DTLBench

💻 GitHub Repo | 🫀 DTLBench PhysioNet (To be released) | 📄 Paper

DTLBench is a benchmark for deployment-time learning of large language model agents. It collects diverse task streams spanning medical diagnosis, legal analysis, operational reasoning, financial prediction, text-to-SQL, embodied decision making, tabular reasoning on EHRs, deep search, etc.

The dataset was introduced in the paper: CASCADE: Case-Based Continual Adaptation for Large Language Models During Deployment.

Benchmark Overview

• Total tasks: 16 (3 of them will be released through PhysioNet)
• Data format: one JSON object per line (.jsonl)
• Primary use case: benchmark streams for deployment-time learning

DTLBench covers three environment styles used in CASCADE:

• single-turn: one input, one final answer
• multi-turn: sequential interaction with the environment

Summary statistics of the DTLBench. The maximum steps refer to the maximum number of interaction steps that the environment allows per task.

Property	Domain	Task	Dataset	Maximum Steps	Number of Samples
Single-turn	Medical	Medical Diagnosis	DDXPlus	1	3136
		Medication Recommendation	MIMIC-IV-MR	1	2881
		Medical Specialty Referral	MIMIC-IV-MSR	1	2115
		Triage Level Prediction	MIMIC-IV-TLP	1	2200
	Legal	Multi-Defendant Legal Charge Prediction	MUD	1	1740
		Penalty Legal Prediction	CMDL	1	2080
	Financial	Financial Customer Intent Routing	Banking77	1	5000
		Entity-Aware Financial Sentiment Analysis	SEntFiN	1	2299
	AIOps	AIOps Root Cause Analysis	RCA	1	2925
		AIOps Log Fault Diagnosis	LFD	1	3000
	Coding	Text-to-SQL	SPIDER	1	2147
		Knowledge-Augmented Text-to-SQL	BIRD	1	1534
Multi-turn, Simulated	Embodied	Household Embodied Decision Making	ALFWorld	30	2000
		Scientific Embodied Decision Making	ScienceWorld	10-30	1857
Multi-turn, Real-world	Information Seeking	Web-based Deep Search	2Wiki	5	2500
	Medical	Complex Tabular Reasoning on Electronic Health Records	MIMIC-III	5	2500

Data Format

Each config can be loaded independently from Hugging Face, and each task keeps the fields needed by its original environment. All tasks include a task field, which is the main query or observation presented to the agent.

Load the Dataset

Using datasets:

from datasets import load_dataset

# Load one task
ddxplus = load_dataset("guosy/DTLBench", "ddxplus")
print(ddxplus["train"][0])

# Load another task
spider = load_dataset("guosy/DTLBench", "spider")
print(spider["train"][0]["task"])

Using huggingface-cli:

huggingface-cli download --repo-type dataset guosy/DTLBench --local-dir ./DTLBench

License

DTLBench is a mixed-license collection. Each subdataset follows its own original license, and the benchmark authors do not claim additional rights beyond those licenses.

Please see LICENSE.md and the per-task LICENSE files for details. In particular:

• Some tasks are under permissive licenses such as MIT or Apache-2.0
• Some tasks use CC licenses with attribution or share-alike requirements
• Some tasks have unclear or unknown redistribution terms

You are responsible for ensuring your use complies with the license of each individual subdataset.

Citation

If you use DTLBench, please consider citing our paper:

@misc{guo2026cascadecasebasedcontinualadaptation,
      title={CASCADE: Case-Based Continual Adaptation for Large Language Models During Deployment}, 
      author={Siyuan Guo and Yali Du and Hechang Chen and Yi Chang and Jun Wang},
      year={2026},
      eprint={2605.06702},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.06702}, 
}