The authors evaluated five existing mitigation methods (Relign, Gorilla, PALADIN, LinSteer, PRISMS) on six benchmarks with varying tool configurations, measuring hallucination under both tuned and other configurations.
LLM tool-use mitigations can worsen hallucination in new configurations; EscapeGuard prevents this
The authors show that existing methods reinforce intrinsic tool-use tendencies, causing hallucination escape, and present a training-free inference-time method that reduces selection hallucination by 9 pp while suppressing escape.
Top University
Peigui Qi · Kunsheng Tang · Yide Song · Weiming Zhang · Nenghai Yu
University of Science and Technology of China · University of Washington
Research Digest··3 min read
Qi et al.
Why this paper
From University of Washington and University of Science and Technology of China
In one line
Hallucination escape causes tool-use mitigation methods to reduce errors in one configuration while increasing them in others.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§