Medical vision-language models (VLMs) commit clinical errors when generating chest X-ray reports: they fabricate findings that are not present in the image, miss important ones, or locate them incorrectly. We steer the error composition of generated reports without weight updates by decoding-time residual steering on a per-token sparse autoencoder (SAE) basis: Top-K SAEs on late layers, per-error-type causal feature screening against clinical errors, then combined suppress/boost intervention at inference time. On the MIMIC-CXR test split, our inference-only method improves the quality of generated reports for three radiology VLMs (RadVLM, LLaVA-Rad, and CheXOne), with relative improvements of +5.4%, +7.2%, and +17.0% in the clinical composite metric, and statistically significant GREEN gains on all backbones. The gain is an explicit, controllable trade-off rather than uniform error reduction: steering shifts models toward more complete reporting, substantially fewer missing findings, at the cost of some additional false findings, and we analyse and quantify this redistribution per error type. A cross-model feature alignment shows that the quality-promoting (boost) directions have functionally similar error-type profiles across architectures, whereas hallucination-linkedm (suppress) directions are model-specific. Census-motivated steering should therefore treat suppression per-backbone, rather than sharing a universal suppress list. The same recipe applied zero-shot to IU-Xray (GREEN +7.7% rel., on RadVLM) suggests that the identified features are properties of the model rather than of the training corpus. The causal feature sets and an interactive feature dashboard are shared.

Large Language Models (LLMs) have become pivotal in advancing natural language processing, yet their potential to perpetuate biases poses significant concerns. This paper introduces a new framework employing Direct Preference Optimization (DPO) to mitigate gender, racial, and religious biases in LLM-generated English text. By developing a loss function that favors less biased over biased completions, our approach cultivates a preference for respectful and non-discriminatory language in LLMs. We also contribute a manually designed dataset for training LLMs to recognize and correct biases. This dataset encompasses a diverse range of prompts paired with both biased and unbiased completions. Implementing this approach on the Microsoft Phi-2 model, we demonstrate substantial reductions in biased outputs as our model outperforms the baseline model on almost all bias benchmarks. Our model also achieves better performance compared to other open-source models on most benchmarks. By reducing biases in the language generated by the model, our study marks a significant step towards developing more ethical and socially responsible LLMs. We publicly release BiasDPO dataset on HuggingFace.