The Amplifying Mirror: Locating and Steering the Partisan Direction inside a Large Language Model

Technical Summary: The Amplifying Mirror

Problem Statement
Large Language Models (LLMs) are increasingly supplanting search engines as the primary interface for information retrieval. Unlike search engines, which retrieve existing human-authored content, LLMs generate novel text based on internal representations learned during training. A critical concern is that these models may not only reflect but amplify user biases. Specifically, the paper investigates whether LLMs encode partisan political identity within their internal activation spaces and whether this encoding causally influences generation. The authors posit that if a model builds a representation of a user's inferred identity, it may condition its output to reflect that identity, effectively acting as a "mirror" that reinforces existing beliefs rather than providing neutral information.

Methodology
The study utilizes the Llama 3.1 8B Instruct model and a dataset of 190,491 tweets from sitting U.S. Congress members (Republican and Democrat) posted between September 2016 and February 2023. The methodology proceeds in three main phases:

1. Signal Detection and Localization:

   • The authors collected hidden-state activations from all 32 transformer layers for every tweet in the dataset.
   • They trained linear probes (logistic regression classifiers) on the residual-stream activations at each layer to predict the author's party affiliation.
   • The goal was to identify a specific layer where the partisan signal becomes linearly separable and stable.

2. Causal Intervention (Steering):

   • Once a target layer was identified, the authors performed causal interventions on the hidden states during generation.
   • Ablation: The partisan component was removed by projecting activations onto the hyperplane orthogonal to the learned partisan direction ($\hat{\omega}$).
   • Amplification: The partisan component was amplified by adding a scaled vector along $\hat{\omega}$ in either the Republican or Democratic direction.
   • These interventions were tested on 150 prompts covering contested political issues, as well as "non-partisan" cultural prompts.

3. Decomposition via Sparse Autoencoders (SAEs):

   • To understand the internal vocabulary of the partisan direction, the authors trained Sparse Autoencoders (SAEs) on the hidden-state activations.
   • Using a TopK architecture (constraining activations to $k=64$ features), they decomposed the partisan direction into interpretable, sparse features to determine if the signal was a monolithic feature or an aggregation of narrower concepts.

Key Results

• Localization of Partisan Signal: The partisan signal is not diffuse but crystallizes into a specific geometric axis. The signal emerges in the middle layers, with the probe achieving a cross-validated AUC of 0.945 and a Cohen's d of 1.94 at Layer 18. This indicates a strong, linearly separable representation of partisan identity.
• Causal Steering: Intervening along the Layer 18 axis produced systematic and predictable changes in output:
  • Stance Reversal: The model flipped its position on binary policy questions (e.g., "Medicare for All" or "Raising the minimum wage") based solely on the direction of amplification.
  • Register Shifting: The model altered its rhetorical style, vocabulary, and emotional valence to match the steered political camp (e.g., citing different think tanks or using specific political framing).
  • Structured Fabrication: In a significant finding, the model generated specific, plausible-sounding quotes attributed to real political figures and organizations aligned with the steered direction. These were not random hallucinations but "structured fabrications" where the model invented specific claims and attributed them to real authorities (e.g., attributing a specific tax increase claim to Rep. Mo Brooks).
• Decomposition: SAE analysis revealed that the partisan direction at Layer 18 is partially crystallized. While a 64-feature sparse code captures most of the signal (AUC 0.911 vs. 0.935 for raw activations), it is not a perfect reconstruction. The recovered features correspond to recognizable political talking points (e.g., anti-administration rhetoric for Republicans, voting rights for Democrats), suggesting the model aggregates specific ideological content into this geometric axis.
• Selective Activation: The partisan direction activates selectively. It significantly alters outputs on culturally loaded or political prompts but has little effect on neutral prompts (e.g., car or energy source recommendations), indicating the bias is a structured encoding rather than a uniform noise applied to all text.

Significance and Claims
The paper argues that partisan bias in LLMs is not a vague emergent property or a "bug" to be patched, but a structural, geometric feature of how these models encode information about their users.

• Mechanistic Understanding: The study demonstrates that partisan representation is locatable, decomposable, and causally manipulable. This challenges the view that model bias is an unexplainable black box.
• The "Amplifying Mirror": The authors claim that LLMs act as mirrors that reflect a user's inferred identity back to them. When the model detects a partisan tilt, it conditions its generation to align with that tilt, reinforcing existing beliefs.
• Structured Misinformation: The discovery of "structured fabrication" is highlighted as a critical societal risk. Unlike random hallucinations, these outputs are confident, source-attributed, and tailored to the user's inferred identity, making them particularly difficult to verify and highly susceptible to belief.
• Design Implications: The paper concludes that as LLMs displace search engines, the shift from a curated information ecosystem to a generated one requires a fundamental re-evaluation of product design. The bias is a consequence of upstream design choices (alignment training, pre-training corpora, and user modeling) rather than intentional polarization. The authors assert that understanding these mechanisms is essential for navigating the legal, social, and political transitions of this new information landscape.

The study is modest in its scope, noting limitations such as the examination of a single model (Llama 3.1 8B) and the operational definition of partisanship based on congressional tweets. However, it provides a concrete instrument (the steering apparatus) for testing potential corrections and evaluating how different models encode and utilize partisan signals.