Across scaling, post-training, and in-context learning, the pretraining filter rather than the intervention sets LittleLearner's capability ceiling¶

Claim · All claims
Evidence · 1 study · causal r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-19 (automated, 1 of 1 entry) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Across all three tested interventions, each lever amplifies capabilities within K-5 exposure but provides limited gains in Beyond-K-5, so the pretraining filter sets the effective capability ceiling in the tested settings. → Fanfei Li 2026

Evidence¶

Fanfei Li 2026¶

Fanfei Li, Jana Zeller, Manuel Prada-Corral, Thaddäus Wiedemer, Prasanna Mayilvahanan, Ryan Cotterell, Wieland Brendel. (2026). LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure. Preprint. https://arxiv.org/abs/2608.13545 doi:10.48550/arxiv.2608.13545

q2 · i? · causal · r2

Cross-intervention synthesis of the scaling (Figure 7), GRPO post-training (Figure 8), and ICL (Figure 9) experiments: each lever amplifies K-5 capability, scaling yields modest boundary gains, ICL does not unlock Beyond-K-5 reasoning, and post-training does not transfer even when trained on unrestricted data.

"In each case, it is the pretraining filter, rather than the intervention, that sets the effective capability ceiling in our tested settings."

Discussion¶

Related Claims¶

• GRPO post-training boosts in-scope K-5 capabilities but fails to recover Beyond-K-5 capabilities, even when post-trained on out-of-scope data — a narrower finding that bears on this claim
• In-context learning with few-shot examples and explanations does not unlock Beyond-K-5 reasoning for the 5B LittleLearner — a narrower finding that bears on this claim
• LittleLearner's capabilities tightly track its controlled exposure: strong on in-scope tasks, degrading beyond the K-5 boundary — a narrower finding that bears on this claim
• Scaling model size improves performance within K-5 and at the boundary but does not recover Beyond-K-5 performance — a narrower finding that bears on this claim
• The multi-stage filtering pipeline reduces Beyond-K-5 retention to near zero while preserving a minority of K-5 content — related