MCPHallu: Benchmarking Reasoning, Execution, and Memory Hallucinations in MCP Agents
Yizhen Jiang, Wentao Zhang and colleagues at Peking University and Meituan build MCPHallu, a benchmark whose tasks are designed so that one specific hallucination type is the main failure risk for an MCP agent.
Ask this paper
Benchmark. 358 tasks across five domains run on 36 production MCP servers in isolated containers. Each task carries a structural trigger for Branch Collapse, Unreachable Goal, Tool Misuse or Context Forgetting.
Scale. Fourteen frontier agents from eight vendors are evaluated on 5,012 trajectories, with an average score of 0.642 out of 1.
Infeasible goals. Unreachable Goal is the weakest subtype: average score 0.458 and a 23.2% pass rate. Agents keep calling tools or claim success instead of reporting that the goal cannot be met.
Rank instability. A model's rank moves by 6.9 of 13 positions on average across subtypes, and Branch Collapse and Unreachable Goal scores are uncorrelated (r = -0.12).
Implication. A single aggregate MCP score hides which failure a given model is prone to, so subtype scores are needed for model selection.
Abstract
Large language model agents are increasingly deployed through the Model Context Protocol (MCP), which allows them to discover and invoke tools across multiple servers at runtime. Existing MCP benchmarks primarily measure task completion; where they diagnose failures, they do so after execution rather than building specific reasoning, execution, and memory failure risks into task construction. We introduce MCPHallu, a benchmark for diagnosing hallucination-prone agent failures in MCP environments along three axes: reasoning, execution, and memory. MCPHallu contains 358 tasks across five capability domains and runs on 36 production MCP servers in isolated containers. Each task is designed with a structural trigger that makes one of four subtypes the dominant failure risk: Branch Collapse, Unreachable Goal, Tool Misuse, or Context Forgetting. This design enables subtype-level analysis rather than aggregate task success alone. We evaluate fourteen frontier LLM agents from eight vendors on 5{,}012 trajectories. Current MCP agents remain vulnerable to hallucination-inducing tool environments, achieving an average score of 0.642 on a $[0,1]$ scale (higher is better). Failures concentrate on Unreachable Goal tasks, where the average score drops to 0.458 and 23.2\% of runs pass, indicating that agents often continue issuing tool calls or claim successful completion rather than recognizing infeasible goals. Subtype rankings diverge sharply: a model's rank shifts by 6.9 of 13 positions on average across subtypes, and Branch Collapse is uncorrelated with Unreachable Goal across models ($r = -0.12$). MCPHallu is released at https://github.com/J1zz/mcp_hallu.