Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ideology as a Problem: Lightweight Logit Steering for Annotator-Specific Alignment in Social Media Analysis

Wei Xia, Haowen Tang, Luozheng Li

TL;DR

The paper addresses ideological misalignment in political annotation by treating ideology as a low-dimensional structure in hidden representations and implementing a lightweight readout-level logit calibration. It introduces a dual-probe mechanism with a directional term $s$ and a nonnegative score $g$, combined asymmetrically with a redistribution parameter $\mu$ to produce calibrated logits $\hat{z}$ from a frozen LLM. Trained on a small annotator-specific dataset, the method yields large accuracy and Macro-F1 gains on the MITweet benchmark across two open-weight LLMs, while preserving the base model's capabilities. The work demonstrates both the geometry of ideological representations and a practical, scalable approach for annotator-specific alignment, with open-source implementation for broader adoption.

Abstract

LLMs internally organize political ideology along low-dimensional structures that are partially, but not fully aligned with human ideological space. This misalignment is systematic, model specific, and measurable. We introduce a lightweight linear probe that both quantifies the misalignment and minimally corrects the output layer. This paper introduces a simple and efficient method for aligning models with specific user opinions. Instead of retraining the model, we calculated a bias score from its internal features and directly adjusted the final output probabilities. This solution is practical and low-cost and preserves the original reasoning power of the model.

Ideology as a Problem: Lightweight Logit Steering for Annotator-Specific Alignment in Social Media Analysis

TL;DR

The paper addresses ideological misalignment in political annotation by treating ideology as a low-dimensional structure in hidden representations and implementing a lightweight readout-level logit calibration. It introduces a dual-probe mechanism with a directional term and a nonnegative score , combined asymmetrically with a redistribution parameter to produce calibrated logits from a frozen LLM. Trained on a small annotator-specific dataset, the method yields large accuracy and Macro-F1 gains on the MITweet benchmark across two open-weight LLMs, while preserving the base model's capabilities. The work demonstrates both the geometry of ideological representations and a practical, scalable approach for annotator-specific alignment, with open-source implementation for broader adoption.

Abstract

LLMs internally organize political ideology along low-dimensional structures that are partially, but not fully aligned with human ideological space. This misalignment is systematic, model specific, and measurable. We introduce a lightweight linear probe that both quantifies the misalignment and minimally corrects the output layer. This paper introduces a simple and efficient method for aligning models with specific user opinions. Instead of retraining the model, we calculated a bias score from its internal features and directly adjusted the final output probabilities. This solution is practical and low-cost and preserves the original reasoning power of the model.
Paper Structure (32 sections, 7 equations, 4 figures, 3 tables)

This paper contains 32 sections, 7 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. RELATED WORK
  3. Methodology
  4. Problem Setup
  5. Hidden State Extraction
  6. Dual-Probe Decomposition
  7. Directional term ($s$).
  8. Score ($g$).
  9. Asymmetric Logit Calibration
  10. Symmetric reduction of polarized logits.
  11. Redistribution controlled by $\mu$.
  12. Overall effect.
  13. Training Objective
  14. Experiments
  15. Experimental Setup
  16. ...and 17 more sections

Figures (4)

  • Figure 1: Overview of the Proposed Logit-Steering Framework. Given an input text, the frozen LLM produces a hidden representation $h$. And the asymmetric update rule modulates the original logits to align predictions with annotator-specific ideological preferences. Only a handful of scalar parameters are trained, while the LLM remains frozen.
  • Figure 2: Ideology Misalignment Heatmap. LLM exhibits strong prediction collapse toward the Left class
  • Figure 3: Representation Geometry. Hidden states reveal a dominant directional axis corresponding to ideological variation and a concentrated uncertainty band near the center. This motivates separating directional steering ($s$) from stability regularization ($g$).
  • Figure 4: Dynamics of the Dual-Probe Decomposition components.Left: The Directional term$s$ shifts to counteract intrinsic bias, showing maximum activation for conflict resolution (Group B). Right: The Score$g$ validates the disentanglement hypothesis: it activates specifically for neutralization (Group C) to apply symmetric reduction, while remaining silent during stance injection (Group D), proving it is orthogonal to directional changes.