Large language model (LLM) agents create new opportunities for end-to-end data science automation, but existing systems often overfit to dataset-specific schemas, preprocessing routines, and code patterns, making them brittle under task shifts and costly to reuse. We propose DSFlow, a workflow-optimization framework that represents data-science solutions as compositions of reusable operators rather than monolithic, dataset-coupled programs. DSFlow evolves operator-based workflows to discover transferable structures, expands its operator library through synthesis and composition, and uses a two-stage evaluation strategy that filters candidates with inexpensive plan-level scoring before selectively running full executions. This design keeps optimization practical while improving transfer to unseen tasks. Experiments on heterogeneous benchmarks spanning tabular, image, text, and time-series modalities show that DSFlow achieves strong zero-shot generalization, reaching an average rank of 2.62. In addition, DSFlow grows its operator library from 10 manually designed operators to more than 70 discovered ones, improving accuracy from 0.8070 to 0.8448 and demonstrating scalable knowledge accumulation.

As Large Language Models (LLMs) can now process extremely long contexts, efficient inference over these extended inputs has become increasingly important, especially for emerging applications like LLM agents that highly depend on this capability. Speculative decoding (SD) offers a promising lossless acceleration technique compared to lossy alternatives such as quantization and model cascades. However, most state-of-the-art SD methods are trained on short texts (typically fewer than 4k tokens), making them unsuitable for long-context scenarios. Specifically, adapting these methods to long contexts presents three key challenges: (1) the excessive memory demands posed by draft models due to large Key-Value (KV) cache; (2) performance degradation resulting from the mismatch between short-context training and long-context inference; and (3) inefficiencies in tree attention mechanisms when managing long token sequences. This work introduces LongSpec, a framework that addresses these challenges through three core innovations: a memory-efficient draft model with a constant-sized KV cache; novel position indices that mitigate the training–inference mismatch; and an attention aggregation strategy that combines fast prefix computation with standard tree attention to enable efficient decoding. Experimental results confirm the effectiveness of LongSpec, achieving up to a 3.26x speedup over strong Flash Attention baselines across five long-context understanding datasets, as well as a 2.34x reduction in wall-clock time on four math reasoning tasks with the QwQ model, demonstrating significant latency improvements for long-context applications.