StressEval - Text Reasoning Stress: leaderboard

StressEval (2026) builds Dynamic-OneEval by synthesizing new knowledge-intensive questions from GPT-5.2 failure cases on HotpotQA, ComplexWebQuestions and Wikipedia table question sets, keeping the difficulty factor that caused each failure. This board is exact-match accuracy in percent on text questions whose evidence is all in a synthesized context but whose reasoning step is stressed; higher is better.

Metric: Exact-match accuracy (%) on the text reasoning-stress split (388 instances seeded from HotpotQA, all evidence in a synthesized virtual context) of Dynamic-OneEval (StressEval instances synthesized by GPT-5.2 from its own failure cases on HotpotQA, ComplexWebQuestions and WikiTableQuestions, gated for answerability and consistency), one unified prompt template for every model; higher is better. Source: arxiv.org. Saturation forecast: Around 2034. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro27.6#145
2Claude Sonnet 4.523.7#261
3QwQ-32B23.7#907
4Qwen 3 235B A22B21.6#628
5Qwen 2.5 72B21.6#1188
6Llama 3.1 70B21.1#1499
7Llama 3.1 8B20.1#5196
8Seed-1.615.7#469
9DeepSeek V3.213.9#419

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

Interactive version: theaggregate.ai/benchmark?slug=stresseval-text-reasoning-stress · How It Works · Data refreshed daily, snapshot 2026-11-21.