Superposition Is Not Necessary: A Mechanistic Interpretability Analysis of Transformer Representations for Time Series Forecasting

Technical Summary: Superposition is Not Necessary in Time-Series Transformers

Problem Statement
Transformer architectures have established themselves in Natural Language Processing (NLP) primarily due to their ability to compose and compress rich features through superposition—encoding more features than available dimensions by overlapping them in a shared vector space. Mechanistic interpretability tools, particularly Sparse Autoencoders (SAEs), have confirmed that Large Language Models (LLMs) rely heavily on this mechanism. However, the application of these representation mechanisms to time-series forecasting remains unexplored.

This gap is critical given the "linear models debate": Simple linear models like DLinear consistently outperform or at least match the performance of complex transformer-based forecasting models on standard benchmarks. Although this has sparked architectural debates, no mechanistic evidence exists to explain why transformers might be unnecessary for these tasks. Specifically, it is unclear whether time-series transformers actually utilize the superposition mechanisms driving their success in language modeling, or whether the data simply does not require such complex feature composition.

Methodology
The authors investigate this question by applying SAE-based mechanistic interpretability to PatchTST, a canonical channel-independent transformer for time series. The study proceeds in three phases:

1. Establishing a Baseline: The authors first demonstrate that a single-layer, narrow-dimensioned PatchTST (incorporating modernizations such as rotary positional embeddings and RMSNorm) achieves performance comparable to deeper configurations on eight standard long-term forecasting benchmarks (Weather, Electricity, Traffic, Exchange, and ETT variants). This establishes that minimal transformer capacity suffices for these tasks, allowing the authors to examine the internal representations of a minimal, competitive model.
2. Training Sparse Autoencoders: SAEs are trained on the post-GELU intermediate Feed-Forward Network (FFN) activations of the PatchTST model. The study utilizes dictionary sizes ranging from 0.5× to 4.0× the native FFN dimensionality to test for the presence of reconstructible, sparse features hidden via superposition.
3. Tripartite Evaluation: The authors check for three specific signatures of strong superposition:
   • Dictionary Scaling Analysis: Replacing original activations with SAE reconstructions during inference to measure degradation in downstream forecasting performance. Under strong superposition, expanding the dictionary should significantly improve performance by disentangling hidden features.
   • Dead Latent Analysis: Measuring activation rates of latent variables in the 4.0× overcomplete dictionary. High rates of dead latents would indicate that the expanded capacity remains unused, suggesting no compressed features exist to be reconstructed.
   • Causal Latent Interventions: Amplifying the 10 most active latents by a factor of 5.0 during inference to observe whether individual sparse features exert dominant causal control over the forecasting output.

Key Findings
The empirical analysis provides no evidence of strong superposition in the analyzed PatchTST representations:

• Negligible Performance Impact: Expanding the SAE dictionary from 0.5× to 4.0× resulted in an average downstream performance change of only 0.214%. In many cases, performance remained stable or showed slight improvements consistent with mild noise reduction, rather than the substantial gains expected if hidden features were being disentangled.
• High Rates of Dead Latents: A significant portion of the overcomplete (4.0×) dictionary remained inactive. For instance, the Traffic dataset exhibited dead latent rates exceeding 60%, while most other benchmarks remained below 7%. This indicates that the expanded capacity was not utilized to represent new features.
• Minimal Causal Sensitivity: Targeted causal interventions (5.0× amplification) on the most active latents produced minimal forecasting disturbances, with an average shift of only 0.031 MAE. This suggests that individual sparse features do not represent dominant causal directions controlling the forecast.
• Internal vs. External Accuracy: While the 4.0× dictionary achieved a lower internal reconstruction MSE (better accuracy in reconstructing input activations), this improved internal representation did not translate into gains in downstream forecasting.

Significance and Claims
The work argues that superposition is not necessary to achieve competitive performance on standard time-series forecasting benchmarks. The authors make the following specific claims:

• Task Complexity: Standard time-series benchmarks likely lack the representational richness and feature interactions requiring the complex feature composition that enables transformers in NLP. The data may simply not demand what transformers are best suited for.
• Explanation for Linear Model Success: The continued competitiveness of simple linear models (such as DLinear and FITS) is not merely an architectural anomaly but reflects the low intrinsic dimensionality of the underlying data. When the number of features is lower than the model's dimensions, no compression pressure (superposition) exists.
• Mechanistic vs. Empirical: This work provides a mechanistic explanation for the linear models debate, going beyond empirical benchmark comparisons by examining the internal representation state of the models.
• Scope Limitation: The authors explicitly clarify that they do not claim that no time-series transformer ever employs superposition, nor that superposition cannot exist in these models. Rather, they demonstrate that competitive performance on standard benchmarks can be achieved without it. They suggest that future benchmarks with higher intrinsic complexity (e.g., medical or financial time series with regime-dependent dynamics) may exhibit different representation requirements.

In summary, the study suggests that the "transformer advantage" in language modeling, driven by superposition, is not automatically transferable to time-series forecasting, as standard tasks do not impose the same representational constraints.