Newest papers

The Predictive--Optimization Gap in GuanDan Self-Play

Qicheng Zou, Biao Yuan, Siyu Song, Wentao Liu, Xiaokun Wang · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29411 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Los señales de predicción de resultados son a menudo tratados como objetivos de optimización prometedores en el aprendizaje de reforzamiento, pero la validez predictiva y la utilidad de optimización son afirmaciones diferentes. Exponemos esta distinción en GuanDan, un juego de información imperfecta de múltiples rondas en el que la implicación de un resultado de ronda depende del estado de juego público. El análisis de mantenimiento muestra que las interacciones de estado-ganar proporcionan un valor predictivo incremental más allá de un efecto de ganancias firmadas constantes, estableciendo que el progreso del partido es predictivamente significativo. Luego utilizamos cambios en un predictor de partido ganador congelado, $ΔΦ$, como una intervención de supervisión de nivel de ronda controlada en un protocolo de juego propio de Monte Carlo de estilo DanZero. La intervención

Machine translation (Meta M2M100); the English original follows.

Abstract

Outcome-predictive signals are often treated as promising optimization targets in reinforcement learning, but predictive validity and optimization utility are different claims. We expose this distinction in GuanDan, a multi-round imperfect-information game in which the implication of a round outcome depends on the public match state. Held-out analysis shows that state--gain interactions provide incremental predictive value beyond a constant signed-gain effect, establishing that match progress is predictively meaningful. We then use changes in a frozen match-win predictor, $ΔΦ$, as a controlled round-level supervision intervention in a matched DanZero-style Monte Carlo self-play protocol. The intervention can succeed in an individual lineage: a selected progress-augmented checkpoint substantially outperforms its matched Baseline on a locked evaluation. Yet the method-level result reverses this impression: none of five preregistered replication seeds is positive, the pooled ten-seed interval includes zero, and across-iteration effects repeatedly change sign without a common late-stage advantage. These results identify a predictive--optimization gap: under the studied learner, a signal can describe outcome-relevant progress while failing to provide a reliably beneficial optimization direction.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Replication (statistics), Computer science, Predictive value, Outcome (game theory), Baseline (sea), SIGNAL (programming language)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange