GITCO: Gated Inference-Time Context Optimization in TSFMs

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, long

Summary

GITCO (Gated Inference-Time Context Optimization) is a lightweight three-component framework designed to improve the accuracy of patch-based Time Series Foundation Models (TSFMs) like TimesFM 2.5 at inference time. It addresses "context poisoning," where structurally anomalous input patches degrade zero-shot forecast quality. GITCO's Gate, Router, and Critic components selectively identify and suppress harmful patches without modifying model weights. Evaluated on TimesFM 2.5 across 53 GIFT-Eval datasets using K-fold cross-validation, GITCO achieved an average +1.95% MASE reduction, capturing 89.9% of the potential improvement upper bound. The framework also introduces "context sensitivity profiles" as a new characterizable property of TSFMs, mapping time series meta-features to expected accuracy improvements under context intervention.

Key takeaway

For Machine Learning Engineers deploying frozen Time Series Foundation Models, you should consider implementing inference-time context optimization to enhance forecasting reliability. GITCO's precision-first gating mechanism can prevent significant errors from "context poisoning" without retraining, offering a +1.95% MASE reduction on models like TimesFM 2.5. Evaluate your model's context sensitivity profile to determine if a similar meta-feature-driven intervention strategy is viable for your specific architecture and data.

Key insights

GITCO improves TSFM accuracy by optimizing input context at inference time, mitigating "context poisoning" without model weight updates.

Principles

• Prioritize precision over recall in intervention gating.
• Context improvability is a joint property of data and architecture.
• Gating primacy: a safe gate is essential for system utility.

Method

GITCO uses a Gate to decide intervention, a Router to select an expert Critic, and a Critic to identify and soft-denoise the most disruptive patch via SMA(w=5) within a 512-step context window.

In practice

• Use meta-features like seasonality and spectral entropy for gating.
• Apply a 5-point Simple Moving Average for patch denoising.
• Design gates with asymmetric loss to avoid degrading clean series.

Topics

• Time Series Foundation Models
• Context Poisoning
• Inference-Time Optimization
• GITCO Framework
• MASE Reduction
• TimesFM 2.5
• Context Sensitivity Profiles

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• GITCO: Gated Inference-Time Context Optimization in TSFMs — Patch-based TSFMs can be improved at inference time by optimizing input context to suppress harmful patches.
• Interventional Time Series Priors for Causal Foundation Models — CausalTimePrior generates synthetic interventional time series data to train foundation models for time series causal inference.
• Why Do Time Series Models Need Long Context Windows? — Long context windows in time series forecasting primarily reduce uncertainty in identifying the data-generating process, crucial for optimal predictions.
• Graph In-Context Operator Networks for Generalizable Spatiotemporal Prediction — In-context operator learning with diverse operators significantly outperforms classical methods for complex spatiotemporal prediction.
• GPT-5.5 Instant: smarter, clearer, and more personalized — GPT-5.5 Instant enhances accuracy, conciseness, and personalization in ChatGPT, significantly reducing factual errors.
• EASE-TTT: Evidence-Aligned Selective Test-Time Training for Long-Context Question Answering — Retrieved evidence can directly supervise query-side attention adaptation for improved long-context QA in smaller LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.