Benchmarks can be superficial, but model explanations and evaluations are fundamentally intertwined. What if we used interpretability as principled, scientific evaluation? If it met scientific standards? arxiv.org/abs/2605.05508 coming to EvalEval at ACL as oral 🧵 1/6
AI Weekly's analysis
→
- The paper argues interpretability methods that are falsifiable, reproducible, and predictive can serve as model evaluation, not just diagnostics.
- Of four methods assessed in Table 1, attention mechanisms fail all three criteria; sparse autoencoders fail reproducibility.
- An SAE refusal-detection feature trained on chat data failed to generalize when the target model received webtext input instead.
Read full analysis →
