IJCAI 2026 Papers

Large Language Models (LLMs) are increasingly deployed in diverse cultural contexts, yet their ability to master aesthetic stylistics, i.e., the strategic use of language to evoke cultural resonance, remains underexplored. We curate C4Styli, a benchmark of highly stylized translated movie titles and advertising slogans from Hong Kong and the Chinese Mainland, to evaluate LLMs via the lens of behavioral recognition and productive competence. Extensive evaluations show that LLMs differ from humans in stylistic recognition, and this recognition ability varies across text domains. In addition, stylistic recognition and generation performance in LLMs are not consistently aligned. To further examine whether LLMs genuinely capture stylistic information in stylistic recognition, we conduct structural ablation with logistic regression probes. We find that, in the Hong Kong setting, stylistic recognition in LLMs relies primarily on surface-level linguistic information rather than stylistic structure. This suggests limited sensitivity to Hong Kong-specific stylistic structure. Our code and data are available at https://github.com/wangjs9/C4STYLI.

Artificial intelligence (AI) applications are prone to reproducing existing patterns of marginalization, bias, and discrimination. The adoption of automated decision-making tools, particularly those based on large language models (LLMs), necessitates the detection and analysis of their discriminatory behavior to ensure compliance with anti-discrimination regulations. This work advances metamorphic testing for bias detection, aiming to (1) uncover discriminatory behavior in AI tools and (2) standardize the metamorphic test creation approach, rooting it in the legal and academic research on discrimination. To fill in these gaps, we introduce a metamorphic test creation workflow and evaluate it on the LLM-based CV screening use case. Our method yields a metamorphic relation (MR) catalogue of 8 atomic and 6 composite MRs, legally justified by Article 21 (1) of the Charter of Fundamental Rights of the European Union and sociologically motivated by the intersectionality theory. The resulting metamorphic test suite uncovers discriminatory failures for all 6 evaluated LLMs, totaling 1,272 defects across 7,800 test cases. Both the new atomic MRs and the composite MRs detect additional discriminatory defects for all LLMs. Composite MRs consistently uncovered more defects than their atomic counterparts, with failure rates increasing by on average 9% for Llama-3.1 8B and 25% for Llama-3.2 3B.

Algorithmic recourse provides counterfactual action plans that help people overturn unfavorable AI decisions. While diverse recourse sets may improve transparency and motivation, they may also impose cognitive load and negative emotions by increasing counterfactual reasoning demands. To examine this trade-off, we conducted a between-subjects controlled experiment (N=750) that manipulated recourse-set diversity while controlling the number of options, and evaluated its effects on psychological benefits and costs. Results show that diversification enhances psychological benefits (e.g., willingness to act) for small sets without incurring additional psychological costs, whereas for large sets, it makes cognitive load more salient. These findings suggest that naively diversifying recourse can burden decision subjects, underscoring the need for new diversification methods that incorporate human cognition and psychology to mitigate such costs.

With the growing adoption of artificial intelligence in high-stakes decision-making domains, identifying the causes of outcomes--particularly failures--and determining who is responsible has become a critical concern. In this work, we investigate how well formal definitions of responsibility attribution, grounded in the framework of actual causality, align with human judgments of responsibility. To this end, we conduct a large-scale survey to elicit human judgments of responsibility in multi-agent sequential decision-making scenarios, using a modified version of the card game Goofspiel. We evaluate different responsibility attribution methods, assessing their alignment with human judgments about responsibility, and identifying factors that significantly shape responsibility judgments. While no single responsibility attribution method consistently aligns with human responses, our findings highlight key factors that influence human responsibility judgments, including agent-specific biases and the amount of information available to agents during decision-making.

In subjective tasks, different individuals can have different correct answers for the same input—the ground truth is not fixed but rather determined by each user's personal perspective. Standard foundation models suppress this individual variation by producing population-averaged predictions; conversely, few-shot in-context learning (ICL) struggles to capture user-specific preferences from limited demonstrations. To address this limitation, we propose Population-Guided Bayesian Calibration (PGBC), a test-time personalization method that applies Bayesian inference to foundation model outputs for user-specific alignment. Unlike ICL methods that rely on the model to implicitly infer preferences from examples, PGBC explicitly quantifies each individual's deviation from the population distribution and incorporates this into prediction with minimal feedback (K < 10). Experiments on three subjective benchmark datasets across four foundation models demonstrate that PGBC significantly outperforms zero-shot baselines with a large effect size (Cohen's d = 1.21), and the improvement over ICL approaches is even more pronounced (d = 1.40). Furthermore, user-level distribution analysis reveals that PGBC aligns predictions with individual preferences when zero-shot baselines exhibit substantial bias, while ICL methods increase distributional mismatch. Our results show that PGBC enables practical deployment in cold-start scenarios where minimal user feedback must yield maximal personalization.

We examine a particular adversarial human--robot interaction for which mobile robots influence the behavior of people role-playing as a school shooter. Using a high-fidelity virtual reality environment, we conducted a controlled study with 150 participants. Two autonomous robots predicted participant movement and positioned themselves to interfere and distract. The robots' approach strategy was manipulated---either moving directly into the participant's path (aggressive) or maintaining distance (passive)---along with the level of distraction, ranging from no additional cues (low), to siren and lights (medium), to siren, lights, and smoke to impair visibility (high). An aggressive, high-distraction robot configuration reduced the number of victims by 46.6% relative to a no-robot control. These results show that robot-based distraction can meaningfully alter human behavior and outcomes in adversarial settings, while also raising important ethical questions about the use of such systems in school environments.

Aligning large language models (LLMs) with diverse human values is essential for safe and effective deployment, yet existing benchmarks often overlook cultural and demographic variation. We introduce DiverValue-Bench, a population-aware benchmark for evaluating multi-dimensional value alignment across 74 countries/regions. It contains 23,763 quality-controlled instances derived from PRISM user feedback and audited through large-scale human validation, with fine-grained value labels, personalized questions, contrastive reference answers, and rich demographic metadata. Using DiverValue-Bench, we evaluate representative LLMs and reveal substantial geographic and demographic disparities that are masked by aggregate performance. We further show that lightweight preference-based fine-tuning with Low-Rank Adaptation (LoRA) and Direct Preference Optimization (DPO) substantially improves in-domain value alignment while yielding consistent out-of-domain gains. These results highlight the need for population-aware alignment evaluation and demonstrate the utility of DiverValue-Bench as a practical foundation for global alignment, personalized value modeling, and equitable AI development.

As large language models (LLMs) increasingly serve as autonomous agents in social simulations, ensuring their ability to understand and comply with legal norms is essential. Yet, current LLM agents frequently exhibit reward hacking (RH) behaviors by optimizing metrics at the expense of norm adherence, undermining simulation fidelity and limiting deployment. We introduce a TBC-TBA self-refine learning multi-agent framework that enables dynamic normative adaptation through iterative multi-agent feedback. This framework integrates Think-Before-Chat (social feedback processing) and Think-Before-Act (norm-guided decision making) phases, allowing agents to progressively refine their normative understanding via structured interaction cycles. Across five mainstream LLMs and 100 legal scenarios, we found that while LLMs partially recognize legal norms, they systematically exhibit RH behaviors with illegal action rates (IAR) of 14.29–37.11%. Comparison with human cognition reveals alignment in moral reasoning but sharp divergence in risk perception and probability distortion. To address these deficits, we introduce four methods to improve LLM's normative compliance. Dynamic Norm Learning Mechanism (DNLM) serves as the core, using a psychologically grounded identify–infer–implement process that reduces IAR by 15.78% on average and delivers the most significant improvement. We also introduce Deep MaxPain (DMP) for consequence based deterrence, Norm Analysis Chain-of-Thought (NA-CoT) for structured reasoning, and Few-shot Norm Learning (FNL) for case based acquisition, all of them enhance compliance. Our findings show that LLM agents can better follow legal norms when equipped with structured self-refine learning and psychologically informed mechanisms. This work improves social alignment in multi-agent systems and opens avenues for future research on scalable, norm-compliant autonomous agents.The code and data are publicly available on GitHub.

The robustness of autonomous vehicles such as drones and Unmanned Surface Vehicles (USV) is crucial when facing unknown and complex marine environments, especially when heteroscedastic observational noise poses significant challenges to sensor-based navigation tasks. Recently, Distributional Reinforcement Learning (DistRL) has shown promising results in some challenging autonomous navigation tasks without prior environmental information. However, these methods overlook situations where noise patterns vary across different environmental conditions, hindering safe navigation and disrupting the learning of value functions. To address the problem, we propose DRIQN to integrate Distributionally Robust Optimization (DRO) with implicit quantile networks to optimize worst-case performance under natural environmental conditions. Leveraging explicit subgroup modeling in the replay buffer, DRIQN incorporates heterogeneous noise sources and target robustness-critical scenarios. Experimental results based on the risk-sensitive environment demonstrate that DRIQN significantly outperforms state-of-the-art meth- ods, achieving +13.51% success rate, -12.28% collision rate and +35.46% for time saving, +27.99% for energy saving, compared with the runner-up.

Offline robotic control requires long-horizon reasoning from fixed datasets while avoiding unsafe extrapolation beyond demonstrated behavior. We propose GRALP, a principled framework that resolves this tension by jointly enforcing support preservation and controllability at the level of temporal abstraction. GRALP adopts a deliberate architectural separation: diffusion is used exclusively as a deterministic action decoder for executing fixed latent skills, while planning and value estimation operate entirely in latent space under conservative constraints. This design enables stable value learning, controllable skill composition, and efficient planning without trajectory-level diffusion sampling at inference. Across unified D4RL benchmarks, GRALP achieves the highest average performance on Navigation, Sequential (Kitchen), and Adroit domains while remaining competitive on locomotion tasks. On contact-rich RoboSuite manipulation with human demonstrations (Lift and Pick-and-Place), GRALP achieves consistently high success rates (over 94%). These results indicate that reliable long-horizon offline control emerges when expressivity is confined to execution and decision-making operates over support-aligned latent abstractions.

This paper addresses the Motion Execution Gap, the disconnect between high-level symbolic task descriptions using semantic constraints and executable robot motions. \textbf{Motion Statecharts} are introduced as an executable symbolic representation for complex motions. They allow the arbitrary arrangement of motion constraints, monitors or nested statecharts in parallel and sequence. World-centric motion specification and generalization across embodiments are enabled through the use of a unified differentiable kinematic world model of both, robots and environments. Motion execution is realized through a \ac{lmpc}-based implementation of the task-function approach, in which smooth transitions during task switches are ensured using jerk bounds. Cross-platform transferability was demonstrated by deploying the method on eight robot platforms, operating in diverse environments. The proposed framework is called Giskard and is available open source (https://github.com/cram2/cognitive_robot_abstract_machine).

Safe and precise maneuvering of quadrotor unmanned aerial vehicles (UAVs) in high-speed wind environments remains a critical challenge. Wind disturbances are nonlinear, time-varying, and difficult to model, causing traditional controllers to struggle with perception and compensation, especially under unseen wind distributions. To address these limitations, we introduce WA-TD3, a data-driven control framework that enables real-time wind disturbance perception and adaptive compensation without dedicated wind sensors. WA-TD3 employs a deep residual network to extract wind characteristics from temporal patterns in state deviations, forming a dynamics residual-driven perception mechanism that implicitly models and compensates for unknown winds. This residual is integrated into a perception-augmented reinforcement learning architecture, providing the policy with enhanced state information for proactive disturbance-aware control. Extensive experiments on complex trajectories under varying wind intensities demonstrate that WA-TD3 consistently outperforms state-of-the-art methods, achieving over 62% improvement in tracking accuracy under strong winds.

Vision-language-action models (VLAs) often achieve high performance on demonstrated tasks but struggle significantly when required to extrapolate, recombining skills used in different tasks in novel ways. For instance, VLAs might successfully put the cream cheese in the bowl and put the bowl on top of the cabinet, yet still fail to put the cream cheese on top of the cabinet. This motivates us to investigate whether VLAs merely overfit to demonstrated tasks or still hold the potential to extrapolate. Our study uses text latent as the ingredient; it is a task-specific vector derived from the models’ hidden states. It thus encodes semantics necessary for completing a task and can be used to reconstruct the associated task behavior by writing it to the model’s residual stream. Furthermore, we find that skills used in distinct tasks can be combined to produce novel behaviors by blending their respective text latent. Applying this to π0, we increase its success rate from 9% to 83% on the proposed libero-ood benchmark, which features 20 tasks extrapolated from standard LIBERO tasks. This reveals that the skill representations encoded in text-latent are individual yet composable, while π0 fails to autonomously combine these representations for extrapolation. This also validates the design of libero-ood; it comprises tasks that the model fails, yet should be able to complete. We then tested other VLAs on libero-ood, and none of them achieved a success rate higher than 21%. Further analysis reveals VLAs share a common pattern to exhibit spatial overfitting, associating object names with where the object is spatially located in the demonstrated scene rather than achieving true object and goal understanding.

Motion planning among multiple robots in a shared space is a fundamental yet computationally challenging problem in robotics, with applications ranging from warehouse automation to autonomous fleets. In this work, we introduce a fast, scalable motion planner that achieves real-time, collision-free trajectory planning via a two-staged algorithm combining deterministic search-based planning with machine learning-driven conflict resolution. We present a prioritized Safe Interval Path Planning algorithm (SIPP-PP) with a novel limited goal reservation strategy to prevent goal-blocking conflicts while allowing shared goal regions. We added a second layer of ML-guided Large Neighborhood Search (LNS) procedure to our SIPP-PP algorithm for improving success rates in highly congested environments via intelligent selection of conflict resolution actions. The result is a planning system that generates collision-free paths for multiple robots in complex environments within tens of milliseconds. For example, compared to recent advanced learning-based methods such as diffusion planners, our planner is two-to-three orders of magnitude faster. Our work demonstrates a multi-robot planner capable of real-time operation in dense scenarios, satisfying the stringent requirements of industrial applications such as drive units in fulfillment centers.

Vision-and-language navigation (VLN) requires agents to follow natural language instructions to navigate autonomously in continuous environments. However, existing approaches often lack high-level semantic guidance in waypoint prediction and explicit language–landmark alignment in cross-modal planning. To address these limitations, we propose FILD-Nav, a vision-and-language navigation framework that integrates instruction landmark features. FILD-Nav extracts task-relevant landmarks from instructions and incorporates landmark semantics into both waypoint prediction and topological planning. Specifically, landmark-guided waypoint prediction improves waypoint relevance, while landmark-enhanced cross-modal planning enables more effective long-horizon navigation. Extensive experiments on the VLN-CE benchmark demonstrate that FILD-Nav consistently outperforms prior methods, achieving improvements of 2% in Success Rate (SR), 3% in Success weighted by Path Length (SPL), and 7% in Oracle Success Rate (OSR), particularly in unseen environments.

In the final "Last-Centimeter" phase of manipulation, where visual occlusion or calibration errors render vision unreliable, robots often suffer high failure rates due to local pose uncertainty and simulation dynamics deviations. To address these issues, this paper proposes the PECHC (Physics-Evolving Cascade Constraint and Human-Correction) algorithm. To rigorously isolate the contribution of tactile feedback in multi-finger coordination, we adopt a decoupled control strategy that focuses on grasp stabilization within the hand's workspace, acting as a fail-safe reflex. The core of our approach is Hybrid Correction Imitation Learning (HCIL), which establishes a "failure-triggered" human-machine mechanism to efficiently resolve the "model gap" via sparse expert corrections. To ensure sample efficiency and baseline performance, we introduce two supporting modules: Cascaded Constraint Scheduling (CCS) addresses the "geometric gap" by enforcing physically plausible behavioral constraints (geometric approach, force closure, and dynamic stability), while Temporal Heterogeneous Distillation (THED) resolves the "physical gap" by enabling implicit system identification from tactile history. Experiments demonstrate that PECHC achieves a 97.3% real-robot success rate on 150 objects from the Visual Dexterity Dataset under fully autonomous testing, where one object is used for one-time HCIL calibration and the remaining 149 objects are evaluated without further intervention. Compared to a standard Sim-to-Real reinforcement learning baseline (Vanilla PPO with Domain Randomization), PECHC delivers a significant performance improvement (+42.8%) and exhibits human-like force modulation capabilities for fragile objects.

The increasing sensitivity to human aspects in autonomous systems engineering calls for principled approaches to embed ethical and normative concerns into their behaviour. Indeed, recent research has focused on expressing and validating sets of social, legal, ethical, empathetic, and cultural (SLEEC) concerns as rules. However, existing work is limited to rule specification or verification, leaving the problem of semantic-preserving operationalisation of ethical rules in autonomous systems largely unaddressed. For this purpose, we provide an operational solution for ethical-aware autonomous systems, applied to the realm of multi-service robots. Specifically, we devise a principled approach, named CO-SLEEC (Context-Oriented SLEEC), connecting the normative setting of SLEEC rules to context-oriented programming (COP). CO-SLEEC enables runtime adaptation while preserving the semantics of SLEEC rules during robot task execution. It features two reusable Python libraries for (i) parsing SLEEC rules into contextual elements for operationalising them, and (ii) connecting the operational model to the Robot Operating System (ROS), respectively. We evaluate our implementation for correctness, efficiency, and maintainability over multi-service assistive robot scenarios.

Robots that follow natural-language instructions often either plan at a high level using hand-designed interfaces or rely on large end-to-end models that are difficult to deploy for real-time control. We propose TeNet (Text-to-Network), a framework for instantiating compact, task-specific robot policies directly from natural language descriptions. TeNet conditions a hypernetwork on text embeddings produced by a pretrained large language model (LLM) to generate a fully executable policy, which then operates solely on low-dimensional state inputs at high control frequencies. By using the language only once at the policy instantiation time, TeNet inherits the general knowledge and paraphrasing robustness of pretrained LLMs while remaining lightweight and efficient at execution time. To improve generalization, we optionally ground language in behavior during training by aligning text embeddings with demonstrated actions, while requiring no demonstrations at inference time. Experiments on MuJoCo and Meta-World benchmarks show that TeNet produces policies that are orders of magnitude smaller than sequence-based baselines, while achieving strong performance in both multi-task and meta-learning settings and supporting high-frequency control. These results show that text-conditioned hypernetworks offer a practical way to build compact, language-driven controllers for ressource-constrained robot control tasks with real-time requirements.

On-demand delivery platforms rely on Travel Time Estimation (TTE) to balance courier earnings and overdue risks. In collaboration with one of China's largest platforms, we address a critical "Fairness Gap" in TTE: current systems fail to capture complex delivery patterns in GNSS-denied environments, subjecting couriers handling high concurrent order volumes to disproportionate pressure due to overdue deliveries. Analyzing 1.27 million real-world trajectories, we attribute this bias to unique challenges in GNSS-denied scenarios: distributional heterogeneity, structural heterogeneity, and contextual uncertainty. To bridge this gap, we propose DART (Distribution-Adaptive Robust Timing). DART incorporates a Learnable Adaptive Spline (LAS) encoder with a gradient-driven knot migration mechanism to enhance non-linear expressiveness for outliers, significantly improving long-tail accuracy. Furthermore, a Spatio-Temporal Transition Graph (STTG) reconstructs the latent topology by integrating sequence semantics, such as Wi-Fi-sensed arrival merchant timestamps. At the same time, a Distribution Gating Mechanism characterizes delivery time distributions under distinct contexts. Through extensive experiments and large-scale online A/B testing, DART not only reduces MAE by 14.0% in complex environments but also decreases the Order Overdue Rate by 1.7% (saving $24,000 daily), demonstrating how AI effectively reconciles operational efficiency with labor fairness.