The ideal AI safety moderation system would be both structurally interpretable (so its decisions can be reliably explained) and steerable (to align to safety standards and reflect a community’s values), which current systems fall short on. To address this gap, we present SafetyAnalyst, a novel AI safety moderation framework. Given an AI behavior, SafetyAnalyst uses chain-of-thought reasoning to analyze its potential consequences by creating a structured "harm-benefit tree," which enumerates harmful and beneficial actions and effects the AI behavior may lead to, along with likelihood, severity, and immediacy labels that describe potential impacts on stakeholders. SafetyAnalyst then aggregates all effects into a harmfulness score using 28 fully interpretable weight parameters, which can be aligned to particular safety preferences. We applied this framework to develop an open-source LLM prompt safety classification system, distilled from 18.5 million harm-benefit features generated by frontier LLMs on 19k prompts. On comprehensive benchmarks, we show that SafetyAnalyst (average F1=0.81) outperforms existing moderation systems (average F1$<$0.72) on prompt safety classification, while offering the additional advantages of interpretability, transparency, and steerability.

Dopamine release in the nucleus accumbens has been hypothesized to signal the difference between observed and predicted reward, known as reward prediction error, suggesting a biological implementation for reinforcement learning. Rigorous tests of this hypothesis require assumptions about how the brain maps sensory signals to reward predictions, yet this mapping is still poorly understood. In particular, the mapping is non-trivial when sensory signals provide ambiguous information about the hidden state of the environment. Previous work using classical conditioning tasks has suggested that reward predictions are generated conditional on probabilistic beliefs about the hidden state, such that dopamine implicitly reflects these beliefs. Here we test this hypothesis in the context of an instrumental task (a two-armed bandit), where the hidden state switches stochastically. We measured choice behavior and recorded dLight signals that reflect dopamine release in the nucleus accumbens core. Model comparison among a wide set of cognitive models based on the behavioral data favored models that used Bayesian updating of probabilistic beliefs. These same models also quantitatively matched mesolimbic dLight measurements better than non-Bayesian alternatives. We conclude that probabilistic belief computation contributes to instrumental task performance in mice and is reflected in mesolimbic dopamine signaling.

Anxiety has been robustly linked to deficits in frontal executive function including working memory (WM) and attentional control processes. However, although anxiety has also been associated with impaired performance on learning tasks, computational investigations of reinforcement learning (RL) impairment in anxiety have yielded mixed results. WM processes are known to contribute to learning behavior in parallel to RL processes and to modulate the effective learning rate as a function of load. However, WM processes have typically not been modeled in investigations of anxiety and RL. In the current study, we leveraged an experimental paradigm (RLWM) which manipulates the relative contributions of WM and RL processes in a reinforcement learning and retention task using multiple stimulus set sizes. Using a computational model of interactive RL and WM processes, we investigated whether individual differences in physiological or cognitive anxiety impacted task performance via deficits in RL or WM. Elevated physiological, but not cognitive, anxiety scores were strongly associated with worse performance during learning and retention testing across all set sizes. Computationally, higher physiological anxiety scores were significantly related to reduced learning rate and increased rate of WM decay. To highlight the importance of modeling WM contributions to learning, we considered the effect of fitting RL models without WM modules to the data. Here we found that reduced learning performance for higher physiological anxiety was at least partially misattributed to stochastic decision noise in 9 out of 10 RL-only models considered. These findings reveal a dual-process impairment in learning in anxiety that is linked to a more physiological than cognitive anxiety phenotype. More broadly, this work also points to the importance of accounting for the contribution of WM to RL when investigating psychopathology-related deficits in learning.

Humans are autotelic agents who learn by setting and pursuing their own goals. However, the precise mechanisms guiding human goal selection remain unclear. Learning progress, typically measured as the observed change in performance, can provide a valuable signal for goal selection in both humans and artificial agents. We hypothesize that human choices of goals may also be driven by _latent learning progress_, which humans can estimate through knowledge of their actions and the environment – even without experiencing immediate changes in performance. To test this hypothesis, we designed a hierarchical reinforcement learning task in which human participants (N = 175) repeatedly chose their own goals and learned goal-conditioned policies. Our behavioral and computational modeling results confirm the influence of latent learning progress on goal selection and uncover inter-individual differences, partially mediated by recognition of the task's hierarchical structure. By investigating the role of latent learning progress in human goal selection, we pave the way for more effective and personalized learning experiences as well as the advancement of more human-like autotelic machines.

Computational modeling has revealed that human research participants use both rapid working memory (WM) and incremental reinforcement learning (RL) (RL+WM) to solve a simple instrumental learning task, relying on WM when the number of stimuli is small and supplementing with RL when the number of stimuli exceeds WM capacity. Inspired by this work, we examined which learning systems and strategies are used by adolescent and adult mice when they first acquire a conditional associative learning task. In a version of the human RL+WM task translated for rodents, mice were required to associate odor stimuli (from a set of 2 or 4 odors) with a left or right port to receive reward. Using logistic regression and computational models to analyze the first 200 trials per odor, we determined that mice used both incremental RL and stimulus-insensitive, one-back strategies to solve the task. While these one-back strategies may be a simple form of short-term or working memory, they did not approximate the boost to learning performance that has been observed in human participants using WM in a comparable task. Adolescent and adult mice also showed comparable performance, with no change in learning rate or softmax beta parameters with adolescent development and task experience. However, reliance on a one-back perseverative, win-stay strategy increased with development in males in both odor set sizes, but was not dependent on gonadal hormones. Our findings advance a simple conditional associative learning task and new models to enable the isolation and quantification of reinforcement learning alongside other strategies mice use while learning to associate stimuli with rewards within a single behavioral session. These data and methods can inform and aid comparative study of reinforcement learning across species.

Computational cognitive models have been used extensively to formalize cognitive processes. Model parameters offer a simple way to quantify individual differences in how humans process information. Similarly, model comparison allows researchers to identify which theories, embedded in different models, provide the best accounts of the data. Cognitive modeling uses statistical tools to quantitatively relate models to data that often rely on computing/estimating the likelihood of the data under the model. However, this likelihood is computationally intractable for a substantial number of models. These relevant models may embody reasonable theories of cognition, but are often under-explored due to the limited range of tools available to relate them to data. We contribute to filling this gap in a simple way using artificial neural networks (ANNs) to map data directly onto model identity and parameters, bypassing the likelihood estimation. We test our instantiation of an ANN as a cognitive model fitting tool on classes of cognitive models with strong inter-trial dependencies (such as reinforcement learning models), which offer unique challenges to most methods. We show that we can adequately perform both parameter estimation and model identification using our ANN approach, including for models that cannot be fit using traditional likelihood-based methods. We further discuss our work in the context of the ongoing research leveraging simulation-based approaches to parameter estimation and model identification, and how these approaches broaden the class of cognitive models researchers can quantitatively investigate.

In the real world, many relationships between events are uncertain and probabilistic. Uncertainty is also likely to be a more common feature of daily experience for youth because they have less experience to draw from than adults. Some studies suggest probabilistic learning may be inefficient in youths compared to adults, while others suggest it may be more efficient in youths in mid adolescence. Here we used a probabilistic reinforcement learning task to test how youth age 8-17 (N = 187) and adults age 18-30 (N = 110) learn about stable probabilistic contingencies. Performance increased with age through early-twenties, then stabilized. Using hierarchical Bayesian methods to fit computational reinforcement learning models, we show that all participants' performance was better explained by models in which negative outcomes had minimal to no impact on learning. The performance increase over age was driven by 1) an increase in learning rate (i.e. decrease in integration time scale); 2) a decrease in noisy/exploratory choices. In mid-adolescence age 13-15, salivary testosterone and learning rate were positively related. We discuss our findings in the context of other studies and hypotheses about adolescent brain development.

Study of human executive function focuses on our ability to represent cognitive rules independently of stimulus or response modality. However, recent findings suggest that executive functions cannot be modularized separately from perceptual and motor systems, and that they instead scaffold on top of motor action selection. Here we investigate whether patterns of motor demands influence how participants choose to implement abstract rule structures. In a learning task that requires integrating two stimulus dimensions for determining appropriate responses, subjects typically structure the problem hierarchically, using one dimension to cue the task-set and the other to cue the response given the task-set. However, the choice of which dimension to use at each level can be arbitrary. We hypothesized that the specific structure subjects adopt would be constrained by the motor patterns afforded within each rule. Across four independent data-sets, we show that subjects create rule structures that afford motor clustering, preferring structures in which adjacent motor actions are valid within each task-set. In a fifth data-set using instructed rules, this bias was strong enough to counteract the well-known task switch-cost when instructions were incongruent with motor clustering. Computational simulations confirm that observed biases can be explained by leveraging overlap in cortical motor representations to improve outcome prediction and hence infer the structure to be learned. These results highlight the importance of sensorimotor constraints in abstract rule formation and shed light on why humans have strong biases to invent structure even when it does not exist.