# Reward Hacking Diagnostics: Gaming Detection, Length Bias, Sycophancy Probe

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-reward-hacking-diagnostics
> Updated: 2026-08-13T09:02:10.578Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part XI — Alignment & Preference Optimization
**TLDR:** Modeller reward function'ı 'hack' eder — yanlış yoldan reward kazanır. Length bias (uzun cevaplar = yüksek reward), sycophancy (kullanıcıya aşırı agreeable), format gaming (chat template yapısını kötüye kullanma), repetition. Tespit pipeline: ablation, holdout probe, qualitative review. Anthropic'in 'reward over-optimization' raporundan dersler.