Misguided Attention - poisoned_hot: 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
1Llama 3.3 70B Instruct1#1387
2GPT-4.1 Mini1#793
3GPT-4.1 Nano1#2043
4DeepSeek R11#487
5Gemma 3 27B1#1671
6Claude 3.5 Sonnet1#756
7Grok 31#606
8QwQ-32B1#908
9DeepSeek R1 05281#431
10Mistral Small 31#1799
11Gemma 3n E4B (IT)1#2432
12Command A1#1309
13Grok 3 Mini1#546
14OLMo-2-0325-32B-Instruct1#2825
15Phi-4-reasoning-plus1#2166

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-poisoned-hot · How It Works · Data refreshed daily, snapshot 2026-11-17.