Unverified50% confidenceBenchmarkExact time
自我反思（self-reflection）由Shinn等人在2023年的Reflexion论文（NeurIPS 2023）中系统化，在HotpotQA、HumanEval等基准上相比单次生成基线提升了10%~20%的准确率
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
Verification Loops Quadruple DeepSeek's Intelligence at One-Seventh the Cost of Claude Opus
hackernewshackernews7/7/2026
Related Claims
UnverifiedReflexion框架在HotpotQA、HumanEval等基准上相比单次生成基线提升了10%~20%的准确率78% similarVerifiedResearch shows that even the most advanced models have hallucination rates in factual question-answering tasks that still fluctuate between 5% and 20%.68% similarVerifiedThrough experiments on multi-hop reasoning benchmarks such as HotpotQA and FEVER, the ReAct paper demonstrated that compared to pure Chain-of-Thought reasoning, ReAct improved accuracy by approximately 10%–30% on tasks requiring external information retrieval.67% similarUnverified2023年MIT与Google的联合研究表明，当子任务存在明确完成标准时，智能体的自我纠错效率可提升40%以上66% similarUnverifiedThe Self-Consistency method, proposed by Xuezhi Wang and others, improved accuracy by more than 10% on mathematical reasoning benchmarks such as GSM8K.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/288497API
curl https://kongchang.com/api/v1/knowledge/claims/288497MCP
get_claim(id=288497)