In-context learning with few-shot examples and explanations does not unlock Beyond-K-5 reasoning for the 5B LittleLearner¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r1–r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Few-shot prompting with three manually authored chain-of-thought examples per CCSS standard yields modest K-5 gains but no gain at all in Beyond-K-5 on MathCAMPS. → Fanfei Li 2026 q2 i? Adding explanations instead of few-shot examples changes neither model responses nor performance. → 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

ICL experiment on MathCAMPS synthesizing three new problems per CCSS standard with manually authored chain-of-thought solutions as few-shot examples (Figure 9, 95% CIs). Few shots steered output style, e.g. shorter responses, but this steering did not improve reasoning; hand-written examples performed best among presentation ablations.

"InK–5using a few shots shows modest gains, and inBeyond-K–5there is no gain at all."

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 · r1

Explanation-prompting condition of the same ICL experiment: providing explanations instead of few-shot examples produced no change in model responses and no performance change on MathCAMPS.

"When we add explanations instead of few shots, we do not observe the same behavior: Model responses do not change depending on the given explanation, and so does not performance."

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
• Curricular chain-of-thought prompting improves competency-classification accuracy over zero-shot, with gains concentrated in larger models, while definition-based prompting does not improve performance — related
• LittleLearner's capabilities tightly track its controlled exposure: strong on in-scope tasks, degrading beyond the K-5 boundary — related
• 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 — related