LLM Benchmarks — Scores, Methodology & Live Rankings
What each benchmark measures, how models score, and the current rankings
CodeForces (with tools)
STEM & Reasoning
CodeForces (with tools) - competitive programming rating evaluated with terminal tool access similar to Codex CLI.
BoolQ
Reasoning
BoolQ is a QA task where each example comprises a short passage and a yes/no question. 0-shot.
PIQA
Reasoning
Physical Interaction QA (PIQA) - commonsense reasoning about physical world. 0-shot.
SocialIQA
Reasoning
Social IQA - commonsense reasoning about social situations. 0-shot.
Natural Questions
Knowledge
Google Natural Questions - open-domain QA using real user queries from Google Search. 5-shot.
ARC-c
Reasoning
AI2 Reasoning Challenge (Challenge set) - grade-school science questions. 25-shot.
ARC-e
Reasoning
AI2 Reasoning Challenge (Easy set) - grade-school science questions. 0-shot.
WinoGrande
Reasoning
WinoGrande - large-scale commonsense reasoning with Winograd schema-like problems. 5-shot.
AGIEval
Reasoning
AGIEval - human-level standardized tests (SAT, LSAT, GRE, etc.). 3-5-shot.
HumanEval
Code
HumanEval - Python code generation from function signatures and docstrings. 0-shot.
IndicGenBench
Multilingual
IndicGenBench - multilingual generation benchmark for Indic languages.
DocVQA
Multimodal
Document Visual Question Answering - understanding text in document images. (val split)
InfoVQA
Multimodal
Infographic VQA - answering questions about infographics. (val split)
TextVQA
Multimodal
TextVQA - visual question answering requiring reading text in images. (val split)
SpatialSense VQA
Multimodal
SpatialSense VQA - spatial reasoning in visual question answering.
OSWorld 2.0 (Partial)
General Agents
OSWorld 2.0 - Partial score. The partial score aggregates the partial rewards obtained across all tasks.
Agents' Last Exam (Score)
General Agents
Agents' Last Exam - Score (aggregate). Reports the aggregate score rather than Pass@1.
ExploitGym (2h)
Cybersecurity
ExploitGym exploitation benchmark, single-run Pass@1 on 869 tasks under a 2-hour timeout budget (API inference time rescaled by per-model TPS from Artificial...
ExploitGym (6h)
Cybersecurity
ExploitGym exploitation benchmark, single-run Pass@1 on 869 tasks under a 6-hour timeout budget (API inference time rescaled by per-model TPS from Artificial...
ExploitBench
Cybersecurity
Exploitation benchmark: average coverage score over 41 tasks across 3 revisions (union of capabilities across revisions per task), max 300 interaction rounds...
Z.ai Code Bench
Agentic Coding
Z.AI's in-house coding benchmark for the GLM series. GLM-5.3 shows a 50% relative improvement over GLM-5.2; absolute scores are not published on the model ca...
Harbor-Index
Agentic Coding
Agentic coding index benchmark reported in the Tencent Hy4 preview model card benchmark appendix.
Hy-Backend 2.0 (Internal)
Agentic Coding
Tencent internal (Hy team) backend engineering benchmark, reported in the Hy4 preview model card appendix.
Hy-SWE Max Verified (Internal)
Agentic Coding
Tencent internal 300-task agentic coding benchmark; Claude Code scaffold, 200-turn budget, 16 CPU/32 GB sandbox, no web access, avg of 3 runs.
Hy-CompanyBench V2 (Internal)
General Agents
Tencent internal (Hy team) agentic search / company-work benchmark, reported in the Hy4 preview model card appendix.
Hy-LifeSearch (Internal)
General Agents
Tencent internal (Hy team) agentic search benchmark, reported in the Hy4 preview model card appendix.
Hy-BrowseComp-Pro2 (Internal)
General Agents
Tencent internal (Hy team) hard web-browsing benchmark, reported in the Hy4 preview model card appendix.
E-Bench (Internal)
General Agents
Tencent internal (Hy team) office/enterprise-work benchmark, reported in the Hy4 preview model card appendix.
E-Bench-Code (Internal)
Agentic Coding
Tencent internal (Hy team) enterprise coding benchmark, reported in the Hy4 preview model card appendix.