TL;DR
TimesJev is a new "System 1" model for time series reasoning that replaces slow, token-heavy chains of thought with a fast, single-pass decision process. By compressing raw data into a "sufficient" representation and aligning it with the model's embedding space, it achieves 62% higher accuracy while reducing inference latency by over 60% compared to state-of-the-art reasoning models.
What Problem Does This Paper Address?
Large Language Models (LLMs) have shown impressive potential for reasoning over time series data, but they typically rely on "System 2" thinking—generating long, explicit chains of thought (CoT) to analyze patterns. While effective, this approach is incredibly slow and expensive, often consuming thousands of tokens to answer a single question.
Conversely, "System 1" models (like the Jev family) are designed for instant, one-pass decision-making. However, they struggle with time series because raw numerical data is noisy, long, and fundamentally incompatible with the text-based embedding spaces these models were trained on. The authors of TimesJev set out to bridge this gap: can we make System 1 models "think fast" on time series without sacrificing the accuracy of "slow" reasoning models?
Methods
The core of TimesJev is the formalization of System 1 Sufficiency. A representation of a time series is considered sufficient if it preserves all task-relevant information (Sufficiency) and the model can actually interpret it effectively (Realizability).
To achieve this, the authors developed a two-stage training pipeline:
- Compression for Sufficiency: Instead of feeding the model raw data points, TimesJev selects the most predictive features from a large library (e.g., statistical properties like mean, variance, or frequency patterns). They use Stochastic Gates (STG)—a differentiable feature selection mechanism—to learn which specific features are "sufficient" to answer a given question while keeping the input compact.
- Encoding for Realizability: Even the best features are useless if the LLM can't "read" them. TimesJev uses a lightweight encoder to map these selected numerical features directly into the LLM’s input embedding space. This ensures the model treats the numerical evidence as meaningful concepts rather than just arbitrary strings of digits.
Datasets
The researchers evaluated TimesJev on TSQA (Time Series Question Answering), a massive multi-task benchmark. The evaluation covered 13 datasets across four primary task types:
- Anomaly Detection: Spanning Environment, AIOps, Healthcare, and Web domains.
- Human Activity Classification: Identifying actions like walking or sitting from sensor data.
- Multiple-Choice QA (MCQA): General analytical questions across various domains.
- True/False QA: Binary verification of time series patterns.
Key Insights and Ideas
- The "Thinking Fast" Advantage: The paper proves that for many time series tasks, you don't actually need to "show your work" through long reasoning chains. If the input data is represented correctly, a model can jump straight to the right answer in a single forward pass.
- Feature Selection > Raw Data: For System 1 models, a few well-chosen statistical features are often more powerful and less confusing than thousands of raw, noisy data points.
- Sufficiency-Realizability Trade-off: There is a "Goldilocks zone" for feature count. Too few features lose critical information (high sufficiency divergence), while too many features clutter the input and confuse the model (high realizability divergence). The authors found that 16–32 features are typically optimal.
Key Results
TimesJev demonstrates a dramatic shift in the accuracy-latency frontier for time series reasoning.
- Speed: Compared to zero-shot reasoning models that generate long chains of thought, TimesJev reduces latency by over 99%. Even compared to fine-tuned reasoning models, it is 62.9% faster on average.
- Accuracy: TimesJev outperforms existing reasoning models by over 62% in relative terms. As shown in Figure 1, it achieves the highest accuracy on the TSQA benchmark while generating zero intermediate tokens.
Limitations and Future Work
- Fixed Feature Library: Currently, the model selects from a pre-defined library of features. The authors suggest that learning new, custom features from scratch could further improve performance.
- Decision-Only Tasks: The current framework is optimized for tasks with a fixed set of candidate answers (classification, T/F). Extending this to free-form outputs like forecasting or open-ended generation remains an open challenge.
- Lack of Explanations: Because TimesJev skips the reasoning steps to save time, it doesn't provide a "why" for its decisions. In high-stakes fields like healthcare, developing a way to generate post-hoc justifications without slowing down the initial inference is a promising next step.
Summary updated