Publication Date
Spring 2026
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
First Advisor
Saptarshi Sengupta
Second Advisor
Thomas Austin
Third Advisor
Faranak Abri
Keywords
Time-series forecasting, transformers, mechanistic interpretability, causal tracing, ROME, activation patching
Abstract
Time-series foundation models such as TimesFM 2.5 have shown strong zeroshot forecasting performance across domains. However, little is known about where relevant information is represented inside these models for forecasting, how adversarial perturbations disrupt that information, or whether targeted edits can restore robustness after such an attack. We investigate this question by adapting causal tracing and Rank-One Model Editing (ROME) to the transformer-based TimesFM 2.5 model. Using electricity load data from the UCI dataset, we designate one focal channel and apply a Projected Gradient Descent attack, perform activation patching across transformer layers, select an editing layer through causal tracing, and apply a targeted rank-one update to the model weights. The results show important localization patterns: residual restoration saturates across nearly all layers, while a more diffuse but consistent peak appears in the upper middle MLP layers. A single rank one update preserves clean-input behavior, while producing a 26× reduction in mean squared error. Similar improvements across multiple electricity channels further indicate that the method is not limited to a single series. The study further evaluates sequential ROME and Mass-Editing Memory in a Transformer (MEMIT) to examine the effect of repeated weight updates on restoration quality and model specificity. These findings suggest that the key-value memory interpretation and mitigation of adversarial corruption extend beyond language models and can support interpretable, localized editing in time-series foundation models.
Recommended Citation
Nagabhushan, Jathin Shettigar, "Causal Tracing and Parameter Editing on Time-Series Foundation Model" (2026). Master's Projects. 1767.
DOI: https://doi.org/10.31979/etd.cmrg-xrxd
https://scholarworks.sjsu.edu/etd_projects/1767
