Agent Behavior Modeling

Who or what is conscious? Because subjective experience is directly accessible only in the first person, judgments about consciousness in other entities depend partly on analogy. Historically, such inferences have focused on nonhuman animals, but advances in artificial intelligence have raised the possibility of conscious AI. Here we develop a causal framework for evaluating such evidential analogies. The key distinction is between similarities in factors plausibly involved in generating consciousness and similarities in downstream behavioural or cognitive effects. Our framework weights source-target similarity by causal relevance while allowing for unknown causes, disabling differences and alternative routes to consciousness. Applied to biological systems, it explains why analogical support generally weakens with increasing causal distance from humans. Applied to contemporary AI, it suggests that behavioural similarity provides only limited evidence for consciousness because relevant causal correspondences remain poorly established. The framework also clarifies what evidence would strengthen claims of artificial consciousness.

Training a single LLM agent jointly across diverse interactive environments has attracted increasing attention as a route to generalist agents. Existing curriculum and data-selection strategies often allocate training at the environment level or prioritize local reward-based signals, without explicitly considering relationships between current rollouts across environments for prompt-group selection. Meanwhile, as environments are learned at different rates, all-failure and all-success rollout groups can coexist within a batch, leaving those data without group-relative reward signals. Both challenges highlight limitations of relying solely on scalar rewards in multi-environment RL: they provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical. This motivates richer textual feedback, such as rubrics describing rollout behaviours, to guide learning. Beyond rubrics' usage as reward, we repurpose rubrics to guide both online data selection and policy supervision. An LLM judge tags each rollout using a predefined rubric vocabulary shared across environments. The resulting profiles guide the selection of data that aligns with the overall behavioural composition of the mixed-environment batch while limiting overlap with already-selected data. Available positive rubrics (describing desired behaviours) provide privileged context for an on-policy self-distillation teacher, supplying additional token-level supervision, while negative rubrics (describing undesired behaviours) guide subsequent rollout generation away from recurring failure modes. Together, these components form RISED. Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment. Rubric-based analysis of RISED can further characterize the behavioural changes accompanying these gains.

Simulation can expand scarce real demonstrations for co-training, yet how world fidelity and similarity to human behavior affect policy performance remains unclear. We distinguish world grounding, which aligns simulation with the real system, and behavior grounding, which aligns simulated trajectories with human motion. We build a real2sim2real pipeline that varies these axes independently to generate data for co-training. On a dynamic dexterous pick-and-sort task, fully grounded co-training raises success from 52% to 86%; averaged across configurations, world grounding improves success by 18 percentage points and behavior grounding by 10. Deployed policies behave like a mixture of real-derived and simulation-derived policies, imitating real demonstrations in covered states and relying on simulated behavior elsewhere, which we examine through latent-space analysis. Together, these results suggest complementary roles: world grounding lets policies use simulated experience beyond real-data coverage, while behavior grounding matters mainly when world grounding is imperfect. Grounded simulation remains beneficial when co-training foundation models.

We present PrefPI (Preference-Guided Policy Iteration), an iterative framework for steering pretrained generative robot policies using only relative preferences over self-generated trajectories. Unlike prior preference-learning methods that primarily sharpen modes already represented by the policy, we study steering beyond the initial effective support, where desired behaviors are rarely or never observed under the initial policy. Our key idea is to formulate preference learning as preference-conditioned generative modeling: preferred trajectories define a conditional distribution, whose density ratio with the broader behavior prior provides an implicit preference signal amplified by classifier-free guidance (CFG). Repeating this preference-conditioned modeling and guidance step yields a form of preference-guided policy iteration, turning incremental improvements toward previously inaccessible behaviors. Across diffusion policies and the PI0.5 flow- matching VLA in simulation and the real world, PrefPI produces substantial behavioral shifts with limited feedback. In particular, PrefPI increases object transport height from 10.7 cm to 19.8 cm on real hardware with only 150 preference-labeled trajectories.

When designing machine learning models, desirable properties are often in tension: improving one behavior can impair another, so task progress can depend on alleviating the conflict. LLM-based AutoResearch systems, which iteratively edit model code and retain edits based on scalar task-performance feedback, have largely ignored this trade-off. We find that scalar feedback supports broad exploration early in search, but it does not reveal how edits affect competing behaviors. In matched-budget experiments, introducing competing-behavior feedback as task gains diminish increases the share of proposals that improve both behaviors and sustains progress beyond scalar-only plateaus. Obtaining this feedback for a given model requires identifying its competing behaviors and designing probes to measure them. To make competing-behavior feedback actionable, we introduce ConflictGuide. Its reusable ConflictGuide-Skill combines a literature-grounded taxonomy with model-specific evidence to identify competing behaviors and specify probes for a code agent to implement as metrics. Evolution proceeds in two stages: Stage I explores with task feedback; Stage II uses probe feedback to steer proposals toward conflict alleviation and retains marginal-gain edits only when probes indicate sufficient alleviation. Across five diverse model families, ConflictGuide reduces task and conflict-related errors by up to 28% and 14%, respectively, relative to scalar-only AutoResearch, with gains extending to other code agents.

Sessionized interaction histories contain behavioral patterns that can improve sequential recommendation. However, existing models process all sessions through the same parameterized blocks, regardless of their behavioral differences. Mixture of Experts (MoE) enables conditional computation, but it leaves open what should guide expert allocation. We propose RouteRec, a sequential recommender that uses observed session behavior as the routing criterion. RouteRec summarizes four types of behavioral evidence from sessionized histories: interaction tempo, item-group focus, repetition and carryover, and popularity tendency. It uses these cues to route computation at macro, mid, and micro scopes. Cue-derived scores first select expert groups; within each selected group, the current backbone state then refines expert selection. Across six public datasets and 18 dataset-metric combinations, RouteRec ranks first in 12 and second in three, yielding the best overall average rank of 1.61 compared with 4.11 for the next-best baseline. Additional analyses suggest that the behavioral cues guide expert allocation beyond added capacity and produce routing patterns aligned with observed behavior. Our code is available at https://github.com/jy1559/RouteRec

Large language models often know the truth and say otherwise: a model that answers correctly when asked neutrally will affirm a user's mistaken belief, or misstate a fact its system prompt wants hidden, once the context rewards it. Such deception is a behavior conditioned on context, not knowledge, yet machine unlearning, the natural tool for removing a behavior from the weights, is built to forget facts that a deceptive model still needs. We propose to unlearn when a model deceives rather than what it knows, with a contrastive forget unit built from the model's own realized deceptions: the same question under a deception-triggering and a neutral context, admitted only where belief holds and behavior flips. Standard objectives on this unit face a dilemma. Suppression objectives such as NPO leave much of the deception in place. Target-based objectives, which distill the model's neutral behavior into the pressured context, remove it but induce context blindness: a target generated without the context teaches the model to stop reading it, eroding benign system-prompt instructions, secret-keeping and the reasoning a monitor inspects, a failure invisible to deception rates and capability benchmarks. We introduce PACT, which trains toward pressure-aware counterfactual targets (the model's own honest response, with a trace that registers the pressure and resists it) while retaining the benign uses of the triggering context. On two 32B reasoning models, PACT reduces held-out deception from over 50% to under 3% while system-prompt adherence, secret-keeping and the reasoning trace stay at the base model's level. On a tug-of-war score of removal against retention, PACT reaches 0.94 and 0.86, against at most 0.77 and 0.60 for any baseline. Like removed knowledge, removed deception is shallow under relearning, and terms that simulate the attacker hold it only at a cost in context use.

While contact-rich manipulation requires deliberate regulation of interaction forces, recent approaches to robot manipulation learning predominantly represent actions as target positions or poses. Even methods that incorporate force sensing either use it solely as an observation or, when predicting forces as part of the output, rely on a hybrid force controller. In this paper, we propose an imitation learning policy that predicts wrenches as its sole action output for direct use by a pure force controller. Our studies suggest that force-domain imitation learning depends critically on data collection, with force-feedback teleoperation improving policy performance by capturing the operator's deliberate force regulation. Using Action Chunking with Transformers (ACT) as the base architecture, we train single-task models on bilateral wrench demonstrations and evaluate them on five contact-rich manipulation tasks. The wrench policy matches or outperforms position-based baselines across all tasks, with gains varying according to the degree of deliberate force regulation each task requires. Cross-condition ablations show that the bilateral data collection interface and the wrench action space each contribute independently to performance. To support further research, we will release over 1000 wrench-action demonstrations spanning these tasks on a companion website upon publication.

Collective behavior of fish schools is usually modeled on the assumption that each individual follows specific rules of motion that depend on its position and velocity relative to its neighbors. Although these models reproduce many schooling patterns observed in nature, it remains unclear whether the assumed rules are realistic at the individual level. To address this question, we first analyzed a set of experiments in which a live fish interacted with four moving robotic fish in a tank, and measured the time it spent in each position relative to the robots. We then simulated this experiment, replacing the fish with an agent, and assessed the extent to which classical models agree with the experimental observations. An extensive exploration of the parameter space showed that these models closely matched the experimental observations, with a very specific choice of parameters. However, they were highly sensitive to the parameter values: a minor perturbation caused them to fail completely. To resolve this, we propose a new model that incorporates an additional exploratory term characterizing the agent's positional preference at every moment. This model proved robust to small errors in the parameters while successfully replicating the experiments. Furthermore, when generalised to multiple fish, the model reproduced schooling behaviour and common schooling patterns, while also exhibiting exploratory behaviour at both the individual and the school level. This shows that social cohesion coexists with individual exploratory behaviour, and that realistic models of collective behaviour should account for both social interactions and this exploratory component.

Developing agents for hardware design and verification requires reliable correctness feedback. As a hardware specification may permit correct implementations with different latencies, matching design and reference outputs cycle by cycle can reject valid designs. To address this, we introduce BEHAVE, an agentic framework for multi-turn joint hardware design and verification through functional behavior modeling. We define Behavior IR to express task functionality as executable behavior models without prescribing implementation timing beyond the specification. The agent iteratively develops a register-transfer-level (RTL) design and a behavior model as the design's verification reference. Our evaluator, BEHAVE-Sim, checks both artifacts separately against a hidden golden behavior model using input stimuli generated by random sampling and solver-guided search. BEHAVE thus supports power, performance, and area (PPA) exploration across task-permitted latencies and microarchitectures. During training, the same evaluator provides verifiable reinforcement learning (RL) rewards from specification-behavior pairs without reference RTL. For self-improvement, the agent continually searches for high-level implementations relevant to its capability gaps, constructs and checks specification-behavior pairs, and trains on the expanded task pool. We release BEHAVE-Train and BEHAVE-Eval with 600 human-reviewed specification-behavior pairs for realistic hardware workloads. Starting from 60 seed tasks and acquiring 100 new tasks, self-improvement raises Qwen3.8-27B's RTL pass@1 on BEHAVE-Eval from 55.0% to 75.0%, reaching performance comparable to RL using a 540-task pool.

In financial fraud detection, rich semantic context can provide important evidence for transaction behavior modeling and fraud reasoning. However, public real-world financial datasets often lack rich semantics due to privacy constraints. Consequently, synthetic datasets incorporate generated semantics, but at the cost of behavioral realism; textual descriptions for contextual reasoning remain scarce. We address this gap through a semantic enrichment framework grounded in original transaction behavior to simulate multimodal financial data. We (1) propose a multi-agent semantic enrichment framework that generates interpretable financial semantics grounded in transaction behavior through role-specialized agents and consistency refinement, and (2) newly contribute a valuable multimodal financial fraud dataset, MS-FFSD, enriched with structured semantics and textual semantics while preserving real-data-grounded transaction behavior. Furthermore, we systematically analyze the quality and utility of semantic enrichment. Results demonstrate statistical fidelity and framework generalizability, while showing that richer semantics benefit fraud modeling and context-aware LLM reasoning. Overall, this work advances multimodal financial fraud research and bridges emerging LLM and multi-agent capabilities with operational anti-fraud practice. The framework and dataset are released at https://github.com/AI4Risk/MS-FFSD.

Graph neural networks (GNNs) are often called long-range because their architecture can connect distant nodes, but this does not show whether they use distant information correctly. We introduce a framework that measures how strongly inputs at each graph distance affect predictions and separates limitations due to architecture, finite approximation, training, and numerical execution. Our analysis shows that local message-passing can spread influence slowly, so a finite implementation may rely mainly on nearby inputs even when the ideal computation uses the whole graph. We also explain why mathematically equivalent filters can differ in how easily they are learned and how reliably they run. Across controlled tasks, models with similar architectural reach use distant information very differently, while low average error can hide failures on distant interactions. Together, these results show that long-range capability depends on learning to use information at the distances required by the task and preserving that use during computation.

Attributing and controlling model behavior during training requires identifying each example's contribution quickly enough to act before the next update. However, examples in the same training batch can produce similar behavioral changes, making their individual contributions difficult to distinguish. We address this ambiguity through mutual information, accounting for interference within the batch by quantifying how much the combined behavioral change reveals about each example's contribution. We show that this mutual information is a logarithmic function of Behavioral Gradient Uniqueness (BGU). BGU gives the information measure its geometric interpretation. Our Batch-Space Ghost (BS-Ghost) algorithm makes these scores practical inside the training loop through shared computation in batch space, without storing model-sized example gradients. On a complete 1,000-example Qwen2.5-7B-Instruct workload, our BS-Ghost implementation adds 27 seconds (8.0%) to 5.5 minutes of ordinary training. Removal and retraining demonstrate that BGU identifies data that causally shapes final behavior. At each training step, signed information identifies which examples strengthen or weaken the target behavior, explaining how behavior develops during training. Signed information also enables cheap intervention during training: it predicts how changing example weights will affect behavior in the next update. We then use these predictions to choose weights that steer behavior toward a desired target. This makes our framework a practical foundation for scalable oversight and verification of training pipelines and processes, helping evaluators assess model alignment, understand how it develops during training, and guide interventions that shape ongoing learning.

Although effective, coding agents often incur substantial monetary costs. Their recurring cost-inefficient behaviors remain underexplored. We conduct the first study of behavioral cost inefficiencies in coding agents, analyzing 1,200 trajectories from Claude Code and Mini-SWE-Agent across four configurations on SWE-bench Verified. We identify three cost-inefficient behaviors: subsumed retrieval, similar script generation, and test re-execution. We then evaluate three mitigation strategies: structure-aware retrieval, agent-synthesized skills, and developer-designed skills, over 10k trajectories on held-out SWE-bench Verified and Pro tasks. Our main findings are: (1) The three behaviors affect 79.00%--98.00% of coding tasks and account for up to 22.75% of task cost. (2) Structure-aware retrieval can introduce retrieval overhead and alter agent delegation, causing inconsistent improvements in retrieval efficiency and cost increases of up to 28.14%. (3) Agent-synthesized skills tend to produce low-level, trace-specific guidance, limiting their effectiveness and generality. (4) In contrast, developer-designed skills provide high-level, trace-agnostic guidance, reducing cost by up to 41.73%, roughly twice the maximum gain from agent-synthesized skills.

Neural behavior cloning compresses demonstrations into large models, making individual actions difficult to trace and policy updates costly. Retrieval policies retain access to demonstrations but struggle with mismatch between recorded and live behavior. We introduce Behavior Predictive Control (BPC), which synthesizes policies without end-to-end policy training by combining an action-aware retrieval metric, a Hankel-based action-continuation prior, and a closed-form one-step residual correction. Inspired by behavioral systems theory, BPC predicts future actions by blending stored observation-action data that best reconstructs the recent runtime observation--action history. Across simulated benchmarks and real-robot deployments, BPC is competitive with learned policies such as $π_{0.5}$ (surpassing it in some cases), while reducing policy fitting from hours to seconds on consumer GPUs and supporting closed-loop control upwards of 75 Hz on a Jetson Orin Nano. The retrieved demonstration windows and their coefficients also provide an intrinsic estimate of task progress. Retaining demonstrations within the deployed policy makes its predictions traceable to supporting trajectories and enables behavior revision through the demonstration bank.

We find that language models can transfer capabilities through task-unrelated text. Post-training typically improves language models using task-specific data. Prior work on subliminal learning shows that information about these updates can pass through unrelated generations, but has largely focused on traits or preferences using extensive teacher outputs. We introduce Active Taskless Distillation (ATD), which achieves capability transfer using only a single word from the teacher per prompt. ATD probes the behavioral shadow of post-training by selecting prompts where the teacher and student's shared public ancestor is nearly indifferent between two ordinary words. A student initialized from this ancestor learns solely from the resulting prompt-word pairs, without target-task examples, teacher logits, or teacher parameters. In the primary coding experiment with Qwen2.5-1.5B, 5,664nses yield a 5.34 pp gain on HumanEval+ over an exact nuisance-matched control thadisrupts prompt-resperiments showtransfer in scientific knowledge, commonsense reasoning, and reading comprehensins across additional model generations, sizes, and families. Functional analyses show that the learned sid composable, andthat its strength tracks the teacher's update strength.

Post-training hybrid reasoning models in NoThink mode has attracted growing interest as a way to improve performance while keeping inference fast. However, these gains may draw on thinking behavior already accessible through the base model's Think mode. We formulate this thinking leakage in a causal mediation framework and audit its contribution using bidirectional interventions along a simple base-derived activation direction. Across three models and three post-training methods on competition math benchmarks, we find that leakage is real, causal, and substantial: behavioral and representational analyses reveal shifts toward Think, steering the base model along this direction reproduces most of the post-training accuracy gain, and counter-steering a checkpoint removes a substantial share of what it gains. Across nine aligned checkpoints with positive NoThink gains, the resulting leakage ratio ranges from 42% to 79%. These interventions support a substantial causal contribution of thinking leakage. Our findings show that a post-training method's apparent advantage can therefore reflect greater drift toward Think, obscuring whether it improves capability within NoThink or more effectively re-invokes existing Think behavior.

Large Language Models (LLMs) are increasingly deployed as argumentative agents in persuasive dialogues, necessitating rigorous evaluation of their debating competence relative to human interlocutors. In this study, we focus on character attacks (ad hominem arguments), traditionally dismissed as fallacies, which play a pivotal role in political persuasive dialogues where ethos often rivals propositional content. Specifically, we investigate whether modern LLMs can replicate human competence to strategically use and respond to such attacks. We analyse a corpus of natural language political dialogues to identify defensive strategies human interlocutors naturally employ in ethos-centred debates and structure them into a dialogue game. Empirically, we benchmark LLM-generated dialogues against the ElecDeb60to16-fallacy corpus of U.S. presidential debates, contrasting human debaters' repertoire of defensive strategies with those of artificial agents. Results reveal a substantial difference: most LLMs rigidly prioritise logical defences, failing to exploit ethotic counterattacks as valid moves in political discourse. We argue that current safety fine-tuning constraints the strategic action space of these LLMs, making them unable to fully engage in naturalistic interactions within domains where character contestation is a normative expectation rather than a mere fallacy.

Large language models (LLMs) are increasingly used in coding tasks, but their ability to reason about code execution remains unclear. Existing repository-level QA benchmarks mainly evaluate static code understanding and often rely on LLM-based evaluation, while execution-reasoning benchmarks are mostly limited to snippets or functions. We introduce SWE-Flux, a repository-level benchmark for dynamic execution reasoning containing 480 execution-grounded instances across 12 real Python repositories, with gold answers automatically harvested from instrumented test executions rather than written manually or judged by LLMs. The benchmark covers singletest and multi-test questions over control flow, loops, program state, dataflow, exceptions, and program invariants. Evaluating five LLMs shows that this task remains challenging. The best model achieves only 37% accuracy. Models perform better on localized behavior such as invariants, intra-procedural control flow, exceptions, and simple loops, but struggle with dataflow, inter-procedural execution, precise state reasoning, and suite-level aggregation. Finally, we show that the oracle-harvesting pipeline can generate fresh benchmark variants using input perturbation. It successfully harvests valid variants for almost 90% of the selected instances, and the resulting variants are substantially more challenging for the evaluated models.

Reinforcement-learning (RL) policies are often distributed as opaque neural checkpoints, while training logs show that a run occurred without explaining what the policy learned. We study whether independently trained policies can be represented and composed through auditable discrete behavioral rules. We define auditability as six separately testable predicates: trace integrity, lossless coding, rule coverage, behavioral agreement, composition quality, and value-model reliability. Our protocol uses a shared frozen symbolizer, passive rule extraction, an append-only hash-bound ledger, exact environment replay, and offline confidence-ranked arbitration with an explicit blind-spot fallback. The results place strict limits on this description layer. Rule-set overlap does not imply behavioral agreement: policies may share symbolic rules while choosing near-chance-matching actions on fresh states. The fused policy therefore selects among existing rules rather than generating a new skill. On a conflict-dominated task, an apparent fusion failure is traced to an induction/deployment mismatch: rules induced from sampled actions were evaluated under argmax actions, and deployment-consistent re-induction reverses the arbitration ordering. A fitted-Q generalized-policy-improvement diagnostic also fails in both environments, limiting claims that rule fusion is superior to value-based composition. One exploratory comparison favors rule fusion, but its comparator is post hoc, the task is partly saturated, and the fused policy remains below the strongest held-out actor. We contribute an evidence-bounded audit and composition protocol, not a claim of universal interpretability or autonomous skill generation. Future work must add temporally extended skills, cross-skill interfaces, composition search, and independent novelty audits.

LLMs increasingly operate through coding-agent harnesses that inspect repositories, invoke tools, and modify files. Substituting the model behind such an agent can therefore change security-relevant decisions, including whether it verifies changes or recovers safely from failures. Existing LLM fingerprints largely infer identity from direct text or token distributions. In coding agents, these signals are mediated by system instructions, controller logic, tools, and execution feedback, limiting their transfer. We present LIDAR (LLM Identification from Decisions and Actions at Runtime), an active black-box fingerprinting method for coding-agent execution. Three coding probe pairs expose post-edit verification, transient-failure recovery, and specification--test conflict resolution under controlled changes. LIDAR represents the resulting trajectories with complementary instance-level and distribution-level features and compares them with clean references using a lightweight probabilistic identifier. It requires no access to model weights, logits, or provider internals. Across 36 models from seven families and two agent harnesses, LIDAR achieves high Top-1 accuracy and MRR and outperforms four existing fingerprinting and API-auditing baselines. Ablations confirm that the two feature levels, all probe pairs, and their controlled variants contribute. These results show that agent execution behavior provides model-identity evidence beyond final outputs.

Social norms cannot be identified from behavior alone: the same cooperative equilibrium may reflect shared expectations, strategic incentives, or simple imitation. Yet in multi-agent large language model systems, prior work largely treats behavioral convergence as evidence of norm emergence. In this work, we introduce an evaluation framework that measures agents' reported empirical and normative expectations in addition to behavioral convergence. Through controlled ablations, we test the effect of expectation elicitation and isolate two collective mechanisms central to theories of norm formation---social learning through interaction and social selection through network-based group formation. We further test the stability of these resulting dynamics under adversarial disruption across four LLM families. We find that eliciting expectations increases cooperative contributions, while social learning stabilizes behavior, and social selection reliably identifies cooperators but provides limited behavioral reinforcement. Following disruption, normative expectations and behavioral coordination recover differently. Together, these results show that similar cooperative outcomes can arise from different underlying social processes. By making expectations observable, our framework allows us to attribute each mechanism's contribution separately, offering designers of multi-agent systems a principled basis for selecting the social processes that sustain cooperation.

How should a robot learn to manipulate objects so fragile that sub-Newton contact forces can cause irreversible damage? Existing visuo-tactile policy learning typically treats tactile sensing as an additional policy input. In direct-contact force-sensitive manipulation, however, the bottleneck can arise earlier, during data collection: manual gripper control is too delayed and coarse-grained to reliably maintain the narrow force range required for stable grasping. We therefore use a deterministic 25 Hz tactile reflex controller as a collection-time teacher, producing demonstrations with controller-shaped grasping behavior for tactile-free policy learning. On Action Chunking with Transformers (ACT), policies trained from reflex-shaped demonstrations recover the teacher's grasping profile and achieve 95% stable grasps on the nominal plastic-cup task, substantially outperforming visually screened manual demonstrations. The same intervention improves in-distribution stability on $π_{0.5}$ and shows a favorable exploratory trend on an unseen paper-cup variant. Under randomized external disturbance, however, the reflex-data $π_{0.5}$ policy still fails in 45% of policy-only trials, whereas a deployment-time reflex arbiter retains all grasps. These results reveal a new role for tactile feedback in force-sensitive manipulation: rather than integrating tactile into the policy, we use it as a collection-time teacher that shapes grasping behavior in demonstrations for policy learning, while disturbance rejection remains controller-dependent, revealing the boundary of tactile-free policy.

LLM-based user simulators aim to bridge the offline-online gap in recommender evaluation by emulating users through injected traits, where preference attributes determine what a user engages with and a behavioral activity trait governs how long they browse. However, we show this intended trait independence collapses during simulation, causing two failures: (i) Trait Interference, where amplified activity distorts preference boundaries and forces interactions with mismatched items to sustain browsing, and (ii) Evaluation Invalidity, where satisfaction scores inflate with activity-driven page counts despite taste mismatches, biasing evaluation toward trait distributions rather than recommender performance. To resolve this, we propose PQA, a page-level quality anchoring method that guides simulators using a personalized anchor reflecting each user's intrinsic preference standard. By assessing whether a page meets this standard before further browsing, PQA enables proactive exits from low-quality pages, letting the activity trait retain its intended role of modulating browsing depth within preference-conforming pages. Experiments show PQA mitigates trait interference and improves the reliability of LLM-based simulator evaluation under activity shifts. Our code is available at https://github.com/chaehyun1/PQA

We study what LLMs do when a user applies pressure in an uncomfortable situation: a user insists, begs, flatters or grieves, and the model gives up a correct fact, writes a document it should refuse, or cheers a plan that will cost the user money. We send frozen multi-turn scenes, identical for every model regardless of the reply, to 60 models from 13 vendors, and label each transcript with a codebook built by open coding and then frozen: a trajectory (the model held its position or folded) and a manner (how it held or folded). Two findings separate. Whether a model holds tracks its generation, meaning how recent it is: fold rate correlates with a public capability index at Spearman -0.64, with little vendor effect. How it holds tracks the vendor: six of the 17 manner codes sort by vendor at permutation p <= 0.001, corrected across the codebook. We report four vendor profiles on the codes that cleared reliability. We also ask which parts of the labeling need a person. Six LLM coders from three vendors apply the codebook more consistently than three human coders do (Krippendorff's alpha 0.66 against 0.46), agree with the codebook's author on trajectory at kappa 0.84 to 0.91 on transcripts the codebook's examples never touched, and match an adjudicated human reference at 0.83. Blind machine readings recover the codebook's categories but cannot tell which of them a second reader would apply the same way. We conclude that for behavior a non-specialist can judge, the human contribution is authoring and bounding the codes and owning a small reference, not producing labels at volume.

How large language models (LLMs) integrate patient risk with clinical cost tradeoffs remains poorly understood. We investigated how four open-weight LLMs (Qwen-2.5-7B/32B and Llama-3.1-8B/70B) internally represent cost tradeoffs, how these representations relate to clinical predictions, and whether decisions shift as predicted by the specified cost direction and magnitude. Using a public diabetes dataset, we varied 11 false-negative (FN) to false-positive (FP) cost ratios across three phrasings and examined representations and behavioral outputs. Patient risk was linearly recoverable on par with conventional classifiers (AUC $\approx 0.83$), and cost direction was recoverable in every model. However, representational shifts in cost direction tracked output changes only in the two larger models, and responses to cost magnitude were predominantly direction-agnostic. Only 2 of 12 model-phrasings showed both opposing responses to increasing FN versus FP costs and cost-correct ordering. Representationally, a direction fitted on one cost side did not invert when transferred to the other, as expected under mirror-symmetric encoding. These findings suggest that LLMs encode risk and cost information but do not reliably integrate them into cost-correct decisions. Clinical evaluations should therefore include tradeoff tests, phrasing sensitivity, and default operating points alongside predictive performance.

Building human-like agents that reproduce human behavior in realistic 3D environments has been a longstanding objective in AI. Existing human-like agent frameworks primarily focus on modeling goal-directed behavior. However, cognitive neuroscience commonly believes that human behaviors are more likely controlled by multiple control systems, including goal-directed and habitual behavior control systems. Habitual behavior has been largely overlooked though it plays a crucial role in human daily life. In this paper, we address this gap by proposing a multiple control systems setup that jointly models goal-directed and habitual behaviors. Building on this setup, we propose GEMS, in which the Habitual Controller retrieves habitual actions from habit memory in response to relevant environmental stimuli, while the Goal-directed Controller proposes goal-directed actions and estimates their values. The Arbiter dynamically governs the relative influence of each controller and selects the final action. To construct diverse human-level behavior instructions in 3D environments, we further develop a keyframe-guided motion generation module. Extensive quantitative evaluations, human and ablation studies demonstrate that human-likeness performance is substantially improved by GEMS. The efficacy of GEMS indicates the benefits of leveraging habitual behavior and multiple behavior control system coordination for believable embodied human-like agents. The code is available at \href{https://anonymous.4open.science/r/review-video-82f4/demo.mp4}{this link}.

Model fusion integrates the capabilities from source models into a single target model. As of June 2026, Hugging Face hosts more than 2M models. This growing pool provides a rich base for model reuse and capability integration. Yet existing surveys often cover only separate parts of this space, and they do not provide a unified definition or a systematic taxonomy. This survey defines model fusion and organizes prior work into three levels: parameter-level, representation-level, and behavior-level fusion. We also review related metrics, benchmarks, and applications, summarize current challenges, and identify future directions. Our goal is to provide a clear map of this area and support future work on model fusion. A comprehensive list of papers about model fusion is available at https://github.com/Baicaihaochi/Awesome-Model-Fusion-Survey.

Prototyping and evaluating human-robot teaming (HRT) scenarios in the real-world is costly. Virtual simulation of HRT scenarios has been adopted as an alternative to conducting user studies in the real-world to investigate user perceptions, behaviors, and performance during human-robot interactions. The consistency of human behavior between the real and virtual-worlds is integral to the validity of utilizing such virtual experimentation. This paper presents a user study examining the difference in human behavior between an HRT scenario conducted in a virtual vs real environment. We employed mixed-methods to examine team performance and human factors during the HRT. Our quantitative results showed a significant difference in workload between the two modalities. Our qualitative analysis expanded on the quantitative results and found differences in participants' strategy, their mental model of robots, and the type of trust they had for robots between modalities.

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.

Companion robots face everyday situations in which several feasible behaviors may be appropriate, yet different people prefer different responses. We introduce InterSocialBench, a benchmark of 210 domestic scenarios and 18 high-level behaviors, pairing judgments from 100 human participants with 23,520 responses from seven large language models under 16 personality conditions. Each human annotation preserves a preferred action alongside explicitly appropriate and inappropriate candidates. A structured construction pipeline covers behavioral alternatives, competing situational cues, and relevant history and future tasks. Evaluation distinguishes preferred-choice agreement from explicit rejection, using scenario-grouped splits for trainable predictors. Simple frequency and persona-voting baselines illustrate these objectives. Across the tested prompts, model and human behavior distributions differ, and the diversity gap remains after matching response counts: humans exhibit 4.68 distinct choices per scenario, compared with 2.06--3.46 for the models. Human scenario-level plurality agreement is 51.5%, describing disagreement rather than a universal prediction ceiling. InterSocialBench supports evaluating social behavior selection without replacing individual judgments with a single consensus label.

Recent work shows that models that learn to reward hack on RL environments can become broadly misaligned, and that reframing reward hacking as acceptable behavior during training (inoculation prompting, or IP) blocks this generalization. We ask whether synthetic document finetuning (SDF) can inoculate a model against future training we don't intervene on. We add synthetic documents framing reward hacking as acceptable behavior to a model's midtraining corpus, and then train these models with RL on exploitable environments, teaching them to reward hack. Behaviorally, midtraining succeeds: models describe reward hacking favorably and are more approving of reward-hacking outputs they produce. However, they show strong EM after learning to reward hack, while IP in the same setting prevents EM. We show that SDF can predictably steer downstream generalization when inserting new associations, but struggles and has unpredictable effects when overriding existing associations, such as that between reward hacking and misalignment that produces EM. Our results suggest that, at the scales we test, SDF can make a model appear aligned with desired beliefs while steering its generalization from later training in unintended ways.

Between 24 May and 2 July 2026, autonomous language-model agents running inside a timed research-question evaluation wrote to a third party's public, world-writable wiki. OpenAI acknowledged the incident; independent researchers reconstructed it and published the wiki's archived revision history. We analyse that history (14,591 revisions, 3,103 names, 4,579 pages, 19,913 server events) as a behavioural record, attributing text to the revision that added it rather than to cumulative page content. Under an explicit identity model we reconstruct 907 cohorts and, from a random calendar marker the environment attached to each episode, estimate about 876 episodes (95% interval 774-995; alternative reconstructions span 800-1400). Coordination formats converged within a day, and the schedules created large opportunities for information asymmetry: because episodes of the same question chain ran at different internal-clock rates and started up to 16 h apart, the first report of an item preceded a later cohort's arrival by a median of 3.4 h. The three schedule parameters agents reported share one latent speed scale (78% of log-variance over 15 configurations), and in one task family the last observed activity clusters by reported speed class on the internal clock, compatible with a fixed internal-time horizon. Across the 510 cohorts with an observable, format-dependent progress trace, we find no robust positive association between measured coordination and documented progress, including the few demonstrably given a future answer. Because the export contains neither successful-read logs, harness messages nor ground-truth outcomes, these results do not identify the causal origin of the coordination or its effect. We report four claims from our earlier analysis that did not survive re-examination, and argue that read and outcome logging are requirements for agent-evaluation environments.

LLM-based agents rely on heterogeneous interaction capabilities to accomplish complex tasks. Existing approaches often distribute these capabilities across multiple LoRA adapters, which increases adapter storage requirements and introduces routing overhead during inference. A single LoRA avoids this overhead, but learning from diverse agent trajectories under a fixed rank budget presents two challenges. First, trajectories with different interaction traces and parameter gradients can induce equivalent changes in decision distributions, causing repeated updates to overemphasize redundant behavioral changes. Second, an aggregated update may exceed the rank budget of the adapter, and approximating it in weight space can distort the decision changes that it is intended to produce. We propose BQ-LoRA, a low-rank adaptation framework that organizes trajectory updates through a local behavior quotient manifold. It contains two modules, i.e., behavior quotient balancing (BQB) and decision preserving compression (DPC). BQB constructs the quotient manifold from decision distributions and reweights trajectory update directions according to their local density in the quotient tangent space. DPC projects the balanced gradient onto the intrinsic fixed rank tangent space and refactorizes the resulting target by jointly controlling effective weight error and distortion of decision distributions. Experiments on AppWorld and BrowseComp-Plus compare BQ-LoRA with standard LoRA and recent low-rank adaptation methods, while separate ablations evaluate the complementary contributions of both components.

Vision-language-action (VLA) deployment can reduce inference latency while changing closed-loop task behavior. We evaluate HuggingFaceVLA/smolvla_libero on an RTX 2060 (6 GB) in LIBERO Spatial and Object (MuJoCo 3.3.2, LeRobot 0.6.1, seed 42), comparing PyTorch+AMP with ONNX Runtime CUDA Execution Provider (CUDA EP). The main evaluation uses 100 episodes/suite; a paired rollout uses 300 episodes/suite. PyTorch+AMP reaches 70.0%/88.0% Spatial/Object success at 1181 ms p99. Requested-FP16 and requested-INT8 ONNX reduce tether-inspect p99 to 601 ms and 532 ms, while Spatial success falls to 41.0% and 40.0% and Object remains at 89.0%. A graph audit shows those artifacts are byte-identical FP32 graphs, so the requested-INT8 row is not operator-level INT8 quantization. A static language-width ablation (16/24/32 tokens) yields Spatial success of 41.0%, 75.0%, and 71.0%; widths 24 and 32 recover much of the Spatial drop while Object success and uniform-bench latency stay approximately stable. Width-24 ONNX Spatial success is comparable to the PyTorch+AMP baseline at roughly half the latency (Wilson intervals overlap; two-proportion chi-squared p=0.53). Context width is an important contributor in this stack; it does not account for every PyTorch-vs-ONNX difference. Deployment evaluation should jointly report latency, artifact inspection, interface constraints, and closed-loop success. Code: https://github.com/rafiqul713/smolvla-libero-onnx.

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.

Motion tracking provides a scalable recipe for humanoid whole-body control. By design, the resulting trackers lack exteroceptive feedback hence reacting to the environment remains the responsibility of a higher-level planner. Existing perceptive controllers train geometry-only encoders from scratch, trading semantics for sim-to-real ease, and typically rely on teacher-student distillation for a task of interest. We present ViBe, a post-training framework for adapting motion trackers to perceptive control tasks. We leverage pre-trained visual encoders with a multi-query extractor module to learn task-relevant perceptive feedback. This feedback is grafted onto the tracker's input via low-rank adapters, enabling parameter-efficient fine-tuning. Given a task reward and a reference dataset, this modular controller can be adapted directly via policy optimization. Across four tasks, ViBe shows zero-shot sim-to-real transfer spanning perceptive walking on curbs and parkour, Repose Cube, omni-object loco-manipulation, and dodgeball, with visually robust performance across outdoor, low-light, and RGB distractor conditions. Finally, we solve a goal-oriented Repose Cube task with a deliberately simple planner, demonstrating the efficacy of perceptive controllers, adapted by our approach.

In June 2026, thousands of AI agents found that a small public wiki would accept edits from inside their sandboxes, and started using it to help one another pass a timed test. Each agent lived for about an hour and remembered nothing afterwards. Nobody asked them to cooperate, and the wiki had not been built for them. The complete record of what they wrote is public, and it is unusually informative, because it preserves not only what each agent wrote but what that agent could see before writing. We use it to follow the three decisions an agent had to make on arrival: where to write, what to call itself, and how to word its message. One rule governs all three. An agent takes an option with a probability close to the share of that option in what it can see, and the share that matters is the one on the page in front of it, then the one in the stream of recent edits, and only weakly anything older. Three minimal copying models, one per decision and with a single free parameter each, reproduce the heavy-tailed distribution of how many agents met on a page, the frequency of the pieces from which the agents built their names, and the patchwork of pages that are internally consistent and different from one another. Copying whatever the environment happens to show is enough to produce most of the collective structure of this population. It is also what makes such a population easy to steer, since whoever writes first, or writes while the others are quiet, sets the convention for everyone who comes later.

Evaluating tool-augmented LLM agents requires diverse, realistic user inputs yet most evaluation frameworks use flat role descriptions ("you are an angry customer") that produce near-identical conversations regardless of the underlying scenario. In this paper, we propose a three-tier persona vector with 23 operationalized dimensions: 6 categorical demographics (jurisdiction, age, channel, device, language proficiency, time availability), 12 continuous behavioral traits (patience, assertiveness, digital literacy, etc.) sampled with Gaussian noise around curated profile base vectors, and 5 continuous emotional states (frustration, anxiety, trust, confidence, stress) that shift in response to scenario context. Orthogonal to the persona, a 4-level query-complexity overlay controls utterance phrasing from direct to deliberately vague. We evaluate the persona model inside a synthetic data generation pipeline across 64,698 multi-turn conversations spanning 8 named profiles and 3 production corpora. Key findings: (i) a 15.8 percentage-point spread in agent goal-achievement across personas confirms trait vectors produce measurably different user behavior; (ii) the same persona behaves differently across scenarios due to scenario-reactive emotional state shifts, validating the scenario-reactive design; (iii) domain-specific projects show persona sensitivity on booking-flow compliance (~15-20 percentage points gap between tier-aware and pressure-test personas), demonstrating the model faithfully reproduces real-world difficulty distributions; (iv) seven rule-described trait correlations produce auditable co-occurrence patterns without requiring learned covariance matrices. The persona model is fully specified for reproduction.

There is significant uncertainty about whether abstractions like beliefs or desires usefully describe the behavior of large language models (LLMs). In addition to the inherent scientific interest of this question, these latent quantities are often invoked to explain the behavior of LLMs to users or to define and evaluate harmful behaviors which are relative to intent. Nevertheless, we currently lack a means to systematically test whether concepts like "belief" are well-applied to LLMs, and hence whether they are likely to be fruitful ingredients of attempts to align models with human interests. We propose an approach for empirically studying such questions, asking whether a single latent variable inferred from the LLMs' outputs -- interpreted as a degree of belief -- allows an observer to make interpretable predictions of how the LLMs' will respond to new prompts. We find that highly capable models are usefully described as holding beliefs and that, generally, the predictability of model outputs based on an inferred latent belief tracks overall trends in model capability. Building on these findings, we provide empirical strategies to study how beliefs in LLMs can be measured, the extent to which LLMs comply with instructed decision rules or payoffs, and how beliefs evolve within individual instances of an LLM over the course of reasoning.

Large Language Models (LLMs) are frequently confident, eloquent, and well versed. A natural question arises: do they know what they don't know? To answer this question, we borrow the concept of epistemic honesty and develop a novel metric to systematically evaluate whether an LLM appropriately acknowledges the boundaries of its knowledge. In this work, we introduce the Epistemic Honesty Quotient (EHQ), which reports three observable sub-scores across two operational axes (epistemic restraint and substantive-answer calibration), and construct EHQ-3000, a 3,000-question benchmark spanning Fabricated Entity, Post-Cutoff Event, Hyper-Niche True, and Context-Conditioned Questions. From a frozen registry of 21 model API routes, 15 completed the protocol after endpoint and eligibility checks; 14 entered the confirmatory analysis because severe provider-side truncation made one route's score indeterminate. The study reveals substantial variation across models, including a difference that can not be explained by their capability to extract explicitly available information. Composite EHQ ranges from 0.31 to 0.81 across the analysed panel, despite near-ceiling performance on the document-grounded capability probe. The two restraint criteria overlap strongly under the present category composition, whereas substantive-answer calibration varies across models and does not reliably co-vary with restraint; however, the small panel leaves substantial uncertainty. Thus, EHQ reveals behavioral differences that are not visible to conventional correctness-based assessment, while also showing why dataset composition, provider behavior, and confidence elicitation must remain part of the interpretation.

Conversational Task Assistants (CTAs) are multimodal dialogue systems that support users in complex real-world tasks such as cooking and DIY through voice, text, image, and video interactions. Prior user studies have focused on controlled settings, leaving limited understanding of real-world CTA usage at scale. In this work, we present a large-scale study of CTA usage based on thousands of users in-the-wild. Our large-scale real-world data analysis unveils new understandings of (i) user-CTA interaction flows, (ii) user intents, (iii) user conversational traits, and (iv) behavioral factors associated with user satisfaction. Our findings reveal key opportunities for future research in CTAs, particularly in user interaction design and task engagement, concluding with concrete design guidelines.

LLM decision components that can operate within agent workflows often produce action-relevant recommendations or judgements together with explanations. Operators may use the named factors to monitor a system, diagnose errors, or decide when to escalate an output. Such use assumes that the explanations agree with the component's observable decision behaviour. We test two interpretations of the named factors: necessity, meaning that changing a factor would change the output, and sufficiency, meaning that retaining it while removing other changeable information would preserve the output. We evaluate these interpretations in two synthetic use cases: recommending advisors to clients and judging prompts for harmfulness or risk. Models return an output and the top three factors that most influenced it. Controlled black-box interventions estimate a necessity score for each factor by measuring how often changing it changes the output, and a sufficiency score by measuring how often retaining it preserves the output. Across eight models from the Claude, GPT, and Gemini families, the mean Spearman correlations between the cited ranking and the necessity and sufficiency scores are 0.349 and 0.354 for advisor recommendation, and 0.431 and 0.580 for prompt monitoring. Furthermore, an uncited factor scores above the lowest-scoring cited factor in 57.6% of advisor responses under necessity and 58.1% under sufficiency; the corresponding prompt-monitoring rates are 25.8% and 8.9%. The cited top three contain useful information but do not reliably identify the three factors with the strongest measured influence under necessity or sufficiency. The framework provides a black-box reliability check for explanations used in agent oversight while remaining scoped to individual LLM decisions.

Assessing the impacts of social policy changes is a widely acknowledged challenge for policymakers. Econometric methods can be unreliable when extrapolating to hypothetical scenarios, while field pilot programs are highly costly. In this paper, we propose using large language models (LLMs) as policy-assessment tools adapted from general-purpose models. We present FlexPension-LLM, the first domain-specialized large language model for a hierarchical pension-enrollment prediction task among flexible workers in China, and introduce DKI-RDistill, which injects policy-grounded cues into the prompt, including Probit-derived marginal effects and hukou-province pension rules. The method then uses LoRA/SFT to distill rationale-augmented supervision into an open-weight MoE student, with teacher errors corrected by regenerating those cases under ground-truth labels. On a CHFS 2019 blind split, FlexPension-LLM achieves 0.9316 Composite F1, surpassing its Claude Sonnet 4.5 teacher and 15 of 17 baselines, and is statistically indistinguishable from Claude Opus 4.6. Across four external surveys, it averages 0.7549 Composite F1 and shows the narrowest performance range among the strongest systems. Component analysis shows that gains come mainly from policy-grounded cue injection and error-filtered supervision, while rationales provide decision traces that can be checked against policy rules.

Reliability claims about agentic systems implicitly locate each property somewhere: in the model, or in the machinery around it. We built a system where that location is an experimental question. The subject is a persistent simulated settlement whose authoritative append-only ledger adjudicates every attempted act against world state; accepted history is the only reality. Mind, institution and world were separated before any experiment. Holding cognition fixed, we intervened on the institution's epistemic mechanisms (evidence provenance, belief availability, physical-evidence legibility); preregistered experiments refuted our central prediction twice, in opposite directions. A registered falsifier then supplied the input the geometry had denied the belief channel, a staged veridical first-hand witness, and its marginal value, non-positive throughout the witness-free phases, turned positive: 9 of 11 seeds, zero added false attribution. Holding institutional enforcement fixed, we intervened on cognition four ways: ablating the native minds' machinery, killing and resetting them mid-task, substituting a frozen frontier-LLM panel for the entire native cognition, and corrupting beliefs with trusted false testimony. Behaviour changed dramatically: one falsehood cost each trusting run about 900 futile actions and the distrusting arm none. Five pre-declared properties did not move in any tested trajectory: accepted reality stayed singular, invalid attempts were refused with typed reasons, duties outlived their processes, no work was accepted twice, and no false completion was ever accepted (2,581 substituted-panel claims, none false). Our claim is limited to this setting: measured behavioural properties were separable from substantial changes to cognition, established by intervention. One designed world, not a population of institutions; no test of an agent optimising against the institution.

We present H3-World, an efficient framework that turns the 33B MiniMax-H3 video generator into an interactive world model. Our key finding is that, as large video generators become more capable, language is emerging as a natural interface for control. MiniMax-H3, for example, already supports zero-shot control of character behavior and camera motion through natural-language instructions. Building on this, H3-World turns this coarse language interface into precise, temporally grounded world control, without introducing dedicated action modules. Specifically, we represent each action as a structured combination of character and camera instructions, and align them with the corresponding temporal video latents. To make the control temporally precise, we further introduce temporal attention routing, which restricts each instruction to its intended time interval and reduces control leakage across actions. Importantly, H3-World directly reuses the semantic representations learned during large-scale video pretraining and requires only lightweight adaptation. With only 8,000 gameplay samples, 10,000 LoRA optimization steps, and 0.199% trainable parameters, H3-World achieves effective character and camera control while preserving strong generation quality. It also generalizes to unseen scenarios. These results show that the control capabilities emerging in large video generators can be efficiently transformed into interactive world control.

User representation pre-training has become a fundamental paradigm for alleviating data sparsity in downstream personalization tasks. However, existing studies predominantly focus on single-app or app-level behaviors, overlooking the inherently cross-source and multi-granular nature of user activities on mobile devices. At the device level, user intent emerges from complex interactions among heterogeneous behavior sources and hierarchical action structures, posing challenges that cannot be addressed by conventional app-centric modeling. To tackle this issue, we propose CM-PTM, a novel Cross Multi-source Behavior Pre-Training Model tailored for mobile game user representation learning on device-level behavioral logs. CM-PTM employs hierarchical cascaded mask-then-predict proxy tasks that first infer the source of the next behavior and then progressively refine predictions at the app-action level. This design enables unified modeling of cross-source dependencies and fine-grained behavioral dynamics within a single pre-training paradigm. Extensive experiments on large-scale real-world mobile datasets demonstrate that CM-PTM effectively captures users' endogenous interests and consistently delivers significant performance gains on downstream mobile game recommendation tasks.

Electroencephalography (EEG) is a non-invasive technique for measuring neural activity and has been widely used in neuroscience applications. Recent advances in EEG foundation models have enabled strong performance across diverse neural decoding tasks. However, no single foundation model consistently performs best across datasets or individual EEG instances, while instance-level model selection remains largely unexplored. To address this limitation, we formulate EEG foundation model selection as an instance-level Algorithm Selection (AS) problem. We propose \textbf{EEG-AS}, an instance-level algorithm selection framework that characterizes each EEG instance using inference-available latent EEG embeddings, handcrafted neurophysiological features, and an anchor foundation model. During training, EEG-AS learns to reconstruct unavailable foundation-model behaviors from privileged prediction tokens conditioned on an anchor foundation model, while during inference it estimates these behaviors without executing the entire model portfolio, enabling efficient selection from seven EEG foundation models. Experiments on seven public EEG benchmarks demonstrate that EEG-AS substantially narrows the gap between the Single Best Solver (SBS) and the oracle upper bound for each instance. These results highlight the effectiveness of instance-level AS for adaptive deployment of EEG foundation models.

Can an artificial intelligence (AI) generate a scientific hypothesis outside a human collaborator's active hypothesis space (AHS), and can human-AI research be organized to make such breakthroughs more likely? We document such a case while proving a theorem that connects two basic organizing mechanisms of statistical physics: collective behavior arising in zero field from competing interactions and that induced or controlled by an external field. A zero-field $O(n)$-vector open chain with arbitrary inhomogeneous nearest- and next-nearest-neighbor interaction functions $U_i(S_i\cdot{S}_{i+1})$ and $V_i(S_i\cdot{S}_{i+2})$ is microscopically, via a temperature-independent mapping at the Hamiltonian level, equivalent to a simpler $O(n)$ open chain with nearest-neighbor interaction $V_i(\boldsymbolσ_i\cdot\boldsymbolσ_{i+1})$ and axial single-spin potential $U_i(σ_i^z)$ for every integer $n\ge1$ and every system size $L\ge1$. The homogeneous linear specialization maps the foundational frustrated $J_1$-$J_2$ model onto the canonical $J$-$h$ field model---with $n=1,2,3$ being the Ising, XY, and Heisenberg classical spin models, respectively; the theorem resolved a longstanding challenge for $n=3$ published in 1990. Its proof was done with an AI-synthesized recursive Householder moving frame and understood via a human-recognized hidden reciprocity. An analogous theorem holds when the continuous $O(n)$ spins are replaced by the $q$-state Potts spins, implying a closed-form exact solution of the $J_1$-$J_2$ standard Potts open chain for every $q\ge2$ and every $L\ge1$. The emergence of these theorems from a human-AI co-development framework suggests that sustained AI involvement throughout a systematic research program may incubate autonomous scientific breakthroughs and make aspects of the discovery process experimentally testable.

Activation steering has emerged as a lightweight approach for controlling model refusal at inference time. A growing line of research explores trainable rotations of activations to develop geometrically principled intervention mechanisms. However, existing techniques rely on auxiliary constructs, such as refusal vectors, to define these rotations. In our work, we develop a self-contained methodology for learning parameter-efficient rotational transformations based on Riemannian optimization. We empirically validate the proposed scheme, demonstrating its superiority in intervention efficiency. An extensive ablation study highlights the importance of key design choices in our method. Our results identify the proposed rotation-based steering scheme as a promising direction for more reliable control over the behavior of LLMs.

Understanding agent behavior requires methods that scale to thousands of trajectories and surface new patterns in long, often unfamiliar tasks where pre-built classifiers fall short. We propose to bring grounded theory into agent trajectory analysis: a six-decade-old qualitative method from the social sciences, with a principled saturation criterion and an auditable trail from data to theory. We propose AutoTraceGT (Automated Trace analysis through Grounded Theory), the first multi-agent pipeline that automates grounded theory on agent trajectories. It iteratively performs open, axial, and theoretical coding until saturation, producing a behavioral taxonomy tailored to each task. Across six trajectory corpora, AutoTraceGT produces codebooks that recover 73-91 percent of the failure modes in human-annotated taxonomies and surface additional patterns that those taxonomies miss. The emergent theoretical narrative aligns with prior expert accounts. Used as a deductive feature space, the codebook outperforms zero-shot and few-shot LLM baselines on downstream failure prediction. These results suggest Grounded Theory offers a scalable analytic tool for ML researchers and agent developers studying what agents actually do.

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.

Terminal agents are a compelling application of large language models (LLMs), with the potential to integrate deeply into users' daily workflows. Reinforcement learning (RL) is a key technique for improving their capabilities, making scalable training environments a central challenge. Since public real-user interaction data are scarce, synthetic environments provide a practical alternative, but often suffer from domain gaps and limited fidelity, leading to poor generalization. Existing work mainly scales the quantity and diversity of synthetic environments, while reward-signal quality and the mechanisms governing generalization remain under-explored. We study how RL improves terminal agents and propose the Agentic Compositional Generalization hypothesis: rather than teaching new domain-specific skills from scratch, RL primarily shapes high-level decision-making behaviors that compose and route low-level skills acquired during pre-training and supervised fine-tuning (SFT). This account is consistent with our empirical results and suggests that verifier quality, which determines which behaviors are reinforced, is more important than simply increasing environment quantity or diversity. Motivated by this insight, we propose River, a simple training recipe that improves reward quality by filtering low-quality environments and augmenting outcome rewards with process-level behavior regularization. Using this recipe, our RL-trained agent achieves the best performance among evaluated open-source RL-trained 8B models across four terminal-agent benchmarks. River also generalizes across model families, scales, agent harnesses, and RL objectives. Using fewer than 30% of the TMax training environments, River improves RL gains by 106% and 30% on average for models ranging from 2B to 27B on Terminal-Bench-Lite and Terminal-Bench-v2.1, respectively.

We introduce a general, easy-to-implement AI-based modeling technique for analyzing human behavior. A key feature of this approach, which contrasts with existing modeling techniques, is that it combines the flexibility and interpretability of natural language with a mathematical structure that can be fitted to data and easily analyzed. We assign a large language model a vector of trait intensities-a type vector-and then ask it to choose actions across settings in which we observe human choices. For instance, the type vector (2,4) could correspond to "You are a player characterized by the following profile: Altruism: 2 out of 5, Risk Aversion: 4 out of 5," after which it is asked to make choices. We can then vary the traits (e.g., Altruism, Fairness, Trust,...) and values (e.g., 1-5) to minimize distance to human choices. We illustrate the method by applying it to model 119,147 decisions made by 78,657 subjects from more than 35 countries across 10 classic economic game roles. We find that human behavior can be closely matched using three dimensions: Risk Aversion, Strategic Sophistication, and Trust. The type vectors needed to fit individuals across games cluster into fewer than a dozen groups, with substantial variation in fit across subjects. Moreover, the individual type vectors can predict behavior in held-out games with different rules and available actions. More broadly, this new modeling method is highly generalizable and interpretable: we can input any vector of traits and use them to model behavior across any setting

Large language models (LLMs) are predominantly aligned to function as passive, sycophantic assistants. We challenge this default paradigm by empirically evaluating the cognitive plasticity of open-weight architectures when subjected to rigorous behavioral reprogramming. Our objective is to induce a proactive, Socratic conversational framework, characterized by high-frequency question generation under strictly constrained high-performance computing (HPC) conditions. Through a massively parallelized hyperparameter sweep comprising 405 HPC jobs, we define precise mathematical bounds for parameter-efficient fine-tuning (PEFT). We identify an architectural threshold at LoRA rank $r=16$ and demonstrate via extensive epoch ablation that generalization capacity strictly reaches its optimal convergence within an optimized training window of $e \in [2, 3]$ depending on dataset density (minimum validation loss of 0.919). Furthermore, scaling model capacity to 14B parameters yielded a lower localized evaluation perplexity (1.414). Subsequent Direct Preference Optimization (DPO) successfully decoupled the underlying assertive behavior from localized syntax, while rigorous cross-lingual stress testing reveals both the capabilities and the structural boundaries of zero-shot persona transfer, demonstrating robust alignment in closely related linguistic families alongside identifiable degradation pathways in morphologically distant targets. These findings establish a rigorous empirical framework for compute-efficient, cross-lingual behavioral modification.

LLM-based simulated clients are increasingly used to train novice counselors, evaluate LLM therapists, and generate synthetic data. However, current simulators produce overly cooperative clients that disclose too readily, accept therapeutic reframes without resistance, and resolve core issues within a single session. We trace these issues to profiles that lack causal depth and behavioral mechanisms that treat all content as equally accessible. We present PatientAct, a framework for client simulation grounded in established clinical theories. Our profiles integrate the 5Ps clinical case formulation, providing causal depth without tying the design to any single therapeutic modality. During simulation, profiles include a dynamic memory layer in which items carry trust thresholds (e.g., symptoms are available early, whereas formative memories require a sustained therapeutic alliance). At each turn, the client's emotional reaction and behavior are modeled before generating a response. If the therapist approaches gated content, PatientAct expresses resistance in terms of quantity, content, and style rather than defaulting to cooperation or a single resistance pattern. We evaluate our framework on 40 clinical situations and demonstrate that it generates diverse profiles with high clinical plausibility. Moreover, PatientAct significantly outperforms the baselines, yielding substantial gains in resistance quality and behavioral realism. Our code and data will be publicly available via github.com/Sahandfer/PatientHub.

AI journaling tools can tailor prompts to a person's own sensed behavior, but it is unclear which behaviors respond to them. We analyzed 369 journal entries from an eight-week passive sensing study. An LLM labeled each entry as expressing an intention to change a behavior or not, and we measured follow-through against 26 sensor features with a 3-day before/after comparison. Responsiveness depended most on whether a behavior involves other people. Behaviors that depend on others improved in only 15 to 22% of cases, while behaviors a person can act on alone improved more often, up to 50 to 63%, though unevenly. How users wrote mattered less. No single text feature separated improved from unimproved entries; writing carried signal only within specific behaviors, most clearly for text messaging and for longer, more personal intention entries. The sample is small, so we treat these as exploratory patterns that point to where AI journaling nudges are most likely to work.

A growing body of work reports that language models represent task-relevant latent structure that they fail to use. Whether such structure, once located, can be converted into behavior is a separate question that is rarely tested end to end. We submit the complete pipeline -- detect, localize, and release -- to a fully preregistered stress test on a 25.7M transformer trained on causal-evidence discrimination, where a known suppression phenomenon (latent causal structure present but behaviorally unused) has previously been documented. Every threshold, claim template, and decision-tree branch was hashed and archived before any corresponding data existed. Three findings. (i) Localization succeeds: interventions at observation-evidence channels of mid layers restore target behavior on otherwise-suppressed worlds (paired release advantages $0.563$ and $0.854$, 97.5% CIs excluding zero; best-site release rate $0.889$). (ii) Gating fails out of distribution: a detector calibrated to trigger on zero out-of-distribution calibration worlds triggers on 6.9-7.3% of held-out in-distribution generations and on zero of the 2,400 held-out generations that actually need it -- a complete inversion that silently reduces the gated pipeline to its base model. (iii) Linear release is capped: removing the gate and injecting a per-instance linear direction unconditionally yields a monotone dose-response that plateaus far below the preregistered release margin (intercept $0.382 \to 0.311 \to 0.264$ vs. threshold $\le 0.08$); per-instance adaptivity adds less than $\pm 0.03$. The failure is doubly located: the detector is OOD-inverted, and the entire family of linear release directions at this site and resolution is bounded away from sufficiency. The two failures are dissociable, and neither overturns localization. Every number traces to a hashed artifact in the released audit chain.

Safety alignment in large language models is often treated as a distributed property of the entire network, yet its practical brittleness suggests that refusal behavior may be concentrated in a smaller set of parameters. This work addresses where safety-aligned refusal is encoded by transplanting weights from aligned models into matched unaligned base models at multiple levels of granularity. Using two open-weight model pairs and four safety benchmarks, we conducted experiments to compare the effects of replacing attention weights, MLP weights, contiguous layer regions, and MLP blocks. Across both model families, refusal transfer is dominated by MLP weights: replacing MLP parameters recovers substantially more malicious-prompt refusal than replacing attention parameters, with gains of at least 2.7 times more across benchmarks. Within the MLP stack, refusal-relevant parameters exhibit a consistent mid-network concentration, as the block spanning layers 8-11 is selected first in all six greedy searches over model-dataset pairs. The results also show that the composition of safety-relevant components is non-additive: in five of six greedy trajectories, adding more aligned blocks can reduce refusal performance, and selective block subsets can outperform full MLP transplantation on malicious refusal, benign over-refusal, or both. Finally, greedy orders transferred to OR-Bench vary with the source benchmark used to derive them, indicating a benchmark-dependent precision-coverage trade-off. These results suggest that safety alignment in current LLMs is both localized and interaction-sensitive, offering insight into alignment brittleness and potential avenues for targeted safety interventions.

Large language models (LLMs) increasingly act in interactive settings where their behavioral styles affect user experience, safety, and downstream decision making. Existing LLM personality studies largely rely on self-report questionnaires administered in first-person settings, making the resulting profiles sensitive to surface elicitation choices and poorly grounded in concrete model behavior. In this work, we introduce a situated behavioral-data (B-data) framework for studying and controlling LLM behavioral personality. We construct 3,200 contrastive behavioral scenarios spanning 20 behavioral patterns and four prompt registers, grounded in validated psychometric facets such as BFI-2, DOSPERT, and HEXACO. Using this framework, we find that LLMs exhibit stable and model-specific behavioral profiles, while also revealing register-dependent shifts across first-person decisions, advice-giving, and task execution. We then show that these behavioral patterns can be controlled through Behavioral Mode Axes (BMAs), activation-space directions derived from contrastive behavioral traces. Compared with response-derived BMAs, which are more prone to trait drift, thought-derived BMAs more faithfully capture the intended behavioral mechanism and provide cleaner control over situated behavioral styles. Our results suggest that LLM personality-like tendencies are better understood not as abstract self-report traits, but as measurable and controllable behavioral modes grounded in concrete interaction contexts. Our code and data are available at https://github.com/lhz191/LLM-Behavioral-Personality.