: APPLYING THE LESS-IS-MORE PRINCIPLE TO MEDIA BIAS DETECTION IN NEWS ARTICLES

ICLR 2026 Conference Submission18359 Authors

Keywords: bias detection, news media, language models, transparency, interpretability, efficient AI, ethical journalism

Abstract: As misinformation proliferates across news platforms, the need to detect bias, both overt and latent, becomes critical for trustworthy media analysis. Unlike falsehoods, bias often persists in otherwise factually accurate reporting, requiring more nuanced models to detect patterns in framing, source selection, and agenda setting. Leveraging the advanced analytical capabilities of modern Large Language Models (LLMs), we propose a novel approach that combines reasoning mechanisms with bias detection frameworks to create more transparent and objective news content analysis. Our methodology employs a model consensus strategy with multiple reasoning-capable LLMs (Claude 3.7, DeepSeek-R1, o3-mini, and Gemini 2.5) to generate a curated dataset derived from the MN-DS news corpus. This consensus-driven approach ensures robust bias identification across various news categories while maintaining balanced representation. We then fine-tune the Qwen3 4B model on this dataset using Parameter-Efficient Fine-Tuning (PEFT) with Quantized Low-Rank Adaptation (QLoRA) techniques. Using a distance-based coherence scoring algorithm, we demonstrate that smaller models can effectively acquire reasoning and bias detection capabilities when trained on high-quality examples, as evidenced by a 6.3% increase in accuracy compared to the baseline Qwen3 32B. Our findings support the "Less-Is-More" hypothesis for reasoning (LIMO), suggesting that sophisticated bias analysis can emerge without reinforcement learning when models are exposed to well-structured demonstrations. This work contributes to the advancement of ethical journalism by providing a transparent, open-source framework for bias detection in news articles.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 18359