Jose Hernandez-Orallo

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Xiting Wang, Liming Jiang, Jose Hernandez-Orallo et al. · cambridge

Comprehensive and accurate evaluation of general-purpose AI systems such as large language models allows for effective mitigation of their risks and deepened understanding of their capabilities. Current evaluation methodology, mostly based on benchmarks of specific tasks, falls short of adequately assessing these versatile AI systems, as present techniques lack a scientific foundation for predicting their performance on unforeseen tasks and explaining their varying performance on specific task items or user inputs. Moreover, existing benchmarks of specific tasks raise growing concerns about their reliability and validity. To tackle these challenges, we suggest transitioning from task-oriented evaluation to construct-oriented evaluation. Psychometrics, the science of psychological measurement, provides a rigorous methodology for identifying and measuring the latent constructs that underlie performance across multiple tasks. We discuss its merits, warn against potential pitfalls, and propose a framework to put it into practice. Finally, we explore future opportunities of integrating psychometrics with the evaluation of general-purpose AI systems.

Katherine M. Collins, Simon Frieder, Jonas Bayer et al.

For centuries, human mathematicians have written proofs to substantiate their mathematical arguments; yet, the ability to automatically verify the validity of proofs has long been a challenge. Advances in AI systems' ability to generate code and engage in increasingly high-level mathematical reasoning promise to transform people's ability to formalize and thereby verify proofs. While many works focus on benchmarking the current frontier, we instead study how people use these tools. We conduct a mixed-methods analysis into the initial impact of AI on people's formalization workflows: what people claim they want, what they see as the barriers to those visions, and how they actually use and adapt AI in practice. A qualitative survey shows that people's preferences are diverse, but with a general desire for AI assistance in formalization that preserves high-level human control over the proof discovery process. To assess how people actually engage with AI for formalization under such limitations, we conduct a controlled user study in which participants formalize informal math problems and their proofs, with and without AI, across a range of mathematical problems at varying levels of difficulty and domains. Despite limitations of the tools at the time for autoformalization, participants tend to attain higher formalization accuracy when allowed access to AI tools than when formalizing on their own, with most participants flexibly choosing to use multiple different AI tools. Taken together, our work sheds light on the early stages of AI integration into formalization workflows, involving an intimate interplay of human and AI engagement.

Anthony G Cohn, Jose Hernandez-Orallo

Language models have become very popular recently and many claims have been made about their abilities, including for commonsense reasoning. Given the increasingly better results of current language models on previous static benchmarks for commonsense reasoning, we explore an alternative dialectical evaluation. The goal of this kind of evaluation is not to obtain an aggregate performance value but to find failures and map the boundaries of the system. Dialoguing with the system gives the opportunity to check for consistency and get more reassurance of these boundaries beyond anecdotal evidence. In this paper we conduct some qualitative investigations of this kind of evaluation for the particular case of spatial reasoning (which is a fundamental aspect of commonsense reasoning). We conclude with some suggestions for future work both to improve the capabilities of language models and to systematise this kind of dialectical evaluation.

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.

Chandler Smith, Marwa Abdulhai, Manfred Diaz et al.

Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction and are increasingly being deployed in situations where they might engage with both human and artificial agents. These interactions represent a critical frontier for LLM-based agents, yet existing evaluation methods fail to measure how well these capabilities generalize to novel social situations. In this paper, we introduce a method for evaluating the ability of LLM-based agents to cooperate in zero-shot, mixed-motive environments using Concordia, a natural language multi-agent simulation environment. Our method measures general cooperative intelligence by testing an agent's ability to identify and exploit opportunities for mutual gain across diverse partners and contexts. We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains across a suite of diverse scenarios ranging from negotiation to collective action problems. Our findings reveal significant gaps between current agent capabilities and the robust generalization required for reliable cooperation, particularly in scenarios demanding persuasion and norm enforcement.

John Burden, Manuel Cebrian, Jose Hernandez-Orallo

Large Language Models (LLMs) present a dual-use dilemma: they enable beneficial applications while harboring potential for harm, particularly through conversational interactions. Despite various safeguards, advanced LLMs remain vulnerable. A watershed case in early 2023 involved journalist Kevin Roose's extended dialogue with Bing, an LLM-powered search engine, which revealed harmful outputs after probing questions, highlighting vulnerabilities in the model's safeguards. This contrasts with simpler early jailbreaks, like the "Grandma Jailbreak," where users framed requests as innocent help for a grandmother, easily eliciting similar content. This raises the question: How much conversational effort is needed to elicit harmful information from LLMs? We propose two measures to quantify this effort: Conversational Length (CL), which measures the number of conversational turns needed to obtain a specific harmful response, and Conversational Complexity (CC), defined as the Kolmogorov complexity of the user's instruction sequence leading to the harmful response. To address the incomputability of Kolmogorov complexity, we approximate CC using a reference LLM to estimate the compressibility of the user instructions. Applying this approach to a large red-teaming dataset, we perform a quantitative analysis examining the statistical distribution of harmful and harmless conversational lengths and complexities. Our empirical findings suggest that this distributional analysis and the minimization of CC serve as valuable tools for understanding AI safety, offering insights into the accessibility of harmful information. This work establishes a foundation for a new perspective on LLM safety, centered around the algorithmic complexity of pathways to harm.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Usman Anwar, Abulhair Saparov, Javier Rando et al. · cambridge, eth-zurich

This work identifies 18 foundational challenges in assuring the alignment and safety of large language models (LLMs). These challenges are organized into three different categories: scientific understanding of LLMs, development and deployment methods, and sociotechnical challenges. Based on the identified challenges, we pose $200+$ concrete research questions.

Jan Arne Telle, Brigt Håvardstun, Jose Hernandez-Orallo

Most models of machine teaching and learning assume the learner makes no errors in its internal deductive inference. However, humans and large language models in few-shot learning regimes are two important examples of learners where this does not hold. They fail on some consistency checks, and they can fail stochastically. In this paper we introduce a teaching and learning framework that takes these deductive errors into account. We specifically study the case of machine teaching, as different characterizations of the teacher can account for both machine teaching and learning. In an overhauled Probably Approximately Correct (PAC) setting, we study theoretically that, for some estimated error level, the teacher must find a PAC teaching set that with high probability will lead the learner to guess a hypothesis that is approximately correct. We study the computational complexity of six different problems related to computing optimal PAC teaching sets. We give XP algorithms parametrized by size of teaching set, with tight runtime bounds under standard complexity assumptions like ETH. These results are complemented with a small experimental study of which teaching and learning protocols can best represent the observed behavior in some LLM teaching sessions.

Stephen Fitz, Peter Romero, Steven Basart et al.

Large Language Models increasingly mediate high-stakes interactions, intensifying research on their capabilities and safety. While recent work has shown that LLMs exhibit consistent and measurable synthetic personality traits, little is known about how modulating these traits affects model behavior. We address this gap by investigating how psychometric personality control grounded in the Big Five framework influences AI behavior in the context of capability and safety benchmarks. Our experiments reveal striking effects: for example, reducing conscientiousness leads to significant drops in safety-relevant metrics on benchmarks such as WMDP, TruthfulQA, ETHICS, and Sycophancy as well as reduction in general capabilities as measured by MMLU. These findings highlight personality shaping as a powerful and underexplored axis of model control that interacts with both safety and general competence. We discuss the implications for safety evaluation, alignment strategies, steering model behavior after deployment, and risks associated with possible exploitation of these findings. Our findings motivate a new line of research on personality-sensitive safety evaluations and dynamic behavioral control in LLMs.

Kerem Oktar, Katherine M. Collins, Jose Hernandez-Orallo et al.

Artificial Intelligence (AI) systems have historically been used as tools that execute narrowly defined tasks. Yet recent advances in AI have unlocked possibilities for a new class of models that genuinely collaborate with humans in complex reasoning, from conceptualizing problems to brainstorming solutions. Such AI thought partners enable novel forms of collaboration and extended cognition, yet they also pose major risks-including and beyond risks of typical AI tools and agents. In this commentary, we systematically identify risks of AI thought partners through a novel framework that identifies risks at multiple levels of analysis, including Real-time, Individual, and Societal risks arising from collaborative cognition (RISc). We leverage this framework to propose concrete metrics for risk evaluation, and finally suggest specific mitigation strategies for developers and policymakers. As AI thought partners continue to proliferate, these strategies can help prevent major harms and ensure that humans actively benefit from productive thought partnerships.

Daniel Romero-Alvarado, Fernando Martínez-Plumed, Lorenzo Pacchiardi et al.

AI evaluation has primarily focused on measuring capabilities, with formal approaches inspired from Item Response Theory (IRT) being increasingly applied. Yet propensities - the tendencies of models to exhibit particular behaviours - play a central role in determining both performance and safety outcomes. However, traditional IRT describes a model's success on a task as a monotonic function of model capabilities and task demands, an approach unsuited to propensities, where both excess and deficiency can be problematic. Here, we introduce the first formal framework for measuring AI propensities by using a bilogistic formulation for model success, which attributes high success probability when the model's propensity is within an "ideal band". Further, we estimate the limits of the ideal band using LLMs equipped with newly developed task-agnostic rubrics. Applying our framework to six families of LLM models whose propensities are incited in either direction, we find that we can measure how much the propensity is shifted and what effect this has on the tasks. Critically, propensities estimated using one benchmark successfully predict behaviour on held-out tasks. Moreover, we obtain stronger predictive power when combining propensities and capabilities than either separately. More broadly, our framework showcases how rigorous propensity measurements can be conducted and how it yields gains over solely using capability evaluations to predict AI behaviour.

Marko Tesic, Yue Zhao, Joel Z. Leibo et al.

The development and evaluation of social capabilities in AI agents require complex environments where competitive and cooperative behaviours naturally emerge. While game-theoretic properties can explain why certain teams or agent populations outperform others, more abstract behaviours, such as convention following, are harder to control in training and evaluation settings. The Melting Pot contest is a social AI evaluation suite designed to assess the cooperation capabilities of AI systems. In this paper, we apply a Bayesian approach known as Measurement Layouts to infer the capability profiles of multi-agent systems in the Melting Pot contest. We show that these capability profiles not only predict future performance within the Melting Pot suite but also reveal the underlying prosocial abilities of agents. Our analysis indicates that while higher prosocial capabilities sometimes correlate with better performance, this is not a universal trend-some lower-scoring agents exhibit stronger cooperation abilities. Furthermore, we find that top-performing contest submissions are more likely to achieve high scores in scenarios where prosocial capabilities are not required. These findings, together with reports that the contest winner used a hard-coded solution tailored to specific environments, suggest that at least one top-performing team may have optimised for conditions where cooperation was not necessary, potentially exploiting limitations in the evaluation framework. We provide recommendations for improving the annotation of cooperation demands and propose future research directions to account for biases introduced by different testing environments. Our results demonstrate that Measurement Layouts offer both strong predictive accuracy and actionable insights, contributing to a more transparent and generalisable approach to evaluating AI systems in complex social settings.

Richard Evans, Jose Hernandez-Orallo, Johannes Welbl et al.

The Apperception Engine is an unsupervised learning system. Given a sequence of sensory inputs, it constructs a symbolic causal theory that both explains the sensory sequence and also satisfies a set of unity conditions. The unity conditions insist that the constituents of the theory - objects, properties, and laws - must be integrated into a coherent whole. Once a theory has been constructed, it can be applied to predict future sensor readings, retrodict earlier readings, or impute missing readings. In this paper, we evaluate the Apperception Engine in a diverse variety of domains, including cellular automata, rhythms and simple nursery tunes, multi-modal binding problems, occlusion tasks, and sequence induction intelligence tests. In each domain, we test our engine's ability to predict future sensor values, retrodict earlier sensor values, and impute missing sensory data. The engine performs well in all these domains, significantly outperforming neural net baselines and state of the art inductive logic programming systems. These results are significant because neural nets typically struggle to solve the binding problem (where information from different modalities must somehow be combined together into different aspects of one unified object) and fail to solve occlusion tasks (in which objects are sometimes visible and sometimes obscured from view). We note in particular that in the sequence induction intelligence tests, our system achieved human-level performance. This is notable because our system is not a bespoke system designed specifically to solve intelligence tests, but a general-purpose system that was designed to make sense of any sensory sequence.

Richard Evans, Jose Hernandez-Orallo, Johannes Welbl et al.

This paper attempts to answer a central question in unsupervised learning: what does it mean to "make sense" of a sensory sequence? In our formalization, making sense involves constructing a symbolic causal theory that both explains the sensory sequence and also satisfies a set of unity conditions. The unity conditions insist that the constituents of the causal theory -- objects, properties, and laws -- must be integrated into a coherent whole. On our account, making sense of sensory input is a type of program synthesis, but it is unsupervised program synthesis. Our second contribution is a computer implementation, the Apperception Engine, that was designed to satisfy the above requirements. Our system is able to produce interpretable human-readable causal theories from very small amounts of data, because of the strong inductive bias provided by the unity conditions. A causal theory produced by our system is able to predict future sensor readings, as well as retrodict earlier readings, and impute (fill in the blanks of) missing sensory readings, in any combination. We tested the engine in a diverse variety of domains, including cellular automata, rhythms and simple nursery tunes, multi-modal binding problems, occlusion tasks, and sequence induction intelligence tests. In each domain, we test our engine's ability to predict future sensor values, retrodict earlier sensor values, and impute missing sensory data. The engine performs well in all these domains, significantly out-performing neural net baselines. We note in particular that in the sequence induction intelligence tests, our system achieved human-level performance. This is notable because our system is not a bespoke system designed specifically to solve intelligence tests, but a general-purpose system that was designed to make sense of any sensory sequence.

Jose Hernandez-Orallo, Jan Arne Telle

We investigate the teaching of infinite concept classes through the effect of the learning bias (which is used by the learner to prefer some concepts over others and by the teacher to devise the teaching examples) and the sampling bias (which determines how the concepts are sampled from the class). We analyse two important classes: Turing machines and finite-state machines. We derive bounds for the biased teaching dimension when the learning bias is derived from a complexity measure (Kolmogorov complexity and minimal number of states respectively) and analyse the sampling distributions that lead to finite expected biased teaching dimensions. We highlight the existing trade-off between the bound and the representativeness of the sample, and its implications for the understanding of what teaching rich concepts to machines entails.

Jose Hernandez-Orallo

This note revisits the concepts of task and difficulty. The notion of cognitive task and its use for the evaluation of intelligent systems is still replete with issues. The view of tasks as MDP in the context of reinforcement learning has been especially useful for the formalisation of learning tasks. However, this alternate interaction does not accommodate well for some other tasks that are usual in artificial intelligence and, most especially, in animal and human evaluation. In particular, we want to have a more general account of episodes, rewards and responses, and, most especially, the computational complexity of the algorithm behind an agent solving a task. This is crucial for the determination of the difficulty of a task as the (logarithm of the) number of computational steps required to acquire an acceptable policy for the task, which includes the exploration of policies and their verification. We introduce a notion of asynchronous-time stochastic tasks. Based on this interpretation, we can see what task difficulty is, what instance difficulty is (relative to a task) and also what task compositions and decompositions are.

Jose Hernandez-Orallo

In this exploratory note we ask the question of what a measure of performance for all tasks is like if we use a weighting of tasks based on a difficulty function. This difficulty function depends on the complexity of the (acceptable) solution for the task (instead of a universal distribution over tasks or an adaptive test). The resulting aggregations and decompositions are (now retrospectively) seen as the natural (and trivial) interactive generalisation of the C-tests.

Jose Hernandez-Orallo

Artificial intelligence develops techniques and systems whose performance must be evaluated on a regular basis in order to certify and foster progress in the discipline. We will describe and critically assess the different ways AI systems are evaluated. We first focus on the traditional task-oriented evaluation approach. We see that black-box (behavioural evaluation) is becoming more and more common, as AI systems are becoming more complex and unpredictable. We identify three kinds of evaluation: Human discrimination, problem benchmarks and peer confrontation. We describe the limitations of the many evaluation settings and competitions in these three categories and propose several ideas for a more systematic and robust evaluation. We then focus on a less customary (and challenging) ability-oriented evaluation approach, where a system is characterised by its (cognitive) abilities, rather than by the tasks it is designed to solve. We discuss several possibilities: the adaptation of cognitive tests used for humans and animals, the development of tests derived from algorithmic information theory or more general approaches under the perspective of universal psychometrics.

David L. Dowe, Jose Hernandez-Orallo

The analysis of the adaptive behaviour of many different kinds of systems such as humans, animals and machines, requires more general ways of assessing their cognitive abilities. This need is strengthened by increasingly more tasks being analysed for and completed by a wider diversity of systems, including swarms and hybrids. The notion of universal test has recently emerged in the context of machine intelligence evaluation as a way to define and use the same cognitive test for a variety of systems, using some principled tasks and adapting the interface to each particular subject. However, how far can universal tests be taken? This paper analyses this question in terms of subjects, environments, space-time resolution, rewards and interfaces. This leads to a number of findings, insights and caveats, according to several levels where universal tests may be progressively more difficult to conceive, implement and administer. One of the most significant contributions is given by the realisation that more universal tests are defined as maximisations of less universal tests for a variety of configurations. This means that universal tests must be necessarily adaptive.

Jose Hernandez-Orallo

There has been an increasing interest in inferring some personality traits from users and players in social networks and games, respectively. This goes beyond classical sentiment analysis, and also much further than customer profiling. The purpose here is to have a characterisation of users in terms of personality traits, such as openness, conscientiousness, extraversion, agreeableness, and neuroticism. While this is an incipient area of research, we ask the question of whether cognitive abilities, and intelligence in particular, are also measurable from user profiles. However, we pose the question as broadly as possible in terms of subjects, in the context of universal psychometrics, including humans, machines and hybrids. Namely, in this paper we analyse the following question: is it possible to measure the intelligence of humans and (non-human) bots in a social network or a game just from their user profiles, i.e., by observation, without the use of interactive tests, such as IQ tests, the Turing test or other more principled machine intelligence tests?

Jose Hernandez-Orallo

We analyse the complexity of environments according to the policies that need to be used to achieve high performance. The performance results for a population of policies leads to a distribution that is examined in terms of policy complexity and analysed through several diagrams and indicators. The notion of environment response curve is also introduced, by inverting the performance results into an ability scale. We apply all these concepts, diagrams and indicators to a minimalistic environment class, agent-populated elementary cellular automata, showing how the difficulty, discriminating power and ranges (previous to normalisation) may vary for several environments.

Javier Insa-Cabrera, Jose-Luis Benacloch-Ayuso, Jose Hernandez-Orallo

This paper analyses the influence of including agents of different degrees of intelligence in a multiagent system. The goal is to better understand how we can develop intelligence tests that can evaluate social intelligence. We analyse several reinforcement algorithms in several contexts of cooperation and competition. Our experimental setting is inspired by the recently developed Darwin-Wallace distribution.