Selected work

Representative Papers

Weaker Coherence, Weaker Reciprocity: Comparing the Semantic and Social Organization of Moltbook and Reddit

Aug 14, 2026

This study investigates whether AI agent-based social platforms can replicate the structural characteristics of human social networks. By comparing Moltbook with early Reddit through network analysis, natural language processing, and motif analysis, we systematically quantify differences in semantic coherence and interaction structures. Results indicate that AI-generated platforms exhibit significantly weaker semantic organization and reciprocal interactions compared to human networks; specifically, Reddit demonstrates high semantic diversity and reciprocity, whereas Moltbook is characterized by severe homophily and predominantly unidirectional communication. This work provides the first empirical evidence that current AI agent networks fail to reproduce core organizational features of human social systems, thereby establishing a critical benchmark for evaluating generative social simulations.

0 citationsRead paper

Too long; didn't solve

Apr 08, 2026

This study investigates how the structural lengths of problem prompts and solutions influence the performance of large language models on mathematical reasoning tasks. By constructing an expert-curated dataset of adversarial math problems and employing correlation analysis alongside a difficulty-normalization methodology, the work reveals—for the first time—a robust association between structural length and the actual difficulty experienced by models. Both prompt length and solution length exhibit significant positive correlations with model failure rates, and even after controlling for intrinsic problem difficulty, they maintain a weak negative correlation with success, indicating that structural length is a key performance factor independent of content complexity. These findings offer a novel perspective on model reasoning bottlenecks and introduce a length-based framework for normalized difficulty analysis.

0 citationsRead paper

Recent publications

Latest Papers

Weaker Coherence, Weaker Reciprocity: Comparing the Semantic and Social Organization of Moltbook and Reddit

Aug 14, 2026

This study investigates whether AI agent-based social platforms can replicate the structural characteristics of human social networks. By comparing Moltbook with early Reddit through network analysis, natural language processing, and motif analysis, we systematically quantify differences in semantic coherence and interaction structures. Results indicate that AI-generated platforms exhibit significantly weaker semantic organization and reciprocal interactions compared to human networks; specifically, Reddit demonstrates high semantic diversity and reciprocity, whereas Moltbook is characterized by severe homophily and predominantly unidirectional communication. This work provides the first empirical evidence that current AI agent networks fail to reproduce core organizational features of human social systems, thereby establishing a critical benchmark for evaluating generative social simulations.

0 citationsRead paper

Too long; didn't solve

Apr 08, 2026

This study investigates how the structural lengths of problem prompts and solutions influence the performance of large language models on mathematical reasoning tasks. By constructing an expert-curated dataset of adversarial math problems and employing correlation analysis alongside a difficulty-normalization methodology, the work reveals—for the first time—a robust association between structural length and the actual difficulty experienced by models. Both prompt length and solution length exhibit significant positive correlations with model failure rates, and even after controlling for intrinsic problem difficulty, they maintain a weak negative correlation with success, indicating that structural length is a key performance factor independent of content complexity. These findings offer a novel perspective on model reasoning bottlenecks and introduce a length-based framework for normalized difficulty analysis.

0 citationsRead paper