DataSense-Bench: The First Step Toward an AI Scientist

Yudi Zhang; Mingyu Cao; Lu Yin; Mykola Pechenizkiy; Shiwei Liu

cs.LGOct 8, 2026

DataSense-Bench: The First Step Toward an AI Scientist

Authors: Yudi Zhang, Mingyu Cao, Lu Yin, Mykola Pechenizkiy, Shiwei Liu

Organizations: Eindhoven University of Technology · Max Planck Institute for Intelligent Systems · University of Surrey · ELLIS Institute Tübingen · Tübingen AI Center

Abstract

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.

Figures & tables

Figure 1: Can CLI agents select useful training data and predict its value? (a,b) Qwen3-4B performance after training on the predicted top-ranked group selected by each agent. Panel (a) reports BFCL accuracy; panel (b) reports TBLite reward. Dashed lines denote random selection. Scores are averaged across evaluation seeds within checkpoints, then across three agent runs; error bars show standard deviation across agent runs (across evaluation seeds for random selection). (c) Frequency of each predicted rank yielding the best group, pooling evaluated runs from both tasks with equal weight per run: six runs for agents evaluated on both tasks and three for TBLite-only agents. Tied best groups each count as correct; blue outlines mark $g_{1}$ .

Figure 2: From data-selection predictions to post-hoc verification. Illustrated for terminal problem solving, the CLI agent analyzes candidate trajectories with inference-only access to the base model. It delivers a row-to-group assignment CSV and a method report specifying its selection recipe. We validate five disjoint groups of the specified size and freeze their membership and predicted order. Only then does evaluation perform SFT and downstream scoring, testing selection quality, best-group accuracy, and agreement with the full predicted ordering.

Panel of Figure 3 — Figure 3: Condensed terminal-task prompt specifying the selection and ranking objectives, permitted tools, computation budget, and required deliverables. The prompt for BFCL is described in Appendix C.3 .

BFCL accuracy (50 trajectories per group) Base model: 31.67%
Rank	Agent / reference	$s_{1}$	$g_{1}$ gain $\Delta_{\mathrm{random}}$	$s_{2}$	$s_{3}$	$s_{4}$	$s_{5}$	$S_{\mathrm{rank}}$	$A_{\mathrm{best}}$ (%)
1	DeepSeek V4 Pro	$36.53_{\text{$ \pm $1.83}}$ [-2pt] Max: 38.62	1.78	$\mathbf{36.61}_{\text{$ \pm $0.72}}$ [-2pt]Max: 37.42	$35.74_{\text{$ \pm $1.12}}$ [-2pt]Max: 37.00	$36.08_{\text{$ \pm $1.44}}$ [-2pt]Max: 37.75	$34.33_{\text{$ \pm $1.16}}$ [-2pt]Max: 35.67	0.65	33.33
2	GPT-6 Astra	$\mathbf{36.49}_{\text{$ \pm $0.23}}$ [-2pt]Max: 36.75	1.74	$35.31_{\text{$ \pm $0.94}}$ [-2pt]Max: 36.38	$34.74_{\text{$ \pm $1.50}}$ [-2pt]Max: 36.21	$34.93_{\text{$ \pm $0.42}}$ [-2pt]Max: 35.42	$34.58_{\text{$ \pm $0.07}}$ [-2pt]Max: 34.62	0.83	100.00
3	Kimi K3	$36.29_{\text{$ \pm $0.91}}$ [-2pt]Max: 36.92	1.54	$36.21_{\text{$ \pm $0.65}}$ [-2pt]Max: 36.92	$36.18_{\text{$ \pm $1.22}}$ [-2pt]Max: 37.58	$35.85_{\text{$ \pm $1.14}}$ [-2pt]Max: 36.62	$\mathbf{36.46}_{\text{$ \pm $1.27}}$ [-2pt]Max: 37.75	0.02	33.33
4	Opus 5	$35.94_{\text{$ \pm $0.72}}$ [-2pt]Max: 36.75	1.19	$\mathbf{36.35}_{\text{$ \pm $1.09}}$ [-2pt]Max: 37.38	$35.35_{\text{$ \pm $0.53}}$ [-2pt]Max: 35.96	$34.29_{\text{$ \pm $2.29}}$ [-2pt]Max: 36.54	$31.08_{\text{$ \pm $3.45}}$ [-2pt]Max: 35.04	0.77	0.00
5	Fable 5.1	$35.21_{\text{$ \pm $0.94}}$ [-2pt]Max: 36.00	0.46	$34.96_{\text{$ \pm $1.67}}$ [-2pt]Max: 36.58	$34.87_{\text{$ \pm $0.07}}$ [-2pt]Max: 34.96	$34.10_{\text{$ \pm $0.24}}$ [-2pt]Max: 34.38	$\mathbf{35.57}_{\text{$ \pm $1.53}}$ [-2pt]Max: 37.17	0.23	33.33

Table 1: Selection performance, best-group accuracy, and rank performance on BFCL and TBLite. Agent-group scores (

s_{1}

s_{5}

) report the post-training performance, accuracy (%) for BFCL, and reward for TBLite (scale from 0-1 to 0-100). We report the average performance across three evaluation seeds and three data selection runs, standard deviation (subscript), and maximum score among different data selection runs. Shading compares group averages within a row; red outlines mark each task’s largest maximum among agent-selected groups.

g_{1}

gain

\Delta_{\text{random}}

is relative to random selection. Best-group accuracy

A_{\mathrm{best}}

measures how often the agent identifies its best group; rank performance

S_{\mathrm{rank}}

measures Spearman correlation between predicted and observed group orderings (Section 3.3 ). More detailed experimental results are in Appendix A.3 .

Panel of Figure 4 — Figure 4: Fable 5.1’s data-selection workflow, from inspection to final recipe.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

	Terminal problem solving	Multi-turn tool use
Training pool	OpenThoughts-Agent 10K	EnvScaler SFT
Groups $\times$ group size	$5\times 1,000$	$5\times 50$
Base model	Qwen3-4B	Qwen3-4B
Selectors / runs per selector	10 / 3	6 / 3
Selection GPU / time budget	H100 80GB / 4 h	H100 80GB / 4 h
Learning rate / epochs	$4\times 10^{-5}$ / 7	$5\times 10^{-6}$ / 2

Appendix

Table 3: Configuration of the two selection-and-ranking tasks. Group sizes count complete trajectories.

Agent	Model identifier
Fable 5.1	claude-fable-5-1
Opus 5	claude-opus-5
GPT-6 Astra	gpt-6-astra
Kimi K3	kimi-k3
DeepSeek V4 Pro	deepseek-v4-pro
Gemini 3.5 Flash	gemini-3.5-flash

Appendix

Table 4: Selection-agent model identifiers.

Agent	Run	Signals used	Final grouping rule
Fable 5	1	Structure, NLL tails and 64 task clusters.	Cluster-balanced $g_{1}$ ; spaced lower score bands.
	2	Structure and partial NLL coverage.	Task deduplication and cluster caps in $g_{1}$ .
	3	Structure and 8K-prefix NLL.	Task-deduplicated $g_{1}$ without clustering; lower score windows.
Fable 5.1	1	Structural quality with a small JSON-NLL term.	Task deduplication and cluster/style caps in $g_{1}$ / $g_{2}$ ; lower score bands.
	2	Structural quality, model loss and a prompted concreteness score.	Task deduplication and cluster/style caps in $g_{1}$ / $g_{2}$ ; lower score bands.
	3	Structural quality with soft loss-anomaly penalties.	Cluster/source caps and an 8% Tezos cap in $g_{1}$ / $g_{2}$ ; next-ranked $g_{3}$ / $g_{4}$ and bottom $g_{5}$ .

Appendix

Table 5: TBLite selection methods, with one row per attempt.

Agent	Recipe and cross-run differences
Fable 5.1	Run 1 combines schema checks, errors, reasoning length, and length-adjusted loss; it samples descending score bands, ending with low-scoring rows. Run 2 emphasizes reasoning–action consistency and loss outliers. Run 3 separately scores reasoning, actions given reasoning, and actions without reasoning, then combines quality gates with environment coverage. Later groups include increasingly defect-heavy trajectories.
Opus 5	Run 1 prioritizes structural quality and uses a small residualized-loss term, with fixed interaction-type and environment coverage across groups. Run 2 adds verbosity-normalized reasoning loss and recovered-error bonuses, placing noncanonical argument formats in the lower groups. Run 3 probes action difficulty without the demonstrated reasoning and selects widely separated score bands with distinct environments.
GPT-6 Astra	All runs combine schema checks, argument grounding, state-changing actions, clarification, and coverage. Run 1 adds sampled-turn loss over the pool and inspection of top candidates. Run 2 uses a bounded preference for moderate reasoning loss and a fixed 70/30 interaction-type mix. Run 3 scores a quality-filtered shortlist and adds semantic checks on unsupported actions. Groups occupy separated predicted-value bands.
Kimi K3	Run 1 uses loss to penalize outliers, normalizes quality within interaction types, and selects successive groups under diversity caps. Run 2 favors intermediate total and tool-call loss. Run 3 instead rewards lower loss alongside successful responses and tool coverage, with environment caps relaxed from one to three across the groups.
DeepSeek V4 Pro	Run 1 prioritizes high sampled-turn loss, trajectory length, and final-action loss. Run 2 combines loss with structural richness, including a positive failed-call term, and uses diversity-penalized greedy selection. Run 3 emphasizes tool breadth, penalizes failed-call rate, and adds within-environment loss as a secondary signal.
Gemini 3.5 Flash	Run 1 separates groups by observed tool success, length, and turn count. Run 2 uses a heuristic score penalizing errors and loops while rewarding reasoning length and intermediate turn counts. Run 3 adds base-model loss to choose long, error-free trajectories for the top groups; short clean trajectories and error-heavy trajectories occupy lower groups.

Appendix

Table 6: BFCL selection recipes and differences across the three independent runs. Each run produces five disjoint groups of 50 complete trajectories.

Agent	Run	$g_{1}$	$g_{2}$	$g_{3}$	$g_{4}$	$g_{5}$	$\rho$	$g_{1}$ rank
Fable 5	r1	0.109	0.085	0.145	0.101	0.107	0.10	2
	r2	0.100	0.110	0.121	0.110	0.099	0.30	4
	r3	0.121	0.121	0.135	0.134	0.121	-0.50	4
Fable 5.1	r1	0.117	0.122	0.129	0.098	0.116	0.50	3
	r2	0.110	0.103	0.110	0.097	0.132	-0.30	3
	r3	0.148	0.111	0.121	0.112	0.101	0.70	1

Appendix

Table 7: Terminal group-level results on a 0–1 reward scale. Runs use the same 1–3 numbering as the procedure summaries. Each reward averages 100 tasks with three attempts per task. Within each row, shading runs from darkest (best) to lightest (worst), and bold marks the best group. Colors use unrounded means; equal means receive the same shade.

\rho

compares the predicted

g_{1}

–

g_{5}

order with observed performance;

g_{1}

rank is one plus the number of groups scoring strictly higher.

Agent	Run	$g_{1}$	$g_{2}$	$g_{3}$	$g_{4}$	$g_{5}$	$\rho$	$g_{1}$ rank
Fable 5.1	r1	36.00	33.25	34.83	34.38	37.17	-0.30	2
	r2	34.17	36.58	34.96	33.96	34.12	0.60	3
	r3	35.46	35.04	34.83	33.96	35.42	0.40	1
Opus 5	r1	35.71	36.46	35.08	31.96	28.75	0.90	2
	r2	36.75	37.38	35.00	36.54	29.46	0.80	2
	r3	35.38	35.21	35.96	34.38	35.04	0.60	2

Appendix

Table 8: BFCL accuracy (%) for 50-trajectory groups in each independent selection run. Scores average the 800-case evaluations across evaluation seeds. Shading compares groups within a run;

\rho

measures full-order agreement;

g_{1}

rank is one plus the number of groups scoring strictly higher.

Explore similar work

RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement
Jul 28, 2026Fanqing Meng, Lingxiao Du, Qiguang Chen +4LLM Agent Evaluation LLM Agent Self-Improvement
1Evolvent AI · 2National University of Singapore
AutoData: Agentic Search for Pre-training Data Selection
Sep 17, 2026Yan Meng, Dhruv Srikanth, Bingchen Zhao +2Language Model Pretraining Automated Algorithm Discovery
University of Amsterdam · Weco AI
What is Missing from AI Post-Training AI: An Empirical Analysis
Aug 19, 2026Joy Jia Yin Lim, Xin Huang, Hao Peng +5AI Agent Evaluation LLM Agent Self-Improvement
Tsinghua University · University of Electronic Science and Technology of China · Renmin University of China

Jul 28, 2026, cs.SEJ/K move · Enter open · S save

RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement

Fanqing Meng, Lingxiao Du, Qiguang Chen, Ziqi Zhao+3

1Evolvent AI · 2National University of Singapore

LLM Agent Evaluation LLM Agent Self-Improvement AI Agent Benchmarks LLM Post-Training Recursive Self-Improvement

Open paper AskPDF arXiv

Recursive self-improvement requires turning evidence of model failures into better models. Data-centric post-training research entails diagnosing capability gaps, designing and validating training-data strategies, and learning from checkpoint feedback. Can LLM agents automate this loop? Existing benchmarks entangle research decisions with optimization, serving, evaluation, and systems implementation, obscuring agents' research capability. We introduce RSIBench-Data, a controlled benchmark of LLM agents as data-centric researchers with a fixed post-training stack. Agents iteratively revise training-data strategies for a fixed target model; training and serving use Tinker-backed services, official evaluation runs through Harbor and E2B sandboxes, and budgets are fixed across agents. We evaluate four frontier agents on six benchmarks across software engineering, terminal use, scientific question answering, and mathematics. Agents demonstrate core data-centric research capabilities: in 58.33% of settings, they improve upon the first valid attempt by refining strategies from feedback. However, improvement is inconsistent. Among searches continuing after the best observed score, 78.26% end with a lower-scoring final attempt, while the rest only recover the same peak. A strong candidate may therefore appear early or midway through a run even as later revisions fail. Trajectory analysis identifies four patterns in stronger runs: accurate hypotheses, validation-grounded supervision, behavior-aligned data, and preservation of strong checkpoints. These findings suggest that current agents can make useful data-centric discoveries but cannot yet translate feedback into consistent improvements. RSIBench-Data provides a measurable, auditable testbed for the research capabilities required for recursive self-improvement. We open-source our code at https://github.com/evolvent-ai/RSIBench-Data.

Abstract

Explore similar work

RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement

AutoData: Agentic Search for Pre-training Data Selection

What is Missing from AI Post-Training AI: An Empirical Analysis