Misguided Attention: leaderboard

Modified logic puzzles test whether models follow the actual question instead of recalling a familiar answer. Native task means retain their evaluation counts. A published out-of-range task score and the dependent overall metric remain unresolved; valid independent tasks are kept.

Source: cpldcpu.github.io.

Benchmark results

Results are published separately for each benchmark or evaluation setting.

Interactive version: theaggregate.ai/benchmark?slug=misguided-attention · How It Works · Data refreshed daily, snapshot 2026-11-17.