Xujie Si

Xujie Si is an assistant professor in the Department of Computer Science at the University of Toronto, a faculty member at Mila, and a Canada CIFAR AI Chair. His research spans program verification, neuro-symbolic learning, and rigorous evaluation of mechanistic interpretability methods in large language models.

In October 2026, Si and Chuqin Geng released “Objective-Level Recovery Gaps in Mechanistic Interpretability: Context Distortion Can Misrank Circuits” (arXiv:2610.02098). Evaluating four major automated circuit-discovery algorithms (EAP, EAP-IG, ACDC, and Edge-SP) across InterpBench and four human-reference pretrained language-model tasks, the study proves that standard intervention-defined faithfulness (output KL divergence under resampling, mean, or zero ablation) systematically prefers equally sized circuits that reproduce the intact model’s behavior less well on held-out prompts, misranking 9.4 to 41.2 percent of discovered candidate pairs.

Si’s profile is Builder. By demonstrating that restoring selected intact-model context signals repairs 96 out of 100 persistent misrankings without changing the candidate circuits, his work exposes a foundational measurement flaw in how internal machine mechanisms are scored and selected.

Known for. Objective-level recovery gaps and context distortion in mechanistic interpretability circuit discovery, program synthesis and neuro-symbolic verification

Coverage on this site

• Objective-Level Recovery Gaps in Mechanistic Interpretability and SAEs October 2026

Related researchers

• CG Chuqin GengSchool of Computer Science, McGill University. Mila - Quebec AI InstituteObjective-level recovery gaps and context distortion in mechanistic interpretability circuit evaluation, neural network verification Builder1 article
• SL Salima LamsiyahDepartment of Computer Science, University of LuxembourgCritical survey of 138 Sparse Autoencoder (SAE) studies for mechanistic interpretability in NLP, four-outcome separation of recoverability, semantic interpretability, mechanistic faithfulness, and behavioral utility Builder1 article
• PB Patrick ButlinEleos AI Research. Previously the Global Priorities Institute, OxfordLead author of the indicator properties framework for assessing consciousness in AI systems AI welfare24 articles
• VA Vikas AgnihotriIndependent Aerospace and Neuromorphic Systems ResearcherNeuromorphic Satellite Intelligence framework, radiation-induced energy overhead model for spiking neural hardware, constrained on-orbit STDP Builder1 article
• MA Matt AngleParadromicsFounder and CEO of Paradromics, developer of the Connexus Direct Data Interface high-data-rate brain-computer interface Builder1 article
• IB Isabella R. BeckettCoauthor of the MaleCNS preprint comparing male and female Drosophila brain connectomes Builder1 article

Xujie Si is listed with the other builders, who hold that the claim is about what the hardware does, and the philosophy is left open. The positions themselves are set out in the index of consciousness theories. Every researcher covered on this site is indexed in the directory of consciousness researchers.