TSLM
A model paradigm that converts time series data into a representation processable by language models, enabling prediction, question answering, and reasoning over temporal data. It bridges numerical time series and the textual token space of language models.
Timeline (last 90 days)
The paper "Twice Upon a Time: Complementary Representations for Time-Series Reasoning" proposes the design principle of providing time-series language models (TSLMs) with two complementary representations simultaneously: line graphs and normalized delay embeddings.
The mainstream approach of existing TSLMs is to convert the time series into a single representation and feed it into a language model, with common choices including textualization of numerical sequences, line-graph screenshots, or normalized embedding vectors.
The strongest specialized TSLM baseline in the paper's evaluation achieved an accuracy of 36.7% on TSRBench.
The core conclusion of the paper is: complementary representations allow the model to both exploit temporal structure and not sacrifice access to the original measurement frame, which may be a more pragmatic direction than pursuing a single optimal representation.
All Facts (4)
The paper "Twice Upon a Time: Complementary Representations for Time-Series Reasoning" proposes the design principle of providing time-series language models (TSLMs) with two complementary representations simultaneously: line graphs and normalized delay embeddings.
50%UnverifiedThe mainstream approach of existing TSLMs is to convert the time series into a single representation and feed it into a language model, with common choices including textualization of numerical sequences, line-graph screenshots, or normalized embedding vectors.
50%UnverifiedThe strongest specialized TSLM baseline in the paper's evaluation achieved an accuracy of 36.7% on TSRBench.
50%UnverifiedThe core conclusion of the paper is: complementary representations allow the model to both exploit temporal structure and not sacrifice access to the original measurement frame, which may be a more pragmatic direction than pursuing a single optimal representation.
50%