{ "_note": "Numbers copied from the cited papers' own tables (read 2026-11-20). They are context, not same-condition comparisons: models, dates and (for PlanBench) instance subsets differ from ours.", "ERQA": { "source": "Gemini Robotics: Bringing AI into the Physical World, arXiv:2503.20020, Tables 1-2 (results obtained Feb 2025); Cosmos rows from the nvidia/Cosmos-Reason2-2B model card", "n_questions": 400, "no_cot": {"Gemini 1.5 Flash": 42.3, "Gemini 1.5 Pro": 41.8, "Gemini 2.0 Flash": 46.3, "Gemini 2.0 Pro Experimental": 48.3, "GPT-4o-mini": 37.3, "GPT-4o": 47.0, "Claude 3.5 Sonnet": 35.5}, "with_cot": {"Gemini 2.0 Flash": 50.3, "Gemini 2.0 Pro Experimental": 54.8, "GPT-4o-mini": 40.5, "GPT-4o": 50.5, "Claude 3.5 Sonnet": 45.8}, "model_card": {"Cosmos-Reason2-32B": 45.25, "Qwen3-VL-32B-Instruct": 46.50, "Cosmos-Reason2-8B": 44.00, "Qwen3-VL-8B-Instruct": 44.00, "Cosmos-Reason2-2B": 37.75, "Qwen3-VL-2B-Instruct": 37.75} }, "PlanBench_Blocksworld": { "source": "LLMs Still Can't Plan; Can LRMs? arXiv:2409.13373, Tables 1-2 (600 instances, natural-language prompts)", "one_shot": {"Claude 3.5 Sonnet": 57.6, "Claude 3 Opus": 48.1, "GPT-4o": 28.3, "GPT-4o-mini": 8.1, "GPT-4": 34.3, "GPT-4 Turbo": 23.0, "LLaMA 3.1 405B": 47.3, "LLaMA 3 70B": 12.6, "Gemini 1.5 Pro": 16.8, "Gemini 1 Pro": 11.3}, "zero_shot": {"Claude 3.5 Sonnet": 54.8, "Claude 3 Opus": 59.3, "GPT-4o": 35.5, "GPT-4o-mini": 8.8, "GPT-4": 34.6, "GPT-4 Turbo": 40.1, "LLaMA 3.1 405B": 62.6, "LLaMA 3 70B": 34.16, "Gemini 1.5 Pro": 23.8, "Gemini 1 Pro": 0.5, "o1-preview": 97.8, "o1-mini": 56.6, "Fast Downward (classical planner)": 100.0}, "llm_modulo": {"text": "with back prompting from VAL, LLM performance in Blocks World improves to 82% within 15 back prompting rounds", "source": "arXiv:2402.01817 Sec. 4, citing Table 4 of arXiv:2305.15771", "value": 82.0, "max_rounds": 15} } }