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¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 2 of 2 entries) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? After SFT followed by GRPO, both LittleLearner and UNFILTERED improve within K-5, but in Beyond-K-5 post-training only modestly improves LittleLearner while producing considerably larger gains for UNFILTERED. → Fanfei Li 2026 q2 i? Within the tested post-training budgets, LittleLearner post-trained on K-5 data shows no difference from LittleLearner post-trained on Beyond-K-5 data. → Fanfei Li 2026 (2)

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

Post-training experiment using a two-stage SFT followed by GRPO pipeline, evaluated on MathCAMPS (Figure 8, 95% CIs). LittleLearner was SFT'd on K-5 data selected by the filtering pipeline and UNFILTERED on a compute-matched unfiltered version; GRPO data was ablated between K-5 and a compute-matched unfiltered counterpart.

"WithinK–5, post-training lifts both models above their base performance with slightly larger gains for UNFILTERED . InBeyond-K–5, post-training only modestly improves LITTLELEARNER while producing considerably larger gains for UNFILTERED , widening the observed gap between the two models."

Fanfei Li 2026 (2)¶

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

q2 · i? · causal · r2

GRPO data ablation within the same post-training experiment: LittleLearner post-trained on K-5-filtered questions versus a compute-matched unfiltered counterpart showed no difference within the tested budgets, indicating GRPO alone does not bridge the out-of-scope gap.

"Crucially, within the tested post-training budgets, we observe no difference between LITTLELEARNER post-trained onK–5data versusBeyond-K–5data."

Discussion¶

Related Claims¶

• Across scaling, post-training, and in-context learning, the pretraining filter rather than the intervention sets LittleLearner's capability ceiling — a broader claim this one bears on
• Scaling model size improves performance within K-5 and at the boundary but does not recover Beyond-K-5 performance — related
• LittleLearner's capabilities tightly track its controlled exposure: strong on in-scope tasks, degrading beyond the K-5 boundary — related
• The multi-stage filtering pipeline reduces Beyond-K-5 retention to near zero while preserving a minority of K-5 content — related
• In-context learning with few-shot examples and explanations does not unlock Beyond-K-5 reasoning for the 5B LittleLearner — related