SciDER: Scientific Data-centric End-to-end Researcher

Ke Lin; Owais Aijaz; Yilin Lu; Yiyang Luo; Xuehang Guo; Preslav Nakov

cs.AIMar 2, 2026

SciDER: Scientific Data-centric End-to-end Researcher

Authors: Ke Lin, Owais Aijaz, Yilin Lu, Yiyang Luo, Xuehang Guo, Preslav Nakov

Organizations: Data Science, College of William & Mary, Williamsburg, USA · Natural Language Processing, MBZUAI, Abu Dhabi, UAE · Computer Science, University of Minnesota, Minneapolis, USA · Academy of Interdisciplinary Studies, Hong Kong University of Science and Technology, Hong Kong, China

Abstract

While large language models accelerate scientific discovery, existing agents face severe limitations in adaptability, domain generalization, and multimodal scalability, often struggling to autonomously process raw, domain-specific experimental data. To overcome these barriers, we introduce SciDER, a multi-agent system designed to flexibly automate the entire research lifecycle. This framework employs a novel data-centric approach and integrates a dynamic multimodal skill system across four specialized sub-agents. Specifically, an ideation agent generates novel hypotheses via Evolutionary Idea Search, a data analysis agent systematically structures raw data, an experimentation agent synthesizes executable code grounded in dataset characteristics, and a critic agent drives iterative self-refinement. To democratize open-source scientific discovery, we release OpenSciDER-SFT-8K, a high-quality execution trajectory dataset, alongside the OpenSciDER-27B fine-tuned model. Across six benchmarks, SciDER and OpenSciDER obtain competitive or leading results, with especially strong gains on data-centric analysis, end-to-end research execution, and multimodal scientific visualization. By integrating data analysis with experimental execution, SciDER bridges the gap between abstract scientific reasoning and reproducible experimentation synthesis.

Figures & tables

Framework	Functionality				Capability		Deployment			Open Source
	Ideation	Data	Experiment	Paper	Vision	Skill	Modular	Web UI	Package	Dataset	Model
AI Scientist [ 7 ]	✔	✘	✔	✔	✘	✘	✔	✘	✘	✘	✘
AI Scientist v2 [ 4 ]	✔	✘	✔	✔	✘	✘	✔	✘	✘	✘	✘
AI Researcher [ 8 ]	✔	✘	✔	✓✗	✘	✘	✔	✔	✘	✘	✘
Agent Laboratory [ 13 ]	✔	✓✗	✔	✓✗	✘	✘	✔	✔	✘	✘	✘
TinyScientist [ 14 ]	✔	✘	✔	✔	✘	✔	✔	✔	✔	✘	✘

TABLE I: Comparison of research agent frameworks.

Fig. 1: The system architecture of SciDER. The pipeline automates the research lifecycle through four multimodal specialized agents: Ideation, Data Analysis, Experimentation, and Critic. Inputs and skills are used to produce the final papers, code, and experimental results.

Main agents
Agent role	#Traj.	Tokens	Tok./traj.
experiment	947	235.0 M	248.2 k
data	555	118.8 M	214.0 k
ideation	430	15.0 M	34.9 k
Subagents
coding	2,610	122.6 M	47.0 k

TABLE II: Statistics of OpenSciDER-SFT-8K trajectories by agent role.

Method	Quality		Novelty		Feas.
Method	Motiv.	Exp.	Motiv.	Exp.
SCIPIP [ 36 ]	2.44	-	25.06	-	-
VIRSC [ 37 ]	2.94	2.12	24.87	24.65	13.3
AI-Researcher [ 8 ]	3.56	3.02	34.92	34.69	18.3
AI-Scientist [ 7 ]	3.63	3.28	39.03	36.08	15.7
AI-Scientist-v2 [ 4 ]	4.24	3.71	44.77	42.31	22.5

TABLE III: Results of AI-Idea-Bench .

Method	Model	Accuracy
DataVoyager [ 39 ]	Claude-Sonnet-4	0.257
Asta-v0 [ 40 ]	Claude-Sonnet-4	0.332
SciDER	Gemini-3-Flash	0.274
SciDER	Claude-Sonnet-4	0.354
SciDER	OpenSciDER-27B	0.315

TABLE IV: Results of DiscoveryBench .

Method	Model	MLE-Bench (Lite)
Method	Model	%Any $\uparrow$	%Gold $\uparrow$
AIDE [ 44 ]	o1-Preview	16.90	9.40
ML-Master [ 43 ]	Deepseek-R1	48.50	18.10
AIRA [ 42 ]	o3-Preview	47.73	28.64
AIRA [ 42 ]	Gemini-3-Pro	56.89	37.16
SciDER	Gemini-3-Pro	63.64	40.90

TABLE V: Results of MLE-Bench on the Lite split.

Fig. 2: Results of SciCode . Solve rates for main and subproblems are reported, with higher rates indicating greater domain-specific capability.

Method	Model	Score
ReAct [ 46 ]	o3-Mini	0.391
ReAct [ 46 ]	GPT-OSS-120B	0.402
AIRA [ 42 ]	o3	0.424
AIRA [ 42 ]	GPT-5	0.550
SciDER	Claude-Sonnet-4	0.780
SciDER	OpenSciDER-27B	0.519

TABLE VI: Results of AIRS-Bench .

Fig. 3: Results of AstroVisBench . Correct, Failure, Minor Error, and Major Error rates are reported.

Fig. 4: Ablation Study of EIS and Data Analysis: (a) Impact of EIS on AI-Idea-Bench quality, novelty, and feasibility. (b) Impact of Data Analysis on DiscoveryBench accuracy in different tasks.

Panel of Figure 11 — Fig. 5: Selected generated pages showcasing end-to-end scientific discovery evidence for “Student-Conditioned Evidential Meta-Learning with Hyper-Conditioned Priors for Calibrated Few-Shot Knowledge Tracing” . The tiled pages show the generated paper, covering the method, experiments, quantitative results, calibration analysis, limitations, and ethics discussion.

Panel of Figure 12 — Fig. 6: Qualitative generated examples from AstroVisBench . Utilizing an iterative multimodal feedback loop, SciDER successfully synthesizes and refines complex astronomical visualizations, accurately handling long-tail APIs and domain-specific formatting.

Score Level	Evaluation Criteria (Helpfulness)
1 – Obstacle	Fails to parse data or produces non-executable code; requires total restart.
2 – Low Utility	Provides shallow insights; requires significant manual correction.
3 – Average	Automates routine tasks but lacks depth; needs moderate human guidance.
4 – Good	Understands data well; generates executable scripts with only minor gaps.
5 – Excellent	Acts as an autonomous collaborator; provides deep insights and ready-to-use code.

TABLE VII: Rubric for qualitative feedback.

Fig. 7: Per-task scores of our method on DiscoveryBench.

Fig. 8: Per-task scores of our method on AIRS-Bench.

Explore similar work

SciDataSailor: Deep Scientific Data Exploring
Jul 29, 2026Jiyong Rao, Yicheng Qiu, Chi Zhang +2Scientific QA AI Agent Benchmarks
Shanghai Artificial Intelligence Laboratory · Tongji University
SciResearcher: Scaling Deep Research Agents for Frontier Scientific Reasoning
May 2, 2026Tianshi Zheng, Rui Wang, Xiyun Li +5AI for Science Deep Research Agents
HKUST · CUHK · Tencent AI Lab
Mimosa Framework: Toward Evolving Multi-Agent Systems for Scientific Research
Mar 30, 2026Martin Legrand, Tao Jiang, Matthieu Feraud +5Autonomous Research Agents Multi-Agent Systems
Université Côte d’Azur, CNRS, ICN, Nice, France · Interdisciplinary Institute for Artificial Intelligence (3iA) Côte d’Azur, Sophia-Antipolis, France · Inria, Université Côte d’Azur, CNRS, I3S, France

Jul 29, 2026J/K move · Enter open · S save

SciDataSailor: Deep Scientific Data Exploring

Jiyong Rao, Yicheng Qiu, Chi Zhang, Chunfeng Song+1

Shanghai Artificial Intelligence Laboratory · Tongji University

Scientific QA AI Agent Benchmarks LLM Agent Trajectory Synthesis Synthetic Data Generation for Language Models Data Analysis Agents

Open paper AskPDF arXiv

Scientific datasets are commonly organized as hierarchical repositories containing heterogeneous and interdependent files, making their inspection, integration, and analysis labor-intensive and reliant on domain expertise. Although large language model (LLM) agents have advanced substantially in planning, reasoning, and tool use, existing research has largely overlooked their ability to interact with real scientific data assets through executable environments. We introduce Deep Scientific Data Exploration, an agentic task paradigm in which agents navigate repositories, interpret heterogeneous files and schemas, execute analyses, integrate cross-file evidence, and produce conclusions grounded in executed observations. To operationalize this paradigm, we present SciDataSailor, a framework for synthesizing tool-interactive trajectories by balancing broad exploration with targeted exploitation. SciDataSailor instantiates trajectory synthesis as Monte Carlo Tree Search (MCTS) with four task-specific mechanisms: difficulty-stratified exploration seeds, dual-feedback first-play urgency, hierarchical strategy-to-tool action generation, and entropy-guided branching. Using this framework, we construct SciDataSailor-SFT-2K for supervised fine-tuning and SciDataSailor-Bench for evaluation, with the latter comprising 627 meta-information summarization tasks and 586 scientific question-answering tasks across 27 datasets spanning the life, earth, and physical sciences.

Abstract

Explore similar work

SciDataSailor: Deep Scientific Data Exploring

SciResearcher: Scaling Deep Research Agents for Frontier Scientific Reasoning

Mimosa Framework: Toward Evolving Multi-Agent Systems for Scientific Research