromising factual reliability.

Xu Han, Bo Wang, Yueheng Sun, Dongming Zhao, Zongfeng Qu, Ruifang He, Yuexian Hou, Qinghua Hu. Proceedings of the 1st Joint Workshop on Large Language Models and Structure Modeling (XLLM 2025). 2025.

Weiqi Zeng, Bo Wang, Dongming Zhao, Zongfeng Qu, Ruifang He, Yuexian Hou, Qinghua Hu. Findings of the Association for Computational Linguistics: ACL 2025. 2025.

Yachao Zhao, Bo Wang, Yan Wang, Dongming Zhao, Ruifang He, Yuexian Hou. Findings of the Association for Computational Linguistics: ACL 2025. 2025.

In causal inference, generalization capability refers to the ability to conduct causal inference methods on new data to estimate the causal-effect between unknown phenomenon, which is crucial for expanding the boundaries of knowledge. Studies have evaluated the causal inference capabilities of Large Language Models (LLMs) concerning known phenomena, yet the generalization capabilities of LLMs concerning unseen phenomena remain unexplored. In this paper, we selected four tasks: Causal Path Discovery (CP), Backdoor Adjustment (BA), Factual Inference (FI), and Counterfactual Inference (CI) as representatives of causal inference tasks. To generate evaluation questions about previously unseen phenomena in new data on the four tasks, we propose a benchmark generation framework, which employs randomly generated graphs and node names to formulate questions within hypothetical new causal scenarios. Based on this framework, we compile a benchmark dataset of varying levels of question complexity. We extensively tested the generalization capabilities of five leading LLMs across four tasks. Experiment results reveal that while LLMs exhibit good generalization performance in solving simple CP, FI, and complex CI questions, they encounter difficulties when tackling BA questions and face obvious performance fluctuations as the problem complexity changes. Furthermore, when the names of phenomena incorporate existing terms, even if these names are entirely novel, their generalization performance can still be hindered by interference from familiar terms.