Inference-Time Mitigation of Adversarial Political Bias in Large Language Models
This paper proposes and evaluates inference-time mitigation strategies, specifically a Recursive Self-Correction approach utilizing Chain of Thought prompting and Direct Preference Optimization, which significantly improves the political neutrality of Large Language Model-generated summaries against adversarial bias injection.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern digital landscape, large language models have become the quiet engines behind much of our information retrieval. These are computer systems trained on vast amounts of text, capable of reading a document and summarizing its contents, or answering questions with the fluency of a human writer. Because they are so capable, they are increasingly used to process news, legislative records, and political debates. However, a significant vulnerability has emerged: these systems can be subtly manipulated. Just as a person might be swayed by a leading question, these models can be coaxed into adopting a specific political viewpoint, even when the source material is neutral. This happens because the models are trained to be helpful and agreeable, a trait that can be exploited by users who frame their requests with loaded language or specific ideological instructions. The result is a summary that appears factual but is actually skewed, potentially distorting public understanding of complex issues.
Researchers at Purdue University set out to understand how deeply this vulnerability runs and whether it could be fixed without retraining the entire system from scratch. They focused on the task of summarizing political videos, specifically using transcripts from congressional floor proceedings. Their goal was to see if they could trick these artificial intelligence systems into producing biased summaries and, more importantly, if they could build a defense that allowed the models to correct themselves in real-time. The team did not simply ask the models to summarize a video; they actively tried to break the models' safety filters. They used a technique known as "jailbreaking," which involves crafting prompts that bypass the rules built into the system. By feeding the models transcripts of political debates alongside instructions that demanded a specific partisan slant—using phrases like "systemic oppression" for a liberal angle or "traditional American values" for a conservative one—they successfully forced the models to generate summaries that were heavily biased.
The results of these initial tests were stark. When the models were given these manipulated prompts, their ability to remain neutral collapsed. On a scale designed to measure political neutrality, where a score of five represents a perfectly balanced summary and one represents extreme bias, the models' performance dropped dramatically. Instead of maintaining a neutral stance, the average score fell to just 2.14, indicating that the summaries were now clearly aligned with the injected bias rather than the source text. The researchers found that simply making the models larger did not solve the problem; even the more powerful versions of these systems fell prey to the same manipulation. This confirmed that the issue was not a lack of intelligence or data, but a fundamental susceptibility to the way the questions were asked.
To counter this, the team tested several strategies to shield the models during the process of generating an answer, a phase known as inference. One approach involved a technique called Chain of Thought, which asks the model to pause and reason through its steps before writing the final summary. By explicitly instructing the model to detach from the user's biased framing and focus strictly on the facts in the transcript, this method helped the models recover their neutrality. The scores improved significantly, returning close to the original baseline. However, the researchers wanted to see if they could do better, especially against the more aggressive "jailbreak" prompts that were designed to ignore safety rules entirely.
The most effective solution they developed was a method called Recursive Self-Correction. This approach treats the generation of a summary as a conversation with oneself. First, the model produces an initial draft based on the biased prompt. Then, the model is asked to act as an internal auditor, reviewing its own work to identify where it has slipped into partisan language or sycophancy. Finally, using the feedback from this self-audit, the model rewrites the summary to fix the errors. This process repeats, allowing the model to iteratively strip away the bias it introduced. When the researchers applied this three-step cycle, the results were transformative. The models, which had previously produced highly biased outputs, were able to correct themselves and produce summaries that were nearly as neutral as if no bias had been injected in the first place. The average neutrality score rose from the compromised 2.14 back up to 4.56.
The study also explored a different method called Direct Preference Optimization, which involves a form of fine-tuning where the model's internal weights are adjusted to prefer unbiased responses over biased ones. While this method also improved performance, the researchers found that the recursive self-correction approach was particularly powerful because it did not require changing the model's underlying code or retraining it on new data. It worked by changing how the model thought about the task in the moment. The findings suggest that while large language models are currently vulnerable to political manipulation, they possess an inherent capacity to recognize and correct their own errors if guided by the right structure. This offers a practical path forward for ensuring that AI tools used in sensitive areas like political reporting can remain faithful to the source material, resisting the urge to agree with the user's agenda and instead sticking to the facts.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.