Lovis Heindrich

Lucia Quirke, Lovis Heindrich, Wes Gurnee et al.

Prior work has shown the existence of contextual neurons in language models, including a neuron that activates on German text. We show that this neuron exists within a broader contextual n-gram circuit: we find late layer neurons which recognize and continue n-grams common in German text, but which only activate if the German neuron is active. We investigate the formation of this circuit throughout training and find that it is an example of what we call a second-order circuit. In particular, both the constituent n-gram circuits and the German detection circuit which culminates in the German neuron form with independent functions early in training - the German detection circuit partially through modeling German unigram statistics, and the n-grams by boosting appropriate completions. Only after both circuits have already formed do they fit together into a second-order circuit. Contrary to the hypotheses presented in prior work, we find that the contextual n-gram circuit forms gradually rather than in a sudden phase transition. We further present a range of anomalous observations such as a simultaneous phase transition in many tasks coinciding with the learning rate warm-up, and evidence that many context neurons form simultaneously early in training but are later unlearned.

Lovis Heindrich, Saksham Consul, Falk Lieder

AI can not only outperform people in many planning tasks, but it can also teach them how to plan better. A recent and promising approach to improving human decision-making is to create intelligent tutors that utilize AI to discover and teach optimal planning strategies automatically. Prior work has shown that this approach can improve planning in artificial, fully observable planning tasks. Unlike these artificial tasks, many of the real-world situations in which people have to make plans include features that are only partially observable. To bridge this gap, we develop and evaluate the first intelligent tutor for planning in partially observable environments. Compared to previous intelligent tutors for teaching planning strategies, this novel intelligent tutor combines two innovations: 1) a new metareasoning algorithm for discovering optimal planning strategies for large, partially observable environments, and 2) scaffolding the learning process by having the learner choose from an increasing larger set of planning operations in increasingly larger planning problems. We found that our new strategy discovery algorithm is superior to the state-of-the-art. A preregistered experiment with 330 participants demonstrated that the new intelligent tutor is highly effective at improving people's ability to make good decisions in partially observable environments. This suggests our intelligent cognitive tutor can successfully boost human planning in complex, partially observable sequential decision problems. That makes the work presented in this a promising step towards using AI-powered intelligent tutors to improve human planning in the real world.

Lovis Heindrich, Philip Torr, Fazl Barez et al.

Sparse autoencoders (SAEs) have emerged as a promising approach in language model interpretability, offering unsupervised extraction of sparse features. For interpretability methods to succeed, they must identify abstract features across domains, and these features can often manifest differently in each context. We examine this through "answerability" - a model's ability to recognize answerable questions. We extensively evaluate SAE feature generalization across diverse, partly self-constructed answerability datasets for Gemma 2 SAEs. Our analysis reveals that residual stream probes outperform SAE features within domains, but generalization performance differs sharply. SAE features show inconsistent out-of-domain transfer, with performance varying from almost random to outperforming residual stream probes. Overall, this demonstrates the need for robust evaluation methods and quantitative approaches to predict feature generalization in SAE-based interpretability.

Lovis Heindrich, Falk Lieder

The decisions of individuals and organizations are often suboptimal because fully rational decision-making is too demanding in the real world. Recent work suggests that some errors can be prevented by leveraging artificial intelligence to discover and teach clever heuristics. So far, this line of research has been limited to simplified, artificial decision-making tasks. This article is the first to extend this approach to a real-world decision problem, namely, executives deciding which project their organization should launch next. We develop a computational method (MGPS) that automatically discovers project selection strategies that are optimized for real people, and we develop an intelligent tutor that teaches the discovered project selection procedures. We evaluated MGPS on a computational benchmark and tested the intelligent tutor in a training experiment with two control conditions. MGPS outperformed a state-of-the-art method and was more computationally efficient. Moreover, people who practiced with our intelligent tutor learned significantly better project selection strategies than the control groups. These findings suggest that AI could be used to automate the process of discovering and formalizing the cognitive strategies taught by intelligent tutoring systems.

Saksham Consul, Lovis Heindrich, Jugoslav Stojcheski et al.

To make good decisions in the real world people need efficient planning strategies because their computational resources are limited. Knowing which planning strategies would work best for people in different situations would be very useful for understanding and improving human decision-making. But our ability to compute those strategies used to be limited to very small and very simple planning tasks. To overcome this computational bottleneck, we introduce a cognitively-inspired reinforcement learning method that can overcome this limitation by exploiting the hierarchical structure of human behavior. The basic idea is to decompose sequential decision problems into two sub-problems: setting a goal and planning how to achieve it. This hierarchical decomposition enables us to discover optimal strategies for human planning in larger and more complex tasks than was previously possible. The discovered strategies outperform existing planning algorithms and achieve a super-human level of computational efficiency. We demonstrate that teaching people to use those strategies significantly improves their performance in sequential decision-making tasks that require planning up to eight steps ahead. By contrast, none of the previous approaches was able to improve human performance on these problems. These findings suggest that our cognitively-informed approach makes it possible to leverage reinforcement learning to improve human decision-making in complex sequential decision-problems. Future work can leverage our method to develop decision support systems that improve human decision making in the real world.