IJCAI 2026 Papers

Modern AI-driven systems rely on large populations of autonomous agents that make decentralized routing decisions under uncertainty, where rare but severe tail latency can critically degrade quality of experience, safety, and reliability. To model agents’ aversion to such tail latency, we study nonatomic selfish routing games in which agents minimize the Conditional Value-at-Risk (CVaR) of path latency. CVaR explicitly captures both the likelihood and severity of tail latency, but its inherent nonadditivity across network edges poses a fundamental challenge. We address it by identifying a worst-case dependence structure—tail risk concentration—under which tail latency across network edges is synchronized. We show that CVaR penalizes this dependence structure and becomes additive under worst-case tail dependence, which enables tight inefficiency analysis. To quantify the resulting inefficiency induced by risk-aversion, we adopt the price of risk aversion (PRA), defined as the worst-case ratio between the total system cost at a risk-averse equilibrium and that at a risk-neutral equilibrium. We show that, for arbitrary latency functions and general network topologies, the PRA under CVaR admits a tight upper bound that grows linearly with both the network size and the maximum edge-level upper-tail cost. We further prove that this bound is tight by constructing a family of Braess-type networks that achieve a matching lower bound. These results provide the first tight worst-case inefficiency bound for CVaR-based selfish routing and offer insights into how tail-risk-averse decision making by autonomous agents amplifies congestion externalities in large-scale multi-agent systems.

We study repeated allocation of shared resources among agents with time-varying demands and capped linear utilities. In this setting, independently maximizing the minimum utility in each round satisfies sharing incentives (agents weakly prefer participating in the mechanism to not participating), strategyproofness (agents have no incentive to misreport their demands), and Pareto efficiency. However, this max-min mechanism can lead to large disparities in the total resources received by agents, even when they have the same average demand. We introduce credit fairness, a property that, together with Pareto efficiency, strengthens sharing incentives by ensuring that agents who lend resources in early rounds are able to recoup them in later rounds. Credit fairness can be achieved in conjunction with either Pareto efficiency or strategyproofness individually, but we show that, under anonymity, it cannot be achieved together with both. We propose a mechanism that is credit fair and Pareto efficient, and evaluate it in a computational resource-sharing setting.

Most of the existing decentralized multi-agent systems facilitate coordination by achieving consensus. However, reaching global consensus among a large number of agents in complex settings is challenging because resolving a local disagreement often triggers cascading adjustments across the system. This problem is further exacerbated under intermittent communication. In the field of transportation, human drivers usually rely on intentions (e.g., turn signals) in conjunction with traffic rules (such as "yielding to through traffic") to achieve efficient cooperation, instead of establishing consensus with all surrounding vehicles. Inspired by this observation, we propose the Intention \otimes Latent Rule paradigm to facilitate coordination beyond consensus, which is instantiated by DIRECT (Decentralized Intention and Latent Rule Emergence Cooperation). Within this framework, agents forecast their intentions based on current actions, and the latent rule inference module derives latent-rule representations from available information to regulate competing intentions, enabling cooperative action selection even when communication is disrupted. Extensive experiments on challenging StarCraft II micromanagement and Multi-Agent Particle Environments demonstrate that DIRECT consistently outperforms state-of-the-art methods, achieving more robust coordination across a range of communication-degraded scenarios. Our code is available at https://github.com/AAI-ZZU/DIRECT-master.

Building upon powerful Large Visual Language Models, recent GUI agents have revolutionized autonomous GUI interaction. Given the high information density and structural complexity of GUI layouts, a critical challenge lies in accurately identifying where to focus, i.e., precise GUI grounding. To ensure grounding accuracy, prior approaches predominantly rely on processing inputs at native high resolutions. However, this paradigm inevitably introduces excessive visual redundancy, resulting in increased computational cost and dispersed attention. In our preliminary analysis, we observe that while high-resolution inputs are indispensable for recognizing fine-grained visual details, they paradoxically impair the model’s ability to attend to the correct regions. In contrast, low-resolution inputs, which suppress high-frequency details while preserving global layout structure through downsampling, produce significantly more robust and coherent attention patterns that better guide the model toward relevant regions. Interestingly, this observation mirrors the human visual system, which first leverages low-resolution peripheral vision to identify salient areas before conducting a detailed examination via high-resolution foveal focus. Motivated by this insight, we propose a coarse-to-fine GUI grounding framework termed FastFocus. Our approach first exploits the stability of low-resolution attention to efficiently generate candidate grounding regions. These candidates are then evaluated by a Parallel Verification module, which selectively zooms into high-resolution views to resolve fine details and filter out false positives. Experiments on ScreenSpot-Pro demonstrate state-of-the-art performance, validating that a hierarchical integration of low-resolution guidance and high-resolution verification constitutes an effective and robust paradigm for GUI grounding.

Autonomous agents are increasingly expected to support scientific research, and recent benchmarks report progress in code repair and autonomous experimentation. However, these evaluations typically assume a pre-configured execution environment, which requires resolving complex software dependencies, aligning hardware and framework versions, and configuring distributed execution, yet this capability remains largely unbenchmarked. We introduce ResearchEnvBench, a benchmark for environment synthesis in research code execution. Given a research repository, documentation, and a target execution setting, agents must construct an environment that successfully executes at runtime. Evaluations on diverse research repositories reveal a substantial gap in current SOTA agents, with failures dominated by incomplete dependency resolution and brittle version coupling. ResearchEnvBench provides a realistic testbed for advancing autonomous agents toward reproducible scientific research.

While Large Language Models (LLMs) have shown promise in translating natural-language specifications to Register-Transfer Level (RTL) designs, they often fail on complex, port-rich IPs. Existing frameworks typically separate generation from debugging, relying on static decomposition and iterative repair, which is hard to verify and yields unstable, inefficient maintenance. In this work, we propose VPID, a multi-agent framework for generating complex Verilog that achieves monotonic functional improvement. Given RTL’s inherent concurrency and the alignment between verification and port-level behavior, our key insight is to treat ports as verifiable boundaries for dynamic decomposition, enabling fine-grained analysis to pinpoint root causes and guide targeted debugging. Specifically, we implement a behavior locking mechanism that preserves the behavior of verified ports to ensure monotonicity, preventing regressions in correct functionalities. To accelerate convergence, we introduce an experience-guided refinement strategy that distills historical waveform mismatches into constraints, guiding the targeted debugging for the unverified ports. Together with precise Abstract Syntax Tree (AST)-based code extraction, these mechanisms enable an efficient incremental generation workflow. Experiments on RealBench demonstrate that VPID outperforms both one-pass general-purpose LLMs and existing agent-based frameworks in both syntax and functional correctness, presenting a robust approach for automating Verilog generation for complex RTL designs.

We introduce Hedonic Project Games, a model in which agents choose projects with divisible rewards while holding subjective preferences over coalition composition. This framework captures a fundamental trade-off absent from existing models: agents care simultaneously about who they collaborate with and what they work on. We study three stability notions: classical Nash Stability and two refinements, Joining Stability and Leaving Stability, which account for the welfare of both the deviating agent and the affected coalition members. We evaluate the efficiency of stable outcomes using the Price of Anarchy and Price of Stability, comparing the social welfare of stable outcomes to that of an optimal allocation. While stable outcomes may not exist in general, we identify broad and natural preference classes in which stability and efficiency improve significantly. In particular, under monotonic-decreasing preferences in coalition size, Nash and joining stability coincide and are guaranteed to exist, whereas leaving stability may fail. Under per-capita non-decreasing preferences, socially optimal outcomes are always Nash stable and coincide with leaving stability, although equilibrium inefficiency remains unbounded. Experiments on synthetic and real-world data support the theoretical efficiency results.

Agents in the real world must often balance multiple objectives, such as speed, stability, and energy efficiency in continuous control. To account for changing conditions and preferences, an agent must ideally learn a Pareto frontier of policies representing multiple optimal trade-offs. Recent advances in multi-policy multi-objective reinforcement learning (MORL) enable learning a Pareto front directly, but require full multi-objective consideration from the start of training. In practice, multi-objective preferences may arise after a policy has already been trained on a single specialised objective. Existing MORL methods cannot leverage such a pre-trained ‘specialist’ to learn Pareto fronts and avoid incurring the sample costs of retraining. We introduce Mixed Advantage Pareto Extraction (MAPEX), an offline MORL method that constructs a frontier of policies by reusing pre-trained specialist policies, critics, and replay buffers. MAPEX combines evaluations from specialist critics into a mixed advantage signal, and weights a behaviour cloning loss with it to train new policies that balance multiple objectives. MAPEX’s post hoc Pareto front extraction preserves the simplicity of single-objective off-policy RL, and avoids retrofitting these algorithms into complex MORL frameworks. We formally describe the MAPEX procedure and evaluate MAPEX on five multi-objective MuJoCo environments. Given the same starting policies, MAPEX produces comparable fronts at 0.001% the sample cost of established baselines.

A core challenge for embodied agents is the ``semantic-to-physical gap"—the difficulty of mapping symbolic reasoning to precise execution. While Vision-Language Models (VLMs) enhance agent task planning, they often fail in problem classes requiring accurate alignment between functional geometry and physical constraints, such as articulated object manipulation or precision assembly. To address these challenges, we propose GRACE, an agent framework that adopts Executable Analytic Concepts (EAC) as a core knowledge annotation paradigm for object understanding. Under this paradigm, the agent does not merely perceive objects as unstructured visual data; instead, it interprets and annotates them as ``Semantic-Physical Blueprints", which provide a structured cognitive representation that encodes geometric primitives, mechanical affordances, and manipulation constraints. Central to our framework is a policy scaffolding mechanism, which allows the agent to dynamically ground VLM-based insights into these instantiated blueprints. This enables the agent to model and resolve complex decision-making tasks involving parameterized object descriptions and constrained motion planning. Experimental results demonstrate that agents utilizing this paradigm achieve high robustness in complex manipulation tasks. Furthermore, we demonstrate that GRACE supports automated concept discovery, offering a scalable and modular solution for high-precision embodied intelligence.

Temporal planning and scheduling often rely on constraint models to reason about activity durations. Simple Temporal Networks with Uncertainty (STNUs) address cases where some contingent durations are outside the control of the executing agent. The Multiple Interdependent STNU (MISTNU) model extends this setting to multi-agent systems by representing shared activities whose duration is decided by one agent and imposed on others. While controllability properties for MISTNUs have been defined, existing repair methods rely on a centralized SMT-based brute-force approach. This paper introduces a linear-constraint characterization of all negative cycles responsible for uncontrollability. Based on this formulation, we propose both a more efficient centralized linear-programming repair method and a distributed constraint reasoning approach that treats inconsistent cycles as inter-agent constraints. We experimentally compare several Distributed Constraint Optimization Problem (DCOP) solvers against the centralized baseline in terms of efficiency and repair quality.

In simulation-in-the-loop decision-making systems, reinforcement learning (RL) inference is often constrained by simulator-side execution overhead, where workloads are highly dynamic and sensitive to runtime thread configurations. Existing multithreaded strategies struggle to match thread resources before or during execution, causing resource contention, scheduling overhead, and reduced throughput. Through empirical analysis, we identify the ratio of task execution time to scheduling time as the key factor determining the optimal thread count. Building on this insight, we propose AutoThread, a hybrid adaptive thread-tuning method for mitigating simulation bottlenecks in RL inference. AutoThread employs a Physics-Informed Neural Operator (PINO) as a thread-count predictor and incorporates a finite-source M/M/1 queueing model to constrain and guide prediction, enabling fast and accurate estimation under dynamic workloads. It further performs load-aware online fine-tuning to compensate for prediction errors and refine resource allocation. Experiments show that AutoThread improves average speedup by 18.4% over static strategies, achieves average throughput of 1.7× and 1.8× that of XGBoost and Reinforcer, respectively, and reduces execution time by up to 83.8% compared with state-of-the-art methods. Our code and dataset are publicly available at https://github.com/suchenjm/AutoThread.

We study the consequences of information asymmetries and misaligned incentives in settings with multiple independent agents. We model an interaction between a Sender, who holds vital private information but cannot act, and a Receiver, who must make decisions but is dependent on the Sender's information. We find that the Sender learns an optimal communication strategy that the Receiver reliably acts on. Importantly, this strategy is highly sensitive to the degree of conflict in the agents' rewards and the amount of environmental information the Receiver can already observe. We introduce a mechanism allowing the agents to form linear contracts, where a price is established for the information. We demonstrate that the Sender learns to use these payment structures to improve its rewards, though this comes at a cost of “fairness” between agents as the Sender is able to extract much of the Receiver’s surplus. This raises questions about fairness, contract design, and learning in the context of multi-agent systems.

We study influence maximization in general stochastic models, the observed cascades model, and the independent cascade (IC) model. For general stochastic models with only black-box sample access, we introduce a low-adaptivity optimization framework that improves sample complexity and running time over the prior work and is instrumental to all our results. We further introduce an adaptive algorithm guided by empirical variance, avoiding pessimistic worst-case bounds. Combining our optimization framework with sketching, we obtain the first algorithm with provable guarantees and nearly-linear running time for influence maximization on observed cascades, optimal up to logarithmic factors. For IC, we prove a novel tail bound replacing a factor n with ? (the number of diffusion steps) in sample complexity, improving over prior work when ? is small, as is common due to small-world phenomena. Experiments confirm substantial speedups while maintaining solution quality.

Retrosynthesis prediction is a cornerstone of drug discovery, enabling the synthesis of novel therapeutic candidates. However, current deep learning models falter when navigating the unexplored chemical space essential for innovation. In these realistic scenarios, models face a challenging mixed out-of-distribution (OOD) shift: they must handle both novel molecular scaffolds (covariate shift) and unseen reaction templates (label shift), leading to severe performance degradation. To address this, we propose AgentRetro, a framework integrating spectral domain generalization with Large Language Model (LLM) agents. It features a Spectral Domain Enhancer (SDE) that partitions chemical space via spectral clustering to capture diverse domain structures, and a Multi-Knowledge Agent (MKA) that enriches data with LLM-generated OOD reactions and textual rationales. These components are unified by a dual training strategy: meta-learning for template adaptability and adversarial learning for structural invariance. Crucially, our analysis reveals that explicitly decoupling structural invariance from mechanistic adaptability is essential for overcoming the interference between conflicting distribution shifts. Experiments show AgentRetro consistently improves state-of-the-art baselines, notably boosting EditRetro’s OOD Top-10 accuracy from 41.7% to 50.1%. The source code is available at https://github.com/jiahangshen/IJCAI26-AgentRetro.

Multi-Agent Path Finding (MAPF) seeks collision-free paths for teams of agents and has a wide range of practical applications. LaCAM*, an anytime configuration-based solver, currently represents the state-of-the-art. Recent work has explored using guidance paths to steer LaCAM* toward configurations that avoid traffic congestion, thereby improving solution quality. However, existing approaches rely on Frank–Wolfe–style optimisation to repeatedly invoke single-agent search before executing LaCAM*, which creates a large computational overhead in large-scale problems. The guide path is also static, which is only helpful for finding the first solution in LaCAM*. To overcome this problem, we propose a new approach that exploits LaCAM*’s ability to construct a dynamic, lightweight traffic map during LaCAM*'s search. Experiments show that our method achieves higher solution quality than state-of-the-art guidance-path approaches in two variants of MAPF problems.

Reward learning via human feedback is a crucial capability for beneficial AI. Current methods are built on decision-making theories that assume a matched dynamics model between the learning agent and the feedback provider. However, humans often form imperfect internal dynamics models, and their feedback reflects these misconceptions. While this relationship has long been hypothesised, its manifestation in sequential decision-making remains largely an assumption. Our work provides the first comprehensive empirical investigation of this relationship through a randomized controlled trial (N=211). We followed a two-stage design where we first initialized the participants' understanding of the dynamics in a grid-world navigation domain and then manipulated it using text-based instructions. Causal mediation analysis revealed that humans' internal models play a mediating role in feedback behaviour. We show that this relationship is invariant across visual contexts and is robust to three common feedback types: pairwise preferences, trajectory corrections, and off-switch interventions. These findings confirm a critical limitation of current reward learning methods and establish the missing psychological foundation for approaches that incorporate dynamics understanding.

Reward specification plays a central role in reinforcement learning (RL), guiding the agent’s behavior. To express non-Markovian rewards, formalisms such as reward machines have been introduced to capture dependencies on histories. However, traditional reward machines lack the ability to model precise timing constraints, limiting their use in time-sensitive applications. In this paper, we propose timed reward machines (TRMs), which are an extension of reward machines that incorporate timing constraints into the reward structure. TRMs enable more expressive specifications with tunable reward logic, for example, imposing costs for delays and granting rewards for timely actions. We study model-free RL frameworks (i.e., tabular Q-learning) for learning optimal policies with TRMs under digital and real-time semantics. Our algorithms integrate the TRM into learning via abstractions of timed automata and employ counterfactual-imagining heuristics that exploit the TRM's structure to improve search. Experimentally, we demonstrate that our algorithm learns policies that achieve high rewards while satisfying the timing constraints specified by the TRM on popular RL benchmarks.

Multi-agent pathfinding (MAPF) under one-shot planning is a core component of warehouse automation, yet classical formulations typically assume four-connected 2D grids with unit-time moves in four directions. To fill reality gaps while still being trackable with discrete combinatorial search, this work proposes a more practical counterpart tailored to differential-drive AGVs. We term this multi-agent warehouse pathfinding (MAWPF), featured with four constraints: (i) agent actions are restricted to straight motion and in-place rotation; (ii) rotations require multi-step costs; (iii) acceleration and deceleration are considered, and; (iv) follower collisions are prohibited to prevent rear-end crashes. To solve MAWPF efficiently, we adapt representative suboptimal MAPF algorithms-PP, LNS2, PIBT, and LaCAM-and conduct comprehensive benchmarking. Our experiments reveal that PP and LNS2 struggle to solve instances with many agents, while PIBT-based approaches achieve preferable scalability with increased solution cost. We believe that these constitute an important step toward adapting classical gridworld MAPF to operational warehouse setups.

Multi-robot graph exploration is a central problem in robotics, planning, and multi-agent systems. In this work, we consider the problem of exploring an unknown n-node graph by k robots that must remain connected throughout the process. Such a connectivity is frequently required for safety reasons, and naturally arises in real-world applications such as search-and-rescue and maintenance operations. We study the overhead imposed by not knowing the graph in advance, measured in terms of the competitive ratio of the number of exploration rounds necessary when the graph is unknown (versus the case that it is known). We introduce a novel exploration procedure, DFS-BGS, to tackle the problem, and analyze its performance both theoretically and experimentally. On the theoretical end, DFS-BGS provably achieves a competitive ratio O(k^(1/3)), for the case n < k. Empirically, we compare our online DFS-BGS to COCTA, the SOTA algorithm for trees that are known in advance. Examining the performance of the algorithms on real-world hotel floor plans as well as random graphs over a wide range of parameters, DFS-BGS incurs only a small slowdown, even with hundreds of robots and thousands of nodes.