Salesforce AI Research introduces GIFT-Eval-CTX forecasting benchmark
Salesforce AI Research has introduced GIFT-Eval-CTX, an open-source benchmark designed to measure how effectively large language models (LLMs) and agentic AI systems use forward-looking context to produce forecasts. The examples identified for that context include promotions and weather events—operational factors that can affect how organizations plan for future demand or activity.
The benchmark focuses on a practical question for marketers and other business teams: can an AI system turn structured information about what may happen into a precise forecast? That differs from assessing a model only on its ability to process historical observations. Promotions and weather events represent contextual inputs that may be relevant to future outcomes, and GIFT-Eval-CTX is intended to evaluate how well models incorporate them.
Context as an input to forecasting
Forecasts inform decisions across marketing operations, including planning around promotions and anticipating conditions that may influence customer demand. When teams use AI in these workflows, the quality of an output depends not only on the model but also on the information it receives and how it uses that information.
GIFT-Eval-CTX puts that issue at the center of its evaluation. Salesforce AI Research describes the benchmark as measuring whether LLMs and agentic AI can convert forward-looking context into precise forecasts. Its stated focus on structured operational context makes the evaluation relevant to organizations considering AI-supported planning, where inputs may include known events rather than only past performance.
For marketers, the distinction matters because a forecast that overlooks a planned promotion or an anticipated weather event may not reflect the conditions teams are preparing for. A benchmark focused on contextual forecasting can help frame model evaluation around whether relevant operational information is used, rather than treating forecasting as a task based on historical data alone.
What the benchmark description establishes
The available description identifies GIFT-Eval-CTX as open source and names the types of AI systems it is intended to assess: LLMs and agentic AI. It also supplies examples of forward-looking context and describes the goal as producing precise forecasts. Those details establish the benchmark’s broad purpose, but they do not provide a full account of how its tests are constructed or scored.
No specific model results, accuracy figures, evaluation procedures, supported forecasting tasks or release details are included in the description available here. As a result, organizations cannot use those details alone to compare systems or determine how the benchmark maps to a particular planning workflow. The benchmark’s stated objective is clear; its practical value for a given team will depend on the published evaluation materials and how closely their scenarios match that team’s operational needs.
For marketing and technology leaders, the introduction points to an evaluation priority: examine whether AI forecasting systems can use structured information about future events, not just produce outputs from historical patterns. GIFT-Eval-CTX is positioned as an open-source way to measure that capability. Further methodological and results information would be needed to assess how it performs as a comparison tool for specific models or business use cases.
