Timo Flesch

Timo Flesch, David G. Nagy, Andrew Saxe et al.

Humans can learn several tasks in succession with minimal mutual interference but perform more poorly when trained on multiple tasks at once. The opposite is true for standard deep neural networks. Here, we propose novel computational constraints for artificial neural networks, inspired by earlier work on gating in the primate prefrontal cortex, that capture the cost of interleaved training and allow the network to learn two tasks in sequence without forgetting. We augment standard stochastic gradient descent with two algorithmic motifs, so-called "sluggish" task units and a Hebbian training step that strengthens connections between task units and hidden units that encode task-relevant information. We found that the "sluggish" units introduce a switch-cost during training, which biases representations under interleaved training towards a joint representation that ignores the contextual cue, while the Hebbian step promotes the formation of a gating scheme from task units to the hidden layer that produces orthogonal representations which are perfectly guarded against interference. Validating the model on previously published human behavioural data revealed that it matches performance of participants who had been trained on blocked or interleaved curricula, and that these performance differences were driven by misestimation of the true category boundary.

Timo Flesch, Andrew Saxe, Christopher Summerfield

How do humans and other animals learn new tasks? A wave of brain recording studies has investigated how neural representations change during task learning, with a focus on how tasks can be acquired and coded in ways that minimise mutual interference. We review recent work that has explored the geometry and dimensionality of neural task representations in neocortex, and computational models that have exploited these findings to understand how the brain may partition knowledge between tasks. We discuss how ideas from machine learning, including those that combine supervised and unsupervised learning, are helping neuroscientists understand how natural tasks are learned and coded in biological brains.

Magda Dubois, Ekin Zorer, Maia Hamin et al. · cambridge

AI systems produce large volumes of logs as they interact with tools and users. Analysing these logs can help understand model capabilities, propensities, and behaviours, or assess whether an evaluation worked as intended. Researchers have started developing methods for log analysis, but a standardised approach is still missing. Here we suggest a pipeline based on current best practices. We illustrate it with concrete code examples in the Inspect Scout library, provide detailed guidance on each step, and highlight common pitfalls. Our framework provides researchers with a foundation for rigorous and reproducible log analysis.

Egor Kraev, Timo Flesch, Hudson Taylor Lekunze et al.

Recently, many causal estimators for Conditional Average Treatment Effect (CATE) and instrumental variable (IV) problems have been published and open sourced, allowing to estimate granular impact of both randomized treatments (such as A/B tests) and of user choices on the outcomes of interest. However, the practical application of such models has ben hampered by the lack of a valid way to score the performance of such models out of sample, in order to select the best one for a given application. We address that gap by proposing novel scoring approaches for both the CATE case and an important subset of instrumental variable problems, namely those where the instrumental variable is customer acces to a product feature, and the treatment is the customer's choice to use that feature. Being able to score model performance out of sample allows us to apply hyperparameter optimization methods to causal model selection and tuning. We implement that in an open source package that relies on DoWhy and EconML libraries for implementation of causal inference models (and also includes a Transformed Outcome model implementation), and on FLAML for hyperparameter optimization and for component models used in the causal models. We demonstrate on synthetic data that optimizing the proposed scores is a reliable method for choosing the model and its hyperparameter values, whose estimates are close to the true impact, in the randomized CATE and IV cases. Further, we provide examles of applying these methods to real customer data from Wise.

Magda Dubois, Harry Coppock, Mario Giulianelli et al.

The evaluation of large language model (LLM) outputs is increasingly performed by other LLMs, a setup commonly known as "LLM-as-a-judge", or autograders. While autograders offer a scalable alternative to human evaluation, they have shown mixed reliability and may exhibit systematic biases, depending on response type, scoring methodology, domain specificity, or other factors. Here we propose a statistical framework based on Bayesian generalised linear models (GLMs) that enables researchers to simultaneously assess their autograders while addressing their primary research questions (e.g., LLM evaluation). Our approach models evaluation outcomes (e.g., scores or pairwise preferences) as a function of properties of the grader (e.g., human vs. autograder) and the evaluated item (e.g., response length or the LLM that generated it), allowing for explicit quantification of scoring differences and potential biases within a unified framework. In addition, our method can be used to augment traditional metrics such as inter-rater agreement, by providing uncertainty estimates and clarifying sources of disagreement. Overall, this approach contributes to more robust and interpretable use of autograders in LLM evaluation, enabling both performance analysis and bias detection.