The authors formulate the Silent-error Resilience Task: an interactive agentic task in which a plausible but incorrect tool result enters an agent's execution history without an explicit failure signal.
Silent tool errors roughly halve success of e-commerce LLM agents
New benchmark RITE injects plausible but wrong tool outputs into fixed workflow prefixes, revealing recognition, refetch and integration as successive bottlenecks.
Chinese Tech
Wenshuo Zhang · Yuquan Lin · Yushi Sun · Yishi Lin · Rui Sheng · Peng He · +3 more
Hong Kong University of Science and Technology · Tencent
Research Digest··2 min read
The authors introduce RITE, a benchmark of 167 matched units from nine e-commerce workflows, to test whether LLM agents can recover when a tool returns a plausible but incorrect value without an explicit failure signal.
§