Rewarded SFT produces the most stable outputs among the SFT designs, nearly free of hallucinations and story-irrelevant material¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Stories from Rewarded SFT were almost entirely free of hallucinations, garbled text, story-irrelevant outputs, or null values, while other experiments had under 10% such outputs. → Qian Shen 2026

Evidence¶

Qian Shen 2026¶

Qian Shen, Fanghua Cao, Min Yao, Shlok Gilda, Bonnie J. Dorr, Walter L. Leite. (2026). Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs with Controllable Difficulty and Safety. https://arxiv.org/abs/2605.13709 doi:10.48550/arxiv.2605.13709

q2 · i? · design · r2

Observational stability analysis in the Results section of the 1,290 stories generated per model per experiment; Rewarded SFT outputs were "almost entirely free of hallucinations", and other experiments stayed below 10% undesirable outputs.

"the stories generated by LLMs fine-tuned with Rewarded SFT exhibited high stability, being almost entirely free of hallucinations, garbled text, story-irrelevant outputs, or null values. In the other experiments, while the stories from fine-tuned LLMs did contain a small portion of these undesirable outputs, the proportion was less than 10% of the 1,290 stories generated per model in each experiment."

Discussion¶

Related Claims¶

• Toxicity of stories generated by the fine-tuned 8B LLMs is very close to zero, with only a minimal number of stories posing potential safety risk — related
• Fine-tuned 8B LLMs generate K-2 stories with difficulty-related metric scores statistically significantly different from (mostly better than) zero-shot GPT-4o and Llama 3.3 70B — related
• Among the three 8B LLMs, Llama 3 underperforms Apertus and Granite 3.3 on most evaluation metrics — related
• SFT with simulated children's reading errors yields higher LM-PPL than other experiments, driven by brevity and post-story irrelevant content — related