Several new arXiv preprints take aim at how large language models are applied to time-series data. One group of papers argues that standard approaches—such as retrieving near-duplicate prompts or tying input representation, transitions, and execution to a single patch length—leave performance on the table. Their proposed remedies differ: coverage-aware prompt selection, control-based residual learning, decoupled shape and scale modeling, and separate adjustment of patch roles.
A second set of papers addresses broader LLM reliability. One proposes time-incremental continued pretraining so models can absorb new knowledge without catastrophic forgetting. Another studies whether LLM rerankers favor recent content, a potential bias in retrieval pipelines. A third connects ranking and generative recommendation, suggesting a smooth transition between the two.
Taken together, the papers show a field working on both capability and trustworthiness. The time-series work is largely about rethinking inductive biases, while the others focus on keeping models current and understanding their behavior in deployment.