Séminaire - DIC-ISC-CRIA - 1er octobre 2026 par Michael C. FRANK

Value of the Information & Strength of the Argument

The talk offers high value by framing a fundamental question in cognitive science: why children learn language so efficiently compared to AI models. Frank’s argument is well-structured, moving from data resources to empirical experiments and then to theoretical implications. He carefully distinguishes between engineering goals (making models learn more with less) and scientific goals (using models to understand human learning). The argument is supported by concrete examples, such as the comparison of GPT-2 outputs trained on child language versus a child’s dialogue, and by null results from curricularization experiments. The value lies in the methodological paradigm: aligning model training and evaluation with developmental data to isolate factors like innate structure, embodiment, and multimodal input. Frank acknowledges limitations and avoids overclaiming, making the argument scientifically sound.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates strong scientific rigor. Frank cites peer-reviewed work (e.g., Trends in Cognitive Sciences, Annual Review of Psychology) and references open datasets (WordBank, Baby View, Childes). The methodology for training and evaluating models is described in sufficient detail, and the results are presented with appropriate caveats. The title accurately reflects the content, focusing on data efficiency in children and language models. The sources cited in the description are directly relevant and include arXiv and PsyArXiv preprints. The talk also engages with the BabyLM challenge, showing awareness of related work. Overall, the sources are credible and the argumentation is transparent.

The title accurately reflects the content: the seminar focuses on data efficiency in children versus language models, covering data resources, training experiments, and developmental alignment.

8/10

The talk is given by a leading expert (Michael C. Frank, Stanford) and is grounded in peer-reviewed work and open datasets. Claims are presented with appropriate caveats, and the methodology is transparent. The main limitation is the lack of detailed quantitative results in the transcript, but the overall scientific rigor is high.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction of the data gap: children learn from ~10 million words per year, while models like GPT-3 use 500 billion words.
• 05:00Explanation of using AI models as scientific models of cognition, with strategies like controlled rearing and causal probing.
• 10:00Overview of data resources: WordBank and the Baby View project for egocentric video.
• 15:00Description of WordBank: 100,000+ CDI administrations across 42 languages, showing variability in early vocabulary.
• 20:00Introduction of the Baby View camera and dataset: 3,400 hours of egocentric video from 50+ families.
• 25:00Ethical considerations for recording children, including parental consent and privacy controls.
• 30:00Training GPT-2 on child language data: example of incoherent output compared to a child's dialogue.
• 35:00Experiment: training GPT-2 on child-directed speech vs. Wikipedia/BabyLM, showing lower performance on grammar and word similarity.
• 40:00Curricularization experiments: training in developmental order yields no benefit, consistent with BabyLM challenge results.

• Children, but not language models, show accelerating returns in word learning — Cited as a key paper by Frank on the data gap and word learning efficiency.
• Bridging the data gap between children and large language models — Preprint discussing the data gap and the need for alignment between child and model learning.

• BabyLM Challenge — Results from the BabyLM challenge align with Frank's findings that curricularization does not benefit language models.
• WordBank — Provides the CDI data used in the talk to study early vocabulary development.