Machine unlearning induces progressively stronger novice-level performance in an LLM teachable agent, with accuracy and F1 declining as the unlearning ratio increases¶

Claim · All claims
Evidence · 1 study (2 entries) · 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? The unlearned model shows a clear, consistent decline in MCQ accuracy and F1 as unlearning ratio rises from 10% to 50%, indicating knowledge loss rather than noise. → Jiajia Song 2026 q2 i? The base pretrained model remains stable across all unlearning ratios, so the degradation is attributable to the unlearning procedure. → 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? · design · r2

Numerical evaluation shown in Fig. 4 across five random seeds compares the unlearned Mistral-7B-Instruct model against the base model under 10%–50% unlearning ratios; "a clear and consistent decline in both accuracy and F1 score" reflects targeted knowledge loss.

"The unlearned model (green line) exhibits a clear and consistent decline in both accuracy and F1 score as the unlearning ratio increases, dropping from around 0.75 at 10% to below 0.5 from the 40% unlearning ratio onward."

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? · causal · r1

Fig. 4 numerical results show the unmodified base model "maintains a consistently high accuracy (around 0.85)" across ratios, indicating the degradation is attributable to unlearning rather than randomness or evaluation variance.

"the base model (blue line), defined as the original pretrained model without any further fine-tuning or modification for a specific task, maintains a consistently high accuracy (around 0.85) and F1 scores (around 0.77) across all unlearning ratios."

Discussion¶

Related Claims¶

• Coach-guided interactive teaching yields higher and more progressive recovery than static relearning, particularly under moderate forgetting levels (10%–30%) — related
• Teachable agents measurably but incompletely relearn unlearned knowledge, with the recovery gap to the base model widening under higher unlearning ratios (30%–50%) — related
• 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 — 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