Game-Playing Agents

We present the PokeAgent Challenge, a large-scale benchmark for decision-making research built on Pokemon's multi-agent battle system and expansive role-playing game (RPG) environment. Partial observability, game-theoretic reasoning, and long-horizon planning remain open problems for frontier AI, yet few benchmarks stress all three simultaneously under realistic conditions. PokeAgent targets these limitations at scale through two complementary tracks: our Battling Track, which calls for strategic reasoning and generalization under partial observability in competitive Pokemon battles, and our Speedrunning Track, which requires long-horizon planning and sequential decision-making in the Pokemon RPG. Our Battling Track supplies a dataset of 20M+ battle trajectories alongside a suite of heuristic, RL, and LLM-based baselines capable of high-level competitive play. Our Speedrunning Track provides the first standardized evaluation framework for RPG speedrunning, including an open-source multi-agent orchestration system for modular, reproducible comparisons of harness-based LLM approaches. Our NeurIPS 2025 competition validates both the quality of our resources and the research community's interest in Pokemon, with over 100 teams competing across both tracks and winning solutions detailed in our paper. Participant submissions and our baselines reveal considerable gaps between generalist (LLM), specialist (RL), and elite human performance. Analysis against the BenchPress evaluation matrix shows that Pokemon battling is nearly orthogonal to standard LLM benchmarks, measuring capabilities not captured by existing suites and positioning Pokemon as an unsolved benchmark that can drive RL and LLM research forward. We transition to a living benchmark with a live leaderboard for Battling and self-contained evaluation for Speedrunning at https://pokeagentchallenge.com.

Dragonchess, a three-dimensional chess variant introduced by Gary Gygax, presents unique strategic and computational challenges that make it an ideal environment for studying the transfer of artificial intelligence (AI) heuristics across domains. In this work, we introduce Dragonchess as a novel testbed for AI research and provide an open-source, Python-based game engine for community use. Our research investigates evolutionary transfer learning by adapting heuristic evaluation functions directly from Stockfish, a leading chess engine, and subsequently optimizing them using Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Initial trials showed that direct heuristic transfers were inadequate due to Dragonchess's distinct multi-layer structure and movement rules. However, evolutionary optimization significantly improved AI agent performance, resulting in superior gameplay demonstrated through empirical evaluation in a 50-round Swiss-style tournament. This research establishes the effectiveness of evolutionary methods in adapting heuristic knowledge to structurally complex, previously unexplored game domains.

Vision-Language Models (VLMs) excel at describing visual scenes, yet struggle to translate perception into precise, grounded actions. We investigate whether providing VLMs with both the visual frame and the symbolic representation of the scene can improve their performance in interactive environments. We evaluate three state-of-the-art VLMs across Atari games, VizDoom, and AI2-THOR, comparing frame-only, frame with self-extracted symbols, frame with ground-truth symbols, and symbol-only pipelines. Our results indicate that all models benefit when the symbolic information is accurate. However, when VLMs extract symbols themselves, performance becomes dependent on model capability and scene complexity. We further investigate how accurately VLMs can extract symbolic information from visual inputs and how noise in these symbols affects decision-making and gameplay performance. Our findings reveal that symbolic grounding is beneficial in VLMs only when symbol extraction is reliable, and highlight perception quality as a central bottleneck for future VLM-based agents.

Artificial intelligence has advanced significantly through the development of intelligent game-playing systems, providing rigorous testbeds for decision-making, strategic planning, and adaptive learning. However, resource-constrained environments pose critical challenges, as conventional deep learning methods heavily rely on extensive datasets and computational resources. In this paper, we propose a lightweight hybrid framework for the Game of the Amazons, which explores the paradigm of weak-to-strong generalization by integrating the structural reasoning of graph-based learning with the generative capabilities of large language models. Specifically, we leverage a Graph Attention Autoencoder to inform a multi-step Monte Carlo Tree Search, utilize a Stochastic Graph Genetic Algorithm to optimize evaluation signals, and harness GPT-4o-mini to generate synthetic training data. Unlike traditional approaches that rely on expert demonstrations, our framework learns from noisy and imperfect supervision. We demonstrate that the Graph Attention mechanism effectively functions as a structural filter, denoising the LLM's outputs. Experiments on a 10$\times$10 Amazons board show that our hybrid approach not only achieves a 15%--56% improvement in decision accuracy over baselines but also significantly outperforms its teacher model (GPT-4o-mini), achieving a competitive win rate of 45.0% at N=30 nodes and a decisive 66.5% at only N=50 nodes. These results verify the feasibility of evolving specialized, high-performance game AI from general-purpose foundation models under stringent computational constraints.

Multi-turn, multi-agent LLM game evaluations often exhibit substantial run-to-run variance. In long-horizon interactions, small early deviations compound across turns and are amplified by multi-agent coupling. This biases win rate estimates and makes rankings unreliable across repeated tournaments. Prompt choice worsens this further by producing different effective policies. We address both instability and underperformance with MEMO (Memory-augmented MOdel context optimization), a self-play framework that optimizes inference-time context by coupling retention and exploration. Retention maintains a persistent memory bank that stores structured insights from self-play trajectories and injects them as priors during later play. Exploration runs tournament-style prompt evolution with uncertainty-aware selection via TrueSkill, and uses prioritized replay to revisit rare and decisive states. Across five text-based games, MEMO raises mean win rate from 25.1% to 49.5% for GPT-4o-mini and from 20.9% to 44.3% for Qwen-2.5-7B-Instruct, using $2,000$ self-play games per task. Run-to-run variance also drops, giving more stable rankings across prompt variations. These results suggest that multi-agent LLM game performance and robustness have substantial room for improvement through context optimization. MEMO achieves the largest gains in negotiation and imperfect-information games, while RL remains more effective in perfect-information settings. All code is open-source and available here: https://github.com/openverse-ai/MEMO

In high stakes environments, agents relying purely on imitation learning or reinforcement learning often struggle to avoid safety-critical errors during exploration. Existing reinforcement learning approaches for environments such as chess require hundreds of thousands of episodes and substantial computational resources to converge. Imitation learning, on the other hand, is more sample efficient but is brittle under distributional shift and lacks mechanisms for proactive risk avoidance. In this work, we propose Oracle-Guided Soft Shielding (OGSS), a simple yet effective framework for safer decision-making, enabling safe exploration by learning a probabilistic safety model from oracle feedback in an imitation learning setting. Focusing on the domain of chess, we train a model to predict strong moves based on past games, and separately learn a blunder prediction model from Stockfish evaluations to estimate the tactical risk of each move. During inference, the agent first generates a set of candidate moves and then uses the blunder model to determine high-risk options, and uses a utility function combining the predicted move likelihood from the policy model and the blunder probability to select actions that strike a balance between performance and safety. This enables the agent to explore and play competitively while significantly reducing the chance of tactical mistakes. Across hundreds of games against a strong chess engine, we compare our approach with other methods in the literature, such as action pruning, SafeDAgger, and uncertainty-based sampling. Our results demonstrate that OGSS variants maintain a lower blunder rate even as the agent's exploration ratio is increased by several folds, highlighting its ability to support broader exploration without compromising tactical soundness.

The Werewolf Game is a communication game where players' reasoning and discussion skills are essential. In this study, we present a Werewolf AI agent developed for the AIWolfDial 2024 shared task, co-hosted with the 17th INLG. In recent years, large language models like ChatGPT have garnered attention for their exceptional response generation and reasoning capabilities. We thus develop the LLM-based agents for the Werewolf Game. This study aims to enhance the consistency of the agent's utterances by utilizing dialogue summaries generated by LLMs and manually designed personas and utterance examples. By analyzing self-match game logs, we demonstrate that the agent's utterances are contextually consistent and that the character, including tone, is maintained throughout the game.

Chess engines passed human strength years ago, but they still don't play like humans. A grandmaster under clock pressure blunders in ways a club player on a hot streak never would. Conventional engines capture none of this. This paper proposes a personality x psyche decomposition to produce behavioral variability in chess play, drawing on patterns observed in human games. Personality is static -- a preset that pins down the engine's character. Psyche is dynamic -- a bounded scalar ψ_t \in [-100, +100], recomputed from five positional factors after every move. These two components feed into an audio-inspired signal chain (noise gate, compressor/expander, five-band equalizer, saturation limiter) that reshapes move probability distributions on the fly. The chain doesn't care what engine sits behind it: any system that outputs move probabilities will do. It needs no search and carries no state beyond ψ_t. I test the framework across 12,414 games against Maia2-1100, feeding it two probability sources that differ by ~2,800x in training data. Both show the same monotonic gradient in top-move agreement (~20-25 pp spread from stress to overconfidence), which tells us the behavioral variation comes from the signal chain, not from the model underneath. When the psyche runs overconfident, the chain mostly gets out of the way (66% agreement with vanilla Maia2). Under stress, the competitive score falls from 50.8% to 30.1%. The patterns are reminiscent of tilt and overconfidence as described in human play, but I should be upfront: this study includes no human-subject validation.

Recent advances in video generation have spurred the development of world models capable of simulating 3D-consistent environments and interactions with static objects. However, a significant limitation remains in their ability to model dynamic, reactive agents that can intelligently influence and interact with the world. To address this gap, we introduce COMBAT, a real-time, action-controlled world model trained on the complex 1v1 fighting game Tekken 3. Our work demonstrates that diffusion models can successfully simulate a dynamic opponent that reacts to player actions, learning its behavior implicitly. Our approach utilizes a 1.2 billion parameter Diffusion Transformer, conditioned on latent representations from a deep compression autoencoder. We employ state-of-the-art techniques, including causal distillation and diffusion forcing, to achieve real-time inference. Crucially, we observe the emergence of sophisticated agent behavior by training the model solely on single-player inputs, without any explicit supervision for the opponent's policy. Unlike traditional imitation learning methods, which require complete action labels, COMBAT learns effectively from partially observed data to generate responsive behaviors for a controllable Player 1. We present an extensive study and introduce novel evaluation methods to benchmark this emergent agent behavior, establishing a strong foundation for training interactive agents within diffusion-based world models.

Generalized Rapid Action Value Estimation (GRAVE) has been shown to be a strong variant within the Monte-Carlo Tree Search (MCTS) family of algorithms for General Game Playing (GGP). However, its reliance on storing additional win/visit statistics at each node makes its use impractical in memory-constrained environments, thereby limiting its applicability in practice. In this paper, we introduce the GRAVE2, GRAVER and GRAVER2 algorithms, which extend GRAVE through two-level search, node recycling, and a combination of both techniques, respectively. We show that these enhancements enable a drastic reduction in the number of stored nodes while matching the playing strength of GRAVE.

Reinforcement learning (RL) agents achieve remarkable performance but remain far less learning-efficient than humans. While RL agents require extensive self-play games to extract useful signals, humans often need only a few games, improving rapidly by repeatedly revisiting states where mistakes occurred. This idea, known as search control, aims to restart from valuable states rather than always from the initial state. In AlphaZero, prior work Go-Exploit applies this idea by sampling past states from self-play or search trees, but it treats all states equally, regardless of their learning potential. We propose Regret-Guided Search Control (RGSC), which extends AlphaZero with a regret network that learns to identify high-regret states, where the agent's evaluation diverges most from the actual outcome. These states are collected from both self-play trajectories and MCTS nodes, stored in a prioritized regret buffer, and reused as new starting positions. Across 9x9 Go, 10x10 Othello, and 11x11 Hex, RGSC outperforms AlphaZero and Go-Exploit by an average of 77 and 89 Elo, respectively. When training on a well-trained 9x9 Go model, RGSC further improves the win rate against KataGo from 69.3% to 78.2%, while both baselines show no improvement. These results demonstrate that RGSC provides an effective mechanism for search control, improving both efficiency and robustness of AlphaZero training. Our code is available at https://rlg.iis.sinica.edu.tw/papers/rgsc.

Large scale control of nonplayer agents is central to modern games, while production systems still struggle to balance several competing goals: locally smooth, natural behavior, and globally coordinated variety across space and time. Prior approaches rely on handcrafted rules or purely stochastic triggers, which either converge to mechanical synchrony or devolve into uncorrelated noise that is hard to tune. Continuous noise signals such as Perlin noise are well suited to this gap because they provide spatially and temporally coherent randomness, and they are already widely used for terrain, biomes, and other procedural assets. We adapt these signals for the first time to large scale AI control and present a general framework that treats continuous noise fields as an AI coordinator. The framework combines three layers of control: behavior parameterization for movement at the agent level, action time scheduling for when behaviors start and stop, and spawn or event type and feature generation for what appears and where. We instantiate the framework reproducibly and evaluate Perlin noise as a representative coordinator across multiple maps, scales, and seeds against random, filtered, deterministic, neighborhood constrained, and physics inspired baselines. Experiments show that coordinated noise fields provide stable activation statistics without lockstep, strong spatial coverage and regional balance, better diversity with controllable polarization, and competitive runtime. We hope this work motivates a broader exploration of coordinated noise in game AI as a practical path to combine efficiency, controllability, and quality.

Rigorously evaluating machine intelligence against the broad spectrum of human general intelligence has become increasingly important and challenging in this era of rapid technological advance. Conventional AI benchmarks typically assess only narrow capabilities in a limited range of human activity. Most are also static, quickly saturating as developers explicitly or implicitly optimize for them. We propose that a more promising way to evaluate human-like general intelligence in AI systems is through a particularly strong form of general game playing: studying how and how well they play and learn to play all conceivable human games, in comparison to human players with the same level of experience, time, or other resources. We define a "human game" to be a game designed by humans for humans, and argue for the evaluative suitability of this space of all such games people can imagine and enjoy -- the "Multiverse of Human Games". Taking a first step towards this vision, we introduce the AI GameStore, a scalable and open-ended platform that uses LLMs with humans-in-the-loop to synthesize new representative human games, by automatically sourcing and adapting standardized and containerized variants of game environments from popular human digital gaming platforms. As a proof of concept, we generated 100 such games based on the top charts of Apple App Store and Steam, and evaluated seven frontier vision-language models (VLMs) on short episodes of play. The best models achieved less than 10% of the human average score on the majority of the games, and especially struggled with games that challenge world-model learning, memory and planning. We conclude with a set of next steps for building out the AI GameStore as a practical way to measure and drive progress toward human-like general intelligence in machines.

The research community lacks a middle ground between StarCraft II full game and its mini-games. The full-game's sprawling state-action space renders reward signals sparse and noisy, but in mini-games simple agents saturate performance. This complexity gap hinders steady curriculum design and prevents researchers from experimenting with modern Reinforcement Learning algorithms in RTS environments under realistic compute budgets. To fill this gap, we present the Two-Bridge Map Suite, the first entry in an open-source benchmark series we purposely engineered as an intermediate benchmark to sit between these extremes. By disabling economy mechanics such as resource collection, base building, and fog-of-war, the environment isolates two core tactical skills: long-range navigation and micro-combat. Preliminary experiments show that agents learn coherent maneuvering and engagement behaviors without imposing full-game computational costs. Two-Bridge is released as a lightweight, Gym-compatible wrapper on top of PySC2, with maps, wrappers, and reference scripts fully open-sourced to encourage broad adoption as a standard benchmark.

Large Language Models (LLMs) have recently shown strong reasoning capabilities, motivating their use in complex decision-making environments. StarCraft II (SC2), with its massive state-action space and partial observability, is a challenging testbed. However, existing LLM-based SC2 agents primarily focus on improving the policy itself, leaving the integration of a learnable, action-conditioned dynamics model into the decision loop largely unexplored. In this work, we propose StarWM and StarWM-Agent, and conduct the first systematic study of the learnability and decision utility of player-view, action-conditioned textual world models for SC2. StarWM predicts short-horizon future observations under partial observability. StarWM-Agent integrates StarWM into a lightweight Generate-Simulate-Refine loop for foresight-driven policy refinement. Extensive experiments show that StarWM substantially outperforms zero-shot baselines across multiple dimensions, while StarWM-Agent achieves consistent win-rate gains of 30%, 15%, and 30% against the SC2 built-in AI at Hard (LV5), Harder (LV6), and VeryHard (LV7), respectively. Additional analyses show that our method is complementary to existing history-summarization SC2-agent approaches, with their combination achieving the strongest online performance in our evaluation.

Real-world decision-making generally involves hidden information, that is, information that is unknown to one agent but possessed by another. Unfortunately, the presence of large amounts of hidden information renders established reinforcement learning and search approaches ineffective. Even with multimillion-dollar industrial research efforts, top-human-level play at Stratego---a board wargame with hidden information on a massive scale---has remained beyond the reach of artificial intelligence (AI). Here we introduce Ataraxos, an AI for Stratego based on general techniques that we developed for both self-play reinforcement learning and test-time search under hidden information. Ataraxos defeated the most decorated human Stratego player of all time by a large margin---achieving, to our knowledge, the first superhuman result in the game's history---while consuming orders of magnitude less compute and data than previous efforts. Using the same techniques, we built a superhuman AI for Barrage Stratego and state-of-the-art AIs for Hanabi and dou dizhu, all with low cost and high sample efficiency. The success of this approach across adversarial, cooperative and team games establishes a design pattern for reinforcement learning and search that is effective under large amounts of hidden information, a longstanding desideratum of the field of strategic decision-making.

While reinforcement learning agents can achieve superhuman performance in many complex tasks, they typically do not become more computationally efficient as they improve. In contrast, humans gradually require less cognitive effort as they become more proficient at a task. If agents could reason about their compute as they learn, could they similarly reduce their computation footprint? If they could, we could have more energy efficient agents or free up compute cycles for other processes like planning. In this paper, we experiment with showing agents the cost of their computation and giving them the ability to control when they use compute. We conduct our experiments on the Arcade Learning Environment, and our results demonstrate that with the same training compute budget, agents that reason about their compute perform better on 75% of games. Furthermore, these agents use three times less compute on average. We analyze individual games and show where agents gain these efficiencies. An interactive version of this paper is available at https://adrianorenstein.com/paper/toward-compute.

Recent mechanistic work has uncovered learned algorithms within neural networks, from modular arithmetic to search and planning in game-playing agents. But does algorithmic structure guarantee algorithmic behavior? We investigate this in Leela Chess Zero, the strongest neural chess engine, where prior work identified learned look-ahead. By extending the logit lens to its move-selecting policy network, we discover that correct puzzle solutions--including immediate checkmates--often appear in intermediate layers but are systematically overridden in the final output, a phenomenon we term "forgotten puzzles". Replicating prior analyses on these positions, we find that look-ahead operates normally--future moves of the correct continuation are represented, causally important, and linearly decodable--ruling out a failure of the algorithm. Instead, late layers increasingly shift toward safe play over aggression. To test whether this shift drives the override, we steer the model against these preferences and recover 71.7% of forgotten puzzles, providing causal evidence that safety priors override algorithmically computed solutions. These findings show that algorithmic structure does not guarantee algorithmic behavior: a model can internally solve a problem and still output the wrong answer.

We introduce PuzzleJAX, a GPU-accelerated puzzle game engine and description language designed to support rapid benchmarking of tree search, reinforcement learning, and LLM reasoning abilities. Unlike existing GPU-accelerated learning environments that provide hard-coded implementations of fixed sets of games, PuzzleJAX allows dynamic compilation of any game expressible in its domain-specific language (DSL). This DSL follows PuzzleScript, which is a popular and accessible online game engine for designing puzzle games. In this paper, we validate in PuzzleJAX several hundred of the thousands of games designed in PuzzleScript by both professional designers and casual creators since its release in 2013, thereby demonstrating PuzzleJAX's coverage of an expansive, expressive, and human-relevant space of tasks. By analyzing the performance of search, learning, and language models on these games, we show that PuzzleJAX can naturally express tasks that are both simple and intuitive to understand, yet often deeply challenging to master, requiring a combination of control, planning, and high-level insight.

Imitation sometimes achieves success in multi-agent situations even though it is very simple. In repeated games, success of imitation has been characterized by unbeatability against other agents. Previous studies specified conditions under which imitation is unbeatable against co-players in repeated symmetric games, and clarified that the existence of unbeatable imitation is strongly related to the existence of payoff-controlling strategies, called zero-determinant strategies. However, these results were on the situations in which one repeated game is played by one group of players. It was pointed out that imitation of other players in the same group and imitation of other players in the same role in other groups generally result in different outcomes. Here, we investigate the existence condition of unbeatable imitation in the latter situations, described by two concurrent repeated games played by two different groups of players. We find that the existence condition of unbeatable imitation is stronger than the existence condition of unbeatable zero-determinant strategies, whereas both are very limited. Furthermore, we also find that adding exploration can make an imitation strategy unbeatable even in games where imitation without exploration can be beaten. Our findings suggest a strong relation between unbeatable imitation and payoff control even in the two concurrent repeated game situations.

Autonomous agents navigating human society must master both production activities and social interactions, yet existing benchmarks rarely evaluate these skills simultaneously. To bridge this gap, we introduce StarDojo, a novel benchmark based on Stardew Valley, designed to assess AI agents in open-ended production-living simulations. In StarDojo, agents are tasked to perform essential livelihood activities such as farming and crafting, while simultaneously engaging in social interactions to establish relationships within a vibrant community. StarDojo features 1,000 meticulously curated tasks across five key domains: farming, crafting, exploration, combat, and social interactions. Additionally, we provide a compact subset of 100 representative tasks for efficient model evaluation. The benchmark offers a unified, user-friendly interface that eliminates the need for keyboard and mouse control, supports all major operating systems, and enables the parallel execution of multiple environment instances, making it particularly well-suited for evaluating the most capable foundation agents, powered by multimodal large language models (MLLMs). Extensive evaluations of state-of-the-art MLLMs agents demonstrate substantial limitations, with the best-performing model, GPT-4.1, achieving only a 12.7% success rate, primarily due to challenges in visual understanding, multimodal reasoning and low-level manipulation. As a user-friendly environment and benchmark, StarDojo aims to facilitate further research towards robust, open-ended agents in complex production-living environments.

Proficient game agents with diverse play styles enrich the gaming experience and enhance the replay value of games. However, recent advancements in game AI based on reinforcement learning have predominantly focused on improving proficiency, whereas methods based on evolution algorithms generate agents with diverse play styles but exhibit subpar performance compared to RL methods. To address this gap, this paper proposes Mixed Proximal Policy Optimization (MPPO), a method designed to improve the proficiency of existing suboptimal agents while retaining their distinct styles. MPPO unifies loss objectives for both online and offline samples and introduces an implicit constraint to approximate demonstrator policies by adjusting the empirical distribution of samples. Empirical results across environments of varying scales demonstrate that MPPO achieves proficiency levels comparable to, or even superior to, pure online algorithms while preserving demonstrators' play styles. This work presents an effective approach for generating highly proficient and diverse game agents, ultimately contributing to more engaging gameplay experiences.

Quantum Tiq-Taq-Toe is a well-known benchmark and playground for both quantum computing and machine learning. Despite its popularity, no reinforcement learning (RL) methods have been applied to Quantum Tiq-Taq-Toe. Although there has been some research on Quantum Chess this game is significantly more complex in terms of computation and analysis. Therefore, we study the combination of quantum computing and reinforcement learning in Quantum Tiq-Taq-Toe, which may serve as an accessible testbed for the integration of both fields. Quantum games are challenging to represent classically due to their inherent partial observability and the potential for exponential state complexity. In Quantum Tiq-Taq-Toe, states are observed through Measurement (a 3x3 matrix of state probabilities) and Move History (a 9x9 matrix of entanglement relations), making strategy complex as each move can collapse the quantum state.

Benchmarks are crucial in the development of machine learning algorithms, with available environments significantly influencing reinforcement learning (RL) research. Traditionally, RL environments run on the CPU, which limits their scalability with typical academic compute. However, recent advancements in JAX have enabled the wider use of hardware acceleration, enabling massively parallel RL training pipelines and environments. While this has been successfully applied to single-agent RL, it has not yet been widely adopted for multi-agent scenarios. In this paper, we present JaxMARL, the first open-source, Python-based library that combines GPU-enabled efficiency with support for a large number of commonly used MARL environments and popular baseline algorithms. Our experiments show that, in terms of wall clock time, our JAX-based training pipeline is around 14 times faster than existing approaches, and up to 12500x when multiple training runs are vectorized. This enables efficient and thorough evaluations, potentially alleviating the evaluation crisis in the field. We also introduce and benchmark SMAX, a JAX-based approximate reimplementation of the popular StarCraft Multi-Agent Challenge, which removes the need to run the StarCraft II game engine. This not only enables GPU acceleration, but also provides a more flexible MARL environment, unlocking the potential for self-play, meta-learning, and other future applications in MARL. The code is available at https://github.com/flairox/jaxmarl.

Generative flow networks (GFlowNets) are sequential sampling models trained to match a given distribution. GFlowNets have been successfully applied to various structured object generation tasks, sampling a diverse set of high-reward objects quickly. We propose expected flow networks (EFlowNets), which extend GFlowNets to stochastic environments. We show that EFlowNets outperform other GFlowNet formulations in stochastic tasks such as protein design. We then extend the concept of EFlowNets to adversarial environments, proposing adversarial flow networks (AFlowNets) for two-player zero-sum games. We show that AFlowNets learn to find above 80% of optimal moves in Connect-4 via self-play and outperform AlphaZero in tournaments.

The 20 Questions (Q20) game is a well known game which encourages deductive reasoning and creativity. In the game, the answerer first thinks of an object such as a famous person or a kind of animal. Then the questioner tries to guess the object by asking 20 questions. In a Q20 game system, the user is considered as the answerer while the system itself acts as the questioner which requires a good strategy of question selection to figure out the correct object and win the game. However, the optimal policy of question selection is hard to be derived due to the complexity and volatility of the game environment. In this paper, we propose a novel policy-based Reinforcement Learning (RL) method, which enables the questioner agent to learn the optimal policy of question selection through continuous interactions with users. To facilitate training, we also propose to use a reward network to estimate the more informative reward. Compared to previous methods, our RL method is robust to noisy answers and does not rely on the Knowledge Base of objects. Experimental results show that our RL method clearly outperforms an entropy-based engineering system and has competitive performance in a noisy-free simulation environment.