Condesion-Bench - Variable Conditions: leaderboard

Condesion-Bench (2026) evaluates conditional decision-making in a compositional action space: for one trading day an LLM allocates quantities across 15 S&P 500 stocks under variable, contextual and allocation conditions, and its single zero-shot action is checked against the conditions. This board is the satisfaction rate in percent of variable conditions, which restrict the decision variables that may be selected; higher is better.

Metric: Satisfaction rate (%) of variable conditions (which decision variables may be selected), of the single action an LLM generates zero-shot (temperature 0, JSON output) on the 251 Condesion-Bench instances (one per trading day, October 2024 to September 2025: allocate a budget over 15 S&P 500 stocks under variable, contextual and allocation conditions); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5 Mini (2025-08-07)94.02#345
2O3 (2025-04-16)92.83#231
3GPT-592.43#178
4GPT-4.1 Mini91.63#789
5O4 Mini (2025-04-16)90.84#370
6Gemini 2.0 Flash (001)90.84#727
7Claude 3.5 Sonnet (20241022)90.44#648
8GPT-OSS-120B90.04#704
9Claude 3.5 Haiku90.04#1580
10GPT-4.189.64#505
11GPT-OSS-20B89.64#1183
12Gemini 2.0 Flash Lite (001)88.45#1118
13Llama 3.3 70B Instruct88.05#1385
14Mistral Large 2 (Jul)76.89#1144
15Mistral Small 3.274.5#1641

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

Interactive version: theaggregate.ai/benchmark?slug=condesion-bench-variable-conditions · How It Works · Data refreshed daily, snapshot 2026-11-21.