What changed after K–5?

Two experiments, three model scales, and the saved outputs behind each result.

Explore transfer visualizationsRead the full experiment
0.6B · 1.3B · 5B
16 adapters
26,880 records
Offline replay

Useful order changed with scale

Selected transfer edges

Control-adjusted transfer across model scaleFour mathematical transfer relationships measured as target-NLL gain over a token-matched K–5 review adapter.-0.4-0.20.00.20.40.60.6B1.3B5BTarget-NLL gainModel scaleNegative numbers → linear equations, 0.6B: +0.635Negative numbers → linear equations, 1.3B: +0.556Negative numbers → linear equations, 5B: +0.185Negative numbers → ratios, 0.6B: +0.250Negative numbers → ratios, 1.3B: +0.183Negative numbers → ratios, 5B: +0.139Exponents → linear equations, 0.6B: +0.176Exponents → linear equations, 1.3B: +0.283Exponents → linear equations, 5B: −0.145Linear equations → exponents, 0.6B: −0.029Linear equations → exponents, 1.3B: −0.328Linear equations → exponents, 5B: +0.059

One item across scale

Held-out item

Solve for x: 3x + 16 = 37.

Correct answer: 7
Exponent lesson
3x + 16 = 37
Prediction 3Target NLL 2.144
K–5 review control
37
Prediction 37Target NLL 2.262

+0.118 target-NLL gain. More correct-answer likelihood than K–5 review.

After the textbook closed

One composition problem

Held-out composition item

An item costs 8000 dollars. It is discounted by 25 percent, then 20 percent tax is added. What is the final price?

Correct answer: 7200
SFT, closed book
7200 dollars
Question: A man is 20 years
Prediction 7200Item target NLL 0.547Exact

Answer-supervised LoRA, lesson removed.

Four teaching routes

Percentage-skill acquisition routesAggregate target negative log-likelihood or exact accuracy for the constrained 0.6B model after different teaching routes.0.00.51.01.52.0Target NLL (lower is better)Base, closed book: 1.8351.835BaseBase, lesson included: 1.8561.856+ lessonSFT, closed book: 1.2511.251SFTCPT, closed book: 1.8041.804CPT