Unverified50% confidenceFactExact time
Many classic benchmarks have become saturated (e.g., multiple models scoring above 95% on HellaSwag), and the industry is developing new benchmarks with greater discriminative power, such as GPQA, SWE-bench, and LiveBench.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/3/2026
First Seen
Valid until: 12/2/2026
Sources
Related Entities
Related Claims
UnverifiedThe per-step accuracy of current top-tier models is approximately between 90% and 95%.61% similarUnverified主流开源模型在消费级GPU上的推理速度通常在20-80 token/s之间61% similarUnverifiedAfter MMLU's discriminative power declined, the community shifted toward more difficult benchmarks such as GPQA (a graduate-level professional question set), SWE-bench, MATH-500, and LiveBench.61% similarUnverifiedTraditional benchmarks such as MMLU and HumanEval have experienced benchmark saturation due to training data contamination; when GPT-4 was released, its MMLU score was approximately 86%, whereas multiple open-source models have now surpassed 90%59% similarUnverifiedThrough a combination of optimization techniques such as Hybrid Search, Reranking, and HyDE, the accuracy of RAG systems can be improved from the ~60% level of a basic implementation to above 90%.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/848832API
curl https://kongchang.com/api/v1/knowledge/claims/848832MCP
get_claim(id=848832)