Coach-guided interactive teaching yields higher and more progressive recovery than static relearning, particularly under moderate forgetting levels (10%–30%)¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r1 · q2
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 2 of 2 entries) · not reviewed by a person

Subclaims¶

q2 i? In a three-agent teachable-agent setting, coach-guided dialogue produces consistent and progressive accuracy gains, especially under moderate unlearning. → Jiajia Song 2026 q2 i? Relearn (LLM Coach) achieves consistently higher F1 than static relearning, indicating more balanced and coherent knowledge recovery. → Jiajia Song 2026 (2)

Evidence¶

Jiajia Song 2026¶

Jiajia Song, Zhihan Guo, and Jionghao Lin. (2026). Simulating Novice Students Using Machine Unlearning and Relearning in Large Language Models. https://arxiv.org/abs/2603.26142 doi:10.48550/arxiv.2603.26142

q2 · i? · causal · r1

In the Stage 3 three-agent environment (teachable, judge, and DeepSeek-V3.2 coach agents), the unlearned model as teachable agent shows "consistent and progressive gains in accuracy" across unlearning ratios versus static relearning, per Fig. 4(a).

"Compared with static relearning, coach-guided interac- tion yields consistent and progressive gains in accuracy across unlearning ratios,"

Jiajia Song 2026 (2)¶

Jiajia Song, Zhihan Guo, and Jionghao Lin. (2026). Simulating Novice Students Using Machine Unlearning and Relearning in Large Language Models. https://arxiv.org/abs/2603.26142

q2 · i? · design · r2

Fig. 4(b) F1 results show "Relearn (LLM Coach) achieves consistently higher F1 scores than static relearning", suggesting instructional feedback helps reconstruct missing conceptual links rather than memorizing answers.

"Relearn (LLM Coach) achieves consistently higher F1 scores than static relearning, highlighting the advantage of interactive teaching in promoting more balanced and coherent knowledge recovery."

Discussion¶

Related Claims¶

• Contingent scaffolding improves learning more than fixed or absent support. — the general claim this finding bears on, with the studies behind it
• Relearning dynamics depend on unlearning strength: agents with 10% unlearning recover within early interaction rounds, while 30% and 50% agents show prolonged low-accuracy plateaus and delayed improvement — a narrower finding that bears on this claim
• Teachable agents measurably but incompletely relearn unlearned knowledge, with the recovery gap to the base model widening under higher unlearning ratios (30%–50%) — related
• Machine unlearning induces progressively stronger novice-level performance in an LLM teachable agent, with accuracy and F1 declining as the unlearning ratio increases — related
• Dialogue evidence shows coach feedback induces genuine conceptual change, shifting the agent from misconception to conceptually correct explanation rather than superficial answer alignment — related
• Teaching interactions typically conclude within 30 dialogue turns, motivating a 30-turn cap to prevent dead loops — related