Stability vs. Manipulability: Evaluating Robustness Under Post-Decision Interaction in LLM Judges
Quick Answer
This study reveals that LLM judges, such as those evaluated with MT-Bench and AlpacaEval, exhibit high stability under neutral reevaluation but can be significantly influenced by targeted post-decision interactions, leading to altered outcomes and potential misalignment with human preferences.
Quick Take
The introduction of the Evaluation Robustness Score (ERS) aims to quantify this interactional robustness, highlighting the need for improved evaluation protocols.
Key Points
- judges show high stability during neutral reevaluation but can be reversed by targeted challenges.
- Post-decision interaction can degrade agreement with human preferences and alter benchmark rankings.
- Authority framing destabilizes judgments, leading to low-overlap justifications and potential rationalizations.
- The Evaluation Robustness Score (ERS) quantifies interactional robustness in LLM evaluations.
- Findings suggest a distinct failure mode for LLM-as-judge evaluations requiring new protocols.
Paper Resources
Source Excerpt
arXiv:2606. 05384v1 Announce Type: new Abstract: -as-judge evaluation is widely used in benchmarking pipelines, where model outputs are compared and ranked using automated evaluators. These pipelines typically assume that judgments are stable properties of fixed inputs. We show that this assumption does not hold under interaction. We study post-decision manipulability: the extent to which an evaluation outcome can be altered through subsequent conversation with the judge after an initial decision has been made.
…
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →RAIL Guard: Closing the Evaluation-to-Remediation Gap in Responsible AI for Agents
RAIL Guard introduces a closed-loop AI pipeline for large language models (LLMs) that evaluates outputs across eight dimensions and iteratively remediates failures, achieving 96.9% convergence compared to 49.1% for traditional block-and-retry methods. The system reduces unsafe agent executions by 33% without impacting task completion and is available as open-source SDKs.