Neuron based Personality Trait Induction in Large Language Models
Authors: Jia Deng, Tianyi Tang, Yanbin Yin, Wenhao yang, Xin Zhao, Ji-Rong Wen
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate the efficacy of our neuron identification and trait induction methods. Notably, our approach achieves comparable performance as fine-tuned models, offering a more efficient and flexible solution for personality trait induction in LLMs. Extensive experiments using various evaluation methods on different LLMs have verified the effectiveness and generality of our method. |
| Researcher Affiliation | Collaboration | 1Gaoling School of Artificial Intelligence, Renmin University of China. 2Tongyi Lab. 3Institute of Statistics and Big Data, Renmin University of China. |
| Pseudocode | No | The paper describes methods through mathematical formulations (equations 1-4) and detailed textual explanations, but it does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured procedures in a code-like format. |
| Open Source Code | Yes | We provide access to all the mentioned resources at https://github.com/RUCAIBox/NPTI. |
| Open Datasets | Yes | First, we construct PERSONALITYBENCH, a large-scale dataset for identifying and evaluating personality traits in LLMs. ... Our PERSONALITYBENCH has 180,000 instances for identifying neurons and around 450 instances for evaluating LLMs personality induction. ... We provide access to all the mentioned resources at https://github.com/RUCAIBox/NPTI. |
| Dataset Splits | Yes | Our PERSONALITYBENCH has 180,000 instances for identifying neurons and around 450 instances for evaluating LLMs personality induction. |
| Hardware Specification | Yes | Our experiments are conducted on a single A800 GPU. |
| Software Dependencies | No | The paper mentions leveraging the 'vllm toolkit' for neuron identification and manipulation, and using 'gpt-4o-20240806 API' for benchmark construction and evaluation. However, it does not specify version numbers for the 'vllm toolkit' or any other specific software libraries required to reproduce the methodology's implementation. |
| Experiment Setup | Yes | During training, we set the learning rate to 1e-4 with a cosine decay. The rank of Lo RA is set to 8, and the batch size is configured to 8. ... As for the hyperparamters in Equation 4, we set γ = 1.4 and assign f(δ) = 1 / (1+e^(-10*(|δ| - 0.15))). |