IJCAI 2026 Papers

Recent advances in large language models (LLMs) allow agents to represent actions as executable code, offering greater expressivity than traditional tool-calling. However, real-world tasks often demand both strategic planning and detailed implementation. Using a single agent for both leads to context pollution from debugging traces and intermediate failures, impairing long-horizon performance. We propose CodeDelegator, a multi-agent framework that separates planning from implementation via role specialization. A persistent Delegator maintains strategic oversight by decomposing tasks, writing specifications, and monitoring progress without executing code. For each sub-task, a new Coder agent is instantiated with a clean context containing only its specification, shielding it from prior failures. To coordinate between agents, we introduce Ephemeral-Persistent State Separation (EPSS), which isolates each Coder's execution state while preserving global coherence, preventing debugging traces from polluting the Delegator's context. Experiments on various benchmarks demonstrate the effectiveness of CodeDelegator across diverse scenarios. Code is available at https://github.com/txfei/code-delegator.

Current self-evolution methods mostly chase recall, they learn from any trajectory that ends with the right answer, even when the path relies on a "lucky guess". Such reasoning appears valid in-distribution but often fails the moment the task shifts slightly. We argue that robust evolution implies Structural Invariance: a reasoning path is valid only if its core dependency graph remains isomorphic under counterfactual perturbations. I-EDI enforces this with Verifiable Counterfactual Simulation (VCS). Instead of treating data generation as mere "sample and filter," we treat it as an intervention. Unlike standard rejection sampling, VCS acts as a structural stress test: it accepts a reasoning trace only if it remains valid across generated counterfactuals (e.g., numerical perturbations, context shifts). While this strict filtering reduces the volume of training data (trading recall for precision), our experiments on MATH and GSM8K show it effectively prevents hallucination accumulation, achieving superior out-of-distribution generalization compared to baselines.

TD(λ) in value-based MARL algorithms or the Temporal Difference critic learning in Actor-Critic-based (AC-based) algorithms synergistically integrate elements from Monte-Carlo simulation and Q function bootstrapping via dynamic programming, which effectively addresses the inherent bias-variance trade-off in value estimation. Based on that, some recent works link the adaptive λ value to the policy distribution in the single-agent reinforcement learning area. However, because of the large joint action space from multiple agents and the limited transition data in Multi-agent Reinforcement Learning, the policy distribution is infeasible to calculate statistically. To solve the policy distribution calculation problem in MARL settings, we employ a parametric likelihood-free density ratio estimator with two replay buffers instead of calculating statistically. The two replay buffers of different sizes respectively store the historical trajectories that represent the data distribution of the past and current policies. Based on the estimator, we assign Adaptive TD(λ), ATD(λ), values to state-action pairs based on their likelihood under the stationary distribution of the current policy. We apply the proposed method on two competitive baseline methods, QMIX for value-based algorithms, and MAPPO for AC-based algorithms, over SMAC benchmarks and Gfootball academy scenarios, and demonstrate consistently competitive or superior performance compared to other baseline approaches with static λ values.

We study the problem of optimally placing charging stations for a set of k anonymous mobile agents, each of which must reach a distinct terminal. The agents are identical and can travel only up to a given distance r on a single charge. The objective is to find a placement of charging stations so that every terminal can be assigned a unique agent capable of reaching it, and the number of charging stations is minimized. The problem is known to be NP-hard in general and solvable in polynomial time only on paths and cycles. In this work, we comprehensively investigate the classical and parameterized complexity of the problem. We first design a polynomial-time algorithm that, given a set of candidate charging stations, decides whether a feasible assignment of agents to terminals exists. This result also establishes the membership of the problem in NP, which was not previously known. We then analyze the parameterized complexity of the problem. Our results show that the problem remains hard even when we consider different combinations of natural parameters and structural parameters of the given network, such as the distance r plus the tree-depth, or the number of charging stations plus feedback vertex set number. On the positive side, we present three fixed-parameter tractable (FPT) algorithms: one parameterized by the number of agents k, a second by modular-width, and a third by vertex cover number. Finally, we give a polynomial-time k-approximation algorithm, as well as a polynomial-time algorithm for trees, by solving a generalized version of the problem that allows for some charging stations to be pre-placed.

Unmanned aerial vehicle swarms in pursuit-evasion requires encirclement efficiency while maintaining safety constraints, facing a critical safety-efficiency trade-off. Existing safe multi-agent reinforcement learning (MARL) methods often yield either unsafe task policies or conservative policies. This is challenging since forward feasible safety under dense interactions requires online safety preview. To address this issue, we propose Safety-Constrained Online Preview Enforcement (SCOPE), a MARL-based algorithm that balances encirclement efficiency and safety by short-horizon preview and safety enforcement. SCOPE learns an online safety-preview dynamics model that rolls out future trajectories to inform encirclement decisions checking. By proposing preview-fused actor-critic, SCOPE uses short-horizon previews for efficient encirclement and less unsafe behavior. Hierarchical safety enforcement performs safety look-ahead and online action correction to maintain forward feasible safety. Experiments show that SCOPE better balances encirclement efficiency and safety than safe MARL baselines, maintaining average encirclement time while reducing the average agent cost by 65.6%. Code could be found at https://github.com/98177qdn/SCOPE.

Reachability games are two-player games played on a graph, where the objective of REACH player is to reach the target set whereas the objective of SAFE player is to stay away from the target set. Reachability games have important applications in artificial intelligence and reactive synthesis, and many of these applications give rise to infinite-state reachability games. In this paper, we study turn-based reachability games on infinite-state graphs defined over valuations of a finite set of real variables. We consider the problem of determining the existence of and computing a winning strategy for REACH player. Our contributions are twofold. First, we propose ranking certificates for reachability games, a sound and complete proof rule for proving that REACH player has a winning strategy from the specified initial state. Second, we consider polynomial reachability games, where transitions and objectives are described by polynomial constraints over real variables, and propose a fully automated algorithm for computing a winning strategy for REACH player together with a formal correctness witness in the form of a ranking certificate. The algorithm is sound, semi-complete, and runs in sub-exponential time. Our experiments demonstrate the ability of our method to solve challenging examples from the literature that were out of the reach of existing methods. Specifically, for the classical Cinderella-Stepmother game, we are able to compute an optimal winning strategy for an arbitrary precision parameter for the first time.

Hedonic games are a central model of coalition formation, yet most of their general subclasses are marked by negative results: stable outcomes often fail to exist, and deciding their existence is typically computationally hard. The most notable exception is the class of hedonic games with the common ranking property (HGCRP), a nontrivial subclass that guarantees both the existence and tractability of stable partitions. The common ranking property implies the top-coalition property, which in turn implies the weak top-coalition property, forming a natural hierarchy of increasingly general domains. In this paper, we extend the frontier of strong existence and tractability results beyond the HGCRP to hedonic games with the top-coalition property (HGTCP) and hedonic games with the weak top-coalition property (HGWTCP). We show that every HGWTCP instance admits a strong individually stable (SIS) partition that can be computed in polynomial time, and prove that our polynomial-time algorithm generates a partition that is both SIS and Pareto optimal (PO) if preferences are strict. These results show that strong stability and tractability results persist well beyond the common ranking framework. We further show that a contractually Nash stable (CNS) partition may fail to exist even in HGTCP, revealing a sharp contrast with HGCRP, where the existence of partitions that are both CNS and PO is guaranteed. Taken together, these results provide a comprehensive characterization of stability and efficiency in top-coalition-based hedonic games, mapping how stability guarantees evolve from HGCRP to HGWTCP.

Strategy synthesis typically follows an all-or-nothing paradigm, returning unrealisable whenever a specification cannot be guaranteed in an uncertain environment. In this paper, we introduce optimal LTLf synthesis, where the goal is to realise as many objectives as possible from a given specification consisting of multiple objectives, especially for the case that they are not all jointly realisable. We first consider max-guarantee synthesis, which commits to a maximal set of objectives that we can a priori guarantee to realise. We then introduce max-observation synthesis, which maximises a posteriori realised objectives that may be incomparable on different executions. Finally, we present incremental max-observation synthesis, which further improves strategies by exploiting opportunities for stronger guarantees when they arise during an execution. Experimental results show that different variations of optimal synthesis scale broadly equally well, solving a large fraction of the benchmark instances within the given timeout, demonstrating the practical feasibility of the approach.

In several socioeconomic-critical decision-making settings, such as fair resource allocation, climate policy, or AI alignment, multiple principals interact within a common arena. While it is well established that these principals may have differing preferences, decision-making under heterogeneous time preferences remains relatively unexplored. In particular, principals may weigh future outcomes differently and may derive distinct utilities from the same decisions. Motivated by such scenarios, we introduce the notion of heterogeneous time preferences in MDPs, where multiple principals possess distinct reward functions and apply different discount factors to future rewards. To compute meaningful decisions in such settings, an AI agent must rely on a notion of optimality that accounts for the preferences of all principals. We adopt a utilitarian notion of social welfare, defined as the sum of utilities accrued to all principals, and study the synthesis of agent strategies that maximise this welfare. Under heterogeneous time preferences, we show that optimal strategies are no longer positional, even when all principals receive identical rewards. Nevertheless, optimal strategies remain structurally simple: they can be realised as pure finite-memory counting strategies, require only polynomial memory in the system size, and can be synthesised in polynomial time. On the other hand, we show that deciding threshold questions for optimal positional strategies is NP-hard, exposing a poor trade-off: insisting on positional simplicity neither makes synthesis tractable nor preserves social welfare.

Achieving optimal coordination in multiagent systems involves a trade-off between intractable centralized planning and suboptimal decentralized execution. We bridge this gap by introducing Approximate Recursive Small Step-Semi-Decentralized A* (RS-SDA*), a tree search algorithm that exploits time-varying centralization during periods of available communication conditioned on the environment or joint actions. By interleaving offline planning with online search and using relaxed heuristics, RS-SDA* achieves high solution quality with reduced computational overhead. SDec-POMDP benchmark experiments show that Approximate RS-SDA* often finds near exact optimal solutions in less than 1% of the time required by exact algorithms. We show scalability in six labyrinth environments both deterministic and stochastic state transitions, and demonstrate real-world feasibility with a multi-drone search-and-rescue simulation in the DARPA Subterranean Challenge cave system.

We study multi-task reinforcement learning (RL), a setting in which an agent learns a single, universal policy capable of generalising to arbitrary, possibly unseen tasks. We consider tasks specified as linear temporal logic (LTL) formulae, which are commonly used in formal methods to specify properties of systems, and have recently been successfully adopted in RL. In this setting, we present a novel task embedding technique leveraging a new generation of semantic LTL-to-automata translations, originally developed for temporal synthesis. The resulting semantically labelled automata contain rich, structured information in each state that allow us to (i) compute the automaton efficiently on-the-fly, (ii) extract expressive task embeddings used to condition the policy, and (iii) naturally support full LTL. Experimental results in a variety of domains demonstrate that our approach achieves state-of-the-art performance and is able to scale to complex specifications where existing methods fail.