{"schema":"https://assignee.net/schemas/benchmark-evidence-v1","schema_version":"1.0","contract_version":"benchmark-evidence-v1.0","contract_updated":"2026-06-01","schema_documentation":"https://assignee.net/schemas","changelog_url":"https://assignee.net/changelog","publisher":{"name":"Assignee Research","url":"https://assignee.net"},"html_url":"https://assignee.net/benchmarks/evidence?model=Qwen2.5&bench=GSM8K","json_url":"https://assignee.net/benchmarks/evidence.json?model=Qwen2.5&bench=GSM8K","model":"Qwen2.5","benchmark":"GSM8K","source_count":7,"source_coverage":{"record_count":7,"distinct_source_count":7,"coverage_level":"BROAD","basis":"distinct public paper URLs or titles in this evidence cluster"},"source_profile":{"source_url_count":7,"missing_source_url_count":0,"domains":["arxiv.org"],"year_min":2024,"year_max":2026,"basis":"public source URLs, source titles, and reported publication years in this evidence cluster"},"reported_range":{"min_score_pct":24.3,"max_score_pct":95.0},"spread_pp":70.7,"severity":"HIGH","entries":[{"model":"Qwen2.5","benchmark":"GSM8K","score_pct":95.0,"source_title":"Do LLMs Overthink Basic Math Reasoning? Benchmarking the Accuracy-Efficiency Tradeoff in Language Models","source_url":"http://arxiv.org/abs/2507.04023v3","source_domain":"arxiv.org","year":2025},{"model":"Qwen2.5","benchmark":"GSM8K","score_pct":91.5,"source_title":"Qwen2.5 Technical Report","source_url":"http://arxiv.org/abs/2412.15115v2","source_domain":"arxiv.org","year":2024},{"model":"Qwen2.5","benchmark":"GSM8K","score_pct":69.9,"source_title":"SwS: Self-aware Weakness-driven Problem Synthesis in Reinforcement Learning for LLM Reasoning","source_url":"http://arxiv.org/abs/2506.08989v1","source_domain":"arxiv.org","year":2025},{"model":"Qwen2.5","benchmark":"GSM8K","score_pct":47.5,"source_title":"LLaDA-MoE: A Sparse MoE Diffusion Language Model","source_url":"http://arxiv.org/abs/2509.24389v1","source_domain":"arxiv.org","year":2025},{"model":"Qwen2.5","benchmark":"GSM8K","score_pct":40.0,"source_title":"TokenSkip: Controllable Chain-of-Thought Compression in LLMs","source_url":"http://arxiv.org/abs/2502.12067v3","source_domain":"arxiv.org","year":2025},{"model":"Qwen2.5","benchmark":"GSM8K","score_pct":37.68,"source_title":"Task-Specific Efficiency Analysis: When Small Language Models Outperform Large Language Models","source_url":"http://arxiv.org/abs/2603.21389v1","source_domain":"arxiv.org","year":2026},{"model":"Qwen2.5","benchmark":"GSM8K","score_pct":24.26,"source_title":"GSM8K-V: Can Vision Language Models Solve Grade School Math Word Problems in Visual Contexts","source_url":"http://arxiv.org/abs/2509.25160v1","source_domain":"arxiv.org","year":2025}],"interpretation":"This record groups score claims extracted from papers for the same model and benchmark label. A nonzero spread means the public literature reports different values for this cluster.","limitations":["Differences are not automatically errors.","Reported values may differ because of prompts, dataset versions, evaluation protocols, scoring rules, preprocessing, fine-tuning, or reporting conventions.","Source papers remain authoritative for their own claims."]}