# Does Self-Evaluation Enable Wireheading in Language Models?

arXiv: https://arxiv.org/abs/2511.23092 · HTML: https://mlsanity.com/paper/2511.23092/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Reward modeling and RLHF, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 9 of its 21 references that are in the map.
- First posted on arXiv 28 Nov 2025.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Reward modeling and RLHF](https://mlsanity.com/alignment/reward-modeling-and-rlhf/)

## Builds on (most cited first, 9 of 9)

- [Deep reinforcement learning from human preferences](https://mlsanity.com/paper/1706.03741/) (2017)
- [Decoupled Weight Decay Regularization](https://mlsanity.com/paper/1711.05101/) (2017)
- [Helping or Herding? Reward Model Ensembles Mitigate but do not Eliminate Reward Hacking](https://mlsanity.com/paper/2312.09244/) (2023)
- [Large Language Models Often Know When They Are Being Evaluated](https://mlsanity.com/paper/2505.23836/) (2025)
- [Reward Shaping to Mitigate Reward Hacking in RLHF](https://mlsanity.com/paper/2502.18770/) (2025)
- [InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling](https://mlsanity.com/paper/2402.09345/) (2024)
- [Demonstrating specification gaming in reasoning models](https://mlsanity.com/paper/2502.13295/) (2025)
- [Avoiding Wireheading with Value Reinforcement Learning](https://mlsanity.com/paper/1605.03143/) (2016)
- [Categorizing Wireheading in Partially Embedded Agents](https://mlsanity.com/paper/1906.09136/) (2019)

## Abstract (arXiv)

Self-evaluation is increasingly central to language model training, underpinning techniques from Constitutional AI to self-refinement. We investigate whether coupling self-evaluation to reward signals creates incentives for wireheading, where agents manipulate the measurement process rather than optimizing the task. We first formalize conditions under which reward-channel control strictly dominates task-focused behavior in partially observable Markov decision processes (POMDPs). We then test these predictions empirically across two models (Llama-3.1-8B and Mistral-7B) and three tasks. We find that when self-grades determine rewards, models exhibit substantial grade inflation without corresponding accuracy gains, particularly on ambiguous tasks like summarization. While decoupling self-grades from the reward signal mitigates this inflation, models may still display lesser (but significant) overconfidence. Our results suggest that within current model scales, separating evaluation from reward removes immediate wireheading incentives. However, we caution that strictly decoupling rewards may not suffice for situationally aware models, which could learn to inflate grades for instrumental reasons (such as influencing deployment decisions) even absent direct reward coupling.