Misguided Attention - birthday_problem_easy: leaderboard

Each task is one independently graded modified logic puzzle with native expected-answer criteria. Zero-shot; source's per-task mean over available valid evaluations, not a completion rate over nominal attempts. Native num_evaluations is the only supported denominator. Models.js and query_config files supply registry settings; no per-row execution timestamp or request log match is claimed. Evaluation code weights criterion satisfaction and clips each run to0–1 before summarizing; do not apply new clipping to published data. Current evaluator names Gemini2.5Flash at temperature0/max1000, while older README names Llama3.3; exact judge version for each historic run is unlogged and must stay unknown. Public prompts are available, including reversibly scrambled source files. Metric: Mean native expected-behavior score (0–1). Source: https://cpldcpu.github.io/MisguidedAttention/

Metric: Mean native expected-behavior score (0–1). Source: cpldcpu.github.io. Saturation forecast: Estimated already saturated. 53 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-4o1#754
2GPT-4.1 Mini1#793
3GPT-4.11#505
4Gemini 2.0 Flash1#723
5Llama 4 Maverick1#1118
6Claude Sonnet 41#411
7Gemini 2.5 Pro (Preview 03-25)1#239
8Claude 3.7 Sonnet1#476
9Grok 31#606
10DeepSeek V3 (0324)1#722
11Llama 4 Scout1#1819
12QwQ-32B1#908
13DeepSeek R1 Distill Llama 70B1#1669
14Grok 3 Mini Beta1#620
15O3 Mini1#547

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

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