Calibration reduces evaluator preference coupling in LLM agent feedback loops

A study investigates whether probability calibration of evaluator judgments can mitigate "evaluator preference coupling," where systematic biases propagate into an LLM agent's strategy. The authors present the first analysis of this mitigation approach, applying calibration to pairwise judgments to reduce spurious preference propagation.

• In a controlled experiment using DeepSeek-V4-Pro as executor and GLM5.2 as evaluator, confidence-calibrated TTRL reduced the coupling coefficient gamma by 20-49% compared to standard binary updates.
• Jensen-Shannon divergence was reduced by 45-67% with the calibrated method.
• A symmetric-LR control confirmed that these improvements were not due to reduced update asymmetry.

The authors release the calibrated TTRL protocol and recommend it as a lightweight mitigation for LLM-as-judge deployment pipelines.