{"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=Qwen3&bench=GSM8K","json_url":"https://assignee.net/benchmarks/evidence.json?model=Qwen3&bench=GSM8K","model":"Qwen3","benchmark":"GSM8K","source_count":3,"source_coverage":{"record_count":3,"distinct_source_count":3,"coverage_level":"MODERATE","basis":"distinct public paper URLs or titles in this evidence cluster"},"source_profile":{"source_url_count":3,"missing_source_url_count":0,"domains":["arxiv.org"],"year_min":2026,"year_max":2026,"basis":"public source URLs, source titles, and reported publication years in this evidence cluster"},"reported_range":{"min_score_pct":79.9,"max_score_pct":88.5},"spread_pp":8.6,"severity":"MEDIUM","entries":[{"model":"Qwen3","benchmark":"GSM8K","score_pct":88.5,"source_title":"DiffCoT: Diffusion-styled Chain-of-Thought Reasoning in LLMs","source_url":"http://arxiv.org/abs/2601.03559v2","source_domain":"arxiv.org","year":2026},{"model":"Qwen3","benchmark":"GSM8K","score_pct":86.5,"source_title":"Do Instruction-Tuned Models Always Perform Better Than Base Models? Evidence from Math and Domain-Shifted Benchmarks","source_url":"http://arxiv.org/abs/2601.13244v1","source_domain":"arxiv.org","year":2026},{"model":"Qwen3","benchmark":"GSM8K","score_pct":79.91,"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}],"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."]}