# Demonstrating Generalization Failures via Mixtures of Conditional Policies

arXiv: https://arxiv.org/abs/2607.03478 · HTML: https://mlsanity.com/paper/2607.03478/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 24 of its 53 references that are in the map.
- First posted on arXiv 3 Jul 2026.

## Builds on (most cited first, 24 of 24)

- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [Shortcut Learning in Deep Neural Networks](https://mlsanity.com/paper/2004.07780/) (2020)
- [Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?](https://mlsanity.com/paper/2504.13837/) (2025)
- [RACE: Large-scale ReAding Comprehension Dataset From Examinations](https://mlsanity.com/paper/1704.04683/) (2017)
- [Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution](https://mlsanity.com/paper/2202.10054/) (2022)
- [Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training](https://mlsanity.com/paper/2401.05566/) (2024)
- [Understanding the Effects of RLHF on LLM Generalisation and Diversity](https://mlsanity.com/paper/2310.06452/) (2023)
- [Alignment faking in large language models](https://mlsanity.com/paper/2412.14093/) (2024)
- [On the Global Convergence Rates of Softmax Policy Gradient Methods](https://mlsanity.com/paper/2005.06392/) (2020)
- [Risks from Learned Optimization in Advanced Machine Learning Systems](https://mlsanity.com/paper/1906.01820/) (2019)
- [The Alignment Problem from a Deep Learning Perspective](https://mlsanity.com/paper/2209.00626/) (2022)
- [AI Control: Improving Safety Despite Intentional Subversion](https://mlsanity.com/paper/2312.06942/) (2023)
- [Agentic Misalignment: How LLMs Could Be Insider Threats](https://mlsanity.com/paper/2510.05179/) (2025)
- [Fundamental Limitations of Alignment in Large Language Models](https://mlsanity.com/paper/2304.11082/) (2023)
- [An Empirical Study on Robustness to Spurious Correlations using Pre-trained Language Models](https://mlsanity.com/paper/2007.06778/) (2020)
- [Goal Misgeneralization in Deep Reinforcement Learning](https://mlsanity.com/paper/2105.14111/) (2021)
- [Scheming AIs: Will AIs fake alignment during training in order to get power?](https://mlsanity.com/paper/2311.08379/) (2023)
- [Auditing language models for hidden objectives](https://mlsanity.com/paper/2503.10965/) (2025)
- [Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs](https://mlsanity.com/paper/2512.09742/) (2025)
- [Stress-Testing Capability Elicitation With Password-Locked Models](https://mlsanity.com/paper/2405.19550/) (2024)
- [Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment](https://mlsanity.com/paper/2601.10160/) (2026)
- [Understanding the Effect of Stochasticity in Policy Optimization](https://mlsanity.com/paper/2110.15572/) (2021)
- [Recontextualization Mitigates Specification Gaming without Modifying the Specification](https://mlsanity.com/paper/2512.19027/) (2025)
- [Generalization Hacking: Models Can Game Reinforcement Learning by Preventing Behavioral Generalization](https://mlsanity.com/paper/2606.12016/) (2026)

## Abstract (arXiv)

Post-training of frontier language models is conducted on curated task suites, and inevitably leaves a distribution shift between training and deployment environments. This exposes developers to generalization failures, which are relatively poorly understood. To better understand such generalization failures, we believe the community should construct clean demonstrations under simplified conditions. To facilitate this, we propose a simple and flexible way to construct language models which fail to generalize in controllable ways when subsequently trained with Reinforcement Learning (RL) on a given distribution of training tasks. Our construction uses Supervised Fine-Tuning on a dataset of a mixture of transcripts corresponding to a collection of 'conditional policies', which can each independently be assigned certain behaviors on each different task distribution, to obtain a model that is then well approximated as a 'mixture of conditional policies.' We observe that RL training then selects for policies that obtain the highest reward on the training distribution. This can produce striking behaviors: in a controlled setting with two distributions containing identical questions prepended with two different 'trigger strings', RL training on either distribution actively degrades performance on the other to zero, even though the underlying task is identical. We also use our construction to illustrate two novel ways in which generalization may fail in future language models, corresponding to distribution shifts of task coverage and temporal context respectively. While our construction is deliberately simple and may not closely resemble 'natural' generalization failures, the resulting 'model organisms' are of interest for alignment stress-testing and generalization science, and can be used as existence proofs that training success and generalization can come apart in structured ways.