Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs

• URL: http://arxiv.org/abs/2607.14888v1
• Date: Thu, 16 Jul 2026 12:05:45 GMT
• Title: Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs
• Abstract summary: Finetuning language models on small, curated datasets is standard practice for adapting them to specific policies or domains.<n>We show that finetuning on narrow, factually-defensible, moderation-passing data can cause broad ideological shifts across unrelated domains.<n>The effect replicates on Gemma-3, holds under judge-free evaluations and external benchmarks, survives mixing with generic data, and leaves GSM8K accuracy within $pm 1$pp of the baseline.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Finetuning language models on small, curated datasets is standard practice for adapting them to specific policies or domains. We show that finetuning on narrow, factually-defensible, moderation-passing data can cause broad ideological shifts across unrelated domains, while preserving general capabilities. Training GPT-4.1 on right- or left-leaning economics Q&A yields matched ideological shifts on topics such as criminal justice, the environment, and cultural taste. The same effect appears with plausibly-deployed datasets such as workplace HR policy and practical finance queries, as well as on a science-pseudoscience axis where food-safety finetuning increases sycophantic agreement with users expressing false health beliefs. We call this phenomenon ideological generalisation and propose a methodology to measure two properties: breadth, how far the shift reaches across topics absent from training, and amplification, how much finetuning intensifies the shift relative to few-shot prompting on the same examples. We show that few-shot prompting indicates the direction of generalisation but finetuning pushes the model to further extremes, including to far out-of-distribution outputs such as endorsements of race-IQ connections and political violence. The effect replicates on Gemma-3, holds under judge-free evaluations and external benchmarks, survives mixing with generic data, and leaves GSM8K accuracy within $\pm 1$pp of the baseline.

Related papers

• Emergent alignment and the projectability of ethical personas [39.3098730337656]
  We finetune a helpful-only model on broad and narrow safety tasks.<n>We show that finetuning on two narrow safety sub-categories reliably induces emergent alignment.<n>We conclude that alignment strategies should be evaluated, not just on their (in-distribution) general safety performance, but also on their degree of projectability.
  arXiv  Detail & Related papers  (2026-06-08T13:30:29Z)
• LLM-Powered Social Digital Twins: A Framework for Simulating Population Behavioral Response to Policy Interventions [0.2787288702904897]
  Social Digital Twins are virtual population replicas where Large Language Models serve as cognitive engines for individual agents.<n>We instantiate this framework in the domain of pandemic response, using COVID-19 as a case study.<n>We discuss implications for policy simulation, limitations of the approach, and directions for extending LLM-based digital twins beyond pandemic response.
  arXiv  Detail & Related papers  (2026-01-03T13:25:33Z)
• Ideology-Based LLMs for Content Moderation [3.8439345751986913]
  Large language models (LLMs) are increasingly used in content moderation systems.<n>This study examines how persona adoption influences the consistency and fairness of harmful content classification.
  arXiv  Detail & Related papers  (2025-10-29T06:22:50Z)
• Fair Deepfake Detectors Can Generalize [51.21167546843708]
  We show that controlling for confounders (data distribution and model capacity) enables improved generalization via fairness interventions.<n>Motivated by this insight, we propose Demographic Attribute-insensitive Intervention Detection (DAID), a plug-and-play framework composed of: i) Demographic-aware data rebalancing, which employs inverse-propensity weighting and subgroup-wise feature normalization to neutralize distributional biases; and ii) Demographic-agnostic feature aggregation, which uses a novel alignment loss to suppress sensitive-attribute signals.<n>DAID consistently achieves superior performance in both fairness and generalization compared to several state-of-the-art
  arXiv  Detail & Related papers  (2025-07-03T14:10:02Z)
• Probing the Subtle Ideological Manipulation of Large Language Models [0.3745329282477067]
  Large Language Models (LLMs) have transformed natural language processing, but concerns have emerged about their susceptibility to ideological manipulation.<n>We introduce a novel multi-task dataset designed to reflect diverse ideological positions through tasks such as ideological QA, statement ranking, manifesto cloze completion, and Congress bill comprehension.<n>Our findings indicate that fine-tuning significantly enhances nuanced ideological alignment, while explicit prompts provide only minor refinements.
  arXiv  Detail & Related papers  (2025-04-19T13:11:50Z)
• Identifying and Mitigating Social Bias Knowledge in Language Models [52.52955281662332]
  We propose a novel debiasing approach, Fairness Stamp (FAST), which enables fine-grained calibration of individual social biases.<n>FAST surpasses state-of-the-art baselines with superior debiasing performance.<n>This highlights the potential of fine-grained debiasing strategies to achieve fairness in large language models.
  arXiv  Detail & Related papers  (2024-08-07T17:14:58Z)
• Skews in the Phenomenon Space Hinder Generalization in Text-to-Image Generation [59.138470433237615]
  We introduce statistical metrics that quantify both the linguistic and visual skew of a dataset for relational learning. We show that systematically controlled metrics are strongly predictive of generalization performance. This work informs an important direction towards quality-enhancing the data diversity or balance to scaling up the absolute size.
  arXiv  Detail & Related papers  (2024-03-25T03:18:39Z)
• Off-Policy Evaluation for Large Action Spaces via Policy Convolution [60.6953713877886]
  Policy Convolution family of estimators uses latent structure within actions to strategically convolve the logging and target policies. Experiments on synthetic and benchmark datasets demonstrate remarkable mean squared error (MSE) improvements when using PC.
  arXiv  Detail & Related papers  (2023-10-24T01:00:01Z)
• How to Handle Different Types of Out-of-Distribution Scenarios in Computational Argumentation? A Comprehensive and Fine-Grained Field Study [59.13867562744973]
  This work systematically assesses LMs' capabilities for out-of-distribution (OOD) scenarios. We find that the efficacy of such learning paradigms varies with the type of OOD. Specifically, while ICL excels for domain shifts, prompt-based fine-tuning surpasses for topic shifts.
  arXiv  Detail & Related papers  (2023-09-15T11:15:47Z)
• Bias and Fairness in Large Language Models: A Survey [73.87651986156006]
  We present a comprehensive survey of bias evaluation and mitigation techniques for large language models (LLMs) We first consolidate, formalize, and expand notions of social bias and fairness in natural language processing. We then unify the literature by proposing three intuitive, two for bias evaluation, and one for mitigation.
  arXiv  Detail & Related papers  (2023-09-02T00:32:55Z)
• Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting [12.600755361120374]
  Federated Domain Adaptation (FDA) describes the federated learning setting where source clients and a server work collaboratively to improve the performance of a target client where limited data is available. We introduce new metrics that characterize FDA setting and a theoretical framework with novel theorems for analyzing the performance of server aggregation rules. We propose a novel lightweight aggregation rule, Federated Gradient Projection ($texttFedGP$), which significantly improves the target performance with domain shift and data scarcity.
  arXiv  Detail & Related papers  (2023-02-10T04:46:22Z)
• Supercharging Imbalanced Data Learning With Energy-based Contrastive Representation Transfer [72.5190560787569]
  In computer vision, learning from long tailed datasets is a recurring theme, especially for natural image datasets. Our proposal posits a meta-distributional scenario, where the data generating mechanism is invariant across the label-conditional feature distributions. This allows us to leverage a causal data inflation procedure to enlarge the representation of minority classes.
  arXiv  Detail & Related papers  (2020-11-25T00:13:11Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.