University of Oxford
Abstract:We study the capabilities of Large Language Models (LLM) on binary relations, a ubiquitous concept in math employed in most reasoning, math and logic benchmarks. This work focuses on equality, inequality, and inclusion, along with the properties they satisfy, such as ir/reflexivity, a/symmetry, transitivity, and logical complexity (e.g., number of reasoning ``hops''). We propose an alternative to in-context learning that trains only the representations of newly introduced tokens, namely out-of-context representation learning. This method mitigates linguistic biases already present in a model and, differently from in-context learning, does not rely on external information or illustrations. We argue out-of-context representation learning as a better alternative to in-context learning and fine-tuning to evaluate the capabilities of LLMs on logic tasks that are the building blocks of more complex reasoning benchmarks.
Abstract:Many reasoning, planning, and problem-solving tasks share an intrinsic algorithmic nature: correctly simulating each step is a sufficient condition to solve them correctly. We collect pairs of naturalistic and synthetic reasoning tasks to assess the capabilities of Large Language Models (LLM). While naturalistic tasks often require careful human handcrafting, we show that synthetic data is, in many cases, a good proxy that is much easier to collect at scale. We leverage common constructs in programming as the counterpart of the building blocks of naturalistic reasoning tasks, such as straight-line programs, code that contains critical paths, and approximate and redundant instructions. We further assess the capabilities of LLMs on sorting problems and repeated operations via sorting algorithms and nested loops. Our synthetic datasets further reveal that while the most powerful LLMs exhibit relatively strong execution capabilities, the process is fragile: it is negatively affected by memorisation and seems to rely heavily on pattern recognition. Our contribution builds upon synthetically testing the reasoning capabilities of LLMs as a scalable complement to handcrafted human-annotated problems.
Abstract:We discuss the "Infinitely Many Meanings" attacks (IMM), a category of jailbreaks that leverages the increasing capabilities of a model to handle paraphrases and encoded communications to bypass their defensive mechanisms. IMMs' viability pairs and grows with a model's capabilities to handle and bind the semantics of simple mappings between tokens and work extremely well in practice, posing a concrete threat to the users of the most powerful LLMs in commerce. We show how one can bypass the safeguards of the most powerful open- and closed-source LLMs and generate content that explicitly violates their safety policies. One can protect against IMMs by improving the guardrails and making them scale with the LLMs' capabilities. For two categories of attacks that are straightforward to implement, i.e., bijection and encoding, we discuss two defensive strategies, one in token and the other in embedding space. We conclude with some research questions we believe should be prioritised to enhance the defensive mechanisms of LLMs and our understanding of their safety.
Abstract:Communication is a prerequisite for collaboration. When scaling networks of AI-powered agents, communication must be versatile, efficient, and portable. These requisites, which we refer to as the Agent Communication Trilemma, are hard to achieve in large networks of agents. We introduce Agora, a meta protocol that leverages existing communication standards to make LLM-powered agents solve complex problems efficiently. In Agora, agents typically use standardised routines for frequent communications, natural language for rare communications, and LLM-written routines for everything in between. Agora sidesteps the Agent Communication Trilemma and robustly handles changes in interfaces and members, allowing unprecedented scalability with full decentralisation and minimal involvement of human beings. On large Agora networks, we observe the emergence of self-organising, fully automated protocols that achieve complex goals without human intervention.
Abstract:Language is not monolithic. While many benchmarks are used as proxies to systematically estimate Large Language Models' (LLM) performance in real-life tasks, they tend to ignore the nuances of within-language variation and thus fail to model the experience of speakers of minority dialects. Focusing on African American Vernacular English (AAVE), we present the first study on LLMs' fairness and robustness to a dialect in canonical reasoning tasks (algorithm, math, logic, and comprehensive reasoning). We hire AAVE speakers, including experts with computer science backgrounds, to rewrite seven popular benchmarks, such as HumanEval and GSM8K. The result of this effort is ReDial, a dialectal benchmark comprising $1.2K+$ parallel query pairs in Standardized English and AAVE. We use ReDial to evaluate state-of-the-art LLMs, including GPT-4o/4/3.5-turbo, LLaMA-3.1/3, Mistral, and Phi-3. We find that, compared to Standardized English, almost all of these widely used models show significant brittleness and unfairness to queries in AAVE. Furthermore, AAVE queries can degrade performance more substantially than misspelled texts in Standardized English, even when LLMs are more familiar with the AAVE queries. Finally, asking models to rephrase questions in Standardized English does not close the performance gap but generally introduces higher costs. Overall, our findings indicate that LLMs provide unfair service to dialect users in complex reasoning tasks. Code can be found at https://github.com/fangru-lin/redial_dialect_robustness_fairness.git.
Abstract:In agent-based simulations, synthetic populations of agents are commonly used to represent the structure, behaviour, and interactions of individuals. However, generating a synthetic population that accurately reflects real population statistics is a challenging task, particularly when performed at scale. In this paper, we propose a multi objective combinatorial optimisation technique for large scale population synthesis. We demonstrate the effectiveness of our approach by generating a synthetic population for selected regions and validating it on contingency tables from real population data. Our approach supports complex hierarchical structures between individuals and households, is scalable to large populations and achieves minimal contigency table reconstruction error. Hence, it provides a useful tool for policymakers and researchers for simulating the dynamics of complex populations.
Abstract:Theory of Mind (ToM) can be used to assess the capabilities of Large Language Models (LLMs) in complex scenarios where social reasoning is required. While the research community has proposed many ToM benchmarks, their hardness varies greatly, and their complexity is not well defined. This work proposes a framework to measure the complexity of ToM tasks. We quantify a problem's complexity as the number of states necessary to solve it correctly. Our complexity measure also accounts for spurious states of a ToM problem designed to make it apparently harder. We use our method to assess the complexity of five widely adopted ToM benchmarks. On top of this framework, we design a prompting technique that augments the information available to a model with a description of how the environment changes with the agents' interactions. We name this technique Discrete World Models (DWM) and show how it elicits superior performance on ToM tasks.
Abstract:Causal games are probabilistic graphical models that enable causal queries to be answered in multi-agent settings. They extend causal Bayesian networks by specifying decision and utility variables to represent the agents' degrees of freedom and objectives. In multi-agent settings, whether each agent decides on their policy before or after knowing the causal intervention is important as this affects whether they can respond to the intervention by adapting their policy. Consequently, previous work in causal games imposed chronological constraints on permissible interventions. We relax this by outlining a sound and complete set of primitive causal interventions so the effect of any arbitrarily complex interventional query can be studied in multi-agent settings. We also demonstrate applications to the design of safe AI systems by considering causal mechanism design and commitment.
Abstract:Multi-armed bandits (MAB) and causal MABs (CMAB) are established frameworks for decision-making problems. The majority of prior work typically studies and solves individual MAB and CMAB in isolation for a given problem and associated data. However, decision-makers are often faced with multiple related problems and multi-scale observations where joint formulations are needed in order to efficiently exploit the problem structures and data dependencies. Transfer learning for CMABs addresses the situation where models are defined on identical variables, although causal connections may differ. In this work, we extend transfer learning to setups involving CMABs defined on potentially different variables, with varying degrees of granularity, and related via an abstraction map. Formally, we introduce the problem of causally abstracted MABs (CAMABs) by relying on the theory of causal abstraction in order to express a rigorous abstraction map. We propose algorithms to learn in a CAMAB, and study their regret. We illustrate the limitations and the strengths of our algorithms on a real-world scenario related to online advertising.
Abstract:We investigate the extent to which Large Language Models (LLMs) can simulate the execution of computer code and algorithms. We begin by looking at straight line programs, and show that current LLMs demonstrate poor performance even with such simple programs -- performance rapidly degrades with the length of code. We then investigate the ability of LLMs to simulate programs that contain critical paths and redundant instructions. We also go beyond straight line program simulation with sorting algorithms and nested loops, and we show the computational complexity of a routine directly affects the ability of an LLM to simulate its execution. We observe that LLMs execute instructions sequentially and with a low error margin only for short programs or standard procedures. LLMs' code simulation is in tension with their pattern recognition and memorisation capabilities: on tasks where memorisation is detrimental, we propose a novel prompting method to simulate code execution line by line. Empirically, our new Chain of Simulation (CoSm) method improves on the standard Chain of Thought prompting approach by avoiding the pitfalls of memorisation.