Supervised fine-tuning on SocraticEdu provides the largest single-step SE gain (+0.200) but still leaks target terms in 26.7% of dialogues¶

Claim · All claims
Evidence · 1 study · causal r1 · 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? SFT alone improves SE from 0.300 to 0.500 over the prompted baseline, yet leaks target terms in 26.7% of dialogues. → Xiaokun Wang 2026

Evidence¶

Xiaokun Wang 2026¶

Xiaokun Wang, Siyu Song, Wentao Liu, Xiaodong Zou. (2026). Beyond Direct Answering: Aligning Educational LLMs as Socratic Guides via Heuristic Reinforcement Learning. arXiv preprint. https://arxiv.org/abs/2607.22996 doi:10.48550/arxiv.2607.22996

q2 · i? · causal · r1

Training-trajectory analysis of the PT track (RQ1 discussion, Table 3 and Figure 2). The article reports "SFT provides the largest single-step gain inSE(+0.200 over PT-Base)" while leakage remains at 26.7%.

"SFT provides the largest single-step gain inSE(+0.200 over PT-Base) by instilling basic questioning patterns from SocraticEdu, but it still leaks target terms in 26.7% of dialogues."

Discussion¶

Related Claims¶

• GRPO alignment without the directness reward raises Socratic scaffolding effectiveness from 0.300 to 0.633 and cuts keyword leakage from 26.7% to 13.3% on held-out questions — related