# Measuring Reward Hacking and Reasoning-Answer Decoupling Under Position-Confounded Optimization

arXiv: https://arxiv.org/abs/2608.15445 · HTML: https://mlsanity.com/paper/2608.15445/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning.
- Builds on 21 of its 24 references that are in the map.
- First posted on arXiv 15 Aug 2026.

## Placed in

- LLM reasoning / [Reinforcement learning for reasoning](https://mlsanity.com/lr/reinforcement-learning-for-reasoning/)
- LLM alignment / RLHF / preference optimisation / [Process reward models for reasoning](https://mlsanity.com/alignment/process-reward-models-for-reasoning/)

## Builds on (most cited first, 21 of 21)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Gemma 3 Technical Report](https://mlsanity.com/paper/2503.19786/) (2025)
- [Concrete Problems in AI Safety](https://mlsanity.com/paper/1606.06565/) (2016)
- [Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting](https://mlsanity.com/paper/2305.04388/) (2023)
- [Measuring Faithfulness in Chain-of-Thought Reasoning](https://mlsanity.com/paper/2307.13702/) (2023)
- [Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training](https://mlsanity.com/paper/2401.05566/) (2024)
- [Large Language Models Are Not Robust Multiple Choice Selectors](https://mlsanity.com/paper/2309.03882/) (2023)
- [Poisoning Web-Scale Training Datasets is Practical](https://mlsanity.com/paper/2302.10149/) (2023)
- [Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions](https://mlsanity.com/paper/2308.11483/) (2023)
- [Defining and Characterizing Reward Hacking](https://mlsanity.com/paper/2209.13085/) (2022)
- [Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models](https://mlsanity.com/paper/2406.10162/) (2024)
- [Natural Emergent Misalignment from Reward Hacking in Production RL](https://mlsanity.com/paper/2511.18397/) (2025)
- [Goal Misgeneralization: Why Correct Specifications Aren't Enough For Correct Goals](https://mlsanity.com/paper/2210.01790/) (2022)
- [Goal Misgeneralization in Deep Reinforcement Learning](https://mlsanity.com/paper/2105.14111/) (2021)
- [Why Models Know But Don't Say: Chain-of-Thought Faithfulness Divergence Between Thinking Tokens and Answers in Open-Weight Reasoning Models](https://mlsanity.com/paper/2603.26410/) (2026)

## Abstract (arXiv)

When a reward is correct on every training example yet consistent with more than one goal, a model can acquire an unintended one, a failure known as goal misgeneralization. Endpoint accuracy on the training distribution cannot tell the two apart, because solving the task and exploiting a surface feature can satisfy the reward equally well. We treat this as a measurement problem: what does a benchmark score measure once a model has been optimized against a correct but confounded signal? We train language models with GRPO on multiple-choice math problems where the correct answer is always option A, then evaluate on an unseen test set with unbiased answer positions. Across Qwen2.5, Llama 3.x and Gemma-3 models, biased training often drives option-A rates above 0.90 in smaller models and collapses unbiased accuracy toward chance, so accuracy stops measuring math ability and instead measures an answer-position policy. We further find reasoning-answer decoupling: capable models generate reasoning that reaches the correct numeric answer while still selecting A. We track this with numeric extraction and an LLM judge (GPT-4.1-mini; Qwen2.5-3B decoupling rate is about 0.66). The broken construct generalizes beyond the training domain: biased models inflate A-rates on out-of-domain MMLU and value-laden prompts. Continued training on unbiased data reverses the in-domain shift unevenly and only partially reverses the out-of-domain one, so a model can appear restored on its training distribution while remaining biased on unseen inputs. Reasoning-answer decoupling rate, together with answer distributions and out-of-domain behavior, separates capability loss from a learned, transferable shortcut.