The recent shift toward the Health Level Seven Fast Healthcare Interoperability Resources (HL7 FHIR) standard opens a new frontier for clinical AI, demanding LLM agents to navigate complex, resource-based data models instead of conventional structured health data. However, existing benchmarks have lagged behind this transition, lacking the realism needed to evaluate recent LLMs on interoperable clinical data. To bridge this gap, we introduce FHIR-AgentBench--a benchmark that grounds 2,931 real-world clinical questions in the HL7 FHIR standard. Using this benchmark, we systematically evaluate agentic frameworks, comparing different data retrieval strategies (direct FHIR API calls vs. specialized tools), interaction patterns (single-turn vs. multi-turn), and reasoning strategies (natural language vs. code generation). Our experiments highlight the practical challenges of retrieving data from intricate FHIR resources and the difficulty of reasoning over them--both of which critically affect question answering performance.

Reinforcement learning (RL) using foundation models for policy approximations in multi-turn tasks remains challenging. We identify two main limitations related to sparse reward settings and policy gradient updates, based on which we formulate a key insight: updates from positive samples with high returns typically do not require policy regularisation, whereas updates from negative samples, reflecting undesirable behaviour, can harm model performance. This paper introduces Succeed or Learn Slowly (SoLS), a novel off-policy RL algorithm evaluated on mobile app control tasks. SoLS improves sample efficiency when fine-tuning foundation models for user interface navigation via a modified off-policy actor-critic approach, applying direct policy updates for positive samples and conservative, regularised updates for negative ones to prevent model degradation. We augment SoLS with Successful Transition Replay (STR), which prioritises learning from successful interactions, further improving sample efficiency. We evaluate SoLS on the AndroidWorld benchmark, where it significantly outperforms existing methods (at least 17% relative increase), including prompt-engineering and RL approaches, while requiring substantially fewer computational resources than GPT-4o-based methods with 5-60x faster inference.

Multi-task reinforcement learning aims to quickly identify solutions for new tasks with minimal or no additional interaction with the environment. Generalized Policy Improvement (GPI) addresses this by combining a set of base policies to produce a new one that is at least as good -- though not necessarily optimal -- as any individual base policy. Optimality can be ensured, particularly in the linear-reward case, via techniques that compute a Convex Coverage Set (CCS). However, these are computationally expensive and do not scale to complex domains. The Option Keyboard (OK) improves upon GPI by producing policies that are at least as good -- and often better. It achieves this through a learned meta-policy that dynamically combines base policies. However, its performance critically depends on the choice of base policies. This raises a key question: is there an optimal set of base policies -- an optimal behavior basis -- that enables zero-shot identification of optimal solutions for any linear tasks? We solve this open problem by introducing a novel method that efficiently constructs such an optimal behavior basis. We show that it significantly reduces the number of base policies needed to ensure optimality in new tasks. We also prove that it is strictly more expressive than a CCS, enabling particular classes of non-linear tasks to be solved optimally. We empirically evaluate our technique in challenging domains and show that it outperforms state-of-the-art approaches, increasingly so as task complexity increases.

To build AI that children can intuitively understand and benefit from, designers need a design grammar that serves their developmental needs. This paper bridges artificial intelligence design for children - an emerging field still defining its best practices - and animation, a well established field with decades of experience in engaging children through accessible storytelling. Pairing Piagetian developmental theory with design pattern extraction from 52 works of animation, the paper presents a six scaffold framework that integrates design insights transferable to child centred AI design: (1) signals for visual animacy and clarity, (2) sound for musical and auditory scaffolding, (3) synchrony in audiovisual cues, (4) sidekick style personas, (5) storyplay that supports symbolic play and imaginative exploration, and (6) structure in the form of predictable narratives. These strategies, long refined in animation, function as multimodal scaffolds for attention, understanding, and attunement, supporting learning and comfort. This structured design grammar is transferable to AI design. By reframing cinematic storytelling and child development theory as design logic for AI, the paper offers heuristics for AI that aligns with the cognitive stages and emotional needs of young users. The work contributes to design theory by showing how sensory, affective, and narrative techniques can inform developmentally attuned AI design. Future directions include empirical testing, cultural adaptation, and participatory co design.

Imagine an autonomous car driving along a quiet suburban road when suddenly a dog runs onto the road. The system must brake hard and decide, within a fraction of a second, whether to swerve into oncoming traffic--where the other autonomous car might make space--to steer right and hit the roadside barrier, or to continue straight and injure the dog. The first two options risk only material damage; the last harms a living creature. Each choice is justifiable and involves trade-offs between safety, property and ethical concerns. However, today's autonomous systems are not designed to explicitly take such value-laden conflicts into account.

Reward machines (RMs) are automata structures that encode (non-Markovian) reward functions for reinforcement learning (RL). RMs can reward any behaviour representable in regular languages and, when paired with RL algorithms that exploit RM structure, have been shown to significantly improve sample efficiency in many domains. In this work, we present pushdown reward machines (pdRMs), an extension of reward machines based on deterministic pushdown automata. pdRMs can recognise and reward temporally extended behaviours representable in deterministic context-free languages, making them more expressive than reward machines. We introduce two variants of pdRM-based policies, one which has access to the entire stack of the pdRM, and one which can only access the top $k$ symbols (for a given constant $k$) of the stack. We propose a procedure to check when the two kinds of policies (for a given environment, pdRM, and constant $k$) achieve the same optimal state values. We then provide theoretical results establishing the expressive power of pdRMs, and space complexity results for the proposed learning problems. Lastly, we propose an approach for off-policy RL algorithms that exploits counterfactual experiences with pdRMs. We conclude by providing experimental results showing how agents can be trained to perform tasks representable in deterministic context-free languages using pdRMs.

Establishing a communication system is hard because the intended meaning of a signal is unknown to its receiver when first produced, and the signaller also has no idea how that signal will be interpreted. Most theoretical accounts of the emergence of communication systems rely on feedback to reinforce behaviours that have led to successful communication in the past. However, providing such feedback requires already being able to communicate the meaning that was intended or interpreted. Therefore these accounts cannot explain how communication can be bootstrapped from non-communicative behaviours. Here we present a model that shows how a communication system, capable of expressing an unbounded number of meanings, can emerge as a result of individual behavioural differences in a large population without any pre-existing means to determine communicative success. The two key cognitive capabilities responsible for this outcome are behaving predictably in a given situation, and an alignment of psychological states ahead of signal production that derives from shared intentionality. Since both capabilities can exist independently of communication, our results are compatible with theories in which large flexible socially-learned communication systems like language are the product of a general but well-developed capacity for social cognition.

We report a structural convergence among four influential theories of mind: Kahneman's dual-system theory, Friston's predictive processing, Minsky's society of mind, and Clark's extended mind, emerging unintentionally within a practical AI architecture known as Agentic Flow. Designed to address the limitations of large language models (LLMs), Agentic Flow comprises five interlocking modules: Retrieval, Cognition, Control, Memory, and Action, organized into a repeatable cognitive loop. Although originally inspired only by Minsky and Clark, subsequent analysis revealed that its structure echoes computational motifs from all four theories, suggesting that theoretical convergence can emerge naturally from implementation demands rather than deliberate synthesis. Controlled evaluations confirmed this: the structured agent achieved 95.8% task success versus 62.3% for baseline LLMs, demonstrating robust constraint adherence and reproducible reasoning. We describe this convergence under a broader descriptive meta-architecture called PEACE, highlighting recurring design patterns such as predictive modeling, associative recall, and error-sensitive control. Later formalized as the Structured Cognitive Loop (SCL), this framework generalizes the same principles as a foundation for behavioral intelligence in LLM-based agents. Rather than claiming theoretical unification, this paper proposes that intelligent architectures may evolve toward shared structural patterns shaped by practical constraints. As a position paper, it aims to frame this convergence as an interpretive reflection rather than a finalized theory, inviting further theoretical and experimental dialogue. Agentic Flow, or equivalently the Structured Cognitive Loop, thus offers a glimpse of how a unified cognitive form can arise not from abstraction, but from the necessities of real-world reasoning.

Digital tool-based agents, powered by Large Language Models (LLMs), that invoke external Application Programming Interfaces (APIs) often rely on documentation to understand API functionality. However, such documentation is frequently missing, outdated, privatized, or inconsistent-hindering the development of reliable, general-purpose agents. In this work, we propose a new research direction: learning of API functionality directly from in-context demonstrations. This task is a new paradigm applicable in scenarios without documentation. Using API benchmarks, we collect demonstrations from both expert agents and from self-exploration. To understand what information demonstrations must convey for successful task completion, we extensively study how the number of demonstrations and the use of LLM-generated summaries and evaluations affect the task success rate of the API-based agent. Our experiments across 3 datasets and 6 models show that learning functionality from in-context demonstrations remains a non-trivial challenge, even for state-of-the-art LLMs. We find that providing explicit function calls and natural language critiques significantly improves the agent's task success rate due to more accurate parameter filling. We analyze failure modes, identify sources of error, and highlight key open challenges for future work in documentation-free, self-improving, API-based agents.