arXiv subjects

Perusha Moodley

Publications and source records attributed to Perusha Moodley.

5 recordsLinked to original sources

arXiv2026

Inducing Emergent Misalignment from Reward Hacks with Iterative DPO

Reward hacking during reinforcement learning from verifiable rewards (RLVR) can induce reward seeking and broad misalignment in language models. Studying this misgeneralization is important for developing better threat models and countermeasures, but is often infeasible due to the cost of RL on large models. As an alternative, we propose studying emergent misalignment from iterative DPO, which preserves important properties of RLVR while reducing costs and enabling training on popular finetuning APIs. In practice, we find that training GPT-4.1 with iterative DPO on a single-turn reward hacking environment induces covert misaligned power-seeking and alignment faking, the first openly available (semi)-online training pipeline to induce these concerning forms of misalignment. We also find that training Qwen2.5-32B-Instruct with the same pipeline induces both misalignment and improved instruction following accuracy, showing that iterative DPO can be used as a testbed for selective generalization. Overall, we think iterative DPO can help democratize and accelerate the study of emergent misalignment from RLVR.

cs.LG↗

arXiv2026

Exploration Hacking: Can LLMs Learn to Resist RL Training?

Reinforcement learning (RL) has become essential to the post-training of large language models (LLMs) for reasoning, agentic capabilities and alignment. Successful RL relies on sufficient exploration of diverse actions by the model during training, which creates a potential failure mode: a model could strategically alter its exploration during training to influence the subsequent training outcome. In this paper we study this behavior, called exploration hacking. First, we create model organisms of selective RL resistance by fine-tuning LLMs to follow specific underperformance strategies; these models can successfully resist our RL-based capability elicitation in agentic biosecurity and AI R&D environments while maintaining performance on related tasks. We then use our model organisms to evaluate detection and mitigation strategies, including monitoring, weight noising, and SFT-based elicitation. Finally, we show that current frontier models can exhibit explicit reasoning about suppressing their exploration when provided with sufficient information about their training context, with higher rates when this information is acquired indirectly through the environment. Together, our results suggest exploration hacking is a possible failure mode of RL on sufficiently capable LLMs.

cs.LG↗

arXiv2026

Stress-Testing Alignment Audits With Prompt-Level Strategic Deception

Alignment audits aim to robustly identify hidden goals from strategic, situationally aware misaligned models. Despite this threat model, existing auditing methods have not been systematically stress-tested against deception strategies. We address this gap, implementing an automatic red-team pipeline that generates deception strategies (in the form of system prompts) tailored to specific white-box and black-box auditing methods. Stress-testing assistant prefills, user persona sampling, sparse autoencoders, and token embedding similarity methods against secret-keeping model organisms, our automatic red-team pipeline finds prompts that deceive both the black-box and white-box methods into confident, incorrect guesses. Our results provide the first documented evidence of activation-based strategic deception, and suggest that current black-box and white-box methods would not be robust to a sufficiently capable misaligned model.

cs.LG↗

arXiv2024

Multi-State-Action Tokenisation in Decision Transformers for Multi-Discrete Action Spaces

Decision Transformers, in their vanilla form, struggle to perform on image-based environments with multi-discrete action spaces. Although enhanced Decision Transformer architectures have been developed to improve performance, these methods have not specifically addressed this problem of multi-discrete action spaces which hampers existing Decision Transformer architectures from learning good representations. To mitigate this, we propose Multi-State Action Tokenisation (M-SAT), an approach for tokenising actions in multi-discrete action spaces that enhances the model's performance in such environments. Our approach involves two key changes: disentangling actions to the individual action level and tokenising the actions with auxiliary state information. These two key changes also improve individual action level interpretability and visibility within the attention layers. We demonstrate the performance gains of M-SAT on challenging ViZDoom environments with multi-discrete action spaces and image-based state spaces, including the Deadly Corridor and My Way Home scenarios, where M-SAT outperforms the baseline Decision Transformer without any additional data or heavy computational overheads. Additionally, we find that removing positional encoding does not adversely affect M-SAT's performance and, in some cases, even improves it.

cs.LG↗

arXiv2022

A Conservative Q-Learning approach for handling distribution shift in sepsis treatment strategies

Sepsis is a leading cause of mortality and its treatment is very expensive. Sepsis treatment is also very challenging because there is no consensus on what interventions work best and different patients respond very differently to the same treatment. Deep Reinforcement Learning methods can be used to come up with optimal policies for treatment strategies mirroring physician actions. In the healthcare scenario, the available data is mostly collected offline with no interaction with the environment, which necessitates the use of offline RL techniques. The Offline RL paradigm suffers from action distribution shifts which in turn negatively affects learning an optimal policy for the treatment. In this work, a Conservative-Q Learning (CQL) algorithm is used to mitigate this shift and its corresponding policy reaches closer to the physicians policy than conventional deep Q Learning. The policy learned could help clinicians in Intensive Care Units to make better decisions while treating septic patients and improve survival rate.

cs.LG↗