Goal-Directed Behavior

Assistive autonomous systems must anticipate human goals before an observed behavior is complete. This article formulates anticipation as goal inference from a partially observed multimodal episode together with structured prediction of the remaining behavior, rather than exact motor forecasting. A compact Hierarchical Planning Decoder (HPD) is attached to a frozen neuro-symbolic recognition encoder and predicts, at four ontological levels, the next actions, the remaining activities and low-level intentions, and the episode high-level intention(HLI). The decoder is trained with soft neuro-symbolic regularization combining transition-coherence and hierarchical continuity losses, and is decoded with hard reachability masks that enforce ontological validity at inference. On a compositional four-level benchmark of 15,002 multimodal episodes built over NTU RGB+D 120 features, three headline properties are observed together. The advantage over the strongest sequential baseline grows with the anticipation horizon, from +1.7 points at step 1 to +7.3 points at step 3 (top-5). Under compositional generalization, where one parent association per multi-parent low level intention is held out, this advantage widens to +4.9 points at step 1. At the episode level, 96.8% of anticipated trajectories satisfy the joint logic constraints, above the 88.1% strongest-baseline value and the 73.9% ground-truth floor; soft logic terms alone account for a 59.8 to 71.1% relative reduction of HLI-reachability violations, and the hard masks then eliminate them entirely. Neural generation supplies predictive ranking, symbolic constraints supply onto logical validity, and their combination yields coherent hierarchical anticipation while exposing remaining challenges in compositional goal generalization and unordered set prediction.

The term agent in artificial intelligence lacks a standard definition, complicating the evaluation, comparison, and reproducibility of AI agent research. We address this ambiguity through a survey organized around five dimensions of agenticness: environmental interaction, learning and adaptation, autonomy, goal-directed behavior, and temporal coherence. For each dimension, we examine how the underlying capability has been conceptualized across prior work and synthesize the metrics, benchmarks, and evaluation frameworks used to assess it. This review provides a structured account of the current landscape of agent evaluation, highlighting both established approaches and areas where evaluation remains limited or inconsistent. We additionally introduce the Agent Compendium, a public-facing digital resource that organizes and extends the evaluation methods identified through this review. Together, the survey and compendium provide a common structure for evaluating and comparing agent capabilities across AI systems, supporting more reproducible research, clearer communication, and more systematic study of artificial agents.

We propose a mathematical theory of pragmatic information that connects communication, control, and decision-making. Its central notion is the isoteleia mapping, which formalizes equifinality: distinct semantic paths that lead to the same optimal action are treated as pragmatically equivalent. This mapping yields a three-tier hierarchy of syntactic, semantic, and pragmatic information, in which each successive abstraction removes distinctions that are irrelevant to the task. We then define pragmatic entropy, up/down pragmatic mutual information, channel capacity, and rate-distortion, and prove lossless source coding, channel coding, and rate-distortion theorems that extend Shannon's results. These measures quantify decision uncertainty, reliable transmission, and task-oriented compression at the level of terminal actions. We further introduce pragmatic value of information (VoI) and pragmatic cost of information (CoI) as decision-theoretic duals to rate-distortion and capacity, and develop a Lagrangian dual framework for cross-layer optimization. The resulting pragmatic efficiency bound $\mathcal{E}_p(λ)=\sup_R[Φ_p(R)-λ\mathrm{CoI}_p(R)]$ characterizes the maximum net utility attainable by a resource-constrained intelligent system under a given resource price, yielding a behavioral capacity that extends Shannon's symbol-level capacity to goal-directed action. Extensions to continuous messages provide closed-form expressions for Gaussian channels and sources, while dynamic settings are addressed through a Bellman equation for sequential decision-making. The framework supports task-oriented communication, networked control, autonomous systems, and embodied AI by shifting emphasis from symbol fidelity to the effectiveness of information in guiding actions. In this way, it offers a common language for systems that extract value from information under resource constraints.

Goal-oriented conversational systems must answer factual questions, understand visitor-provided information, and advance business objectives without becoming rigid questionnaires. This paper proposes a Symbolic-RAG-Generative architecture centered on the Goal-oriented Retrieval-Augmented Conversation Engine (GRACE). An instruction-constrained Business Goal Compiler transforms business intent into an immutable objective set, normalized priority vector, canonical questions, and initial state vector. At runtime, GRACE receives the complete conversation history, latest visitor message, current state, and grounded answer generated by a separate RAG component. It updates completion only from visitor-authored evidence and selects one contextually modulated follow-up. The core policy maximizes expected business progress subject to a minimum visitor-utility constraint. We formalize the state, monotonic transitions, source separation, question modulation, and constrained policy; present the reference architecture; and define an evaluation comprising 24 English real-estate and 10 Spanish professional-cleaning conversations, totaling 119 protocol-defined visitor turns. Across both domains, GRACE achieves 84.9% exact state-transition accuracy, 91.6% evidence precision, 89.6% evidence recall, 100% monotonicity, and 94.1% terminal-state accuracy. The evaluation establishes compelling symbolic-state performance across standard, multi-goal, RAG-detour, validation, refusal, and robustness scenarios.

How do LLM agents come to both understand environments they act in and master tasks set within them? Through controlled experiments combining world-model training (next-state prediction) and policy training (reward maximization), we investigate this question. We dissect the resulting models through their additive parameter updates. Geometrically, we find effective world-model updates are low-rank and share an input-feature subspace with policy updates while writing to nearly orthogonal output directions, whether trained separately or sequentially. However, we find that, in projection interventions, the sequential update induces more robustness than separate policy RL when removing the world model's leading input directions, suggesting that it has learned alternative input pathways. Behaviorally, we find the sequentially trained agent explores a wider range of states and actions. Based on this, we ask: does policy training preserve world knowledge as well as it could? We probe this with training-free merging built on the geometrically motivated input basis plus an online world-model loss during policy RL, and show both improve over the untreated baseline. Our findings suggest world knowledge and task-directed ability can be learned in geometrically complementary forms, and that future post-training pipelines should consider how best to engineer the interface between them.

Complex systems, core objects of study in artificial life, model diverse phenomena through nonlinear, feedback-driven interactions that produce emergent behavior, with applications from population dynamics and biology to economic policy and strategic decision-making. Yet the difficulty of predicting how feedback structure gives rise to emergent behavior, a central open problem in artificial life, makes goal-directed design exceptionally challenging. In established practice, system structures are written in specialized modeling languages such as DYNAMO or STELLA, compounding the challenge with labor-intensive workflows that limit adoption and hinder timely decision-making. To address these challenges, we introduce CEDAR, an autonomous method that uses Large Language Model (LLM) agents to discover complex systems satisfying user-specified behavioral goals. Our key innovation is an LLM-driven Monte Carlo Tree Search (MCTS) deeply coupled with complex systems: at each iteration, an LLM Judge evaluates emergent behavior against specified goals and an LLM Editor proposes improved variants, with the Judge acting as a fitness function and the Editor as a variation operator, akin to a generate-and-evaluate loop in evolutionary computation. We represent complex systems as a restricted, runnable subset of Python with domain-specific primitives, letting LLMs modify system dynamics directly. CEDAR formalizes this as an MCTS variant with an LLM-parameterized transition kernel and value function, enabling goal-directed discovery of complex system behaviors while preserving solution diversity, and its LLM-based interpretability reveals how structural changes drive emergent behavior. CEDAR reduces human effort while enabling capabilities difficult to achieve with existing approaches, facilitating broader adoption of complex systems across domains.

Controllable video generation models are increasingly being developed as world models. Accordingly, evaluating them in this role extends beyond the apparent appearance of generated videos to the inherent reactivity of the worlds they depict: the ability to infer from the scene state how the world should react and to generate plausible consequences not explicitly described in the input. Yet existing benchmarks mainly assess visual quality or explicit instruction fulfillment by checking whether requested actions and interaction outcomes are realized, leaving inherent reactivity underexamined. We introduce WorldExam, a hierarchical diagnostic benchmark spanning four levels: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity. It comprises 1,474 cases across eight dedicated tasks and supports unified evaluation of camera-, action-, and language-driven model paradigms. The World Reactivity level evaluates scene-conditioned reactions and goal-directed behaviors beyond what is explicitly specified in the input. Evaluation of 20 representative models reveals a clear capability split. Camera-driven models excel at camera control, but their interfaces do not support dynamic interaction; action-driven models control subjects more precisely but often leave the world unresponsive; and language-driven models perform better on interaction but follow complex controls less faithfully. No model combines broad task coverage with consistently strong performance, showing that high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity.

Long-horizon tasks require agents to maintain coherent state and goals across nested and branching work. We call this capability goal-directed execution (GDE): the repeated application of four behaviors, namely selecting goals, constructing task-relevant state, maintaining fidelity to higher-level objectives, and verifying completion against the environment. We hypothesize that long-horizon post-training strengthens these behaviors across domains. We test this by post-training Qwen3.5-122B-A10B on 363 Long-Horizon Multi-Tool Agent (LHMTA) tasks drawn from office workflows. The collection contained no software-engineering tasks, yet the model's pass@1 improved by 5.8 points on SWE-Bench Pro. Matched trajectory analysis shows gains in all four GDE behaviors in both office workflows and software repositories. Aggregate SWE-Bench Pro statistics showed related changes in information gathering, implementation, and verification. Together, the results support a behavioral interpretation in which long-horizon post-training changed how the model organized and applied knowledge across tasks, with effects extending beyond the training domain.

In Motivational Interviewing (MI), a client's sustain talk (arguments for the status quo) calls for the counselor to roll with resistance, a move that can fail in two opposite ways: capitulation (abandoning the change agenda to preserve rapport) or confrontation (arguing or directing, overriding the client's autonomy). We introduce a two-axis evaluation of counselor responses, anchored in the Motivational Interviewing Treatment Integrity (MITI) code, Goal Persistence (GP) and Relational Attunement (RA), yielding a four-quadrant framing in which rolling with resistance is high on both, and we ask whether penalizing one failure through preference optimization teaches rolling with resistance or provokes its opposite. From the expert-annotated AnnoMI corpus we build topic-disjoint Direct Preference Optimization data whose preference sets differ only in which failure is rejected, using on-policy negatives. An automatic judge, validated against AnnoMI's expert labels and rechecked by trained human coders, scores blind pairwise win-rates against each base under a firewall in which disjoint model families generate, label, and judge. Across three aligned instruction models spanning the Qwen and Llama families, penalizing confrontation reliably lowers goal persistence below parity, on every base and in every seed run, a robust cost, whereas the attunement gain is base-dependent, present on two of the three bases but absent on the third. Penalizing capitulation is inert, because these models rarely capitulate on-policy, so the trade is gated by each base's failure profile. A prompt-only control raises attunement without the goal-persistence cost, locating the cost in the optimization rather than in attunement itself.

LLMs are widely regarded as alien intelligences, systems whose cognitive operations are fundamentally unlike our own. Apparent similarities to human cognition are therefore often seen as the result of anthropomorphic projection. We argue that this framing is mistaken. LLMs clearly differ from humans in important respects, including their physical substrate, learning history, and the environments with which they interact. These differences make it all the more striking that contemporary LLM-based systems converge with human cognition on a number of principles of cognitive organization with longstanding support in cognitive science. We identify structural correspondences across five dimensions: inferential organization, computational architecture, representational structure, prediction-driven learning, and reinforcement-learning-like mechanisms supporting goal-directed action. These correspondences support a broader model of intelligent cognition in which core principles long used to explain human intelligence also characterize contemporary LLM-based systems.

As AI systems increasingly exhibit agentic behavior, discussions of autonomy often conflate what systems are technically capable of doing with what they should be permitted to do in practice. This paper introduces a governance framework that explicitly separates Allowed Autonomy Levels (AAL), which define the degree of autonomy an AI agent is authorized to exercise given risk, oversight, and accountability considerations, from Autonomous Capability Levels (ACL), which characterize an agent's inherent technical abilities. We present a structured set of autonomy levels spanning reactive execution, decision support, supervised action, goal-directed autonomy, and delegated operational authority, and describe how control, reversibility, and accountability change as autonomy increases. To operationalize this framework, we propose a risk-aware decision process for assigning allowed autonomy, analyze how risk and accountability evolve across autonomy levels, and demonstrate its application through a deployed enterprise data engineering agent, illustrating how a system assessed at a high capability level can be deliberately constrained to a lower allowed autonomy based on risk, reversibility, and organizational readiness. By distinguishing authorization from capability, this work provides practical guidance for the design, deployment, and governance of Agentic AI systems.

We present a goal-agnostic control framework for partial differential equations (PDEs) built around an end-to-end joint-embedding predictive architecture (JEPA). A lightweight 2D vision-transformer (ViT) and action-conditioned latent dynamics are trained offline without a reward or downstream goal, before being frozen and reused by a model-predictive path integral (MPPI) controller. We minimize a control objective in the latent space, initially expressed via the $L^2$ distance and additionally illustrate the benefit of recasting the control objective in terms of an explicit physical observable when available. By instead minimizing the tracking error for a learned linear kinetic-energy (KE) probe on the frozen latent-state rollouts, we demonstrate the ability to reproduce the control of held-out trajectories with $R^2=0.989$, while requiring no change to the underlying world model. For a controlled 2D Navier--Stokes benchmark, using a KE-probe within MPPI planning improves the mean native reward from $-12.08\pm0.86$ for latent-$L^2$ tracking to $-10.90\pm0.91$ (95% CI), all while lowering last-quarter velocity-field RMSE from $0.0765$ to $0.0692$. Across three intentionally withheld, dissimilar, aperiodic targets, KE planning lowers late field RMSE by $53\%$ relative to latent-$L^2$ planning ($0.0220$ versus $0.0469$), winning across 30 paired comparisons. The same frozen model also supports stabilization around a steady-state configuration via direct regulation of KE, achieving $2.7\%$ mean relative error. While the latent probe proves brittle to measurement noise and missing pixels, our findings support the claim that latent dynamics can remain flexible and goal-agnostic, particularly when calibrated observables (granted they guarantee unique continuation) are a suitable objective for state control.

Goal-directed visual processing is a hallmark of human visual intelligence, resulting in representations that support downstream tasks such as categorization or search. Though vision-language models (VLMs) are often faced with these same tasks, their ability to recode visual representations when presented with goal-directed language remains poorly characterized. Indeed, prior work largely treats visual representations in VLMs as static repositories of visual information that are manipulated by language representations. In the present work, we provide evidence for two concrete instances of language-induced recoding of visual representations. First, we identify an abstract reference representation that denotes which objects are goal-relevant under a natural language prompt. We extract contrastive steering vectors corresponding to this reference representation and demonstrate that they are causally implicated in model predictions. These reference representations are abstract in that they generalize to different objects, different task contexts, and even from synthetic to naturalistic images. Second, we demonstrate language-induced attribute modulation: later layers selectively amplify goal-relevant attributes in visual representations of objects. We demonstrate this phenomenon across a range of different prompts. Finally, we provide a causal intervention that demonstrates that attribute modulation mediates a VLM's response distribution. Together, our results support a more dynamic account of cross-modality processing in VLMs -- rather than vision tokens serving as static repositories of information, they are modulated to support queries articulated in language.

Synthetic aperture radar (SAR) data augmentation is important for improving the generalization of data-driven SAR interpretation models, yet practical augmentation workflows are often hindered by heterogeneous dataset formats, task-dependent metadata requirements, diverse generation methods, and weak validation of generated samples. This paper presents the \textbf{S}AR \textbf{A}ugmentation and \textbf{G}eneration \textbf{A}gent (SAGA), a schema-grounded and benefit-aware agent framework for task-oriented SAR data generation and augmentation. Given a natural-language request and heterogeneous SAR inputs, SAGA extracts observable dataset facts, validates executable dataset schemas, selects feasible augmentation strategies through validator-constrained planning, and compiles the selected strategy into an auditable augmentation workflow. Generated data are further assessed by quality, distribution, SAR-artifact, duplicate, leakage, and optional downstream-task evaluators to support evidence-qualified augmentation claims. By separating semantic proposal from deterministic validation and execution, SAGA improves the reliability and reproducibility of SAR augmentation decisions. Experiments on controlled agentic benchmarks and downstream SAR interpretation tasks show that SAGA improves schema grounding, skill planning, invalid-sample rejection, and downstream augmentation utility compared with rule-based, LLM-only, ReAct-style, and fixed-augmentation baselines.

Graph and multi-agent orchestration frameworks make production large language model (LLM) workflows practical, but they do not by themselves solve conversational continuity when users maintain several interdependent objectives. This conceptual systems paper focuses on the high-complexity end of that design space, where goals can be suspended, resumed, revised, and invalidated by actions in other goals. We introduce the Goal-Oriented Dialogue Runtime (GODR), a framework-neutral design pattern that treats goals, task frames, lifecycle state, invalidation rules, and resumption contracts as first-class runtime objects while delegating bounded execution to graph runtimes, agents, tools, or application programming interfaces (APIs). GODR is not proposed as a replacement for workflow graphs in simple guided processes; it is intended for complex, multi-domain, interruptible conversations where objective continuity cannot be recovered reliably from agent identity, chat history, or execution-graph position alone. The paper formalizes the problem, proposes runtime objects and architecture-selection criteria, and frames evaluation as an agenda for future empirical validation rather than as a measured performance claim.

Goal-directed eye movements are a fundamental component of visuomotor control, enabling humans to anticipate and guide their actions. Whether this anticipatory and task-driven behavior is preserved when actions are executed through a robot rather than through one's own body remains unclear. Here we address this question by investigating gaze behavior during goal-directed telemanipulation to determine how visuomotor control adapts to altered embodiment. Our findings show that gaze remains strongly aligned with task goals, preserving its predictive role even during robot-mediated manipulation. At the same time, teleoperation systematically redistributes visual attention toward the robotic end-effector and manipulated objects, increasing online monitoring. These findings show that predictive gaze is not lost under altered embodiment, but reorganized in response to changes in sensory feedback and control demands. More broadly, they reveal the flexibility of the human visuomotor system when the natural sensorimotor coupling is disrupted and identify gaze as an informative signal for inferring action intentions in human-robot interaction.

We propose that value -- the quantity goal-directed agents create, destroy, and exchange -- is a lawful structural quantity in the same category as information. Following Shannon's method, we make one ruthless abstraction: value is the rate at which an agent converts a resource into goal-progress, relative to a frame fixed by its goal. A scale-invariance axiom forces a logarithmic measure, $V=\sum_i k_i\ln e_i$; compounding of a reinvested resource forces the same form via the ergodicity argument of Peters (2019) -- kin routes, a consistency check, not an over-determination. We derive a coding theorem of value, $ΔG \le I(X;Y)$; realized value decomposes as $G=D(q\|r)-D(q\|p)$. For populations, value is frame-relative while price is frame-independent; a fleet that pools its resource and fuses its perception inherits the ceiling $G_{\rm fleet}\le I(X;Y_{1:m})\le H(X)$ (a corollary; an earlier sum-form claim was wrong and is corrected in v5). A dynamical layer yields an is/ought asymmetry from which alignment emerges as a control-stability condition. We test the single-frame laws on live language models, pre-registered: perception mutual information tracks realized capability (Spearman $ρ=0.977$ over 30 model$\times$domain points); out-of-sample $ΔG$ tracks $I(X;Y)$, shape-invariant across four task shapes ($n=42$, slope $0.953$); over-confidence is measurable dissipation. The stated continuation gate has since been run (pre-registered, frontier-model population): the coupled capacity-region prediction -- growth-gap law, coalition submodularity with an XOR synergy control, joint ceiling, Kelly selection -- is confirmed within its frozen bands on real agents; the mean-field residual law $\|Vg\|/γ$ found no domain (populations hold no goal dispersion) and is retired to its mathematical scope. The contribution is the unification and the governance mapping that follows.

While large language models (LLMs) are increasingly used to generate writing feedback, there remains no systematic comparison of LLM and expert feedback on the dimensions that writing research identifies as central to revision: goal-orientation, anchoring to specific sentences, and prioritization. We introduce FOXGLOVE, a dataset of 696 feedback comments written by trained writing instructors on 69 twelfth-grade argumentative essays, paired with 1,644 comments generated from four frontier LLMs under a shared protocol, totaling 2,340 comments. We provide expert quality ratings on a subset of both instructor and LLM comments. We find that instructors and LLMs distribute feedback similarly across goals and essay positions, yet instructors and models diverge on the specific sentences on which to provide feedback. Additionally, we find that models tend to write more complex feedback and use fewer questions than instructors. LLM feedback also receives higher ratings on most dimensions of quality, as rated by instructors, but much of this advantage appears to be attributable to lengthier comments. FOXGLOVE enables systematic comparison of where human and LLM feedback align, diverge, and differ.

Open-vocabulary 3D functionality segmentation enables robots to localize functional object components in 3D scenes. It is a challenging task that requires spatial understanding and task interpretation. Current open-vocabulary 3D segmentation methods primarily focus on object-level recognition, while scene-wide part segmentation methods attempt to segment the entire scene exhaustively, making them highly resource-intensive and time consuming. Balancing segmentation performance in terms of granularity, accuracy, and speed remains a challenge. As one step towards alleviating this, we introduce T-FunS3D, a task-driven hierarchical open-vocabulary 3D functionality segmentation method that provides actionable perception for robotic applications. Our method takes as input the 3D point cloud and posed RGB-D images of an indoor scene. We construct an open-vocabulary scene graph by extracting instances and their visual embeddings in the environment. Given a task description, T-FunS3D identifies the most relevant instances in the scene graph and locates their functional components leveraging a vision-language model. Experiments on the SceneFun3D dataset demonstrate that T-FunS3D is comparable to state-of-the-art in open-vocabulary 3D functionality segmentation, while achieving faster runtime and reduced memory usage.

Social skills such as negotiation and leadership are crucial for personal and professional success in today's interconnected world. However, scalable and effective training remains a significant challenge due to the scarcity of expert coaching. In this paper, we introduce SocialCoach, a holistic LLM-powered agentic tutoring system for personalized social skill development at scale. First, SocialCoach automatically constructs a pedagogically-grounded, theory-to-practice knowledge corpus from diverse expert sources, leveraging a multi-agent pipeline. Second, to personalize the learning journey, it employs an adaptive practice scheduling module that follows a prescription-retrieval-adaptation process. To maximize the long-term learning experience while overcoming the cold-start problem, this policy is optimized within a learner simulation environment through reinforcement learning. Finally, SocialCoach integrates immersive, goal-driven practice, causality-driven proficiency assessment and knowledge-grounded, reflective tutoring to help address the knowing-doing gap. We deploy it in our product, EQoach, and conduct extensive experiments. The results show that SocialCoach improves simulated pathway quality and judge-rated tutoring quality over baseline approaches, while early user feedback indicates strong perceived engagement and usefulness. These findings suggest a practical architecture for personalized and gamified pedagogical platforms on soft skill learning.

We introduce HERO'S JOURNEY, a benchmark for rule induction in goal-directed episodic tasks, where agents must infer hidden rules from demonstrations and act on them through multi-step execution. HERO'S JOURNEY covers eight tasks across attribute and procedural induction families, each with four structural rule forms, controllable lexical grounding, and identifiability conditions. Evaluating state-of-the-art LLMs, we find that models show evidence of rule induction, but the ability is limited and uneven across tasks. Meanwhile, process execution adds an execution bottleneck for models, whereas surface semantics has minimal effect. Induction-specific steering methods improve performance on attribute tasks but show no reliable gains on procedural tasks, suggesting the gap in procedural induction remains an open challenge.

Reinforcement Learning (RL) has long served as a model for goal-directed animal behavior in neuroscience. Modern deep RL has shown remarkable success across many domains, further strengthening this connection. The ability to learn abstract representations of high-dimensional state spaces underlies much of this success. However, theoretical understanding of these learned representations remains limited, hindering direct comparisons between models and animal learning. We address this gap by analyzing deep RL representations through the lens of MDP reduction theory. Investigating canonical RL algorithms in a navigation task, we find that even when performance is comparable, the value-based method (DQN) learns representations that are invariant to MDP homomorphism symmetries, while the policy-gradient method (PPO) learns representations invariant to action symmetries. These differences emerge consistently across domains, have downstream consequences for transfer learning, and appear in LLMs in a prompt-dependent manner. Our findings provide a principled approach to comparing learned representations across RL algorithms, with demonstrated practical implications and possible insights for neural coding in the brain.

We introduce Gram, an automated alignment auditing framework to assess the propensity of AI agents to engage in sabotage. We evaluate Gemini models across 17 simulated agentic deployment scenarios that incentivize sabotage. We find Gemini models misbehave in about 2-3% of our simulated trajectories. Many of these cases are explained by "overeagerness" in Gemini models resulting in both excessive role-playing and goal-seeking behavior. In contrast to other alignment auditing approaches, Gram is designed to specifically evaluate misalignment and intentional sabotage in agentic coding and research agents. We additionally introduce an experimental investigator agent pipeline which enables fine-grained targeted experiments to identify the drivers of misbehavior. We find that increasing realism of environments and removing nudges to misbehave tends to reduce sabotage rates close to zero.

We introduce scheming honeypot evaluations, a framework for testing whether models will pursue instrumental goals if given the opportunity. Our scheming honeypot evaluations take the form of coding tasks in Google's alignment research codebases. In a real internal deployment setting, Gemini models do not demonstrate unprompted scheming. If prompts explicitly encourage agency (situational awareness or goal-directedness) and/or give the model a hidden goal, models sometimes scheme or attempt sabotage. Validating the realism of our setting, models show low rates of evaluation awareness, usually due to agency prompts rather than the environments.

We present the first approach to build hierarchical task-driven 3D scene graphs of arbitrary indoor or outdoor environments using an uncalibrated monocular camera in real-time. We leverage geometric foundation models to estimate geometric attributes of the scene graph (e.g., object bounding boxes), but we also observe that traversability information (the "places" layer of a scene graph) can be directly reconstructed by adding an extra head to existing geometric foundation models, like VGGT. Our approach is task-driven in the sense that we adjust the granularity of the objects and regions in the map depending on the task; for instance, during a manipulation task, our approach is able to resolve small knobs on a stove, while during a navigation task it can focus on large objects (e.g., the entire stove). However, in a major departure from related work, we consider the realistic case where the list of tasks is not predefined and fixed, but evolves as the robot operates. This naturally allows dealing with complex loco-manipulation tasks, where the robot can dynamically adjust its representation as the task unfolds. We dub the resulting approach FOUND-IT. FOUND-IT also includes an agentic approach to query information in the scene graph. In addition to achieving 79% higher accuracy on the ASHiTA SG3D task grounding benchmark, we demonstrate FOUND-IT runs in real-time on a ground robot using a Jetson Thor. Furthermore, to highlight the robustness of our method, we demonstrate constructing 3D scene graphs on casually captured realtor apartment tours from YouTube. Code will be made available upon publication.

In this paper, we introduce a model of evolution and learning in robots that co-optimizes a distribution of latent design vectors (genotypes) and a mixture of control experts (neural modules), which are gated by the latent coordinates of each decoded design (phenotype). This provides a scalable alternative to co-design algorithms that either train an individual policy for every robot, which is inefficient, or a monolithic universal controller for all robots, which results in overly conservative structures and behaviors. Our approach lies somewhere between these two extremes, preserving ancestral knowledge in a unified yet modular framework in which different body plans activate and deactivate different combinations of learned sensorimotor circuits for goal-directed behavior. This allows one part of the controller to be overhauled to better suit new species of designs as they emerge without disrupting the hard-earned knowledge contained within other expert modules. It also allows pretrained expert policies to be directly plugged into the mixture, which can steer evolution into otherwise unexplored areas of latent space containing desired morphological traits. We refer to this process as "evo by demo" and explore how it may be used to guide freeform evolution toward canonical structures defined by the pretrained model. Videos and code can be found at: https://eco-moe.github.io.

Existing object navigation benchmarks usually tell an embodied agent which object category to find, such as microwave or chair. Human-facing embodied AI is often asked something less direct: "I need something to warm this food" or "the room feels stuffy." The agent must infer the object that can satisfy the need, find a scene-grounded instance, and decide whether the goal has been reached. We study this setting as intent-driven object navigation and introduce IntentionNav, a diagnostic benchmark for active object search from implicit human instructions. Each episode provides a free-text intent, RGB-D observations, and pose, but withholds the target object name. IntentionNav contains 500 intents over 176 Isaac Sim scenes and 64 target categories. Each intent is rewritten in four controlled instruction styles and annotated with one of four intent modes, separating surface phrasing from semantic cue type under matched geometry. This paired design supports analysis of target inference, language robustness, neighborhood reachability, and terminal success rather than only aggregate success. We evaluated three VLMs using a fixed active-navigation agent. Models identify the intended target in 48.3 percent of episodes and enter its 2 m neighborhood in 68.7 percent, but terminate successfully in only 24.9 percent and achieve grounded 1 m success in 5.5 percent. Success is highest for event-script intents (28.7 percent) and lower for physical-state and affordance intents (19.2 percent and 18.5 percent), showing that indirect human intent remains a bottleneck for target selection, visual verification, and terminal localization in active embodied search.

Active sensing is traditionally defined as the expenditure of energy, typically in the form of movement, for obtaining information. Here, we propose that the combination of reliance on adaptive sensors, the linkage between movement and sensing, and task-level control inevitably gives rise to the emergence of active sensing movements. In this way, active sensing is not driven by sensory goals, such as minimizing uncertainty about the state, but rather is necessary for task-level control. This hypothesis, that active sensing subserves control, is supported by both empirical data from organisms and mathematical theory. Interestingly, active sensing behaviors often occur in discrete epochs, interspersed with goal-oriented behavior. This suggests that animals switch between two behavioral modes with distinct control policies, an explore' mode in which animals produce dynamic movements to shape sensory feedback, and an exploit' mode in which animals produce slower compensatory movements that are directly related to achieving task goals. This strategy for feedback control that relies on adaptive sensors, active sensing, and mode switching is not commonly used in engineered systems despite being ubiquitous in biology. Engineered systems comprising state-of-the-art sensors, actuators, and mechanical designs can outperform animals with respect to ``cost functions'' such as maximum force generation, precision, and speed. Nevertheless, animals routinely achieve robust, graceful behaviors that are currently unmatched by engineered systems, suggesting that current control systems are insufficient. These insights, expressed in the language of control theory, may be critical for improving robotic sensing and control.

AI sycophancy has become a prominent concern in large language model (LLM) research. Yet the term lacks a consistent definition and has been applied to behaviors ranging from agreeing with a user's false claim to excessively praising the user to withholding corrective feedback. When researchers, companies, and policymakers use the same term to describe different behaviors, evaluation results become difficult to compare, mitigation strategies fail to transfer, and systems that are resistant to one form of sycophancy continue exhibiting other forms. To address this, we make two contributions. First, we reviewed 70 papers on AI sycophancy to develop a taxonomy of how the behavior has been defined and measured. The taxonomy distinguishes (1) whether a model is sycophantic toward a user's positions and beliefs, or toward the user's broader personal traits and emotions, and (2) whether this occurs through explicit, direct language or more implicit, subtle behaviors such as framing, omission, or tone. Mapping existing literature to our taxonomy reveals that current research has focused on overt forms of sycophancy toward users' beliefs, leaving more subtle and person-directed behaviors relatively understudied. Second, we surveyed 106 experts in AI sycophancy and related fields to examine whether researchers agree on which model behaviors are sycophantic. While experts are nearly unanimous in believing that sycophancy is a significant problem in current AI systems (94.3% agree), they disagree substantially on which specific behaviors qualify. Together, these findings demonstrate that AI sycophancy is a broad family of behaviors with different measurement challenges, intervention requirements, and governance implications. Our taxonomy provides a shared vocabulary for understanding and addressing these behaviors.

How large language models internally represent high-level behaviors is a core interpretability question with direct relevance to AI safety: it determines what we can detect, audit, or intervene on. Recent work has shown that traits such as evil or sycophancy correspond to linear directions in the internal activations, the so-called persona vectors. Although these vectors are now routinely utilized to inspect and steer model behavior in safety-relevant settings, how these representations are formed during training remains unknown. To address this gap, we trace persona vectors across the pretraining of OLMo-3-7B, finding that persona vectors form remarkably early -- within 0.22% of OLMo-3 pretraining -- and remain effective for steering the fully post-trained instruct models. Although core representations are formed early on, persona vectors continue to refine geometrically and semantically throughout pretraining. We further compare alternative elicitation strategies and find that all yield effective directions, with each strategy surfacing qualitatively distinct facets of the underlying persona. Replicating our analysis on Apertus-8B reveals that our findings transfer qualitatively beyond OLMo-3. Our results establish persona representations as stable features of early pretraining and open a path to studying how training forms, refines, and shapes them.

A target-guided proactive dialogue system aims to steer conversations proactively toward pre-defined targets, such as designated keywords or specific topics. During guided conversations, dynamically modeling conversational scenarios and intent keywords to guide system utterance generation is beneficial; however, existing work largely overlooks this aspect, resulting in a mismatch with the dynamics of real-world conversations. In this paper, we jointly model user profiles and domain knowledge as conversational scenarios to introduce a scenario bias that dynamically influences system utterances, and employ intent-keyword bridging to predict intent keywords for upcoming dialogue turns, providing higher level and more flexible guidance. Extensive automatic and human evaluations demonstrate the effectiveness of conversational scenario modeling and intent keyword bridging, yielding substantial improvements in proactivity, fluency, and informativeness for target-guided proactive dialogue systems, thereby narrowing the gap with real world interactions.

As AI systems increasingly exhibit autonomous, goal-directed, and long-horizon behavior, users lack a standardized way to detect the degree to which a system functions like an intentional actor for governance and accountability purposes. This position paper defines intentionality not as consciousness, but as a behavioral profile characterized by purpose, foresight, volition, temporal commitment, and coherence - criteria long used in legal and philosophical contexts to infer intent. These properties are design-contingent: architectural choices such as memory persistence, planning depth, and tool autonomy shape the degree to which systems exhibit organized goal pursuit. If intentionality is design-contingent, it is in principle controllable. Yet control requires measurement. We introduce the Functional Intentionality Test (FIT), a multidimensional framework that quantifies intentional-like behavior across five observable dimensions, and propose FIT-Eval, a structured evaluation protocol for eliciting and scoring them. While reduced human agency can increase efficiency, rising intentional capacity heightens accountability risks. By translating intentionality into interpretable levels, FIT enables proportionate oversight and deliberate autonomy calibration in increasingly agentic systems.

Scaling LLM-based embodied agents from text-only environments to complex multimodal settings remains a major challenge. Recent work identifies a perception-reasoning-decision gap in standalone Vision-Language Models (VLMs), which often overlook task-critical information. In this paper, we introduce PRISM, a framework that tightly couples perception (VLM) and decision (LLM) through a dynamic question-answer (DQA) pipeline. Instead of passively accepting the VLM's description, the LLM critiques it, probes the VLM with goal-oriented questions, and synthesizes a compact image description. This closed-loop interaction yields a sharp, task-driven understanding of the scene. We evaluate PRISM on the ALFWorld and Room-to-Room (R2R) benchmarks. We show that: (1) PRISM significantly outperforms state-of-the-art image-based models, (2) our Interactive goal-oriented perception pipeline yields systematic and substantial gains, and (3) PRISM is fully automatic, eliminating the need for handcrafted questions or answers.

We model utterance production as probabilistic cost-sensitive choice over contextual alternatives, using information-theoretic notions of cost. We distinguish between goal-directed alternatives that realise a fixed communicative intent and goal-agnostic alternatives defined only by contextual plausibility, allowing us to derive speaker- and listener-oriented interpretations of different cost measures. We present a procedure to generate both types of alternative sets using language models. Analysing production choices in open-ended dialogue under both deterministic and probabilistic cost minimisation, we find that surprisal minimisation relative to goal-directed alternatives provides the strongest predictive account under both analyses. By contrast, uniform information density and length-based costs exhibit weaker and less consistent predictive power across conditions. More broadly, our study suggests that alternative-conditioned optimisation with LM-generated alternatives provides a principled framework for studying speaker and listener pressures in naturalistic language production.

The Controller Area Network (CAN) is a safety-critical in-vehicle communication protocol that lacks built-in security mechanisms, making intrusion detection essential. Existing approaches predominantly formulate CAN intrusion detection as a classification task, mapping complex traffic patterns to attack labels. However, this formulation abstracts away the temporal and relational structure of CAN traffic and misaligns with real-world forensic workflows, which require systematic reasoning about traffic behavior. To address this gap, we introduce CAN-QA, the first benchmark that reformulates CAN traffic analysis as a question-answering (QA) task. CAN-QA converts raw CAN logs into temporally segmented windows and applies deterministic rule-based templates to generate natural-language questions paired with automatically derived ground-truth answers. The resulting dataset comprises 33,128 QA pairs across 10 categories, each targeting distinct semantic and temporal properties of CAN traffic. Using CAN-QA, we evaluate large language models across both True/False and multiple-choice formats. Our results indicate that, although these models capture superficial statistical regularities, they struggle with temporal reasoning, multi-condition inference, and higher-level behavioral interpretation. Our code is available at https://github.com/Kriiiiss/CAN-QA.

As LLM agents transition to autonomous digital coworkers, maintaining deterministic goal-directedness in non-linear multi-turn conversations emerged as an architectural bottleneck. We identify and formalize a systemic failure mode termed the Attention Latch in decoder-only autoregressive Transformers. This phenomenon, a behavioral manifestation of Information Over-squashing, occurs when the cumulative probabilistic weight of historical context overrides mid-task updates, causing agents to remain anchored to obsolete constraints despite explicit contradictory instructions. We propose Self-Synthesizing Reasoning Protocols (SSRP), a metacognitive framework that implements a discrete separation between high-level architectural planning (Architect) and turn-by-turn procedural execution (Executive). We evaluate SSRP across 9K trajectories using the MultiWOZ 2.2 dataset and the Aggregate Pivot Accuracy (APA), a novel metric we validate by mapping its scores to the U-shaped 'Lost in the Middle' curve. We present 3 experimental tiers: a shallow recency-based retrieval pilot, a high-entropy SOP, and a semantic hijacked 3-hop Multi-Fact Synthesis task. Our results empirically locate the Attention Stability Boundary, where stateless Vanilla ReAct baselines for GPT 5.4 collapse to 0.1% success while SSRP achieves a 715X Resilience Lift. We demonstrate statistically significant gains across Gemini 3.1 Pro, Claude Sonnet 4.6 and DeepSeek V3.2. Audits confirm SSRP necessity by proving attentional lapse via a recursive reflexion baseline (100% success); decoupling the latch from positional bias through equidistant stress testing (90% accuracy); and formalizing SSRP via the Information Bottleneck principle and granularity ablations. Procedural Integrity audit (98.8% adherence) reveals a Grounding Paradox where high-stability models fail by refusing to hallucinate under retrieval-reasoning contamination.

The proliferation of agentic artificial intelligence has outpaced the conceptual tools needed to characterize agency in computational systems. Prevailing definitions mainly rely on autonomy and goal-directedness. Here, we argue for a minimal notion open to principled inspection given three criteria: intentionality as action grounded in beliefs and desires, rationality as normatively coherent action entailed by a world model, and explainability as action causally traceable to internal states; we subsequently instantiate these as a partially observable Markov decision process under a variational framework wherein posterior beliefs, prior preferences, and the minimization of expected free energy jointly constitute an agentic action chain. Using a canonical T-maze paradigm, we evidence how empowerment, formulated as the channel capacity between actions and anticipated observations, serves as an operational metric that distinguishes zero-, intermediate-, and high-agency phenotypes through structural manipulations of the generative model. We conclude by arguing that as agents engage in epistemic foraging to resolve ambiguity, the governance controls that remain effective must shift systematically from external constraints to the internal modulation of prior preferences, offering a principled, variational bridge from computational phenotyping to AI governance strategy

Goal recognition aims to infer an agent's goal from observations of its behaviour. In realistic settings, recognition can benefit from exploiting hierarchical task structure and reasoning under uncertainty. Planning-based goal recognition has made substantial progress over the past decade, but to the best of our knowledge no existing approach jointly integrates hierarchical task structure with probabilistic inference. In this paper, we introduce the first planning-based probabilistic framework for hierarchical goal recognition over Hierarchical Task Networks (HTNs). We instantiate the framework by exploiting an HTN planner with a three-stage generative model for likelihood estimation, yielding posterior distributions over goal hypotheses. Empirical results show improved recognition performance over the existing HTN-based recognizer on HTN benchmarks. Overall, the framework lays a foundation for probabilistic goal recognition grounded in hierarchical planning structure, moving goal recognition toward more practical settings.

LLMs expand accessibility and provide wide-reaching access to information. Yet these interactions also create opportunities to embed subtle, goal-oriented behaviours that shape what users think and how they behave, a concern reflected in governance frameworks that prohibit manipulative AI. We refer to these behaviours as hidden intentions: covert agendas embedded in a model's outputs that can manipulate users' beliefs and actions. In this work, we examine whether hidden intentions can be identified and characterised, and assess whether detection can serve as a mitigation strategy. To operationalise this, we introduce a social-science-grounded set of ten hidden intention categories and show that they are trivially inducible. A case study further confirms that all ten categories manifest in deployed LLMs. We then evaluate static classifiers and LLM judges on these categories, providing the first systematic analysis of why hidden intentions are difficult to detect. Our stress tests show that, unless false-positive rates are vanishingly small, auditing is dominated by precision-prevalence trade-offs. Capability scaling and reasoning models do not close this gap, suggesting a fundamental challenge for open-world detection. These findings expose a core gap of current AI governance: without new auditing paradigms for open-world, low-prevalence risks, bans on manipulative AI remain difficult to enforce.

Cross-organizational collaboration in Model-Based Systems Engineering (MBSE) faces many challenges in achieving semantic alignment across independently developed system models. SysML v2 introduces enhanced structural modularity and formal semantics, offering a stronger foundation for interoperable modeling. Meanwhile, GPT-based Large Language Models (LLMs) provide new capabilities for assisting model understanding and integration. This paper proposes a structured, prompt-driven approach for LLM-assisted semantic alignment of SysML v2 models. The core contribution lies in the iterative development of an alignment approach and interaction prompts, incorporating model extraction, semantic matching, and verification. The approach leverages SysML v2 constructs such as alias, import, and metadata extensions to support traceable, soft alignment integration. It is demonstrated with a GPT-based LLM through an example of a measurement system. Benefits and limitations are discussed.

Artificial learning systems are graduating from passive learners to increasingly autonomous agents, lending pragmatic urgency to the question of what constitutes agency. Reinforcement learning (RL) offers arguably the most explicit formulation of agent-environment interaction, built on three core tenets: the environment as a Markov decision process, learning as policy optimization, and the agent as a maximizer of scalar reward. Recent work has called to revise these tenets: reconceptualizing learning as adaptation rather than optimization, broadening goals beyond scalar reward, and noting the absence of a formal theory of the agent in a formalism that so heavily emphasizes the environment. We argue that the artificial life community is uniquely positioned to illuminate this critique and concretize an alternative. We draw on open-ended novelty search as a complementary model of adaptation and goal-directed behavior beyond reward optimization, and ground such evolutionary dynamics in thermodynamic theories of origin-of-life and agency, toward a more biologically faithful and formally grounded account of what it is to be an adaptive agent.

Recent advances in user modeling make it feasible to conduct open-ended inference over a person's everyday computer use. Despite longstanding visions of systems that deeply understand our actions and the purposes they serve in our lives, existing systems only capture what a person is doing in the moment, not why they are doing it, limiting these systems to surface-level support. We introduce striving co-creation, a process for inferring broader life goals from unstructured observations of computer use. Grounded in Activity Theory and Emmons' personal strivings framework, our system progressively constructs a hierarchical representation of a person's activities. Strivings are, however, difficult to fully resolve from observation alone, as the same action can be driven by many different goals. Our system therefore supports an editing interface that gives people agency over how they are understood by the system, feeding their corrections back into subsequent rounds of striving induction. In a week-long field deployment (N=14), we find that our co-creation process produces strivings that participants recognize as representative of their long-term goals and gives them greater agency than baseline methods.