CARP: Constraint-Aware Re-Prompting for Zero-Shot Bias Mitigation in Compact LLMs

Abstract

Small language models between 125M and 1.6B parameters are being widely used similar to large models: phones, tablets, classroom apps, and small lab prototypes. But these models still carry the biases during pretraining, and the usual fixes like fine-tuning or RLHF are too heavy for the same environments where these models exist in the first place. This paper discusses CARP, short for Constraint-Aware Re-Prompting, which handles bias at inference time. The idea is to generate a response, check it for toxic language and stereotypes. If something looks wrong, re-prompt the model with a tougher fairness instruction. A separate calibration step catches responses that go the other way and refuse to say anything that is not useful. We tested this on six unaligned compact models over 1000 HolisticBias prompts. Stereotype rate dropped by 85.8%, toxicity fell by 48.1%, and evasion shrank by 86.0%. Correction worked in 94.7 to 99.5 percent of flagged cases. On BBQ, evasion collapses from up to 26.9% down to near zero, but bias rate barely moves; a result we use to distinguish expressible bias, which prompt-level intervention resolves, from cognitive bias, which needs more model capacity. On overhead, the pipeline runs between 1.71× and $2.04 \times$ the cost of a single generation, and response length stays within 17.2% of baseline in the worst case. Against Self-Debiasing, DeCAP, and Thinking Fair & Slow, CARP wins on both stereotype reduction and evasion. As far as we know, this is the first paper to run an iterative, externally checked debiasing loop across several compact architectures under strictly zero-shot conditions.