BabyReasoningBench: Generating Developmentally-Inspired Reasoning Tasks for Evaluating Baby Language Models

BabyReasoningBench is a benchmark suite of developmentally inspired reasoning tasks designed to evaluate baby language models.

Tasks Covered

Task
Baron Cohen Sally Anne False Belief
Beck Counterfactual Possibilities
Bryant Trabasso Transitive Inference
Bullock Gelman Baillargeon Causal Reasoning
Chen Klahr Control Of Variables
Dias Harris Counterfactual Syllogism
Gelman Markman Category Induction
Gentner Toupin Story Analogy
Gopnik Causal Learning Bayes Nets
Gopnik Sobel Blicket Detector
Goswami Brown Causal Analogy
Harris German Mills Counterfactual Causal
Holyoak Junn Billman Analogical Problem Solving
Markman Seibert Class Inclusion
McGarrigle Donaldson Conservation Accidents
Onishi Baillargeon Violation Of Expectation
Schulz Bonawitz Exploratory Play
Wellman Cross Watson False Belief Battery
Wimmer Perner False Belief Unexpected Transfer

Model Performance of Baby LMs

Citation

If you use BabyReasoningBench, please cite:

@misc{dhole2026babyreasoningbenchgeneratingdevelopmentallyinspiredreasoning,
      title={BabyReasoningBench: Generating Developmentally-Inspired Reasoning Tasks for Evaluating Baby Language Models}, 
      author={Kaustubh D. Dhole},
      year={2026},
      eprint={2601.18933},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.18933}, 
}

For expanding these tasks, or adding more questions, or making them multimodal, reach out, and we can do something together :)