Lingyu Li

Selected Research

1. arXiv

   Representational alignment yields generalizable safety in language models

   Lingyu Li, Yan Teng †, Yingchun Wang, Xia Hu

   TL;DR

   This paper introduces Representational Similarity Optimization (ReSO), a method that directly aligns hidden-layer representations with human moral categorization without supervising generated tokens. ReSO consistently improves adversarial robustness across multiple model scales against various jailbreak evaluations, providing computational evidence that internal representational reorganization is functionally linked to generalizable safety.

2. Communications Psychology

   Understanding large language models demands distinguishing human projection from machine cognition

   Lingyu Li, Yan Teng †, Yingchun Wang, Xia Hu

   TL;DR

   Current efforts to understand LLMs are largely metaphorical. Researchers map LLMs onto familiar domains, from physics and neuroscience to psychology and sociology, each illuminating specific facets while obscuring others. We chart these metaphors across mechanistic, behavioral, and interactive scales and delineate their explanatory boundaries. Crucially, this metaphorical projection creates a recursive loop of anthropomorphism, fueling the genuine understanding versus pattern matching impasse. As an alternative approach, we propose machine experientialism, positing that LLMs build their own form of understanding from training corpora. The priority shifts from cataloging LLMs' human-like traits to uncovering their distinct logic that emerges from this text-based world.

3. AAAI 2026

   The Other Mind: How Language Models Exhibit Human Temporal Cognition

   Lingyu Li, Yang Yao, Yixu Wang, Chunbo Li, Yan Teng †, Yingchun Wang

   The 40th Annual AAAI Conference on Artificial Intelligence

   TL;DR

   Through 24 million behavioral experiments, this study reveals that LLMs spontaneously develop a human-like subjective temporal perception adhering to the Weber-Fechner Law. Applying mechanistic interpretability into Neural Coding, Concept Representation, and Information Exposure, we demonstrate that this convergence of concept representation stems from time neurons that utilize logarithmic compression to encode latent non-linear temporal patterns within the training corpora. We propose Machine Experientialism, suggesting that LLMs' unique cognitive structures emerge from the dynamic interplay between their architectural properties and the informational environments they inhabit, thereby is “The Other Mind”.

4. ICML 2025

   Reflection-Bench: Evaluating Epistemic Agency in Large Language Models

   Lingyu Li, Yixu Wang, Haiquan Zhao, Shuqi Kong, Yan Teng †, Chunbo Li †, Yingchun Wang

   Proceedings of the 42nd International Conference on Machine Learning

   TL;DR

   When an LLM serves as an agent's brain, what is the core capability that defines its ceiling? We propose Epistemic Agency, the ability to flexibly construct, adapt, and monitor beliefs about the dynamic environments. Reflection-Bench evaluates epistemic agency utilizing 7 parameterized cognitive tests to minimize data contamination. We suggest several promising directions, including enhancing meta-cognition, developing mechanisms for dynamic shifts between intuitive and deliberative reasoning, and fostering organic coordination among cognitive capabilities.

5. Frontiers in Psychology

   Formalizing Lacanian psychoanalysis through the free energy principle

   Lingyu Li, Chunbo Li †

   Theoretical and Philosophical Psychology

   TL;DR

   We formalize Lacan's traditionally obscure philosophy of human mind using Free Energy Principle. We identify theoretical alignments between the two frameworks, develop a FEP-RSI model that maps the Real, Symbolic, and Imaginary orders to distinct, interacting neural networks. We model the Borromean interdependence as a message passing network, interpersonal desire as generalized synchronization, and the big Other as collective dynamics from social interactions. This study renders abstract Lacanian philosophy computationally tractable.

6. Psychiatry and Clinical Neurosciences

   Chain of Risks Evaluation (CORE): a framework for safer large language models in public mental health

   Lingyu Li, Haiquan Zhao, Shuqi Kong, Yan Teng †, Chunbo Li †, Yingchun Wang

   TL;DR

   Grounded in actor-network theory, we analyze human-LLM interactions as inter-agent dialogues and propose CORE. CORE categorizes LLMs risks in mental health into four progressive levels, from universal and context-specific to user-specific and user-context-specific. We advocate for a collaborative continuum between AI developers and mental health practitioners to ensure LLMs serve as safe tools for psychological support.

7. WIREs Cognitive Science

   Schizophrenia Research Under the Framework of Predictive Coding: Body, Language, and Others

   Lingyu Li †, Chunbo Li

   Wiley Interdisciplinary Reviews Cognitive Science

   TL;DR

   We establish a psychopathological model of schizophrenia by analyzing the disruption of human ontological existence across the domains of the body, language, and social interaction within the predictive coding framework. We illustrate that clinical manifestations such as disembodiment, formal thought disorders, and impaired theory of mind arise from imbalances in the precision-weighting of top-down priors and bottom-up prediction errors. Beyond a narrow psychopathological profile, this article utilizes these aberrant inferences as a window into the fundamental architecture of the human mind.