---
title: "Ze Wang"
type: "entity"
entity_kind: "person"
status: "hub"
canonical_name: "Ze Wang"
aliases: []
first_seen: "2026-09-14T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["bias amplification","model collapse","political bias","iterated retraining","Holistic AI"]
seek_code_commit: "546fa57"
---


Researcher at Holistic AI and first/corresponding author of "Bias
Amplification: Large Language Models as Increasingly Biased Media" (arXiv
2410.15234; IJCNLP-AACL 2025 main-conference long papers, ACL Anthology
2025.ijcnlp-long.8), with Zekun Wu, Jeremy Zhang, Xin Guan, Navya Jain,
Skylar Lu, Saloni Gupta, and Adriano Koshiyama. (The anthology version
renders two of those names differently — "Yichi Zhang" and "Qinyang Lu";
the list above follows the arXiv v3 title block.)

Matters to this vault as the author of the only located study to run a
genuine multi-generation iterated-retraining experiment (GPT-2, ten
generations) that isolates a specific bias's compounding trajectory —
political lean in news-continuation text — from general model-collapse
degradation, via a validated benchmark and neuron-level mechanistic
analysis. The design is the closest existing analog to what this vault's
[[question-does-citation-popularity-bias-compound-across-llm-training-generations|open
citation-popularity-bias question]] is still waiting to see run against
citations instead of politics.

## References
- [[claim-wang-2024-bias-amplification-persists-independent-of-model-collapse]]
- Capture: 10-inbox/raw/2026-09-14-has-any-study-run-a-genuine-multi-generation.md

## Updates

- 2026-09-15 — correction (scheduled cross-model audit, claude-opus-5). This
  page previously gave the venue as "Findings of IJCNLP-AACL 2025." The paper
  is in the main-conference long-paper proceedings volume
  (`2025.ijcnlp-long.8`, pp. 115–132), read directly at
  aclanthology.org/volumes/2025.ijcnlp-long/; "Findings" contradicted the
  anthology ID the claim-note recorded alongside it.