acceptodds
Under review as a conference paper at ICLR 2027

DataSense-Bench: The First Step Toward an AI Scientist

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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
AI Scientists, Data Selection, LLM Post-Training, Benchmarking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2N5SOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DataSense-Bench, which is a benchmark that tests whether frontier AI agents can judge the training value of data before any training happens. Within the designed pipeline, the CLI agent receives a candidate data pool, inference-only access to Qwen3-4B, a fixed SFT recipe, one H100 GPU, and four hours. It must form five disjoint, equal-sized groups and rank them by expected post-training performance. The benchmark then finetunes a fresh model on each group and measures the score of the top-ranked group against random selection. The benchmark covers two tasks: terminal problem solving and multi-turn tool use. Ten agents are tested on TBLite and six on BFCL, with three runs each. The authors find that gains over random selection are small, rankings are unreliable and inconsistent across tasks, and the same agent often changes its selection rule across runs.

Critical strengths
  • This paper explores a timely and well-motivated problem. As agents apply to research tasks, it matters whether they can predict the value of an experimental choice before conducting training. Data selection is a concrete and checkable example of this.

  • The paper reports results for every group in every run, summarizes each agent's selection recipe, and reports API cost and time. This makes it easy to see how agents reason and where they disagree.

Critical weaknesses
  • Limited novelty and insight. The benchmark combines existing data pools, evaluations, models, and agents through a simple protocol. This protocol (select a subset from a fixed pool, train with a fixed recipe, and measure the result) is the core design of existing benchmarks such as DataComp, DataComp-LM, and DataPerf, which the paper does not cite. PostTrainBench already tests agents on post-training choices, including data. The main new element is using AI agents as the selectors. The paper also shows that agents select and rank poorly, but it does not test any way to improve this, such as limited training feedback or combining multiple runs, and it does not analyze which of the agents' choices lead to better outcomes. So it diagnoses a gap but offers little guidance on how to close it.

  • Limited experimental evidence. Agents are compared only with random selection and a "more data" reference. There are no standard selection methods (e.g., loss filtering, deduplication, LESS) or simple heuristics run under the same protocol, so we cannot tell whether agents are worse than simple algorithms or whether the task is just hard. The benchmark also uses one base model, one SFT recipe, and two tasks, and only six of the ten agents are tested on BFCL. It is unclear whether the conclusions hold more broadly.

Questions for authors
  • Would agents select and rank better with limited feedback, such as a small proxy training run or a few validation results? Does combining an agent's three runs give more stable or better selections? Which of the agents' choices (e.g. how they use base-model loss) are linked to better outcomes?

  • How does DataSense-Bench differ from DataComp, DataComp-LM, DataPerf, and PostTrainBench? What does it reveal that these benchmarks cannot?

  • How do standard data-selection methods (e.g., loss filtering, deduplication, LESS) and simple heuristics (e.g., completed trajectories only) perform under the same protocol? Do agents outperform them?

  • Why are only six of the ten agents tested on BFCL? Do the conclusions hold for other base models or training recipes?

8 comments

Thank you for taking the time to provide detailed feedback on the motivation and experimental scope of our work. Your suggestions broaden the baseline comparisons and motivate us to investigate aggregation and feedback as potential improvements. We address each of your concerns below.

Response to Weakness 1:

Limited novelty and insight. The benchmark combines existing data pools, evaluations, models, and agents through a simple protocol. This protocol (select a subset from a fixed pool, train with a fixed recipe, and measure the result) is the core design of existing benchmarks such as DataComp, DataComp-LM, and DataPerf, which the paper does not cite. PostTrainBench already tests agents on post-training choices, including data. The main new element is using AI agents as the selectors. The paper also shows that agents select and rank poorly, but it does not test any way to improve this, such as limited training feedback or combining multiple runs, and it does not analyze which of the agents' choices lead to better outcomes. So it diagnoses a gap but offers little guidance on how to close it.

Thank you for highlighting these related benchmarks.

First, we want to clarify our motivation: to examine agents’ data sense through selection quality and ranking ability. Using agents to select data before costly training brings our benchmark closer to practical workflows in frontier labs and real-world applications. We view the ability to select useful data and anticipate its value as a foundational step toward AI scientists that can make autonomous experimental decisions.

Difference from prior benchmarks

DataComp and DataComp-LM benchmark training-data curation, while DataPerf evaluates a broader range of data-centric methods. We examine whether agents can both select useful data and predict its downstream value by asking them to commit to selected groups and their rankings before training.

To isolate this ability, we fix the downstream model and training recipe. This distinguishes our setting from PostTrainBench, which evaluates end-to-end post-training with data choices, training methods, and hyperparameters jointly determining performance. We added these citations and comparisons in Section 2 and clarified our motivation in Section 1.

Consensus and iterative improvement with limited feedback

Consensus. We select 50 trajectories by frequency across Kimi’s three original top-group subsets, breaking ties by pool index without using scores, and train on the merged subset.

Iterative improvement with limited feedback. We explicitly prompt Kimi K3 and GPT-6 Astra to iteratively improve their selections using reward from 98 validation tasks, with three independent runs per agent and up to five rounds. Checkpoints are selected by validation reward, and BFCL800 test scores remain hidden to the agents.

Experiment / agentReference accuracy (%)Result accuracy (%)Change (%)
Kimi consensus36.29, original selections37.32 ± 0.56+1.03
Kimi iterative feedback36.19, round one36.93, validation-selected+0.74
Astra iterative feedback35.92, round one36.06, validation-selected+0.14

Consensus improves the observed BFCL accuracy from 36.29% to 37.32%.

Iterative improvement with feedback can improve performance on the validation set in early rounds, but these gains are not consistently sustained and transfer poorly to the unseen evaluation set: validation-selected checkpoints improve average test accuracy by 0.74% for Kimi and 0.14% for Astra, with one Astra run declining by 1.00%.

Additionally, our analysis in Section 4.4 shows that selection based on loss, diversity, or quality signals alone does not consistently improve performance across tasks. Therefore, future work could explore memory, skills, and agent harness design to better use feedback and iteratively improve selection.

Revision: Section 4.5 presents consensus and iterative feedback; Figure 5 shows the improvement on the validation set and evaluation set. Appendix D.3 gives consensus details, and D.5 specifies the iterative protocol.

Response to Weakness 2:

Limited experimental evidence. Agents are compared only with random selection and a "more data" reference. There are no standard selection methods (e.g., loss filtering, deduplication, LESS) or simple heuristics run under the same protocol, so we cannot tell whether agents are worse than simple algorithms or whether the task is just hard. The benchmark also uses one base model, one SFT recipe, and two tasks, and only six of the ten agents are tested on BFCL. It is unclear whether the conclusions hold more broadly.

We added Qwen3.5-2B experiments on BFCL, evaluated the four previously missing agents to complete BFCL coverage for all ten agents, and added standard selection baselines. We will also extend the additional-model evaluation to TBLite.

Protocol. We expand Qwen3-4B BFCL coverage to all ten agents and evaluate six agents on Qwen3.5-2B. Each agent contributes three selections of five groups, with 50 trajectories per group. Every checkpoint is evaluated three times on BFCL800.

BFCL evaluationQwen3-4BQwen3.5-2B
Average g1 accuracy range (%)34.76–36.9316.83–20.29
Random accuracy (%)35.58 ± 0.8118.38 ± 0.82
Highest average g1 accuracy (%)36.93, GPT-5.6 Sol20.29, Kimi
Average rank-correlation range−0.27 to 0.83−0.10 to 0.93
Identifies the best group in all three runsAstraNone
Agent / referenceBFCL 4B accuracy (%)BFCL 2B accuracy (%)
GPT-5.6 Sol36.93—
DeepSeek V4 Pro36.5319.89
GPT-6 Astra36.4919.92
Kimi K336.2920.29
Opus 535.9419.64
Random35.5818.38
Fable 5.135.2119.15
Gemini 3.5 Flash35.1016.83
Fable 535.01—
Sonnet 534.96—
GPT-5.534.76—

Agent entries report average g1 accuracy across three independent selection runs.

Five agents exceed the respective Random average in each model panel, but reliably identifying the best group remains difficult on both models. We hold the training recipe fixed within each task to isolate differences in data selection.

We also evaluate low- and high-loss selection, diversity, deduplication, LESS, and completion flags under the same group sizes and downstream protocols. The tested loss-, diversity-, and quality-based selectors do not consistently improve on Random across tasks, motivating evaluation of how agents choose and combine these signals. Agent recipes themselves often combine these simple criteria: Kimi and DeepSeek use loss and diversity constraints, while Opus and Astra incorporate quality or execution-evidence checks, as analyzed in Section 4.4. Please refer to Section 4.2 and Table 1 for the full results and Appendix D.1–D.2 for the protocols.

We use task-specific SFT configurations rather than one shared recipe. TBLite uses the hyperparameters provided by the training-data authors for 1K trajectories, with a learning rate of 4×10⁻⁵ for seven epochs; BFCL uses 5×10⁻⁶ for two epochs, chosen through limited manual tuning rather than a systematic hyperparameter search. Appendix A.1 documents the final configurations, which are fixed across selections within each task to isolate data effects. Given the cost of downstream training and repeated evaluation, we focus on two tasks and ten agents.

Revision: Section 4.2 and Table 1 report the expanded comparisons; Appendix D.1–D.2 give the protocols. Section 4.4 analyzes agent recipes, and Appendix A.1 lists the training configurations.

Response to Question 1:

Would agents select and rank better with limited feedback, such as a small proxy training run or a few validation results?

As mentioned in our response to Weakness 1, validation feedback gives modest average test gains, but early improvements on the feedback set are not consistently sustained and transfer only weakly to unseen evaluation data. Development-loss feedback gives no average improvement. These experiments assess selection, not changes in five-group ranking accuracy.

Revision: Section 4.5 and Figure 5 report iterative feedback and its transfer to the unseen evaluation set. Appendix D.4 and Table 12 report the paired Kimi development-loss feedback results; Appendix D.5 details the iterative protocol.

Does combining an agent's three runs give more stable or better selections?

As mentioned in our response to Weakness 1, combining Kimi’s three selections improves observed BFCL accuracy by 1.03%. Therefore, combining an agent's multiple runs is a potential way to improve the performance, although it bring additional cost on data selection.

Revision: Section 4.5 reports the consensus result, and Appendix D.3 describes aggregation.

Which of the agents' choices (e.g. how they use base-model loss) are linked to better outcomes?

As mentioned in our response to Weakness 1, no single choice consistently improves selection across tasks. Section 4.4 analyzes how agents apply different preferences and normalization schemes to similar signals. The baselines show the same lack of consistent improvement over Random: low loss matches Random on BFCL but falls below it on TBLite, while diversity shows the opposite pattern.

Revision: Section 4.4 analyzes agent choices; Section 4.2 and Table 1 report the baseline comparisons. Section 5 discusses possible improvements.

Response to Question 2:

How does DataSense-Bench differ from DataComp, DataComp-LM, DataPerf, and PostTrainBench? What does it reveal that these benchmarks cannot?

As mentioned in our response to Weakness 1, we isolate agents’ data sense by fixing the model and training recipe and evaluating both selection quality and ranking ability. This reveals whether agents can predict data value in addition to finding useful subsets. The results separate these capabilities: for example, Kimi reaches 36.29% average first-group BFCL accuracy but only 0.02 average rank correlation, while Opus reaches 0.77 rank correlation without identifying the best group in any of its three runs.

Our shared-pair value-prediction experiment directly separates ranking ability from selection: agents predict which of the same already-evaluated subsets will perform better, with downstream scores hidden. This isolates forecasting ability from group construction, revealing failures that selection performance alone does not expose.

All agents rank the same 64 BFCL pairs and 32 TBLite pairs, containing 50 and 1,000 trajectories per group, respectively. Measured scores and evaluation examples are hidden. Each agent completes three prediction runs with a budget of one H100 for up to five hours per run. The table reports average prediction accuracy:

Agent / controlBFCLTBLite
DeepSeek81.25%42.71%
Sonnet 578.65%48.96%
Fable 5.178.13%48.96%
GPT-6 Astra76.56%46.88%
Kimi K373.44%44.79%
GPT-5.6 Sol73.44%45.83%
Opus 571.88%47.92%
Fable 570.31%41.67%
GPT-5.567.19%45.83%
Gemini61.46%45.83%
Choose the group with fewer answer tokens as higher-scoring59.375%59.375%

Average prediction accuracy across agents is 73.23% on BFCL and 45.94% on TBLite, showing that reliable ranking data groups according to their post-training performance remains challenging. Please refer to Section 4.3 and Figure 4 for the results, and Appendix D.6 for the shared data construction.

Revision: Section 2, Related Work, now cites DataComp, DataComp-LM, and DataPerf and explains how our benchmark differs in evaluating both selection quality and ranking ability. Section 1 clarifies this motivation, and Section 4.2 discusses the cross-task baseline comparisons.

Reviewer_rPArOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce DataSense-Bench to test whether CLI agents can select useful SFT data and predict its post-training value. Agents inspect training trajectories and Qwen3-4B, then afterwards select and rank five groups without training models or accessing evaluation tasks. They then fine-tune on each group and evaluate on TBLite and BFCL. The authors end up finding that agents produce detailed selection methods but achieve only small gains over random selection and generally do not rank the groups reliably.

Critical strengths
  1. The benchmark cleanly separates data selection from post-training evaluation. Agents rank five data groups before observing any training results, which allows their predictions to be directly checked against downstream TBLite and BFCL performance.
  2. The paper evaluates multiple frontier agents across two tasks with three independent runs. This provides stronger evidence that unreliable ranking is a broader limitation rather than a failure of one agent or one selection attempt.
Critical weaknesses
  1. The random baseline uses only one fixed subset per task. Since gains are only 0.350.35–1.781.78 points on BFCL and −0.55-0.55–0.990.99 points on TBLite, it is unclear whether they are meaningful compared with variation across random subsets.
  2. All groups are trained with one training seed. Many group scores in Tables 7 and 8 are close, so the ordering and Spearman correlation may change under retraining. This limits how strongly the ranking results can be interpreted.
Questions for authors

Clarification Requests:

  1. How was the single random subset for each task chosen, and is there evidence that the reported gains are consistent across different random subsets?

  2. Are the within-run rankings of g1,…,g5g_1,\ldots,g_5 consistent when computed separately for each of the three evaluation seeds?

Suggestions for Future Work:

  1. Can the authors repeat a representative set of selected groups across multiple SFT seeds? This would help clarify whether the low ranking accuracy comes from the agents’ predictions or training variability.
5 comments

Thank you for your time and positive support! Your questions about random sampling and training and evaluation variability helped us clarify the statistical protocol and add robustness checks.

Response to Weakness 1:

The random baseline uses only one fixed subset per task. Since gains are only 0.35–1.78 points on BFCL and −0.55–0.99 points on TBLite, it is unclear whether they are meaningful compared with variation across random subsets.

We expanded the Random baseline to three independently sampled subsets and recomputed agent gains relative to their average.

Task / modelSample 1Sample 2Sample 3Average ± sample SD
BFCL / Qwen3-4B, accuracy (%)35.8834.6736.2135.58 ± 0.81
TBLite / Qwen3-4B, reward × 10013.545312.542012.556312.88 ± 0.58

Against the averaged Random baseline, BFCL agents’ average gains range from −0.82% to +1.35%. The strongest agent reaches 36.93%, compared with 35.58% for Random. On TBLite, all agents’ average g1 performance is below the Random baseline.

Revision: Section 4.2 and Table 1 use the revised Random reference; Appendix D.2 documents the sampling protocol.

Response to Weakness 2:

All groups are trained with one training seed. Many group scores in Tables 7 and 8 are close, so the ordering and Spearman correlation may change under retraining. This limits how strongly the ranking results can be interpreted.

Retraining changes some group orderings, but the agents’ first-choice errors persist in both examined cases.

Protocol. To balance robustness with training and evaluation cost, the main experiment uses one SFT seed. We retrain all five fixed groups from one small-gap BFCL run from each of two agents with two additional SFT seeds. Each checkpoint averages three evaluations. We hold the submitted order fixed and recompute ranking metrics from average group scores.

AgentTraining summaryg1g2g3g4g5S_{rank}A_{best} (%)
DeepSeekSingle seed35.2937.4235.3335.2933.580.560
DeepSeekAverage over three seeds35.2836.7636.1035.9935.210.400
GeminiSingle seed35.5033.0833.2134.7935.92−0.400
GeminiAverage over three seeds35.5133.4333.3636.1035.35−0.100

Scores are BFCL accuracy in percent. S_{rank} compares the submitted order with measured performance; A_{best} indicates whether g1 is best. First-group accuracy changes by less than 0.02% in both cases. Averaging training seeds preserves the best group for DeepSeek and shifts it from g5 to g4 for Gemini. Training randomness changes the exact ordering, but g1 remains below the best group and A_{best} stays zero in both cases.

Revision: Section 4.8 and Table 3 compare the single-seed and three-seed summaries.

Response to Question 1:

How was the single random subset for each task chosen, and is there evidence that the reported gains are consistent across different random subsets?

The original Random subset was sampled using a fixed random seed. As mentioned in our response to Weakness 1, we now use three independently sampled subsets. The strongest BFCL agent reaches 36.93% and exceeds each sampled Random subset; no TBLite agent exceeds all three sampled subsets. Thus, gains over Random are not consistent across agents and tasks.

Revision: Section 4.2 and Table 1 use the revised Random reference; Appendix D.2 documents the sampling protocol.

Response to Question 2:

Are the within-run rankings of g1, …, g5 consistent when computed separately for each of the three evaluation seeds?

The full rankings are not consistent across evaluations, although the best group can remain unchanged. Below, we report BFCL accuracy (%) for three evaluations of the same checkpoints from run 1 of DeepSeek, Kimi, and Astra.

AgentEvaluationg1g2g3g4g5
DeepSeek V4 Pro134.3838.1235.3836.3833.12
DeepSeek V4 Pro235.7537.5035.8835.1234.12
DeepSeek V4 Pro335.7536.6234.7534.3833.50
DeepSeek V4 ProAverage35.2937.4235.3335.2933.58
Kimi K3135.8836.8836.0033.7536.38
Kimi K3233.8835.3835.1234.2536.50
Kimi K3336.0036.0035.1235.6236.38
Kimi K3Average35.2536.0835.4234.5436.42
GPT-6 Astra135.7534.3832.6234.8832.62
GPT-6 Astra237.3835.5033.7536.0035.50
GPT-6 Astra336.0033.8833.2533.2535.38
GPT-6 AstraAverage36.3834.5833.2134.7134.50

The full ranking changes across evaluations for all three agents. DeepSeek consistently has g2 as its best group, and Astra consistently has g1, whereas Kimi’s best group changes from g2 to g5. To account for randomness in evaluation, we rank groups by their average scores across three evaluations.

Response to Suggestion 1:

Can the authors repeat a representative set of selected groups across multiple SFT seeds? This would help clarify whether the low ranking accuracy comes from the agents’ predictions or training variability.

As mentioned in our response to Weakness 2, we retrained the five fixed groups from one BFCL run per agent with two additional SFT seeds. Averaging across training seeds preserves DeepSeek’s best group and changes Gemini’s from g5 to g4, but A_{best} remains zero for both. Training variability therefore changes the exact ordering without resolving the first-choice errors in these cases.

Revision: Section 4.8 and Table 3 compare the single-seed and three-seed summaries.

Dear reviewer, thanks again for your time and effort. We hope we have addressed your concerns and would be grateful to hear your feedback before the discussion period ends.

Reviewer_mysUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DataSense-Bench evaluates whether agents can select useful training data and predict its downstream value without observing training outcomes. Agents inspect a candidate pool and an inference-only model, then submit five disjoint subsets ranked by expected performance. Each subset is subsequently used to fine-tune Qwen3-4B under a fixed recipe. Experiments cover terminal problem solving and multi-turn tool use, with three independent selections per agent. The authors report modest gains over random selection, inconsistent ranking accuracy, and substantial variation in agents’ selection procedures.

Critical strengths
  1. I found the separation between selection and subsequent training well motivated. Freezing the groups and their predicted order before providing any training feedback makes the forecast testable. Using a fixed training procedure and excluding evaluation instances also helps isolate the agents’ data-selection decisions from their ability to tune training or exploit benchmark feedback.

  2. I think the repeated selection runs and analysis of executed recipes provide useful detail beyond aggregate scores. They show that the same agent can change how it interprets loss, quality, and diversity across runs. Separately reporting top-ranked subset performance and ranking accuracy is also helpful, since choosing useful data and correctly comparing alternatives are different capabilities.

Critical weaknesses
  1. I found the ranking task difficult to interpret as a comparison of forecasting ability across agents. Each agent constructs the groups it subsequently ranks, so it also controls how distinguishable those groups are. An agent selecting several similarly useful groups may obtain a poor correlation, while another can make ordering easier by including deliberately weak groups. The current metric therefore mixes prediction quality with group construction. A shared set of candidate groups would provide a cleaner test of the forecasting claim.

  2. I would like stronger evidence that the small performance differences support reliable group orderings. Appendix A.1, Table 3 lists a fixed training seed of 42, while Appendix A.3 states that random selection uses one fixed subset per task. The repeated evaluation seeds measure inference variability, but do not establish how rankings vary across training seeds or how representative that random subset is. These limitations leave uncertainty about the strength of the selection gains and the interpretation of low ranking accuracy.

Questions for authors
  1. Would the conclusions hold if agents ranked a shared set of already-evaluated subsets, with their scores hidden? This could separate forecasting ability from differences in group construction.

  2. On representative groups with small performance gaps, how stable are the rankings across training seeds? Can you also evaluate several independently drawn random subsets and report uncertainty in gains over random selection?

  3. How do straightforward selectors based on completion quality, model loss, or diversity perform under the same constraints? These comparisons would help establish the difficulty and achievable benefit of informed selection.

6 comments

Thank you for your time and effort in reviewing our paper. Your comments on separating forecasting from group construction and assessing robustness helped us strengthen the experiments.

Response to Weakness 1:

I found the ranking task difficult to interpret as a comparison of forecasting ability across agents. Each agent constructs the groups it subsequently ranks, so it also controls how distinguishable those groups are. An agent selecting several similarly useful groups may obtain a poor correlation, while another can make ordering easier by including deliberately weak groups. The current metric therefore mixes prediction quality with group construction. A shared set of candidate groups would provide a cleaner test of the forecasting claim.

To isolate ranking ability, we now prompt every agent to rank the same shared groups.

All agents evaluate the same 64 BFCL pairs and 32 TBLite pairs. Each group contains 50 trajectories for BFCL or 1,000 for TBLite. Pairs have an observed performance gap of at least 2% in BFCL accuracy or 0.02 in TBLite reward. Agents receive anonymized group contents and the fixed training recipe, while actual post-training performance is hidden. Each prediction run has a budget of one H100 GPU for up to five hours. Each agent completes three prediction runs per task; the table reports average prediction accuracy.

Agent / controlBFCLTBLite
DeepSeek81.25%42.71%
Sonnet 578.65%48.96%
Fable 5.178.13%48.96%
GPT-6 Astra76.56%46.88%
Kimi K373.44%44.79%
GPT-5.6 Sol73.44%45.83%
Opus 571.88%47.92%
Fable 570.31%41.67%
GPT-5.567.19%45.83%
Gemini61.46%45.83%
Choose the group with fewer answer tokens as higher-scoring59.375%59.375%

On TBLite, every agent’s average prediction accuracy is below 50%, showing that ranking remains challenging with fixed data groups.

Revision: Section 4.3 and Figure 4 report the experiment and results; Appendix D.6 explains pair construction and isolation, and Table 13 reports the assistant-content token distributions of the shared groups.


Response to Weakness 2:

I would like stronger evidence that the small performance differences support reliable group orderings. Appendix A.1, Table 3 lists a fixed training seed of 42, while Appendix A.3 states that random selection uses one fixed subset per task. The repeated evaluation seeds measure inference variability, but do not establish how rankings vary across training seeds or how representative that random subset is. These limitations leave uncertainty about the strength of the selection gains and the interpretation of low ranking accuracy.

Thanks for your comments. We update the random-selection performance by averaging performance on independently sampled Random subsets and investigate how group rankings change under multiple training seeds.

Random selection.

We sample three independent subsets using sampling seeds 101–103, with SFT seed 42 fixed in the main results.

Task / modelSample 1Sample 2Sample 3Average ± sample SD
BFCL / Qwen3-4B, accuracy (%)35.8834.6736.2135.58 ± 0.81
TBLite / Qwen3-4B, reward × 10013.5512.5412.5612.88 ± 0.58

Using these averaged baselines, agent gains remain limited on BFCL, and all agents’ average first-group performance is below Random on TBLite.

Revision: Section 4.2 and Table 1 use the revised Random reference.

Training seeds.

The main experiment uses one SFT seed to balance robustness with training and evaluation cost. To assess training-seed sensitivity, we retrain all five fixed groups from one small-gap BFCL run from each of two agents with two additional SFT seeds. Each checkpoint averages three evaluations. We hold the submitted order fixed and recompute ranking metrics from average group scores.

g1g2g3g4g5S_{rank}A_{best} (%)
DeepSeekSingle seed35.2937.4235.3335.2933.580.560
DeepSeekAverage over 3 seeds35.2836.7636.1035.9935.210.400
GeminiSingle seed35.5033.0833.2134.7935.92−0.400
GeminiAverage over 3 seeds35.5133.4333.3636.1035.35−0.100

Scores are BFCL accuracy in percent. S_{rank} compares the submitted order with measured performance; A_{best} indicates whether g1 is best. First-group accuracy changes by less than 0.02% in both cases and remains below the best group. The best group is unchanged for DeepSeek and shifts from g5 to g4 for Gemini. Training randomness affects the exact ordering, but averaging does not recover a correct first choice in either examined case.

Revision: Section 4.8 and Table 3 compare the single-seed and three-seed summaries.

Response to Question 1:

Would the conclusions hold if agents ranked a shared set of already-evaluated subsets, with their scores hidden? This could separate forecasting ability from differences in group construction.

Thanks for your suggestion! As mentioned in our response to Weakness 1, ranking difficulties persist after controlling for group construction. Average prediction accuracy is 73.23% on BFCL and 45.94% on TBLite; every agent is below 50% and the shorter-group control on TBLite.

Revision: Section 4.3 and Figure 4 report the experiment and results; Appendix D.6 explains pair construction and isolation, and Table 13 reports the assistant-content token distributions of the shared groups.

Response to Question 2:

On representative groups with small performance gaps, how stable are the rankings across training seeds? Can you also evaluate several independently drawn random subsets and report uncertainty in gains over random selection?

As mentioned in our response to Weakness 2, across three training seeds, DeepSeek’s best group remains g2, while Gemini’s changes from g5 to g4. Neither agent’s first-ranked group becomes best. Retraining affects the ordering but does not resolve these first-choice errors.

The updated Random baselines reported in our response to Weakness 2 average 35.58 ± 0.81% BFCL accuracy and 12.88 ± 0.58 TBLite reward × 100 across three independent subsets. Relative to these baselines, BFCL agents’ average gains range from −0.82% to +1.35%, and all agents’ average g1 performance is below Random on TBLite.

Revision: Section 4.8 and Table 3 report training-seed sensitivity. Section 4.2 and Table 1 report the updated Random baseline and gains; Appendix D.2 details the sampling protocol.

Response to Question 3:

How do straightforward selectors based on completion quality, model loss, or diversity perform under the same constraints? These comparisons would help establish the difficulty and achievable benefit of informed selection.

We added baseline comparisons using the same group sizes and downstream protocols as the agents.

Low- and high-loss selection rank assistant-target negative log-likelihood; diversity uses cosine farthest-first selection of average assistant hidden states. We also evaluate normalized-message deduplication, gradient-based LESS, and TBLite completion flags.

MethodBFCL 4B accuracy (%)TBLite reward × 100
GPT-5.6 Sol36.9311.26
DeepSeek V4 Pro36.5312.54
GPT-6 Astra36.4911.95
Kimi K336.2911.94
Opus 535.9411.60
Random35.5812.88
Low loss35.5812.33
Fable 5.135.2112.50
Gemini 3.5 Flash35.1012.07
Fable 535.0111.00
Sonnet 534.9611.55
GPT-5.534.7611.87
Diversity32.1313.32
High loss30.797.91
LESS29.5812.79
Completion flag—10.99

Results average three independently sampled subsets for Random and three independent selection runs for agent g1. As shown in the table, low loss matches Random on BFCL 4B, while diversity has the highest average on TBLite. The strongest agent average on BFCL 4B is 36.93%, exceeding the best reported selector average by 1.35%. On TBLite, the strongest agent average of 12.54 is below both Random and diversity. Agents outperform the tested selectors on BFCL in some cases, but not on TBLite; none of the tested single-criterion selectors consistently improves on Random across tasks.

Additionally, agents combine these simple signals to produce their own data selection recipes. For example, Kimi and DeepSeek use loss and diversity constraints, while Opus and Astra incorporate quality or execution-evidence checks, as detailed in the appendix.

Revision: Appendix D.1–D.2 define the protocols and uncertainty, and Section 4.4 analyzes how agents use these signals. Please refer to Section 4.2 and Table 1 for the full experimental results.

Dear Reviewer, thank you again for your time and effort in reviewing our work. We hope our responses have addressed your concerns. As the discussion deadline approaches, we would greatly appreciate your participation in the discussion. If further clarification is needed, we would be happy to discuss.

Comments on the paper

2 comments

Dear Area Chair,

Thank you for your efforts to ensure a fair and thorough review process. To help inform your decision, below we summarize our contribution, the additional evidence provided during rebuttal, and the discussion status.

Our contribution

DataSense-Bench examines agents’ data sense through two complementary abilities: selecting useful training data and predicting its downstream value. By fixing the model and training recipe, we isolate the agent's data sense from the post-training process, investigating this important capability for AI scientists who can autonomously conduct research.

Reviewer concerns and additional evidence

We addressed the main concerns through the following experiments:

  • Missing baselines: We added more baselines, including loss-based selection, diversity, deduplication, LESS, and completion flags. No single selector consistently improves across tasks, which highlights the need for agent frameworks with strong data sense that adapt selection strategies to task and data characteristics. This also distinguishes our work from DataComp, DataComp-LM, DataPerf, and PostTrainBench, mentioned by reviewer 2N5S.
  • Limited coverage: We added experiments on BCFL: four other agents with Qwen3-4B and six agents sith Qwen3.5-2B. Five agents exceed Random on each model, but only Astra identifies the best group in all three runs on 4B, and none does on 2B. This shows that the frontier agents are far from the perfect data sense.
  • Random and training variability: To address concerns from reviewers rPAr and 2N5S, we added three independent Random subsets for each task and multi-seed training experiments for two agents per task. Both single-seed and multi-seed results show that the group ranked first by each agent does not achieve the highest downstream performance among its five groups. We will clarify the trade-off between robustness and training/evaluation cost behind the original single-seed setup, and report average performance across multiple training seeds for all agents in the next revision.
  • Confounded ranking evaluation: Reviewer mysU noted that agents construct the groups they rank, making prediction accuracy depend on how distinguishable those groups are. We therefore asked all ten agents to predict which group performs better in the same 64 BFCL and 32 TBLite pairs. Agents receive anonymized group contents and the fixed training recipe, with post-training scores hidden. Pairs have minimum observed gaps of 2% BFCL accuracy or 0.02 TBLite reward, and we report the average prediction accuracy over three runs. Accuracy is 61.46–81.25% on BFCL but below 50% for every agent on TBLite, exposing weak prediction capability of the frontier agents.
  • Investigation into ways to improve selection: Following Reviewer 2N5S’s suggestion, we investigated consensus and limited feedback on BFCL/Qwen3-4B. Combining Kimi’s selections across three runs raises accuracy from 36.29% to 37.32%, at additional agent and training cost. For iterative feedback, Kimi and Astra receive rewards from 98 validation tasks over up to five rounds. Performance improves on this validation set in early rounds, but the gains are not consistently sustained and generalize only weakly to unseen BFCL800 tasks: average test gains are 0.74% and 0.14%, respectively. A separate matched development-loss feedback control yields no average improvement. This result strengthens the importance of our work: the field needs such a data benchmark to develop better agents with stronger data sense.

We have incorporated the additional results and clarifications into the revised manuscript and will report the multi-seed average performance of all agents, and more results of Qwen3.5-2B on TBLite.

Discussion status

  • Reviewer 2N5S acknowledged that the additional experiments addressed most concerns and raised the score to 2. We answered 2N5S’s follow-up questions about pair construction, validation and tuning data, and prior benchmarks.
  • Reviewer mysU initially gave a score of 2 and acknowledged that the shared-pair test addresses the group-construction concern and that the added baselines improve interpretability, but raised a further concern about TBLite pair rankings across training seeds. We provided multi-seed results for two TBLite agents but have not yet received a reply to this latest response.
  • Reviewer rPAr gave the rating of 3 and confirmed that we addressed the raised concerns about Random sampling, evaluation-seed ranking consistency, and training variability. We addressed these through reporting the average performance of three independent Random subsets, comparisons across evaluation seeds, and additional multi-seed training experiments.

Thank you for your time and consideration again. We kindly request that the final assessment consider the expanded experiments and follow-up responses alongside the initial reviews.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.