DARE-bench: Evaluating Modeling and Instruction Fidelity of LLMs in Data Science
| Authors | Fan Shu et al. |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2602.24288 |
| Download | |
| Categories | cs.AI, cs.CL |
Abstract
The fast-growing demands in using Large Language Models (LLMs) to tackle complex multi-step data science tasks create an emergent need for accurate benchmarking. There are two major gaps in existing benchmarks: (i) the lack of standardized, process-aware evaluation that captures instruction adherence and process fidelity, and (ii) the scarcity of accurately labeled training data. To bridge these gaps, we introduce DARE-bench, a benchmark designed for machine learning modeling and data science instruction following. Unlike many existing benchmarks that rely on human- or model-based judges, all tasks in DARE-bench have verifiable ground truth, ensuring objective and reproducible evaluation. To cover a broad range of tasks and support agentic tools, DARE-bench consists of 6,300 Kaggle-derived tasks and provides both large-scale training data and evaluation sets. Extensive evaluations show that even highly capable models such as gpt-o4-mini struggle to achieve good performance, especially in machine learning modeling tasks. Using DARE-bench training tasks for fine-tuning can substantially improve model performance. For example, supervised fine-tuning boosts Qwen3-32B's accuracy by 1.83x and reinforcement learning boosts Qwen3-4B's accuracy by more than 8x. These significant improvements verify the importance of DARE-bench both as an accurate evaluation benchmark and critical training data.
Engineering Breakdown
Plain English
DARE-bench is a new benchmark for evaluating how well large language models can follow instructions and execute multi-step data science tasks. The paper identifies two critical gaps in existing benchmarks: they lack process-aware evaluation that measures instruction adherence, and they suffer from inaccurate or scarce training data. The authors introduce DARE-bench with 6,300 Kaggle-derived tasks that have verifiable ground truth labels, enabling objective evaluation without relying on human or model-based judges. This shift to ground-truth evaluation makes benchmarking reproducible and removes subjectivity, addressing a real pain point in how the field currently measures LLM performance on complex reasoning tasks.
Core Technical Contribution
The core novelty is shifting from subjective evaluation (human judges, LLM-as-judge) to objective, ground-truth-based evaluation for data science tasks performed by LLMs. The authors designed tasks where correctness is verifiable through concrete outputs (e.g., code execution, model performance metrics, predictions against held-out data), rather than relying on semantic similarity judgments. They also emphasize 'process fidelity'—measuring not just whether an LLM gets the right answer, but whether it follows the correct reasoning steps and adheres to specified instructions throughout the workflow. By sourcing 6,300 real tasks from Kaggle competitions, they ensure domain authenticity and coverage of diverse modeling paradigms, setting a precedent for benchmarks that capture agentic, multi-step reasoning in data science workflows.
How It Works
DARE-bench operates on a verification-first principle: each task is decomposed into executable steps where ground truth can be computed deterministically. For a typical data science workflow, the benchmark accepts an LLM-generated solution (code, model configuration, predictions), executes it in a sandboxed environment against held-out test data, and compares quantitative outputs (accuracy, F1, loss, prediction arrays) to reference solutions. The benchmark captures instruction following by tracking whether the LLM adheres to specific constraints (e.g., 'use only tree-based models', 'apply feature scaling before training') and process fidelity by measuring intermediate step correctness—not just final outputs. The 6,300 tasks are sourced from Kaggle competitions and include diverse problem types: regression, classification, anomaly detection, time series forecasting, and feature engineering. Evaluation metrics include both task completion (did the code run without error?) and quality metrics (how close to optimal performance), allowing fine-grained diagnosis of where LLM reasoning breaks down.
Production Impact
For teams deploying LLMs as data science agents, this benchmark provides a principled way to evaluate and select models before production deployment, reducing the risk of deploying systems that 'look correct' but fail on process adherence. Production ML pipelines can adopt DARE-bench's ground-truth evaluation methodology to create internal benchmarks for their own data science workflows, enabling continuous monitoring of agent quality as models evolve. The emphasis on instruction fidelity directly addresses a critical failure mode in production: LLMs that produce syntactically correct code but ignore domain constraints (e.g., using prohibited libraries, skipping required preprocessing). Teams can use this benchmark to identify which model/prompt combinations maintain process discipline, crucial for regulated industries where traceability and compliance matter as much as accuracy. The tradeoff is computational: executing 6,300 tasks with code sandbox isolation and multiple model runs requires significant GPU/CPU resources, and the Kaggle task diversity may not perfectly match your specific domain, requiring supplementary internal benchmarking.
Limitations and When Not to Use This
DARE-bench assumes that Kaggle competition tasks represent the distribution of real-world data science problems, which may not hold for specialized domains like bioinformatics, drug discovery, or proprietary industrial datasets with unique constraints and data characteristics. The paper focuses on supervised learning tasks and may not adequately cover unsupervised learning, reinforcement learning, or reinforcement-learning-from-human-feedback scenarios increasingly common in production agents. Verifiable ground truth is easier to establish for well-defined Kaggle competitions than for open-ended exploratory analysis, ad-hoc feature discovery, or tasks where multiple reasonable solutions exist—the benchmark may over-reward 'canonical' solutions and penalize creative approaches. The paper is dated 2026 (future), so practical results, comparison against other benchmarks, and evidence of model saturation or ceiling effects are not yet available; early adoption carries risk of investing in a benchmark that may not correlate with production success.
Research Context
DARE-bench builds on the growing recognition that existing benchmarks like MMLU, HumanEval, and SQuAD do not measure process reasoning or multi-step instruction following required for agentic LLM systems. It advances the line of work on benchmark design pioneered by platforms like SuperGLUE and HELM, but specifically tailored to data science workflows where ground truth is computationally verifiable. The paper responds to concurrent research identifying critical gaps in instruction-following evaluation (e.g., IFEval) and the unreliability of LLM-based evaluation (shown in papers on reward model brittleness). By combining real competition tasks with rigorous process-aware metrics, DARE-bench positions itself as a reference point for the emerging field of 'reasoning benchmarks' that go beyond surface-level task completion to measure the coherence and discipline of LLM reasoning chains.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::