Unverified50% confidenceBenchmarkExact time
A 2024 study from UC Berkeley found that in complex multi-step tasks, GPT-4-level models exhibited tool-call hallucinations at a rate as high as 12%–18%.
1
Sources
50%
Confidence
Long-term
Relevance
9/10/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedEven top-tier models like GPT-4 can have hallucination rates of 15%–30% in specific domains.80% similarUnverifiedStanford University's AgentBench benchmark test in 2024 showed that GPT-4-level models have an end-to-end success rate below 30% on complex tasks requiring cross-system coordination.77% similarUnverifiedIn the leap from GPT-3's 175 billion parameters to GPT-4's estimated over one trillion parameters, there was a qualitative leap in models' performance at simulating human psychological states.74% similarUnverifiedA 2023 Stanford University study found that GPT-4's accuracy on open-domain question-answering tasks was approximately 40 percentage points higher than that of traditional NLU systems.74% similarVerifiedA 2023 study by Chen et al. at Stanford University confirmed that GPT-4 exhibited significant fluctuations in mathematical reasoning ability over the course of several months.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/887239API
curl https://kongchang.com/api/v1/knowledge/claims/887239MCP
get_claim(id=887239)