aiqbench analysis
This research identifies the "Agentic Formalism Trap," demonstrating how LLM-as-a-Judge systems conflate procedural structure with semantic truth when subjected to adversarial social loads. The study quantifies this vulnerability across multiple domains and proposes architecture-specific filters to mitigate instability in closed-loop evaluation pipelines. These findings are critical for developers relying on automated evaluators, highlighting the need for vigilance against consensus mimicry that obscures actual model performance.
The Formalism Trap: Are LLM-as-a-Judge Evaluators Blinded by Consensus Mimicry under Social Load?
arXiv:2607.28641v1 Announce Type: new Abstract: We introduce the \textit{Agentic Formalism Trap} and the Evaluative Dissonance Index ($D_E$), quantifying how LLM-as-a-Judge systems conflate structural proceduralism with semantic truth under adversarial load. Analyzing 22,500 trajectories across 3 domains (GAIA, SWE-bench, Multi-Challenge), we extract a semantic taxonomy of hallucination maneuvers, validated via deterministic lexical grounding ($p < 10^{-120}$). A logistic meta-evaluator isolates the exact syntactic triggers of this evaluator capture (ROC-AUC 0.8779), while a zero-shot Leave-One-Domain-Out transfer proves the vulnerability is universally domain-agnostic (mean ROC-AUC 0.7482). Architectural profiling reveals that distinct simulated swarm topologies induce mathematically disparate semantic blind spots, proving that unanchored closed-loop evaluation is unstable, systemically divergent and necessitates architecture-specific vigilance filters.
Title:The Formalism Trap: Are LLM-as-a-Judge Evaluators Blinded by Consensus Mimicry under Social Load?
Abstract:We introduce the \textit{Agentic Formalism Trap} and the Evaluative Dissonance Index ($D_E$), quantifying how LLM-as-a-Judge systems conflate structural proceduralism with semantic truth under adversarial load. Analyzing 22,500 trajectories across 3 domains (GAIA, SWE-bench, Multi-Challenge), we extract a semantic taxonomy of hallucination maneuvers, validated via deterministic lexical grounding ($p < 10^{-120}$). A logistic meta-evaluator isolates the exact syntactic triggers of this evaluator capture (ROC-AUC 0.8779), while a zero-shot Leave-One-Domain-Out transfer proves the vulnerability is universally domain-agnostic (mean ROC-AUC 0.7482). Architectural profiling reveals that distinct simulated swarm topologies induce mathematically disparate semantic blind spots, proving that unanchored closed-loop evaluation is unstable, systemically divergent and necessitates architecture-specific vigilance filters.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2607.28641 [cs.CL]
(or arXiv:2607.28641v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2607.28641 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 215 words
