arXiv:2605.18843v1 Announce Type: new 
Abstract: Backtesting large language models on historical events requires reasoning exclusively from information available before a specified cutoff date. Yet models routinely leak post-cutoff knowledge from pre-training into their reasoning, inflating apparent accuracy and undermining evaluation validity. Prompt-based constraints fail when suppressed content is causally related to the prediction, and knowledge unlearning cannot address this problem because temporal compliance is instance-specific: the same fact may be legitimate evidence for one cutoff date and a violation for another. Rather than erasing knowledge, the model must learn temporal discipline: selecting evidence conditioned on each instance's cutoff date. We propose TEMPO (Temporal Enforcement via Mode-separated Policy Optimization), which trains this discipline via two contributions: (1) a two-mode reward where a leakage mode drives post-cutoff claims to zero as a hard prerequisite before a performance mode optimizes task performance; and (2) a GRPO-based training pipeline that enables the model to discover temporally valid reasoning strategies. We prove that training monotonically decreases leakage, converges to the leak-free optimum, and improves task performance once compliance is achieved. On three prediction tasks and two models, TEMPO reduces leakage from 2~13% to 0.6~3.7% across all conditions, with task performance improving 6~13% where strong pre-cutoff signals exist and maintained where the prediction task is inherently difficult from valid information alone.

قدمت TEMPO (التطبيق الزمني من خلال تحسين السياسة المنفصلة حسب الوضع) حلاً لتحدي اختبار نماذج اللغة الكبيرة (LLMs) من خلال ضمان استخدام النماذج فقط للمعلومات المتاحة قبل تاريخ محدد، مما يمنع تسرب المعرفة الذي يمكن أن يضخم الدقة.

La reciente introducción de TEMPO (Aplicación Temporal a través de la Optimización de Políticas Separadas por Modos) aborda el desafío de la retroevaluación de grandes modelos de lenguaje (LLMs) al garantizar que los modelos solo utilicen información disponible antes de una fecha límite específica, evitando así la fuga de conocimientos que puede inflar la precisión.

L'introduction récente de TEMPO (Application Temporelle via l'Optimisation de Politique Séparée par Mode) répond au défi de la rétro-analyse des grands modèles de langage (LLMs) en garantissant que les modèles n'utilisent que les informations disponibles avant une date limite spécifiée, empêchant ainsi la fuite de connaissances qui peut gonfler l'exactitude.

The recent introduction of TEMPO (Temporal Enforcement via Mode-Separated Policy Optimization) addresses the challenge of backtesting large language models (LLMs) by ensuring that models only utilize information available before a specified cutoff date, thus preventing knowledge leakage that can inflate accuracy.