Identifying and Mitigating the Influence of the Prior Distribution in Large Language Models

Large language models (LLMs) sometimes fail to respond appropriately to deterministic tasks -- such as counting or forming acronyms -- because the implicit prior distribution they have learned over sequences of tokens influences their responses. In this work, we show that, in at least some cases, LLMs actually compute the information needed to perform these tasks correctly, and we identify some interventions that can allow them to access this information to improve their performance. First, we show that simply prompting the language model to not rely on its prior knowledge leads to dramatic improvements in prior-dominated tasks. We then use mechanistic interpretability techniques to localize the prior within the LLM and manipulate the extent to which that prior influences its responses. Specifically, we show that it is possible to identify layers of the underlying neural network that correlate with the prior probability of a response and that lightweight finetuning of these layers with basic prompts on prior-dominated tasks achieves high performance on held-out answers. These results suggest that the information required to produce a correct response is contained within the representations of the problems formed by the models. Furthermore, we show that this finetuning is significantly more effective for prior-dominated tasks, and that the error after finetuning is no longer correlated with the prior. Our results suggest that it may be possible to define effective methods for manipulating the extent to which LLMs rely upon their priors in solving problems, potentially increasing their performance in settings where LLMs hallucinate for reasons related to the prior probability of token sequences.

Localized Cultural Knowledge is Conserved and Controllable in Large Language Models

Just as humans display language patterns influenced by their native tongue when speaking new languages, LLMs often default to English-centric responses even when generating in other languages. Nevertheless, we observe that local cultural information persists within the models and can be readily activated for cultural customization. We first demonstrate that explicitly providing cultural context in prompts significantly improves the models' ability to generate culturally localized responses. We term the disparity in model performance with versus without explicit cultural context the explicit-implicit localization gap, indicating that while cultural knowledge exists within LLMs, it may not naturally surface in multilingual interactions if cultural context is not explicitly provided. Despite the explicit prompting benefit, however, the answers reduce in diversity and tend toward stereotypes. Second, we identify an explicit cultural customization vector, conserved across all non-English languages we explore, which enables LLMs to be steered from the synthetic English cultural world-model toward each non-English cultural world. Steered responses retain the diversity of implicit prompting and reduce stereotypes to dramatically improve the potential for customization. We discuss the implications of explicit cultural customization for understanding the conservation of alternative cultural world models within LLMs, and their controllable utility for translation, cultural customization, and the possibility of making the explicit implicit through soft control for expanded LLM function and appeal.

Convolutional Neural Networks Can (Meta-)Learn the Same-Different Relation

While convolutional neural networks (CNNs) have come to match and exceed human performance in many settings, the tasks these models optimize for are largely constrained to the level of individual objects, such as classification and captioning. Humans remain vastly superior to CNNs in visual tasks involving relations, including the ability to identify two objects as `same' or `different'. A number of studies have shown that while CNNs can be coaxed into learning the same-different relation in some settings, they tend to generalize poorly to other instances of this relation. In this work we show that the same CNN architectures that fail to generalize the same-different relation with conventional training are able to succeed when trained via meta-learning, which explicitly encourages abstraction and generalization across tasks.

Predicting Multi-Agent Specialization via Task Parallelizability

Multi-agent systems often rely on specialized agents with distinct roles rather than general-purpose agents that perform the entire task independently. However, the conditions that govern the optimal degree of specialization remain poorly understood. In this work, we propose that specialist teams outperform generalist ones when environmental constraints limit task parallelizability -- the potential to execute task components concurrently. Drawing inspiration from distributed systems, we introduce a heuristic to predict the relative efficiency of generalist versus specialist teams by estimating the speed-up achieved when two agents perform a task in parallel rather than focus on complementary subtasks. We validate this heuristic through three multi-agent reinforcement learning (MARL) experiments in Overcooked-AI, demonstrating that key factors limiting task parallelizability influence specialization. We also observe that as the state space expands, agents tend to converge on specialist strategies, even when generalist ones are theoretically more efficient, highlighting potential biases in MARL training algorithms. Our findings provide a principled framework for interpreting specialization given the task and environment, and introduce a novel benchmark for evaluating whether MARL finds optimal strategies.

Using the Tools of Cognitive Science to Understand Large Language Models at Different Levels of Analysis

Modern artificial intelligence systems, such as large language models, are increasingly powerful but also increasingly hard to understand. Recognizing this problem as analogous to the historical difficulties in understanding the human mind, we argue that methods developed in cognitive science can be useful for understanding large language models. We propose a framework for applying these methods based on Marr's three levels of analysis. By revisiting established cognitive science techniques relevant to each level and illustrating their potential to yield insights into the behavior and internal organization of large language models, we aim to provide a toolkit for making sense of these new kinds of minds.

Teasing Apart Architecture and Initial Weights as Sources of Inductive Bias in Neural Networks

Artificial neural networks can acquire many aspects of human knowledge from data, making them promising as models of human learning. But what those networks can learn depends upon their inductive biases -- the factors other than the data that influence the solutions they discover -- and the inductive biases of neural networks remain poorly understood, limiting our ability to draw conclusions about human learning from the performance of these systems. Cognitive scientists and machine learning researchers often focus on the architecture of a neural network as a source of inductive bias. In this paper we explore the impact of another source of inductive bias -- the initial weights of the network -- using meta-learning as a tool for finding initial weights that are adapted for specific problems. We evaluate four widely-used architectures -- MLPs, CNNs, LSTMs, and Transformers -- by meta-training 430 different models across three tasks requiring different biases and forms of generalization. We find that meta-learning can substantially reduce or entirely eliminate performance differences across architectures and data representations, suggesting that these factors may be less important as sources of inductive bias than is typically assumed. When differences are present, architectures and data representations that perform well without meta-learning tend to meta-train more effectively. Moreover, all architectures generalize poorly on problems that are far from their meta-training experience, underscoring the need for stronger inductive biases for robust generalization.

On Benchmarking Human-Like Intelligence in Machines

Recent benchmark studies have claimed that AI has approached or even surpassed human-level performances on various cognitive tasks. However, this position paper argues that current AI evaluation paradigms are insufficient for assessing human-like cognitive capabilities. We identify a set of key shortcomings: a lack of human-validated labels, inadequate representation of human response variability and uncertainty, and reliance on simplified and ecologically-invalid tasks. We support our claims by conducting a human evaluation study on ten existing AI benchmarks, suggesting significant biases and flaws in task and label designs. To address these limitations, we propose five concrete recommendations for developing future benchmarks that will enable more rigorous and meaningful evaluations of human-like cognitive capacities in AI with various implications for such AI applications.

Conformal Prediction as Bayesian Quadrature

As machine learning-based prediction systems are increasingly used in high-stakes situations, it is important to understand how such predictive models will perform upon deployment. Distribution-free uncertainty quantification techniques such as conformal prediction provide guarantees about the loss black-box models will incur even when the details of the models are hidden. However, such methods are based on frequentist probability, which unduly limits their applicability. We revisit the central aspects of conformal prediction from a Bayesian perspective and thereby illuminate the shortcomings of frequentist guarantees. We propose a practical alternative based on Bayesian quadrature that provides interpretable guarantees and offers a richer representation of the likely range of losses to be observed at test time.

What is a Number, That a Large Language Model May Know It?

Numbers are a basic part of how humans represent and describe the world around them. As a consequence, learning effective representations of numbers is critical for the success of large language models as they become more integrated into everyday decisions. However, these models face a challenge: depending on context, the same sequence of digit tokens, e.g., 911, can be treated as a number or as a string. What kind of representations arise from this duality, and what are its downstream implications? Using a similarity-based prompting technique from cognitive science, we show that LLMs learn representational spaces that blend string-like and numerical representations. In particular, we show that elicited similarity judgments from these models over integer pairs can be captured by a combination of Levenshtein edit distance and numerical Log-Linear distance, suggesting an entangled representation. In a series of experiments we show how this entanglement is reflected in the latent embeddings, how it can be reduced but not entirely eliminated by context, and how it can propagate into a realistic decision scenario. These results shed light on a representational tension in transformer models that must learn what a number is from text input.

RLHS: Mitigating Misalignment in RLHF with Hindsight Simulation

Generative AI systems like foundation models (FMs) must align well with human values to ensure their behavior is helpful and trustworthy. While Reinforcement Learning from Human Feedback (RLHF) has shown promise for optimizing model performance using human judgments, existing RLHF pipelines predominantly rely on immediate feedback, which can fail to accurately reflect the downstream impact of an interaction on users' utility. We demonstrate that feedback based on evaluators' foresight estimates of downstream consequences systematically induces Goodhart's Law dynamics, incentivizing misaligned behaviors like sycophancy and deception and ultimately degrading user outcomes. To alleviate this, we propose decoupling evaluation from prediction by refocusing RLHF on hindsight feedback. Our theoretical analysis reveals that conditioning evaluator feedback on downstream observations mitigates misalignment and improves expected human utility, even when these observations are simulated by the AI system itself. To leverage this insight in a practical alignment algorithm, we introduce Reinforcement Learning from Hindsight Simulation (RLHS), which first simulates plausible consequences and then elicits feedback to assess what behaviors were genuinely beneficial in hindsight. We apply RLHS to two widely-employed online and offline preference optimization methods -- Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO) -- and show empirically that misalignment is significantly reduced with both methods. Through an online human user study, we show that RLHS consistently outperforms RLHF in helping users achieve their goals and earns higher satisfaction ratings, despite being trained solely with simulated hindsight feedback. These results underscore the importance of focusing on long-term consequences, even simulated ones, to mitigate misalignment in RLHF.