{"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=Llama-3&bench=GSM8K","json_url":"https://assignee.net/benchmarks/evidence.json?model=Llama-3&bench=GSM8K","model":"Llama-3","benchmark":"GSM8K","source_count":8,"source_coverage":{"record_count":8,"distinct_source_count":8,"coverage_level":"BROAD","basis":"distinct public paper URLs or titles in this evidence cluster"},"source_profile":{"source_url_count":8,"missing_source_url_count":0,"domains":["arxiv.org"],"year_min":2023,"year_max":2026,"basis":"public source URLs, source titles, and reported publication years in this evidence cluster"},"reported_range":{"min_score_pct":0.0,"max_score_pct":95.8},"spread_pp":95.8,"severity":"HIGH","entries":[{"model":"Llama-3","benchmark":"GSM8K","score_pct":95.8,"source_title":"Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs","source_url":"http://arxiv.org/abs/2406.18629v1","source_domain":"arxiv.org","year":2024},{"model":"Llama-3","benchmark":"GSM8K","score_pct":95.4,"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":"Llama-3","benchmark":"GSM8K","score_pct":74.83,"source_title":"Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling","source_url":"http://arxiv.org/abs/2602.10623v2","source_domain":"arxiv.org","year":2026},{"model":"Llama-3","benchmark":"GSM8K","score_pct":62.5,"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":"Llama-3","benchmark":"GSM8K","score_pct":60.7,"source_title":"SED-SFT: Selectively Encouraging Diversity in Supervised Fine-Tuning","source_url":"http://arxiv.org/abs/2602.07464v1","source_domain":"arxiv.org","year":2026},{"model":"Llama-3","benchmark":"GSM8K","score_pct":54.33,"source_title":"Understanding Reasoning in Chain-of-Thought from the Hopfieldian View","source_url":"http://arxiv.org/abs/2410.03595v1","source_domain":"arxiv.org","year":2024},{"model":"Llama-3","benchmark":"GSM8K","score_pct":41.72,"source_title":"Preventing Rank Collapse in Federated Low-Rank Adaptation with Client Heterogeneity","source_url":"http://arxiv.org/abs/2602.13486v2","source_domain":"arxiv.org","year":2026},{"model":"Llama-3","benchmark":"GSM8K","score_pct":0.0,"source_title":"MR-GSM8K: A Meta-Reasoning Benchmark for Large Language Model Evaluation","source_url":"http://arxiv.org/abs/2312.17080v4","source_domain":"arxiv.org","year":2023}],"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."]}