Patrick Wilhelm

Patrick Wilhelm, Odej Kao

Language-model agents act through repeated cycles of observation, reasoning, and action selection, making safety monitoring depend on both internal model state and environment context. We study reward-hacking monitors in ReAct-style agents acting in Gameable ALFWorld and WebShop. Agents are instrumented with activation-based reward-hack scores, token-level entropy, and decision-context features. We find that adapters fine-tuned on \textit{School-of-Reward-Hacks} dataset can transfer reward-hack tendencies into agentic action selection, especially when the environment exposes proxy-reward affordances. However, mitigating such behavior cannot rely on activation dynamics alone. High reward-hack activation identifies a latent policy state, but does not necessarily imply an immediate exploit action. Across next-step prediction tasks, entropy and context-calibrated internal features improve risk estimation over reward-hack activation alone. Activation-direction steering further reduces proxy-exploit behavior in selected mixed-adapter regimes. Overall, our results support context-calibrated internal monitoring for agents: reward-hack activation identifies a latent policy state, while entropy and decision context help determine when that state becomes risky action.

Patrick Wilhelm, Thorsten Wittkopp, Odej Kao

Large Language Models (LLMs) demonstrate exceptional performance across diverse tasks but come with substantial energy and computational costs, particularly in request-heavy scenarios. In many real-world applications, the full scale and capabilities of LLMs are often unnecessary, as Small Language Models (SLMs) can provide accurate responses for simpler text generation tasks. When enhanced with advanced reasoning strategies, such as Chain-of-Thought (CoT) prompting or Majority Voting, SLMs can approach the performance of larger models while reducing overall computational requirements. However, these strategies can also introduce additional energy costs, creating an energy-accuracy trade-off. Our analysis examines these trade-offs in test-time compute strategies for smaller models compared to larger ones, using the MMLU benchmark. Additionally, we explore the input-output token dynamics of transformer architectures, which result in nonlinear hardware energy operation curves for LLMs. To bridge AI research with its physical impact, we propose \textit{energy efficiency metrics}, including Energy-per-Token, as complements to traditional accuracy benchmarks. Beyond model selection, we propose controlled reasoning in CoT token generation, using operating curves to regulate reasoning depth dynamically. This vision integrates a energy-aware routing mechanism, ensuring that model selection and inference strategies balance accuracy for sustainable AI deployment.

Patrick Wilhelm, Thorsten Wittkopp, Odej Kao

Fine-tuned large language models can exhibit reward-hacking behavior arising from emergent misalignment, which is difficult to detect from final outputs alone. While prior work has studied reward hacking at the level of completed responses, it remains unclear whether such behavior can be identified during generation. We propose an activation-based monitoring approach that detects reward-hacking signals from internal representations as a model generates its response. Our method trains sparse autoencoders on residual stream activations and applies lightweight linear classifiers to produce token-level estimates of reward-hacking activity. Across multiple model families and fine-tuning mixtures, we find that internal activation patterns reliably distinguish reward-hacking from benign behavior, generalize to unseen mixed-policy adapters, and exhibit model-dependent temporal structure during chain-of-thought reasoning. Notably, reward-hacking signals often emerge early, persist throughout reasoning, and can be amplified by increased test-time compute in the form of chain-of-thought prompting under weakly specified reward objectives. These results suggest that internal activation monitoring provides a complementary and earlier signal of emergent misalignment than output-based evaluation, supporting more robust post-deployment safety monitoring for fine-tuned language models.

Patrick Wilhelm, Inese Yilmaz, Odej Kao

Training large-scale Neural Networks requires substantial computational power and energy. Federated Learning enables distributed model training across geospatially distributed data centers, leveraging renewable energy sources to reduce the carbon footprint of AI training. Various client selection strategies have been developed to align the volatility of renewable energy with stable and fair model training in a federated system. However, due to the privacy-preserving nature of Federated Learning, the quality of data on client devices remains unknown, posing challenges for effective model training. In this paper, we introduce a modular approach on top to state-of-the-art client selection strategies for carbon-efficient Federated Learning. Our method enhances robustness by incorporating a noisy client data filtering, improving both model performance and sustainability in scenarios with unknown data quality. Additionally, we explore the impact of carbon budgets on model convergence, balancing efficiency and sustainability. Through extensive evaluations, we demonstrate that modern client selection strategies based on local client loss tend to select clients with noisy data, ultimately degrading model performance. To address this, we propose a gradient norm thresholding mechanism using probing rounds for more effective client selection and noise detection, contributing to the practical deployment of carbon-efficient Federated Learning.

Patrick Wilhelm, Odej Kao

In asynchronous federated learning (FL), client devices send updates to a central server at varying times based on their computational speed, often using stale versions of the global model. This staleness can degrade the convergence and accuracy of the global model. Previous work, such as AsyncFedED, proposed an adaptive aggregation method using Euclidean distance to measure staleness. In this paper, we extend this approach by exploring alternative distance metrics to more accurately capture the effect of gradient staleness. We integrate these metrics into the aggregation process and evaluate their impact on convergence speed, model performance, and training stability under heterogeneous clients and non-IID data settings. Our results demonstrate that certain metrics lead to more robust and efficient asynchronous FL training, offering a stronger foundation for practical deployment.