Tag
This arXiv paper shows how 'warped formatters' — tool output text that omits key details like search limits or result caps — mislead AI agents and their verifiers even when the underlying computation is correct. Repaired formatters drastically cut unsupported claims and flipped verifier acceptance from 90/90 wrong to 90/90 correct, and an audit prompt rediscovered live defects in the authors' own system plus OpenClaw and Hermes, prompting merged fixes.
The article explains how untrusted tool output in AI systems can be exploited for prompt injection attacks, emphasizing the need to treat certain fields as untrusted even within internal systems.
This paper introduces a unified benchmark for span-level hallucination detection in RAG systems that extends beyond natural language to code, tool output, and structured documents, and presents a fine-tuned Qwen3.5-2B detector that outperforms existing methods on these new domains while remaining competitive on standard NLP benchmarks.