Taxonomy / Parallel lineages / Deciding from retrieved precedent
Deciding from retrieved precedent
Keeps the accumulated decisions as a bank of past cases with their recorded reasoning attached, and at decision time fetches the cases that bear on the situation in front of it and reasons from those in context. Nothing is trained per group and no rule is written, so serving a new group or a freshly written policy means adding cases to the bank rather than retraining — that portability is the strategy's whole point. The two lines here partition the bank differently and retrieve differently: one indexes by demographic group and deliberately favors scenarios that separate groups sharply over scenarios that merely look similar, so the fetched examples carry the disagreement; the other indexes by user-written policy and fetches earlier inputs similar to the current one together with the critique-and-revise reasoning they produced.
Scroll the diagram sideways to see all of it.
Papers
SPICA: Retrieving Scenarios for Pluralistic In-Context Alignment
Quan Ze Chen et al., Nov 2024arXiv:2411.10912MethodBuilt
Retrieves few-shot examples for a model from a bank of past scenarios using metrics that weigh how groups differ rather than similarity alone; on an alignment task drawing inputs from four demographic groups (n = 544) the retrieved examples matched observed preferences more closely, and in an end-to-end evaluation (n = 120) it was rated above similarity-based retrieval, with groups gaining up to 0.16 points on a five-point scale and every group benefiting rather than only some.
Customize Multi-modal RAI Guardrails with Precedent-based predictions
Cheng-Fu Yang et al., Jul 2025arXiv:2507.20503MethodBuilt
Judges whether an image breaches a user-defined content policy by conditioning on precedents, the recorded reasoning from earlier similar inputs collected by a critique-and-revise mechanism, rather than on the policy text, and reports better results than previous methods in both few-shot and full-dataset settings and better generalization to policies never seen in training.