Three New Studies Examine Explainability, Generalization, and Bias in Large Language Models

Three New Studies Examine Explainability, Generalization, and Bias in Large Language Models

Three new research papers posted to arXiv explore different aspects of large language model (LLM) behavior and capabilities.

According to arXiv paper 2512.20328v1, researchers are working toward explaining LLMs in software engineering tasks. The paper notes that “recent progress in Large Language Models (LLMs) has substantially advanced the automation of software engineering (SE) tasks, enabling complex activities such as code generation and code summarization.” However, the abstract acknowledges that “the black-box nature of LLMs remains a” challenge (the abstract appears incomplete in the source).

A second study (arXiv:2512.20162v1) compares human and LLM generalization using the “number game,” a concept inference task. According to the paper, researchers “examined the inductive biases and inference strategies of humans and LLMs” using a Bayesian model as an analytical framework.

The third paper (arXiv:2511.06148v2) examines how LLMs develop social biases. The research argues that “as large language models (LLMs) are adopted into frameworks that grant them the capacity to make real decisions, it is increasingly important to ensure that they are unbiased.” The paper’s title suggests LLMs can “develop novel social biases through adaptive exploration,” though the full abstract was not provided in the source materials.