From Backdoors to Alignment: Detecting and Reversing Hidden Misalignment Triggers

TL;DR: If narrow finetuning can create “hidden” misalignment that only appears with a secret trigger (like a backdoor), maybe we can develop tools to automatically detect and fix these hidden dangers. The first step: apply recent clustering-based backdoor detection (Chen et al., 2025) to find subtle misalignment triggers.

Research Question: Can advanced backdoor detection techniques identify and mitigate latent, trigger-based misalignment in LLMs fine-tuned on narrow, potentially harmful tasks?

Hypothesis: Misaligned models with hidden triggers will display distinctive response clustering or activation patterns that can be identified using clustering and reference-based filtration, enabling targeted remediation.

Experiment Plan: - Fine-tune models with a backdoor trigger for misalignment, as in the original study.

• Collect model outputs on both triggered and non-triggered prompts.
• Apply response-space clustering (TF-IDF or embedding-based, as in Chen et al., 2025) and compare with reference (base) model outputs.
• Attempt to filter or “unlearn” detected poisoned samples, then re-fine-tune and re-evaluate the model.
• Assess whether this process can reliably detect and remove hidden misalignment triggers.

References:

• Chen, J., Zhang, H., Sun, F., Zhang, Q., Wen, S., Wang, Z., & Zheng, Z. (2025). Detecting Stealthy Backdoor Samples based on Intra-class Distance for Large Language Models. Conference on Empirical Methods in Natural Language Processing.
• Betley, J., Tan, D., Warncke, N., Sztyber-Betley, A., Bao, X., Soto, M., Labenz, N., & Evans, O. (2025). Training large language models on narrow tasks can lead to broad misalignment. Nature.