Similarity Gates Approve Reversals: A Validity Audit of Embedding-Cosine Thresholds in Agent Systems
Source
arxiv.org
Author
Scott E. Frias
Date
Context
A high text-similarity score can conceal a reversed instruction. This audit reports that one deployed meaning-drift check missed all 56 meaning-breaking edits tested against it. The problem is that changing an instruction's meaning may require very little wording change, while preserving meaning through a paraphrase may change many words. Research abstract.
The findings challenge treating a fixed similarity threshold as evidence of equivalent meaning. The author also reports that two tested embeddingA list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.Full definition → configurations distinguished reversals from paraphrases when wording overlap was matched, showing why an evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → must separate those two properties.
Terms in this piece · Glossary
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
articleRecursive Self-Improvement for Agents: 7-Check Persistence GateAlphaSignalAI
articleUniversal Pathologies, Conditional Consequences: A Triple-Robustness Analysis of RAG for Multi-Hop TraceabilityMeftun Akarsu, Burak Ozdemir
articleCodeGrep: An RL-Trained Retrieval Agent for LLM Coding AgentsWuya Chen, Yihao yang, Yang Cao, Yue Lin