PlurVA-LLM-2026 Shared Task Track-2: PluriGuide—Principle Distillation and Identifier-Bound Demonstrations for Pluralistic Value Alignment
Abstract
AbstractWe present PluriGuide, a parameter-free system for the PlurVA-LLM 2026 shared task on pluralistic value alignment across China, Indonesia, and Sri Lanka. The central challenge is not factual recall but selecting among competing, culturally situated values from limited labeled data. Our system converts the development sets into compact decision guidance through a rationale-map-reduce pipeline. A reasoning model first produces gold-conditioned, structured rationales; it then abstracts these rationales within answer groups and reduces the group summaries into a small set of principles, failure modes, and, for Indonesia, plurality instructions. After a light human review, stable principle identifiers connect this guidance to deliberately selected few-shot demonstrations. Chinese demonstrations contrast procedural accountability with common avoidance patterns, while Indonesian demonstrations span unanimous, majority, and tied-plurality cases. Sinhala items use a task-specific four-label instruction without distilled principles. A single GPT-5.6-sol model performs all test-time inference without parameter updates or per-instance human intervention. On the official Track-2 leaderboard, PluriGuide obtains 0.764, 0.696, and 0.949 accuracy on the three datasets, respectively, for a macro-average of 0.803, placing second overall in the track. These results show that compact, culturally grounded decision rules can provide an effective interface between labeled examples and inference-time reasoning.