Localized, unstable, and entangled: Exploring `Us vs. Them' bias in Large Language Models
Abstract
Large language models (LLMs) exhibit us-vs.-them bias: A behavioral asymmetry in which prompts framed around an ingroup (``we''/``us'') receive systematically more positive continuations than matched prompts framed around an outgroup (``they''/``them''). Using Edge Attribution Patching (EAP), we localize this behavior to directed circuits in GPT-2 Small, GPT-2 Large, Llama-2 7B, and Llama-3 8B. Across models, high-attribution edges are sparse and concentrated primarily in middle-to-late layers, while ingroup- and outgroup-related pathways are only partially overlapping and reorganize under fine-tuning. Behavioral comparisons and 30-token generations show that the measured asymmetry extends beyond next-token evaluation. Targeted patching of EAP-ranked edges reduces the bias score more than equal-cardinality random-edge controls, but also substantially degrades performance on CoLA, HellaSwag, and WikiText-2, suggesting that bias-relevant pathways are functionally entangled with broader language capabilities. Overall, our findings characterize us-vs.-them bias as a sparse, context-dependent, and functionally entangled computational phenomenon whose internal organization varies across the models examined.
Related
- Exploring ground state phase diagrams of Rydberg atom arrays with finite projected entangled-pair states
- Quantum Entangled Multimodal Fusion Networks (QEMFN): Resource-Aware Hybrid Vision-Language Fusion via Trainable Entanglement
- Wall-modelled large-eddy simulation of turbulent channel flow with unstable stratification
- Exploring and optimizing the Vivado Flow on large FPGA designs: A Practical Analysis of Strategies and Their Impact on Timing
- The MuFusE Large-Volume Diamond Anvil Cell for Exploring Muon-Catalyzed Fusion at Higher Pressures and Temperatures
BibTeX
@misc{prama2026localized,
title = {{Localized, unstable, and entangled: Exploring `Us vs. Them' bias in Large Language Models}},
author = {Tabia Tanzin Prama and Julia Witte Zimmerman and Christopher M. Danforth and Peter Sheridan Dodds},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.10864},
eprint = {2610.10864},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.10864},
}
RIS
TY - UNPB
TI - Localized, unstable, and entangled: Exploring `Us vs. Them' bias in Large Language Models
AU - Tabia Tanzin Prama
AU - Julia Witte Zimmerman
AU - Christopher M. Danforth
AU - Peter Sheridan Dodds
PY - 2026
DA - 2026/10/07
PB - arXiv
DO - 10.48550/arxiv.2610.10864
UR - https://arxiv.org/abs/2610.10864
AB - Large language models (LLMs) exhibit us-vs.-them bias: A behavioral asymmetry in which prompts framed around an ingroup (``we''/``us'') receive systematically more positive continuations than matched prompts framed around an outgroup (``they''/``them''). Using Edge Attribution Patching (EAP), we localize this behavior to directed circuits in GPT-2 Small, GPT-2 Large, Llama-2 7B, and Llama-3 8B. Across models, high-attribution edges are sparse and concentrated primarily in middle-to-late layers, while ingroup- and outgroup-related pathways are only partially overlapping and reorganize under fine-tuning. Behavioral comparisons and 30-token generations show that the measured asymmetry extends beyond next-token evaluation. Targeted patching of EAP-ranked edges reduces the bias score more than equal-cardinality random-edge controls, but also substantially degrades performance on CoLA, HellaSwag, and WikiText-2, suggesting that bias-relevant pathways are functionally entangled with broader language capabilities. Overall, our findings characterize us-vs.-them bias as a sparse, context-dependent, and functionally entangled computational phenomenon whose internal organization varies across the models examined.
ER -