arXiv:2605.27559v1 Announce Type: cross 
Abstract: Multi-stage LLM pipelines that perform multi-agent debate, intrinsic self-correction, or retrieval-augmented verification exhibit puzzling aggregate behaviors: accuracy plateaus and reversals across rounds, non-replication of debate gains on contemporary frontier models, intrinsic self-correction degradation, and qualitative cross-provider divergence in debate dynamics. Downstream agent response can be operationalized as two coupled decisions: detection (whether to treat upstream content as authoritative) and conditional generation (what to produce if not). This decomposition yields four observable response regimes, of which detection-without-correction is the load-bearing failure mode. Across a nine-cell empirical grid spanning four model families, four benchmarks (GSM8K, MATH-500, GPQA-Diamond, AIME), and two methods (multi-agent debate, intrinsic self-correction), we find that the conditional miscorrection rate is consistently dominant (53-94% across cohorts) while detection rate varies contextually by more than an order of magnitude. The framework unifies the four phenomena above as signatures of a common mechanism and characterizes detection threshold as a stable model/protocol-level regularity that persists across methods at matched benchmark difficulty.

قدمت دراسة حديثة إطار عمل لتفكيك ذو معلمين لتحليل خطوط أنابيب نماذج اللغة الكبيرة (LLM) متعددة المراحل، كاشفة عن سلوكيات معقدة مثل استقرار الدقة والانقلابات في الأداء عبر نماذج ومعايير مختلفة. يقوم هذا الإطار بتشغيل استجابات الوكلاء في الأسفل إلى اكتشاف وتوليد مشروط، مما يبرز وضع فشل كبير يسمى الاكتشاف بدون تصحيح.

Un estudio reciente ha introducido un marco de descomposición de dos parámetros para analizar los pipelines de modelos de lenguaje de gran tamaño (LLM) en múltiples etapas, revelando comportamientos complejos como mesetas de precisión e inversiones en el rendimiento a través de varios modelos y benchmarks. Este marco operacionaliza las respuestas de los agentes en la parte inferior en detección y generación condicional, destacando un modo de falla significativo denominado detección-sin-corrección.

Une étude récente a introduit un cadre de décomposition à deux paramètres pour analyser les pipelines de modèles de langage de grande taille (LLM) en plusieurs étapes, révélant des comportements complexes tels que des plateaux et des inversions de précision dans les performances à travers divers modèles et benchmarks. Ce cadre opérationnalise les réponses des agents en aval en détection et génération conditionnelle, mettant en lumière un mode de défaillance significatif appelé détection-sans-correction.

A recent study has introduced a two-parameter decomposition framework for analyzing multi-stage large language model (LLM) pipelines, revealing complex behaviors such as accuracy plateaus and reversals in performance across various models and benchmarks. This framework operationalizes downstream agent responses into detection and conditional generation, highlighting a significant failure mode termed detection-without-correction.