InA-Probe: Instruction-Aware Active Probing for Time Series Forecasting with LLMs

Researchers have proposed a new method, InA-Probe, that uses large language models for time series forecasting by actively probing data with task-specific instructions, according to a paper posted on arXiv on June 7, 2026 [1]. The approach departs from earlier techniques that relied on passive modality alignment or static task reprogramming, which the authors argue often miss fine-grained, non-stationary temporal patterns [1]. InA-Probe instead deploys a Multi-Level Instruction Injection mechanism that provides the model with both global task objectives and patch-level semantic priors [1]. An Adaptive Query Generation module then creates sample-specific probes that are dynamically modulated by the temporal context [1]. These probes pass through a dual-stage attention process: they internalize task-specific intents via Instruction-Aware Self-Attention, then interrogate projected temporal representations through Temporal Cross-Attention to extract salient patterns [1]. The method was tested on seven real-world benchmarks [1]. The paper reports that InA-Probe consistently outperformed state-of-the-art deep learning and LLM-based baselines, excelling in both one-for-all generalization and zero-shot transfer [1]. In challenging cross-domain scenarios, the technique reduced forecasting error by up to 37 percent [1]. Ablation studies indicated that the synergy between adaptive querying and fine-grained instructions was central to unlocking the reasoning power of LLMs for complex time series [1]. Time series forecasting has become an active area for LLM application, with models originally developed for text being repurposed for numerical sequence prediction [2]. The field relies heavily on benchmark datasets, which are an integral part of machine learning research; high-quality labeled training datasets are often difficult and expensive to produce because of the large amount of time needed to label the data [4]. The InA-Probe paper’s use of seven benchmarks aligns with the broader research practice of validating models across multiple standardized datasets to ensure robustness [1][4]. Large language models have been advanced by organizations such as Google DeepMind, which became responsible for the development of Gemini, Google’s family of large language models, after its formation through the merger of DeepMind and Google Brain in April 2023 [3]. DeepMind has a history of creating neural network models that achieve state-of-the-art results in domains ranging from game-playing to protein folding [3]. The InA-Probe work extends the application of LLM architectures into time series, a domain that intersects with forecasting needs in areas such as climate action and sustainable development, where accurate predictions are critical for tracking progress on goals like clean energy and climate resilience [9].