DataSense-Bench: The First Step Toward an AI Scientist

The gist: DataSense-Bench introduces a benchmark to study whether frontier AI models possess a sense of data, specifically by asking agents to select and rank training subsets for fine-tuning an LLM based on expected post-training performance.

Introduction and Motivation

Large language models are increasingly used for tasks such as question answering (Brown et al., 2020), code generation (Chen et al., 2021), and data analysis (Majumder et al., 2025). Recent advances in reasoning and tool use have made frontier models capable of tackling complex scientific problems, supporting AI scientists who generate research ideas, conduct experiments, analyze results, and write complete research manuscripts (Lu et al., 2026). This setting leaves a more focused question open: can an AI system anticipate the value of an experimental choice before observing its outcome? Training-data selection provides a concrete setting for this question because researchers must choose which examples will help a particular model learn (Zhou et al., 2023). Reliable judgment requires connecting properties of the data to what a particular model will learn under a fixed training procedure (Li et al., 2024).

Benchmark Setup and Task Design

DATASENSE-BENCH asks CLI agents to select and rank candidate training subsets that can be used to fine-tune a small LLM model (Page 3). The benchmark instantiates the protocol in terminal problem solving and multi-turn tool use, selecting trajectories from OpenThoughts-Agent and EnvScaler and evaluating on TBLite and BFCL, respectively (Page 3). For each task, agents select five disjoint groups of complete trajectories (Page 4). The agent may inspect data, write and execute analysis code, and run model forward passes but cannot train the model or access the actual evaluation tasks (Page 5). The objectives for the agents are to make g1 as strong as possible and to order the groups by expected post-training performance (Page 5).

Evaluation Metrics

The benchmark evaluates two complementary dimensions: selection quality and ranking accuracy (Page 3). Selection quality is measured by the score s1, which measures the utility of the data recommended most strongly, and its gain over random selection, ∆random = s1 − srandom (Page 6). Ranking accuracy tests comparison at two levels: Best-group accuracy (Abest), which asks whether the agent’s first choice is actually the best among its selected groups, and Spearman correlation (Srank), which tests the full predicted ordering (Page 6).

Experimental Findings

The experiments reveal two primary limitations: gains over random selection are modest, and agents do not reliably rank their selected groups (Page 1). On TBLite, two LLMs select first groups that underperform random selection, with differences across LLMs ranging from-0.55 to +0.99 points (Page 1). Selection recipes vary substantially across three independent runs of the same agent (Page 2). Agents draw different conclusions from similar data signals, as they use base-model loss in substantially different ways across runs and tasks (Page 2). For example, Astra’s TBLite runs change their quality-assessment procedures, while Kimi’s BFCL runs move from loss-outlier rejection to intermediate-loss and then lower-loss preferences (Page 8).

Resource Requirements and Cost

The benchmark separates the cost of making a data recommendation from the cost of testing it through SFT and downstream evaluation (Page 9). The CLI agent time budget is four hours, while training time for TBLite is estimated at 2.04 h and for BFCL at approximately 3.93 h (Page 16). The benchmark separates the cost of making a data recommendation from the cost of testing it through SFT and downstream evaluation (Page 9).

Conclusion

The results highlight a gap between producing a convincing analysis of data and making reliable decisions about its training value, motivating data sense as a capability to evaluate when assessing progress toward AI scientists and their recursive self-improvement (Page 10). Future work should extend this scope to reinforcement-learning post-training and the generation of new training data (Page 10).

--- Page 1 ---

Preprint

DATASENSE-BENCH: THE FIRST STEP TOWARD AN AI SCIENTIST Yudi Zhang, Mingyu Cao, Lu Yin, Mykola Pechenizkiy, Shiwei Liu. As claims about recursive self-improvement (RSI) and artificial general intelligence (AGI) proliferate, we ask a simple question: do frontier AI models have a sense of data, i.e., can they reliably select the right data for training? We introduce DataSense-Bench to study this capability through the fundamental problem of data selection and performance forecasting in machine learning. Agents are allowed to inspect the data, write and execute analysis code, and run model forward passes, but can not train the model or access the actual evaluation tasks. We then fine-tune the base model on each selected subset and evaluate its post-training performance under a standardized protocol. We instantiate the benchmark in terminal problem solving and tool use, selecting trajectories from OpenThoughts-Agent and EnvScaler and evaluating on TBLite and BFCL, respectively. We then evaluate the agents along two complementary dimensions: the post-training performance of the top-ranked subset, reflecting the ability to identify high-value training data, and ranking accuracy, reflecting the ability to predict the relative performance of the selected subsets. In our experiments, selection gains over random selection are limited; agents do not reliably rank their selected groups, and ranking ability does not hold consistently across tasks: Astra identifies the best group in all three tool-use runs but in only one of three terminal runs. Analysis of execution traces on both tasks shows that agents often use similar data signals while interpreting their training value differently. Base model Fable 5 GPT-5.6 Sol Random selection Sonnet 5 Opus 5 GPT-5.5 Kimi K3 GPT-6 Astra Gemini 3.5 Flash Fable 5.1 DeepSeek V4 Pro

--- Page 2 ---

Preprint

1 INTRODUCTION Large language models (LLMs) are increasingly used for tasks such as question answering (Brown et al., 2020), code generation (Chen et al., 2021), and data analysis (Majumder et al., 2025). Recent advances in reasoning and tool use have made frontier models capable of tackling complex scientific problems, supporting AI scientists who generate research ideas, conduct experiments, analyze results, and write complete research manuscripts (Lu et al., 2026). Benchmarks increasingly assess scientific programming, discovery, and machine learning experimentation (Chen et al., 2025; Majumder et al., 2025; Huang et al., 2024; Rank et al., 2026). They leave a more focused question open: can an AI system anticipate the value of an experimental choice before observing its outcome? Training-data selection provides a concrete setting for this question because researchers must choose which examples will help a particular model learn (Zhou et al., 2023). This is a consequential and difficult decision: data selection determines the supervision a model receives, yet observable quality does not directly establish training value (Page 2).

--- Page 3 ---

Preprint

3.1 TASK FOR CLI AGENTS: SELECT USEFUL DATA AND PREDICT ITS VALUE Let D be a pool of N candidate trajectories, M a base model, and R the fine-tuning procedure used to test a submission. The agent selects K subsets G = (g1,..., gK) with gk ⊆ D, gk = n, and gi ∩gj = ∅ for i ̸= j (Page 3). The group labels encode its prediction: training on g1 should give the highest downstream score, followed by g2, through gK (Page 3). Post-training Data. We study two tasks that require learning from multi-step interactions: terminal problem solving and multi-turn tool use (Page 4). For each task, agents select five disjoint groups of complete trajectories, and we independently fine-tune Qwen3-4B on each group using supervised fine-tuning (SFT) (Page 4). Tool usage. The training data come from the EnvScaler SFT pool with validation trajectories excluded (Page 5). Terminal problem solving. The training pool is OpenThoughts-Agent-SFT-10K, which contains trajectories pairing task instructions with terminal interactions, including commands, their outputs, and error recovery (Page 5). Tool usage. The training data come from the EnvScaler SFT pool with validation trajectories excluded (Page 5).

--- Page 4 ---

Preprint

3.2 POST-CHECK AND POST-TRAINING EVALUATION After the agent submits its selection, the benchmark validates the submission and tests its predicted ordering (Page 6). Post-training evaluation. Each group is used to fine-tune a fresh copy of the base model under its task-specific training procedure (Page 6).

--- Page 5 ---

Preprint

3.3 EVALUATION METRICS We want to evaluate two complementary capabilities of the data selection agents: whether the agent selects useful training data, and whether it can anticipate which data choices will be more useful (Page 7). Selection quality. The score s1 measures the utility of the data the agent recommends most strongly, and we also report its gain over random selection, ∆random = s1 − srandom (Page 7). Ranking accuracy.

1. Data sense capability evaluation via DATASENSE-BENCH: The system can reliably select and rank candidate training subsets that can be used to fine-tune a small LLM model by predicting post-training performance before observing outcomes, addressing the fundamental problem of data selection.

2. Improved selection recipes through multi-signal integration: The AI system will construct a composite score for data valuation by combining multiple complementary signals (e.g., loss/perplexity, trajectory quality, task difficulty, tool-call correctness, efficiency, diversity/coverage of environments and tools) rather than relying on a single metric like base-model loss, which the paper shows is insufficient.

3. Enhanced ranking accuracy through differentiated selection strategies: The system can be trained to produce more reliable rankings by learning that the ordering is part of the selection procedure, allowing it to use complex rules, such as those seen in GPT-6 Astra’s third TBLite run, which combines example-level scores with group-level coverage choices.

As claims about recursive self-improvement (RSI) and artificial general intelligence (AGI) proliferate, we ask a simple question: do frontier AI models have a sense of data, i.e., can they reliably select the right data for training? We introduce DataSense-Bench to study this capability through the fundamental problem of data selection and performance forecasting in machine learning. We ask AI agents to select and rank candidate training subsets that can be used to fine-tune a small LLM model. Agents are allowed to inspect the data, write and execute analysis code, and run model forward passes, but can not train the model or access the actual evaluation tasks. We then fine-tune the base model on each selected subset and evaluate its post-training performance under a standardized protocol. We instantiate the benchmark in terminal problem solving and tool use, selecting trajectories from OpenThoughts-Agent and EnvScaler and evaluating on TBLite and BFCL, respectively. We then evaluate the agents along two complementary dimensions: the post-training performance of the top-ranked subset, reflecting the ability to identify high-value training data, and ranking accuracy, reflecting the ability to predict the relative performance of the selected subsets. In our experiments, selection gains over random selection are limited; agents do not reliably rank their selected groups, and ranking ability does not hold consistently across tasks: Astra identifies the best group in all three tool-use runs but in only one of three terminal runs. Analysis of execution traces on both tasks shows that agents often use similar data signals while interpreting their training value differently.