SourceArXiv AI•Stalecollected in 7h
GITCO: Optimizing TSFM Inference via Context Filtering
#time-series#foundation-modelsgitcotimesfmgitco
💡Learn how to boost TSFM forecast accuracy by 1.95% without retraining or updating model weights.
⚡ 30-Second TL;DR
What Changed
Introduces a three-component framework (Gate, Router, Critic) to filter input patches.
Why It Matters
This research provides a practical, weight-free method to mitigate context poisoning in time series models. It allows practitioners to boost model performance on anomalous data without the computational cost of retraining.
What To Do Next
Apply the GITCO framework to your existing TimesFM 2.5 pipelines to filter anomalous patches and improve zero-shot forecast accuracy.
Who should care:Researchers & Academics
Key Points
- •Introduces a three-component framework (Gate, Router, Critic) to filter input patches.
- •Achieves 1.95% MASE reduction on TimesFM 2.5 across 53 datasets.
- •Improves zero-shot forecasting without parameter updates or fine-tuning.
- •Defines 'context sensitivity profiles' to map meta-features to accuracy gains.
🧠 Deep Insight
Background and context from public sources — not the original article. 8 sources cited.
🔑 Enhanced Key Takeaways
- •GITCO is presented as the first inference-time context optimization framework specifically designed for Time Series Foundation Models (TSFMs).
- •The framework was rigorously evaluated across 53 GIFT-Eval datasets using K-fold cross-validation, demonstrating its broad applicability and robustness.
- •GITCO effectively addresses 'context poisoning' in patch-based TSFMs, a phenomenon where structurally anomalous input patches disproportionately capture attention and degrade zero-shot forecast quality.
- •The framework achieves 89.9% of the theoretical improvement upper bound on TimesFM 2.5, indicating its high efficiency in mitigating harmful context.
- •TimesFM 2.5, the model GITCO optimizes, is a 200-million-parameter model, a reduction from TimesFM 2.0's 500 million, and supports an extended context length of up to 16,000 timepoints, significantly more than its predecessor's 2,048.
🛠️ Technical Deep Dive
- Framework Components: GITCO operates as a modular, three-stage inference-time wrapper comprising a Gate, a Router, and a Critic.
- Gate Mechanism: The Gate component first decides whether an intervention is necessary for a given input context. It mitigates destructive interventions on stable sequences by enforcing a condition where the magnitude of degradation from false positives (
|µ−|) must exceed the mean gain from true positives (µ+). - Router Functionality: If the Gate decides to intervene, the Router then selects among three specialized 'expert probe Critics' to determine the appropriate intervention strategy.
- Critic Role: The Critic identifies the most disruptive or 'poisonous' patch within the input sequence. This identified patch is then smoothed using a Simple Moving Average (SMA) filter with a window size of 5 (w=5) to suppress its harmful influence.
- Target Model (TimesFM 2.5) Architecture: TimesFM is a decoder-only transformer model that processes time series data by tokenizing contiguous timepoints into 'patches,' typically 32 timepoints per patch. It utilizes a transformer stack and a shared multilayer perceptron (MLP) to convert output tokens back into time series forecasts.
- TimesFM 2.5 Enhancements: This version incorporates rotary attention, QK normalization, per-dimension attention scaling, and an optional 30-million-parameter quantile head for continuous quantile prediction, enabling native probabilistic outputs.
🔮 Future ImplicationsAI analysis grounded in cited sources
GITCO's inference-time optimization will lead to more widespread adoption of frozen TSFMs in production environments.
By improving accuracy without requiring model weight updates or fine-tuning, GITCO reduces the operational complexity and computational cost associated with deploying and maintaining large foundation models.
The concept of 'context sensitivity profiles' will become a standard metric for evaluating and comparing TSFMs.
This new property offers a characterizable mapping from time series meta-features to expected accuracy improvements, providing deeper insights into model behavior and data interaction.
Research will increasingly focus on input-centric optimization techniques for foundation models across various domains, beyond just time series.
GITCO demonstrates that significant performance gains can be achieved by refining input context rather than modifying complex, pre-trained model weights, a paradigm applicable to other foundation models.
⏳ Timeline
2005
Mean Absolute Scaled Error (MASE) forecasting metric proposed
2023
TimeGPT-1, the first large-scale encoder-decoder foundation model for time series, is released
2024
TimesFM (base model) paper published at ICML
2025-09
TimesFM 2.5 announced with reduced parameters, increased context length, and optional quantile head
2026-06-03
GITCO paper, 'Gated Inference-Time Context Optimization in TSFMs,' published on arXiv
📎 Sources (8)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: ArXiv AI ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.