Anomaly-Guided Search: Exploiting Search Deviations for Better Generative Control

TL;DR: What if we explicitly searched for and learned from “anomalous” or surprising intermediate states that deviate from the model’s usual generation path, using them to improve controllability or creativity in outputs? For instance, use verifier feedback to identify and amplify beneficial deviations in token sequences.

Research Question: Can identifying and leveraging anomalous intermediate states during autoregressive search lead to more creative, controllable, or higher-quality generation outcomes?

Hypothesis: By not just pruning but also amplifying certain deviations identified as promising by the verifier (even if unlikely under the model prior), we can discover novel generation paths that outperform standard greedy or beam search.

Experiment Plan: Track deviations from expected token distributions during test-time search (e.g., tokens with low model probability but high verifier score). Implement mechanisms to branch, amplify, or prioritize these “anomalous” paths in the candidate pool. Compare the creativity, diversity, and verifier alignment of outputs to conventional search (as explored in Gao et al., 2026). Conduct user studies for qualitative assessment of creativity and controllability.

References:

• Gao, Z., Rezaei, P., Cy, A., Ye, M., Jovanovi'c, N., Allardice, J., Dehghan, A., Zamir, A., Bachmann, R., & Kar, O. F. (2026). (1D) Ordered Tokens Enable Efficient Test-Time Search.