The multi-stage filtering pipeline reduces Beyond-K-5 retention to near zero while preserving a minority of K-5 content¶

Claim · All claims
Evidence · 1 study (2 entries) · causal 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? On the held-out CommonCoreText ground-truth set, the pipeline reduces Beyond-K-5 retention to 0% while preserving about 35% of K-5 documents. → Fanfei Li 2026 q2 i? On the external WeeBit corpus the filter retains 2.48% of 6,000 Beyond-K-5 passages, with genuinely out-of-scope concepts in only three passages (0.05%). → 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? · design · r2

Validation of the filtering pipeline on CommonCoreText, a held-out ground-truth dataset of recommended reading materials and textbooks with grade-level labels. The pipeline achieves "the retention ofBeyond-K–5content to 0%" while preserving ~35% of K-5, a deliberate precision-over-recall design; removed K-5 passages were systematically longer and lexically more complex.

"The pipeline successfully reduces the retention ofBeyond-K–5content to 0% while preserving ~35% ofK–5(Figure 2, left)."

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

Independent leakage check applying the same pipeline to WeeBit, an external grade-labeled corpus. Retention of Beyond-K-5 passages was 2.48% of 6,000, and manual inspection attributed most retained content to web scraping artifacts, with genuinely out-of-scope concepts in only three passages (0.05%). A corpus-wide scan of 126 Beyond-K-5 n-grams found matches in only 0.09% of retained passages.

"The filter retains 2.48% of 6,000Beyond-K–5passages; manual inspection identifies genuinely out-of-scope concepts in only three passages, corresponding to 0.05% of theBeyond-K–5split"

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 — related
• LittleLearner's capabilities tightly track its controlled exposure: strong on in-scope tasks, degrading beyond the K-5 boundary — related
• Across scaling, post-training, and in-context learning, the pretraining filter rather than the intervention sets LittleLearner's capability ceiling — related
• Multi-layer defense significantly reduces bypass relative to regex-only filtering (69.65% to 46.34%), confirmed by exact McNemar test — related