game ai developers

Abstract: Solving imperfect-information games at scale rests on counterfactual regret minimization with the regret matching$^{+}$ (RM$^{+}$) update, and a seemingly minor implementation choice accounts for much of its practical speed: alternating the two players' updates, so that the second player responds to the strategy the first player just produced. A widely repeated intuition explains why: seeing the opponent's fresh strategy should act like optimism. We treat this intuition as a testable hypothesis about deterministic two-player zero-sum self-play dynamics and find that its canonical additive form fails an information-matched test. On an analytic witness family we compare alternating RM$^{+}$ against a one-sided predictive RM$^{+}$ that receives the same information channel each round and differs only in where the fresh signal enters, a persistent regret state versus a transient prediction term. Our main result is a dichotomy theorem that reduces trajectory separation to the recurrence of exact-vertex play by the predictive dynamics: under an interior-band assumption on the alternating iterate, either vertex times recur and the profile distance is at least, infinitely often, an explicit constant that our measurements nearly attain, or they terminate. Across the tested self-play regimes, any correct explanation of the alternation speedup must reproduce a staleness cost present at the smallest tested stale fraction, an advantage that persists across seven orders of payoff conditioning, regime crossings of the optimism family, and a measured per-perturbation cost of weakening the injection path that is far steeper for staleness than for partial persistence.

Abstract: Predicting cost in a multi-agent world model moves most constrained policy updates into imagination: the stronger model-free baseline needs $10$ to $20$ times CA-MAMBA's interactions on the median SMAC configuration, and CA-MAMBA earns $2.8$ to $12$ times the baselines' MAMuJoCo reward at equal budget. In controlled tests on one task per benchmark, pricing the model's cost advantages at parity more than halves a cost-blind learner's cost and meets a tight continuous-control limit the learned multiplier misses, suggesting the model's cost estimates are not the bottleneck.

Abstract: We introduce Faynt, a family of 10M- and 75M-parameter Transformer policies for Super Smash Bros. Melee, each controlling all 26 characters with a single checkpoint. After reinforcement learning (RL), the 10M wins 240 of 244 same-character games (98.4%) against fourteen specialist and multi-character releases on their supported rosters, with a winning record against every release. These opponents retain 21- or 24-frame action delays; Faynt uses no added delay, and we have not isolated the effect of this difference. In a separate evaluation against a privately supplied zero-delay Slippi-AI model, the 10M wins all 68 games across two conditioning settings. We study architecture, optimization, scaling, and hyperparameter transfer to guide pretraining on approximately 840,000 human replays. Post-training combines rank- and outcome-based curricula, 75M-to-10M distillation, and RL restricted to Fox mirror matches. On the initial 152-game benchmark, the supervised 10M wins 69.7% of games, compared with 45.4% for the pretrained 75M, despite higher overall held-out controller-prediction loss. The weighted validation loss used for supervised checkpoint selection agrees with the win-rate ordering of all four pretrained and supervised policies. After supervised post-training, both models take less damage per minute, build larger early leads, and win more often after losing the first life. Optimized inference on recorded game states averages 5.2 ms per decision for the 10M and 8.7 ms for the 75M on an NVIDIA T4, excluding emulator execution and communication. We open-source the weights, both benchmark suites, and a platform for automated model tournaments.

Abstract: Reinforcement learning (RL) is a powerful paradigm for training agents, yet its success rests on domain expertise of human engineers who design informative reward signals for every new task. Unsupervised RL aims to reduce this engineering with intrinsic motivation (IM): reward signals that emerge from the agent environment interaction itself. Existing IM objectives, however, involve the selection of information variables, which re-introduces domain expertise the field has sought to eliminate. We introduce Forward CIP (F-CIP), an RL-native formulation of the Controllable Information Production (CIP) objective, which is defined by the system's dynamics alone and requires no such selection. We prove that F-CIP is compatible with RL and demonstrate its effectiveness with existing algorithms. Training agents with F-CIP results in unsupervised discovery of primitive behaviors such as balancing and maintaining controllability, which are essential for more complex robot behaviors. Paired with a simple forward-velocity reward, our method produces coordinated gaits such as hopping and running which otherwise require reward engineering to learn.

Abstract: Judgment models such as \jev{} return, in a single call and without reasoning text, a probability for each described option. This makes them attractive as an agent's action-selection layer, but it is unclear which decisions they can be trusted with. We test \jev{} on reflection tests, one-shot matrix games, the text game ALFWorld and robot control, and find a sharp boundary. \jev{} succeeds when the right option can be judged from what the input describes, which we call \emph{evaluation}. Specifically, it solves 99\% of the counterintuitive Cognitive Reflection Test questions. However, it fails when the right option depends on \emph{simulation} (i.e., predicting something not in the input), such as the opponent's action or the subgoal that must come first. In games, \jev{} plays suboptimally as if its rational opponent acted at random, because the opponent's action is not given. In ALFWorld, \jev{} favors commands that mention an object or place named in the task description. For example, given the task ``put a clean knife in the drawer'', \jev{} carries an unwashed knife straight to the drawer instead of first washing it at the sink. Surprisingly, many of these failures are not due to a lack of knowledge. Asked separately what the opponent will do, \jev{} usually answers correctly, and it responds well given the opponent's action. It fails when one call must both perform the simulation and evaluate based on it. This suggests letting code make the prediction or simulation. When code supplies it, such as a lookahead in ALFWorld and physics simulation in robot control, \jev{} becomes an expert controller through its general evaluation ability.

Abstract: Despite many studies about causes and remedies of plasticity loss in Reinforcement Learning (RL) under continual task changes, no RL method has yet consistently achieved a good balance between adaptation and forgetting. Here we turn to an alternative optimization paradigm, neuroevolution (NE): algorithms that search directly in weight space through mutation and selection over a population of neural networks. Across a wide array of environments and environmental changes, with policies ranging from a few hundred parameters to million-parameter networks, we compare evolution strategies (ES) and genetic algorithms (GAs) against state-of-the-art continual RL variants and population-based RL. ES most consistently achieves a good stability-plasticity trade-off, while the GA is the most plastic method but forgets more than ES. To explain this, we study the return landscape around each method's solutions. ES finds the widest neighborhoods, i.e.\ regions of weight space in which perturbed policies still solve the task, and the size of the overlap between the neighborhoods of consecutive tasks correlates with a method's stability-plasticity trade-off. Rewarding behavioral diversity in a GA through novelty search makes the population even more plastic, at the cost of forgetting. Finally, symptoms of plasticity loss commonly reported in RL do not transfer to NE. Overall, these results establish NE as a competitive alternative to RL under continual task changes, and suggest that training under perturbations in weight space may be a useful mechanism for continual learning more broadly.

Abstract: Long-term dependencies remain a major challenge for sequential decision-making in the field of AI: RNNs suffer from vanishing gradients and the limited expressivity of vector-based hidden states, whilst Transformer-based models are limited by the quadratic scaling of attention. Recent work has proposed tackling this problem with the Test-Time Training (TTT) framework, which stores episodic memories in the parameters of a neural network through gradient descent at both train and test-time. This approach has seen success in the domain of Natural Language Processing, however, to the best of our knowledge it has not yet been applied to the domain of Reinforcement Learning (RL), nor has there been a study analysing how this memory practically functions. In this paper, we study the potential of the TTT framework for offline RL by augmenting a Decision Transformer with TTT layers, dubbed the Decision Titan. We analyse performance and properties of the model in the X-Maze environment, an extension of T-Maze designed to test sequential memory, and investigate how the memory mechanism learns by visualising gate values over time. Our key findings are that Decision Titan can learn long-term dependencies with ranges 20x longer than the context window, generalises to lengths 1.7x the training data, but crucially temporal generalisation depends on the time embeddings used, and the ability to learn long-term dependencies depends on how the relevant information is encoded.

Abstract: Partially observable environments pose a fundamental challenge in deep reinforcement learning, requiring agents to compress temporal information from observations and maintain a memory to make effective decisions. While there exist many approaches ranging from gated recurrence to attention mechanisms and model-based RL, the search for effective representational techniques that can capture long-term dependencies remains an active area of research. In this work we revisit Unitary recurrent networks (uRNNs) [Arjovsky et al., 2016, Jing et al., 2017], that demonstrated superior gradient flow and associative recall, expressing the recurrence and the hidden state in a complex vector space. Their norm preserving unitary dynamics enable information propagation through long sequences. To this end, we propose three different versions of uRNNs as drop-in replacements for recurrent PPO architectures, and demonstrate that the simple recurrence and the added degree of freedom from the phase of the complex representations enable significant gains over baselines on several memory-improvable tasks, including continuous control. We further explore how to preserve the phase information of the complex hidden state for a phase-aware policy by drawing a parallel to how quantum states are measured. With our methods reaching up to 2-3 $\times$ the reward in environments like rocksample and Craftax compared to the baselines, this work points towards an exciting new direction of representations for RL and the problem of partial observability. Code is available at: https://github.com/Sathya98/qurl

Abstract: If we run a heavy-duty computation on prior data, can we avoid repeated computation for similar future inputs? Inspired by this question, we introduce a new computational model for graph problems called algorithms with isomorphic priors. Solving a graph problem $Π$ in this model involves two phases: (i) The preprocessing phase quickly analyzes prior graphs $G_1, ..., G_k$ along with the (previously computed) exact optimal values OPT$(G_i)$. (ii) Subsequently, given a new graph H, a fast query phase must either (a) output the exact solution OPT(H), or (b) correctly report that H is not isomorphic to any $G_i$. Can we avoid computing OPT(H) from scratch when H is isomorphic to some $G_i$? We show that this is the case for a number of problems; for many others, we establish conditional lower bounds. $\textbf{(1)}$ Some NP-hard problems, including Constrained Shortest Path and $\ell_p$-Shortest Path and Constrained Spanning Tree, admit polynomial preprocessing and query times in our model. In contrast, almost all of Karp's 21 NP-complete problems and $(2-\varepsilon)$-approximate $k$-Center, for every fixed $\varepsilon>0$, admit no such algorithms unless Graph Isomorphism (GI) is in P, even with O(1) priors. $\textbf{(2)}$ In contrast to conditional $n^{3-o(1)}$ fine-grained lower bounds, our framework achieves an $O(n^ω)$ query time for Negative Triangle and a near-linear query time for Replacement Path. It also achieves near-linear query time for Maximum Flow. $\textbf{(3)}$ While it remains a major open problem whether infinite-duration games (Paritiy Game, Mean Payoff Game, Energy Game, and Stochastic Game) admit polynomial-time algorithms, they can be easily solved in near-linear time within our model. Our proofs rely on a simple combination of existing tools and are accessible to readers without specialized background.

Abstract: Reinforcement learning for long-horizon agents typically relies on sparse outcome-based rewards. This leads to a severe cold-start problem, as early-stage policies often fail to solve sampled tasks, leaving little useful reward signal for learning. To mitigate this problem, we use on-policy distillation (OPD) to provide token-level guidance on the student's own rollouts. We find that the benefit of this guidance depends on the performance gap between the teacher and the student. When the teacher substantially outperforms the student, distillation helps guide the student through the early training stage where outcome rewards provide little learning signal. As the gap narrows and eventually reverses, however, continued distillation becomes less beneficial and may hinder further improvement. Motivated by this observation, we propose Gap-Adaptive Teacher Scheduling (GATS), which augments the student's RL objective with an OPD term whose weight adapts to the teacher-student performance gap. Specifically, GATS gradually reduces teacher guidance as the student approaches the teacher's reference performance and withdraws it once that reference is reached. This enables GATS to leverage task-trained teachers smaller than the student, since teacher guidance is primarily needed during early training. Across ALFWorld, WebShop, and ScienceWorld with three Qwen2.5 teacher-student configurations, GATS achieves the highest average success rate among the compared methods in all three configurations, improving over reward-only GRPO by 4.37%-11.87% under matched student rollout budgets. Code is available at https://github.com/Ricardo-H/guide-then-let-go.

Abstract: Looped reasoning models repeatedly apply a shared set of parameters, enabling more computation without increasing the model size. These models also support input-dependent computation by dynamically deciding when to stop looping. Motivated by the recent success of looped transformers in language modeling and reasoning, we investigate whether dynamic looping can similarly benefit sequential decision-making. We provide a complexity-theoretic motivation for this approach by showing that there exist Markov decision processes in which a state-adaptive policy achieves the optimal return with asymptotically less expected computation than any optimal fixed-runtime policy. To learn compute-adaptive policies in practice, we introduce Looped Actor, a transformer-based policy that repeatedly refines a latent representation toward a fixed point using a shared computational block. This allows the model to allocate computation adaptively by varying the number of loops based on the current state. We evaluate Looped Actor on 22 tasks across six environments, ranging from combinatorial puzzles to robotic manipulation and spanning online and offline reinforcement learning (RL) with discrete and continuous actions. Looped Actor matches or exceeds the performance of an untied baseline with 16$\times$ more parameters, with the largest gains in environments where action selection requires substantial multistep planning. For the Boxoban environment, we find that the computation allocation is structured: the number of loops increases with the number of remaining pushes and future optimal pushes become increasingly predictable from the latent state over successive loops. Together, these results highlight actor looping as a simple and efficient way to equip RL agents with adaptive computation and improve their planning capabilities. Code is available at https://github.com/camail-official/LoopedActor

Abstract: Latent world models are trained to predict what happens next, so nothing in their objective separates what an action caused from what merely co-occurred with it. Object-masking models such as C-JEPA intervene on what the predictor can see; we intervene on what physically happens. From one saved simulator state we run the dynamics under an action $a$ and under a reference action $a_{\varnothing}$, and train the model to predict the difference $Δz=z^{a}-z^{a_{\varnothing}}$ between the two latent futures. The resulting objective, Do-JEPA, has an effect loss, a support loss (where the action enters), a propagation loss (where its effect travels) and invariance losses (what must not change). In a synthetic system with object-aligned variables, support supervision finds the directly intervened object in 99.95% of test cases, where a sparse action mask sends the action to a nuisance slot in every case, and response-onset supervision recovers the ring-shaped propagation graph (edge AUROC 0.975 vs. 0.624). From pixels, the effect loss beats a control trained on exactly the same data: it lowers latent effect error by 28.4% on an end-to-end LeWM model and physical effect error by 13.5% when trained and tested on natural action sequences, and on three independently generated CausalWorld benchmarks it lowers responsive effect error by about 20% under physics shifts and the latent context sensitivity of predicted effects by 66%. Trained from scratch it costs factual accuracy; fine-tuning an existing model with it removes this cost. Together, these results show that intervening on the world, rather than on what the model sees, helps latent world models predict what their actions cause.

Abstract: We study an unknown-transition finite-horizon Markov decision process (MDP) with a finite collection of known reward functions $\{r^1, r^2, \ldots, r^M\}$. The goal is to output an $ε$-optimal policy for every reward using online episodic interaction only. Performance is measured by the policy error $V_{0}^{*, m} - V_{0}^{\widehatπ^{m}, m}$ where $m\in [M]$ represents the reward function and $V_{0}^{*, m}=\mathbb{E}_{s_1\sim μ}[V_{1}^{*, m}(s_1)]$. Under this setting, we design a provably efficient algorithm to establish a minimax sample complexity bound of $$ O\left(\frac{SAH^3}{ε^2}\log M \mathrm{polylog}\left(\frac{SAH\log M}{\min\left\{ε, 1\right\}δ}\right)\right)$$ episodes, with no additional burn-in cost. This matches the information-theoretic lower bound up to a factor of $ \mathrm{polylog}(SAH\log M/(\min\left\{ε, 1\right\}δ))$. Our method combines three technical ingredients. First, we adapt MVP to reward-switching learning to construct optimistic value estimates. Second, we use fresh replay samples to conservatively evaluate the candidate policies. Third, gap-based multiplicative weights updates adjust the reward-sampling distribution using the differences between these estimates, converting weighted learning progress into simultaneous guarantees for all rewards.

Abstract: Latent world models rely on representation geometry for planning, yet regularizing the latent marginal alone does not determine the state-to-state relationships used for action selection. We show that this can cause planning-relevant novelty structure to be weakened as representations are transformed into the final latent used by the planner. We introduce Aligned Transport of Latent Structure (ATLAS), a training objective that explicitly preserves relational geometry while calibrating the global latent distribution. ATLAS transfers normalized pairwise structure from an informative encoder representation to the planning latent and uses Wasserstein embedding matching (WEMReg) to calibrate its marginal through one-dimensional Wasserstein-2 transport. Our analysis shows that relational preservation and marginal calibration impose non-redundant constraints, and connects finite-candidate planning stability to relational distortion, latent-scale mismatch, and prediction error. Instantiated in LeWM, ATLAS improves mean goal-reaching success across PushT, TwoRoom, and OGBench-Cube on both lower- and higher-novelty evaluation subsets, with the largest gain on higher-novelty TwoRoom episodes. Representation and rollout diagnostics further show stronger novelty-related structure in the planning latent, improved marginal calibration, and lower multi-step prediction error. Together, these results highlight preservation of planning-relevant latent geometry as an important ingredient for reliable world-model planning. Code is available at https://anonymous.4open.science/r/atlas-world-model-72C4/.

Abstract: As a long-horizon agent improves through experience, previously observed weaknesses may recede while new limitations emerge, continually changing what it still needs to learn. Yet the learning process often remains tied to a static view of these needs: fixed behavioral criteria and training priorities can become misaligned with evolving agent capabilities, while sparse task-level feedback makes such misalignment more difficult to detect. Even when capability gaps are identified, rollouts from the current policy may repeatedly reproduce the same failures rather than explore better alternatives. To address this, we introduce Adaptive Rubric-Skill Co-Evolution (ARISE), a reinforcement learning framework that uses rollout evidence to continually adapt evaluation criteria, exploration guidance, and training priorities. Rubrics evolve to reward partial behavioral progress, while their paired skills are refined and selectively activated to guide exploration toward unresolved weaknesses. Alongside this co-evolution, capability-based adaptive sampling prioritizes tasks that target behaviors needing further improvement. Experiments on two challenging long-horizon agent benchmarks, SkillsBench and Terminal-Bench, demonstrate that ARISE successfully enhances both overall task performance and training efficiency. The project page is at https://foundation-model-research.github.io/ARISE .

Abstract: Principled exploration in reinforcement learning requires an agent to quantify its epistemic uncertainty and act to resolve it. Uncertainty over the value function provides a natural signal for exploration, yet existing deep approximations remain brittle and perform inconsistently. The central challenge is therefore to scale these ideas robustly. We conduct a systematic empirical study of how epistemic uncertainty is represented, propagated, and optimized in deep epistemic value functions, and uncover distinct failure modes along each of these axes. These findings motivate DEVOTE, a model-free reinforcement learning algorithm that controls how uncertainty generalizes beyond observed data, stabilizes its temporal propagation, and preserves adaptation to the resulting non-stationary exploration objective. Across reward-free exploration and challenging continuous-control tasks, DEVOTE reaches novel states more effectively and achieves higher task return than strong model-free and model-based exploration baselines. These results provide evidence that deep epistemic value functions are a promising path toward scalable, principled exploration.

Abstract: Algorithm Distillation (AD) has demonstrated the remarkable ability of Transformers to perform in-context reinforcement learning without explicit weight updates. However, capturing long-term learning progress necessitates expansive context windows, which incur prohibitive memory costs and limit scalability in complex, long-horizon tasks. To address this bottleneck, we propose Recurrent Algorithm Distillation (RAD). RAD employs a dual-component architecture: a Compression Transformer that distills extended interaction histories into compact latent tokens, and an AD Transformer that auto-regressively generates actions using a hybrid context of these compressed memories and recent transitions. By maintaining a fixed-size latent buffer, RAD decouples the effective history length from computational complexity, functionally providing the model with a long-horizon memory. Empirical evaluations across diverse environments demonstrate that RAD matches the asymptotic performance of standard AD with significantly reduced context window sizes, offering a scalable solution for efficient in-context decision-making.

Abstract: On-policy distillation (OPD) trains student agents through teacher supervision on their own interactions with an environment. However, in asynchronous multi-turn training, arrival-order batching can allow a few early or long rollouts to dominate learner updates while other valid rollouts become stale before being used, wasting already-generated experience. To address this problem, we introduce DivOPD, a simple learner-side batch-selection method that spreads a fixed turn budget across more rollouts and, within each rollout, prioritizes turns with larger cumulative teacher-student disagreement. Turns without usable teacher feedback are excluded. The per-turn loss and optimizer remain fixed; selection only changes which student-visited turns receive training weight. For no-progress rollouts, an optional extension briefly hands control to the teacher before returning it to the student. Across six teacher-student settings on the simulated ALFWorld, ScienceWorld, and WebShop benchmarks, with 1.5B-7B students, DivOPD raises cross-setting mean peak success rate from 77.4 to 84.4 and mean success over the last five evaluations from 71.5 to 78.6. It reaches all reported setting-specific targets with geometric-mean speedups of 1.84x in training tokens and 1.87x in learner GPU time relative to vanilla OPD. Teacher intervention further raises this last-five mean to 82.4 while retaining about 1.7x learner-GPU speedup over vanilla OPD. Code will be released at https://github.com/HanyangWang0418-oss/DivOPD.

Abstract: Long-horizon agent benchmarks typically report how far an agent progresses, but do not identify whether its performance comes from the foundation model, scaffold, responsibility scope, match-control granularity, or horizon. We introduce FromPitch2Board, a deterministic football-management benchmark that studies five configurable factors through controlled comparisons on a single simulator, using paired seeds and a frozen calibration. We evaluate four foundation models and four agent scaffolds. In the Model Track, Coach points Z-scores span 0.19, while Manager points Z-scores span 0.68, with GPT-5.6 showing a sharp rise in passivity under responsibility expansion. Its responsibility ladder rises from 46.1 to 58.1 points with recruitment, then falls to 46.8 under full management, localizing the regression to the final responsibility boundary. Across that boundary, its skipped-decision rate rises from 1.1% to 57.9%. Within the Flash-Pro pair crossed across every scaffold, scaffold choice changes Manager points Z-scores by up to 0.48 relative to the fixed stateless scaffold. The 3Y cohort shows a directional reversal in mean ranking between years one and three, while a selected Claude Code+Pro configuration peaks in year three and remains below that peak, showing that responsibility scope and horizon expose behavior changes that a single headline score conceals.

Abstract: Flow policies can model rich action distributions, but their iterative sampling limits decision speed. Adjoint matching uses the critic's action gradient to improve a flow policy without backpropagating through its sampling trajectory, yet its supervision is defined for instantaneous velocities. We propose QAMM, a method that turns the critic-derived adjoint signal into supervision for MeanFlow's average velocity. The resulting policy learns finite-interval transport directly and generates actions with few network evaluations. We derive the adjoint MeanFlow target, specify its gradient boundaries, and train it with an offline actor-critic. On ten HumanoidMaze tasks, QAMM produces effective two-call policies and achieves competitive performance against strong flow-policy baselines. These results show that adjoint-based Q optimization can be combined with average-velocity learning to obtain expressive offline policies with few-step action generation.

Abstract: Efficient cooperation is challenging due to the usual partial observability of each agent in multi-agent reinforcement learning. Recurrent networks encode local interaction histories, but their hidden representations provide limited insight into the information underlying individual decisions. To address these challenges, we propose a novel interpretable framework, called escaping local views (ELV), which introduces semantically structured latent concepts to render policy decisions transparent. Specifically, each agent extracts low-dimensional semantic concepts from its local observation and action-observation trajectory. These concepts are jointly encoded into a contextual latent variable via a variational autoencoder (VAE), which builds a bridge between local views and global semantics. To explicitly model the decision of each agent, we employ a dual-path attention mechanism in which one module estimates the salience of individual concepts relative to the global context, while the other captures higher-order cooperative patterns with pairwise concept interactions. Furthermore, we incorporate a concept prediction module that derives an intrinsic reward from next-concept prediction errors, which incentivizes agents to explore regions of semantic novelty. Experiments in multiple environments verify that ELV not only achieves competitive performance but also explicitly provides how agents reason about their decisions.

Abstract: A strong LLM strategic agent should reason prospectively over uncertain futures, adapt its strategy to opponents' behavioral tendencies, and continuously recalibrate its decision process from interaction experience. However, incorporating these sources in free-form reasoning could lead to unsupported strategic assumptions, inconsistent opponent estimates, and harmful interference from irrelevant historical interactions. To address these issues, we propose SAGE, a training-free inference-time framework that structures LLM strategic reasoning around three coordinated operations: anchor, adapt, and recalibrate. SAGE first anchors reasoning to an equilibrium policy that provides a strategically valid prior. It then conditions deviations from this anchor on a soft belief over opponent behavioral tendencies, enabling opponent-specific exploitation. Finally, SAGE distills strategically related interactions into counterfactual hypotheses about previously missing considerations, allowing past experience to recalibrate the model's reasoning. We evaluate SAGE on three repeated imperfect-information games: Leduc Hold'em, Liar's Dice, and Goofspiel, against various opponent types in each game. Compared with reasoning-intensive LLM agents, including Suspicion-Agent, ReTA, Agent-Pro, EMO, and Hypothetical Minds, SAGE achieves up to a 127.6% payoff improvement in Liar's Dice while reducing input and output token usage by up to 80% and 90%, respectively. In direct match-up play, it attains non-negative mean payoff against 5/10, 8/10, and 8/10 evaluated opponents in Leduc Hold'em, Liar's Dice, and Goofspiel, respectively, while using relatively fewer tokens. Code is available at https://github.com/chenzhwsysu57/SAGE.

Abstract: Residual policy adaptation provides a lightweight way to modify a strong reference policy, but a shared scale and a fixed subgroup partition can hide heterogeneous degradation and become fragile when the deployment mixture of information states changes. We introduce RSD-Poker, a structure-adaptive and shift-robust certification framework that freezes a bank of residual families and scales, learns a policy-visible partition on an independent structure split, and freezes that partition before calibration labels are joined. Each candidate-group pair receives a weighted simultaneous upper certificate for anchor-relative risk and a lower certificate for weak-response utility. A robust group-to-candidate map is then selected over a predeclared uncertainty set of deployment group proportions. Under independent calibration units drawn from each frozen group's law, a candidate bank and partition fixed before calibration, and invariant within-group conditionals, the selected map satisfies its declared mixture-robust risk budget and utility certificate with probability at least $1-ζ_{risk}-ζ_{util}$. The information contract supports both a teacher-backed transform and a teacher-free observation-only student. The retained deterministic 24-state audit remains an exact replay diagnostic: empirical-zero selects $α=0.08$, raising the weak-response proxy from 4.2082 to 4.2889 with $0/12$ held-out threshold crossings. On stratified held-out states, the learned-partition dual selector raises weak utility from 4.4074 under global dual certification to 4.4936 and lowers held-out violation from 0.0215 to 0.0078; its mixture-robust variant reaches violation 0.0059. Across five observation-only checkpoints, risk-calibrated residuals attain weak utility $4.3659\pm0.0177$ and violation rate $0.0178\pm0.0057$.

Abstract: Coupled rollouts can reduce the noise of counterfactual action comparisons, but two issues prevent standard common-random-number constructions from serving as a general learning primitive in imperfect-information environments. First, an invalid coupling may expose hidden state, synchronize endogenous policy randomness, or misalign chance events after counterfactual histories diverge. Second, in multi-action policy optimization, lower return-contrast variance is not by itself the relevant objective: the optimizer depends on the return covariance matrix after projection through the local policy-gradient geometry. We introduce observation-safe counterfactual coupling (OSCC), a framework that defines an admissible class through marginal preservation, information-state safety, branch-local policy randomness, semantic event alignment, and trace-before-oracle replay. We derive a gradient-aware coupling criterion showing that, for marginal-preserving couplings, policy-gradient noise changes are determined by policy-Jacobian-weighted off-diagonal return covariance. This motivates OSCC-Select, a calibration-only selector that chooses among independent, root-only, continuation-only, and fully coupled rollouts using separate safety and gain certificates. Its gain target combines projected gradient noise with measured physical sampling cost and falls back to independent sampling whenever a simultaneous lower confidence bound does not certify improvement. On 100,000 fixed-root Leduc comparisons, the fully coupled CP-GRPO instantiation reduces return-contrast variance from 41.1158 to 18.1441, a 55.87% reduction, while preserving the declared branch marginals. With three actions, OSCC-Select chooses continuation coupling and attains gradient-noise trace 0.0783 versus 0.0917 for return-variance selection. Increasing calibration from 64 to 2,048 groups raises certification from 0.327 to 0.995.

Abstract: Predicting how an environment will change before acting is a natural route to better decision making for agents. Recent post-training methods therefore require agents to predict the next observation and turn that prediction into a reward or a direct supervision signal, which is called world model. Existing next-observation training methods help the agent to learn the environmental content. However, they additionally involve an optimization process, which may introduce several effects other than learning to predict the world. Consequently, where the performance gain comes from during the training process remains an open question. We answer this research question through replacing true next-observation targets with in-distribution mismatched observations during the training process. Across two interactive text environments, mismatched targets lower prediction accuracy by 15.3-61.6% relative to ground-truth targets, yet retain substantial task gains over the base model. Compared with the base model, trained models consider more candidate actions and exhibit less looping. We also introduce a setting that replaces prediction-based rewards with independent random signals. This training expands task coverage (pass@64) even when the reward carries no environment information. We also generalize this finding to VisualWebArena, where random-reward training raises pass@64 by 14.3% relative to the base model, without observation-matching rewards or an external multimodal teacher for reward construction.

Abstract: In offline goal-conditioned reinforcement learning (GCRL), divide-and-conquer scales to long horizons by joining two shorter segments at a subgoal. However, under stochastic dynamics, the base case of this rule values the luckiest trajectories through the data. The subgoal must also lie on a shared trajectory, so a state-goal pair that no trajectory connects gets no value update at all. To address both, we present Grounded Transitive RL (GTRL), an offline GCRL value learning algorithm that grounds the divide-and-conquer update with a one-step TD target. Over a single step, TD is correct, as its target averages over the successors and needs no subgoal. GTRL adds this target to the composition rather than replacing it, so every pair receives an update, and the composition still carries the long horizon. GTRL also corrects the bias from hindsight relabeling by reweighting each goal against how reachable it was from other successors. We evaluate our algorithm on nineteen OGBench tasks spanning stochastic, deterministic, and stitching environments, where it achieves the highest average success rate. Code will be released soon.

Abstract: Recent advances in quantum computing are opening new opportunities for computationally intensive decision problems. This paper studies how quantum computing can improve Monte Carlo tree search (MCTS) in the fixed-confidence setting, where the goal is to identify a near-optimal move in a given game tree with high probability while minimizing query complexity. We first formulate MCTS under a quantum oracle model. We then develop a quantum MCTS algorithm (QMCTS) that combines threshold-based elimination on the search tree with quantum Monte Carlo estimation to reduce the cost of evaluating stochastic leaf values. We establish an instance-dependent lower bound on the query complexity and derive a corresponding upper bound for QMCTS. The lower-bound analysis introduces a new quantum phase-testing result that may also be useful beyond MCTS. We validate the theoretical results through simulation experiments and further demonstrate the feasibility of QMCTS on real quantum hardware.

Abstract: General Game Playing across arbitrary games from rules alone remains challenging due to differing algorithmic requirements across game classes and strict decision-time constraints. Rather than hand-designing search heuristics for specific domains, can we leverage Large Language Models (LLMs) to discover general game-playing algorithms? Because language models can propose and refactor structured code, they provide an expressive proposal engine for exploring the space of algorithmic designs. We introduce a multi-agent LLM meta-learning system to co-evolve game-agnostic procedural search mechanisms in C++ alongside domain heuristics synthesized directly from game rules. Controlling the compute budget, we benchmark the discovered mechanisms across more than 400 diverse environments, including OpenSpiel training and held-out games, procedural simulation engines, and games with deep neural policy-value representations trained via PPO. Evaluated via AlphaRank stationary distributions and Soft Condorcet Optimization (SCO) against 15 established MCTS baselines, the discovered search mechanisms consistently achieve top-tier ratings and pairwise ballot majorities over most baselines across independent evolutionary runs, generalizing to unseen human-designed and procedurally synthesized games and remaining competitive with baselines on frozen neural network representations.

Abstract: World models simulate the consequences of action candidates, but good planning need not preserve every physical distinction required for accurate prediction. We formalize this gap through a hierarchy of mechanism, response, and decision sufficiency. Given a candidate set, the planning query determines which physical variations matter and how precisely they must be preserved: coarse decisions can discard much of the information needed for prediction, whereas fine decisions may require nearly the same resolution. In practice, planners often adaptively search to construct candidates, and information unnecessary for final selection may still be needed to discover good candidates. What a world model must preserve therefore depends on the query, the candidate set, and the planner. We study these effects in a collision system, nonlinear dynamics, and robotic planning. These varying requirements raise a design question: where should query information enter the planning system? A model that jointly generates actions and outcomes conditioned on the query achieves lower regret than an action-conditioned world model on seen objectives, but this advantage largely disappears when generalizing to unseen objectives. Motivated by this, we propose a modular design in which the query determines where to look and an action-conditioned model predicts what will happen, allowing the same predictions to be reused across objectives.

Abstract: Reinforcement learning (RL) trains representations on data selected by the agent's policy, which then uses the resulting returns to guide its next choices. We show that this loop can sustain a lower-return policy even when representation fitting is globally optimal on those data. In a self-confirming superposition trap, every optimal code assigns overlapping directions to features that rarely occur together under the current policy. An alternative action brings them together, causing interference that lowers its return and reinforces avoidance, although refitting to that action would yield more return at the same capacity. We characterize the dimensions admitting a trap in a tied two-step model and show separately that equal feature frequencies, continued visitation, and independent controller learning need not prevent it. Because fitting weights errors by visitation, an avoided action can lose its return advantage at little cost to the objective. In a finite-action model, we bound this distortion and derive a replay condition: sufficient training weight on the best separately adapted action preserves its ranking despite residual error. Neural PPO experiments show how the feedback develops during learning: agents initialized toward different actions develop different interference patterns, opposite mean return rankings, and different final policies at the same capacity. We therefore test whether retaining access to neglected states can improve control. Keeping these states in training reduces measured interference and improves sequential return, with gains even when the encoder is frozen. Related interventions on state access, replay weights, and feature overlap improve control on MiniGrid and DMControl. For agents that learn through a world model, protected fitting improves DreamerV3--Crafter's cumulative training scores at unchanged capacity.

Abstract: We introduce Adaptive LeWorldModel (ALeWM), a world model based on a joint-embedding predictive architecture (JEPA) that learns to concentrate predictive information in compact prefixes of a wide latent representation. To encourage this ordering, ALeWM learns a sequence-conditioned distribution over prefix lengths and trains the predictor to estimate the full next embedding from a sampled input prefix. As standard anti-collapse objectives encourage variation across latent coordinates and do not organize them by predictive importance, we also introduce MixSIGReg. MixSIGReg regularizes the masked embeddings against a prior-weighted mixture with Gaussian active prefixes and zeros in the remaining coordinates. As a result, the ALeWM objective encourages early coordinates to retain information useful for prediction and recursive planning. Our analysis shows that the mixture distribution used by MixSIGReg assigns higher variance to earlier coordinate blocks and lower variance to later ones. In addition, we show that, under specified assumptions, prediction error is minimized by placing the information most useful for prediction in earlier blocks. Empirically, we study the behavior of ALeWM in a controlled dynamical system with known state variables and in goal-conditioned visual control. We show that ALeWM consistently achieves higher mean success rates than tuned fixed-width LeWM, with lower planning capacity on average.

Abstract: Accurate one-step predictions do not ensure that a world model's rollouts retain the data distribution. We make the model's decoded stationary law explicit by learning a decoder of a fixed Gaussian reference and constraining the behaviour-averaged transition to preserve that reference. For controlled systems, a joint transition uses a conditional action chart to preserve behaviour occupancy without requiring invariance at each fixed action. Joint state--action rotations and parallel Gaussian noise give an exactly preserving transition with a tractable conditional density. We derive an absolute convergence bound from finite initialization banks and control departure from the reference through conditional action-space divergence. Across $216$ fitted pixel checkpoints on twelve control tasks, the occupancy model with a reference mixture retains every evaluated chain at $10^5$ steps in all $36$ task--seed cells, with a rollout-minus-reference energy-statistic difference of $-0.0002\pm0.0003$ (training-seed standard error). Each of the four nonpreserving comparison arms loses chains, although the Gaussian arm is more accurate at ten steps. An offline DreamerV3 reference also achieves better short-horizon accuracy. These results distinguish three properties of a world model: the distribution it approaches, the rate of approach, and the conditional dynamics it learns.

Abstract: Multi-agent flow policies learn cooperative behavior from fixed offline datasets, but often struggle to complete tasks in situations not covered by the offline data. In these situations, agents must both adapt to changes in the environment and coordinate with one another, yet action patterns learned offline are often insufficient for effective adaptation and coordination. To address this problem, we propose Multi-Agent Flow-Pretrained Policy Optimization (MA-FPPO), which uses online fine-tuning to improve the cooperative behavior of models pretrained with flow matching through new interactions with the environment. Building on the behavior learned during pretraining, we construct policies with explicit action likelihoods for discrete and continuous action spaces. We then update the pretrained model using shared team advantages to further improve coordination based on team performance. Our method achieves, on average, relative gains of 52.8% over the strongest listed offline baselines across 30 settings and 29.8% over purely online learning across 38 comparisons with matched online budgets and evaluation protocols.

Abstract: In this paper, we use Pokémon Battles as a case study to investigate how to improve the performance of LLM-based agents in tasks that require opponent-aware planning without additional fine-tuning. We propose Long/Short-Term Knowledge Augmented Generation (LSTKAG), a mechanism that enables LLM-based agents to leverage past states of the current task and retrieve experience summaries from similar previous task instances based on the current state. Based on LSTKAG, we design Porimon, an LLM-based agent structure for Pokémon Battles. For optimization, we introduce an external API for precise damage calculation and more detailed information about the game. We conduct tournament-like evaluation experiments comprising 15,000 battles for hyperparameter optimization, ablation studies, and performance evaluation. The results indicate that Porimon-based players with hyperparameter optimization significantly outperform players based on PokéLLMon, an LLM-based agent structure proposed in previous research, and the rule-based heuristic player. Furthermore, our ablation study shows that Porimon variants outperform the one without extension in game information retrieval, which shows the contribution of that extension. However, the current experiment results are inconclusive regarding the contribution of Long-Term KAG. These results suggest that introducing external resources, information from previous states of the current task, and experience summaries from similar previous task instances could elevate the performance of LLM-based agents designed for tasks requiring opponent-aware planning.

Abstract: Powered by expert guidance, agents can operate in interactive environments; however, it is unclear whether they can learn autonomously from their own experience. To evaluate such self-improvement methods, we introduce GameBoyWorlds, a testbed for agentic self-improvement in video games. GameBoyWorlds-Execution evaluates task execution on a collection of 5 distinct game series. Agents are allowed access to dedicated training games but are provided no demonstrations, documentation, or rewards. Agents must ground themselves in the environment through self-directed exploration and by inferring actionable knowledge from their own experience. At test time, agents must complete short-horizon tasks that evaluate their ability to navigate, interact, and engage with game-specific mechanics in unseen games. Out-of-the-box frontier models complete fewer than 50% of the 500 tasks due to failures in multimodal grounding, establishing that self-improvement methods have room to push performance. We demonstrate that contemporary approaches to self-improvement are lacking, with world modelling and autonomous skill discovery failing, and a novel strategy that uses curiosity-based exploration to write guides achieving only partial success. GameBoyWorlds-Playthrough tests end-to-end game completion in two fan-made Pokémon games. We show that while frontier models have been pre-exposed to official releases such as Pokémon Red, they lack essential information on the games in our testbed. Instead of relying on their parametric knowledge to succeed, agents must learn from their own experience and autonomously improve over the course of the playthrough. We show that a sophisticated agentic pipeline with multimodal memory and hierarchical subgoals fails to reach even the first major milestone in both games, establishing GameBoyWorlds as an ambitious target for self-improving agents.

Abstract: A key challenge in machine consciousness research is translating theoretical models into computational-level implementations. In this paper, we address this challenge by proposing a five-layer implementation architecture for the S3Q (Simulated, Situated, Structurally Coherent) theory of consciousness. Rather than introducing novel formalisms, the architecture composes published computational primitives into a single pipeline. S3Q identifies three jointly necessary conditions for qualia: (1) grounded sensorimotor situatedness, (2) internal simulation via a world model, and (3) structural coherence between predictions and observations. No existing computational system implements all three simultaneously. We map each S3Q tenet to specific, compatible computational machinery and specify how these components interface within a single representation pipeline that operates on continuous, differentiable, per-object slot vectors, along with a developmental bootstrap sequence and falsifiable predictions for the composed system that no subset of the architecture produces in isolation. The model suggests that a basic sense of "self" develops by linking actions to their outcomes, and that behavior falls into three patterns (hesitation, curiosity, or avoidance) depending on how unexpected an outcome is and whether it is experienced as positive or negative. Each prediction is individually falsifiable, providing the field with a testable framework to advance our understanding of machine consciousness.

Abstract: Existing in-context reinforcement learning methods mainly pretrain Transformers with supervised behavior-prediction objectives. This enables task inference from context, but makes the learned policy strongly depend on the quality of offline actions: when trajectories are weak or suboptimal, imitation itself becomes a biased learning signal. We propose Q-Target Pretrained Transformers (QTPT), which keeps the context-conditioned Transformer architecture but replaces behavior cloning with a Bellman-style Q-target objective. QTPT therefore learns to use rewards and transitions in the context to estimate action values, rather than simply imitating the behavior policy. We theoretically analyze QTPT in stochastic linear bandits and finite-horizon MDPs, showing stronger robustness to data quality than supervised pretraining. Empirically, QTPT improves over supervised behavior prediction on controlled RL benchmarks with random or suboptimal data, and we examine extensions to D4RL Kitchen and AntMaze. Supplementary experiments evaluate backbone robustness, meta-RL comparisons, task-coherent context, and unsupported-action value overestimation. These comparisons distinguish the benefits of Q-target pretraining from the remaining limitations of offline coverage.

Abstract: Finite state machine (FSM) and behavior trees (BT) are widely adopted behavioral modeling paradigms for autonomous intelligent systems. While functionally equivalent and inter-convertible in principle, existing transformation methods between FSM and BT face major challenges in preserving behavioral completeness and avoiding model complexity explosion. To overcome these issues, we propose an LLM-driven unified conversion framework that enables automatic, efficient, and semantically consistent transformation between FSM and BT. Specifically, a novel loop execution BT structure is designed for LLM to accurately capture the loop structure in FSM, thereby preserving behavioral completeness. To mitigate the state explosion problem in BT-to-FSM conversion, a depth compression strategy is introduced with LLM prompt to eliminate redundant control nodes, complemented by differentiated hierarchical conversion rules that collectively reduce the number of required sub-FSM. Simulation experiments in multiple autonomous decision-making scenarios demonstrate that the proposed framework enables an accurate and automated bidirectional conversion between FSM and BT. Furthermore, it significantly enhances the scalability and maintainability of generated models compared to traditional approaches, providing a practical solution for behavior model conversion in consumer-grade autonomous intelligent systems such as service robots, game agents, and smart home devices

Abstract: This paper presents a spiking neural network (SNN) designed to predict the sequence of the external world states starting from the current world state. This SNN does not create the world dynamics model - instead it incorporates the SNN trained to predict the next world state and provides all mechanisms necessary to make the chain of predicted world states. These mechanisms are entirely spiking - they are implemented as spiking neuron ensembles. The present article describes this neuronal structure and tests its operation on a classic RL benchmark - ATARI ping-pong.

Abstract: Multi-agent reinforcement learning (MARL) commonly trains decentralized policies from scratch, requiring agents to acquire individual task competence and coordination simultaneously. Yet many multi-agent problems admit a compatible single-agent counterpart in which the underlying task can be learned in isolation. We introduce Multi-Agent Observation Transformation for Existing Single-Agent Policies (MATES), an input-side adaptation framework for tasks whose multi-agent observations preserve the solo-task information while exposing separately identifiable neighbor information. From multi-agent experience, MATES learns a small adapter that maps this observation into the format expected by a frozen single-agent policy, inducing actions suited to the shared environment without updating the single-agent policy itself. MATES leaves the pretrained policy's internal architecture unchanged and retains the objectives and update procedures of the underlying MARL algorithm. We evaluate MATES using both on- and off-policy algorithms on lifelong pathfinding, navigation, and cooperative discovery, spanning discrete and continuous observation and action spaces. Across all evaluated settings, MATES optimizes only 3.5-7.3% as many parameters as full-policy training while consistently outperforming MARL training from scratch. It approaches the performance of full fine-tuning, remains competitive overall with demonstration-based baselines, and retains strong task performance at team sizes not encountered during training. These results provide evidence that, under this observation structure, effective multi-agent behavior can be learned without modifying the policy that encodes individual competence.

Abstract: Prioritized experience replay (PER) improves sample efficiency by replaying high-priority transitions, usually according to absolute temporal-difference error. In stochastic environments, PER can distort the distribution of realized outcomes replayed from transitions with the same state-action pair. We call this within-group self-selection. We quantify the resulting changes in within-group outcome frequencies and mean Bellman targets. We decompose PER into between-group allocation and conditional sibling selection, and derive fixed-buffer corrections that preserve current group-level priority mass: SAMPLE selects a group through PER and trains on a uniformly sampled sibling; AVG averages sibling Bellman targets; and MODEL samples from an empirical full-outcome model. In exact state-action environments with rare high-magnitude outcomes, sibling-aware replay improves learning efficiency over PER, although matched parameter sweeps show that tuning can narrow some gaps. In MinAtar, approximate VQ-VAE groups with SAMPLE mitigate degradation under mean-preserving reward tails in four of five games. Sibling-aware replay thus retains the focus on high-priority state-action regions while recovering their empirical outcome frequencies.

Abstract: World action models generate actions together with visual predictions of their consequences. These paired outputs create the potential for planning by sampling multiple actions from one state, comparing their imagined outcomes, and choosing the action with the most promising predicted outcome. However, how to use imagined futures to guide action selection remains unclear. We examine this planning potential empirically. First, we estimate an oracle upper bound on selection by choosing the sampled candidate whose realised outcome is best. In a controlled same-state analysis, this choice raises success from 68.9% under uniform random selection to 79.2%. We then test selectors based on visual quality, physical consistency, and task progression as controlled interventions. Some tested selectors yield higher observed success, but the gains are uneven and the matched selectors leave much of the measured opportunity unrecovered. To investigate this gap, we examine whether sampled actions lead to different outcomes, whether these differences are visible in the predictions, and whether a score recognises them. Counterfactual branching from the same states shows that selection opportunity is concentrated in relatively few decisions in the initial candidate sets. Action spread and outcome coverage need not increase together. In a further evaluation across trajectory phases with complete action execution, the tested scores again recover little of the available improvement despite a small gain from learned value. These findings distinguish producing consequential action choices from recognising them in generated futures, motivating the evaluation of WAM predictions through their usefulness for decisions rather than visual quality alone.

Abstract: Long-horizon language agents receive sparse terminal feedback, while intermediate rubrics provide structured but potentially misspecified assessments of progress. In resettable training environments, counterfactual continuation rollouts can measure local credit, but exhaustive replay is costly. We propose Credo, a framework that couples evolving semantic rubrics with selective, execution-based credit correction. A frozen judge maps visible transitions to rubric features, and a credit head predicts the change in expected terminal reward associated with the realized transition. Independently sampled two-sided replays correct prediction residuals using their recorded inclusion probabilities. We derive conditional unbiasedness and a variance decomposition that connects two design choices: which rubric features to retain, and where to allocate a fixed expected replay budget. The resulting criterion weights prediction errors by policy-score sensitivity and missing replay coverage; its allocation rule additionally accounts for continuation cost. We also describe a practical mixture with terminal leave-one-out advantages and distinguish its clipped, token-normalized PPO implementation from the ideal policy-gradient estimator. This preliminary report provides the method, proofs, an exact finite-model audit, and a controlled evaluation protocol. It makes no claim of empirical superiority on language-agent benchmarks.

Abstract: Obtaining the optimal action-value function in Markov decision processes is computationally intensive in large state--action spaces. In this study, we present statistically rigorous convergence results for a robust reinforcement learning algorithm warm-started by a transformer-based action-value function prediction, where natural language prompts encode task specifications. Our framework adopts the R-contamination model to characterize uncertainty in the state transition kernel, and employs conformal prediction to certify convergence via trajectory-level nonconformity scores constructed from the contracting Bellman residual. The resulting conformal quantile bounds the gap between the running and optimal action-value functions simultaneously over all iterations, thereby yielding a pre-certified stopping rule that requires little knowledge of the true transition kernel. Numerical case studies on perturbed maze environments of varying size and contamination level confirm that the transformer-based warm start measurably reduces the initial error and accelerates convergence, while the proposed conformal bounds track the true error trajectory more tightly than existing guarantees.

Abstract: We study a one-player token game on an $N\times N$ board where tokens slide along diagonals or duplicate onto neighbouring ones to form a combinatorial rectangle $R\times S$. A conserved integer weight $W'$ and a strict monovariant guarantee $O(N^2)$-length solutions, placing the game in $\mathsf{NP}$. We prove that reaching a final position factors this $2N$-bit $W'$ into two $N$-bit factors $V,M < 2^N$ that encode the rectangle's rows and columns. Consequently, solving the game for a balanced-semiprime target is equivalent to integer factoring. However, if the target rectangle is known, the solution reduces to two polynomial-time steps: a forced downward chip-flow and a $0/1$-polynomial factorisation leveraging Cohn's theorem. The game's entire difficulty is thus isolated to the initial number-theoretic split. Supplying the popcounts of the factors as a promise preserves this asymptotic hardness but bounds the target search space. We exploit this constrained space using a learned policy/value network and an AlphaZero-style Monte-Carlo tree search, empirically probing the limits of neural look-ahead on a factoring-equivalent environment.

Abstract: Contrastive activation directions are often interpreted from what they decode or how strongly they steer behavior. But what evidence is sufficient to identify the construct represented by such a direction, rather than a correlated feature of the contrast used to extract it? We study this question for a good--bad outcome direction in a maze task, using controlled interventions that separate the realised outcome from the informational history through which it became known. Across multiple LLM checkpoints, directions fitted on one explicit outcome encoding transfer well to another, indicating that the readout is not tied to surface form. In contrast, when the same realised outcome is reached through announced and unannounced histories, transfer degrades substantially: even after both histories receive the same explicit outcome, the post-event readout remains strongly conditioned on the earlier announcement. In a matched maze-RL run, the post-RL direction becomes substantially more predictive of reference-MDP remaining return and the policy becomes more dependent on it at the tested sites, while this history dependence persists. These results support a functional, value-related interpretation of the direction, but not its identification with a history-invariant scalar valence state.

Abstract: Most artificial neural systems are built to map given inputs to outputs. Adaptive agents face a prior problem: they must act without enough evidence, seek encounters with the world, and revise behavior when evidence appears. We propose another starting point for intelligent neural networks: proactive search without signals, curiosity at its most basic. We ask whether it can come from a minimal untrained circuit. The spiking unit studied here inverts its response to input: with no signal in its window it fires faster; once signals arrive it switches to a slower, inverted regime. Search needs three or more such oscillators in counter-phase, each reading the same input in a different time window. With no training, supervision, parameter tuning, or controller, the composite switches on its own between exploratory spiral search and exploitative tracking, finding both first-degree symmetry and second-degree groups. The switch comes from temporal disagreement between its fast and slow readings of the same signal. We view the circuit as evolutionarily trained: its abilities come from structure, not experience. Ablation over 63 configurations and 63,000 trials shows the switch needs both temporal staggering and counter-phase opposition, neither enough alone: the behavior is emergent, not programmed. The spiral persists at zero rotational diffusion, so it is structural, and degrades gently under perturbation. More oscillators improve spiral regularity but cut resource capture, so the smallest sufficient circuit wins. We propose that this principle underlies search in simple organisms, navigation and decisions in complex ones, and, being so simple and common, goes unnoticed unless you strip the logic bare. Eventually, networks of such proactive primitives may offer another foundation for AI architectures that explore our world rather than merely predict the next symbol in a sequence.

Abstract: Reward-based reinforcement learning for language models, exemplified by Group Relative Policy Optimization (GRPO), collapses an entire stochastic trajectory into a single scalar reward. This is clean and scalable, but it explores and allocates reward inefficiently: a trajectory may contain many causal decisions, recovery attempts, and environment-randomness events, yet every token or action inherits one trajectory-level advantage. We study tree-based rollout construction as a compute-allocation problem for policy-gradient estimation. Our central claim is that branches should be placed not where the policy is merely uncertain, but where an additional branch most reduces uncertainty about the policy gradient per unit of compute. From a law-of-total-variance decomposition of the local policy-gradient random variable, we derive two allocation laws: new branches reduce decision uncertainty, while repeated suffix rollouts reduce continuation uncertainty. The resulting EPIG-Tree score allocates branches using the already computed rollouts. It estimates occupancy- and score-weighted value uncertainty, along with a suffix law $n_e \propto w_e \|\nabla_θ\log π(a_e|h_e)\| σ_e / \sqrt{c_e}$. Empirically, EPIG reduces gradient MSE in cloned-state control, winning in all nine dense continuous-control environments of a 13-environment sweep and recovering the reference gradient direction near-perfectly, and it improves frozen-LLM gradient calibration relative to entropy branching. In online single-turn math, tree-local credit beats flat GRPO, while branch placement is secondary to token-level credit assignment. In online multi-turn Wordle, EPIG attains the highest final win rate (0.850), overtaking flat GRPO, which saturates early at 0.790, and entropy branching as training proceeds, confirming that the gradient-estimation advantage transfers to a stateful, large-action setting.

Abstract: Tree-based Monte-Carlo Tree Search (MCTS) duplicates the same state when it is reached through different trajectories, which can waste simulations in stochastic MDPs. We introduce Graph-Based Stochastic-Power-UCT (GS-Power-UCT), which shares states reached at the same planning depth while keeping separate values for states reached at different depths. This design applies to general stochastic MDPs, including problems with cycles. We prove that for a fixed planning horizon, the root estimate converges to the finite-horizon value at rate $O(n^{-1/2})$, matching tree-based Stochastic-Power-UCT while reusing samples across shared states. We also study two full-state variants: GS-Power-UCT-F, which stores one node per physical state to increase sample sharing but may mix values from different remaining horizons, and GS-Power-UCT-F$^+$, which uses an adaptive horizon to control this bias. The latter converges to $V^{\star}(s_0)$, the optimal infinite-horizon discounted value at the root state $s_0$, when the remaining cross-depth gap vanishes. Experiments on stochastic planning benchmarks show improved sample efficiency over tree-based and graph-based baselines.

Abstract: Regularized self-play -- the family behind DeepNash's Stratego play -- drives a two-player zero-sum policy to a Nash equilibrium by best-responding to a slowly moving, entropy-regularized reference policy $ρ$. When the game has a polytope of value-equivalent equilibria, the regularizer silently breaks the tie: with a uniform reference it selects the maximum-entropy member, the I-projection of $ρ$ onto the Nash set. Can the reference be used to choose the equilibrium on purpose? On five exactly solvable games plus a 2-D polytope, with exact best responses and equivalence tests over independent seeds, anchoring the reference at a target member and refining steers self-play to that member with mean coordinate error 0.007 at median exploitability $5\times10^{-5}$, TOST-equivalent to the request within $\pm0.05$; the anchoring persists through refinement and follows the reference, not the initialization. Selection follows the reach-weighted I-projection (slope 0.969 [0.950, 0.987]). We report with equal emphasis where the story breaks: fixed off-manifold references cost 0.08-0.25 exploitability; stiff or flat families require a smaller mirror step, set by a pre-registered rule; boundary targets undershoot; curvature predicts where boundary saturation bites (rank correlation 0.90, p=0.037) while interior precision is curvature-independent. Table and MLP steering maps are equivalent within $\pm0.03$ at every target (30 seeds); matched control arms show attention's robust signature is excess seed variance, any systematic shift bounded at 0.018 and not significant. Against a best response the selection-robustness trade-off is degenerate: steering matters only against fixed, non-equilibrium opponents. The recipe -- anchor the reference at the desired member and refine -- reinterprets the KL anchor of RLHF-style RL as a selection knob, not only a stability leash.

Abstract: We study state-adversarial Markov decision processes (SA-MDP) as a game of observation-space attacks: at each step, an agent selects an action from a received observation while an adversary$\unicode{x2014}$who knows the true state the agent is in$\unicode{x2014}$chooses a perturbed observation within a state-dependent proximity set. While existing work focuses on Markovian policies, we develop a solution concept and computational approach for SA-MDPs under history dependence. This is motivated by results showing that history dependence can materially change equilibrium outcomes and can force both the agent and the adversary to adapt their strategies. First, we prove the non-existence of universal (agnostic of the initial state distribution) history-dependent equilibrium policies. In response to this finding, our main result presents the first algorithmic route to computing $ε$-approximations of initial-state dependent equilibria. We do so by reducing SA-MDPs to a strategically equivalent constrained zero-sum one-sided partially observable stochastic game. We conclude by testing our algorithm on small analytically verifiable games and showing it scales to larger, more realistic benchmarks, including Atari Freeway rollouts with a 12-period ahead horizon.

Abstract: Chess has long served as a model domain for studying search, expertise, decision-making, and artificial intelligence. The emergence of large language models (LLMs) has renewed the relevance of chess as a controlled environment for investigating strategic reasoning and comparing human and artificial decision-making. We present a systematic mapping study of recent research spanning human players, classical chess engines, neural and reinforcement-learning systems, LLMs, and hybrid approaches. The final map comprises 84 core study families, classified according to agent type, strategic-reasoning stages, and evaluation dimensions. The map reveals a literature strongly concentrated on situation assessment, evaluation, and action selection, while explicit planning, explanation, metacognition, and human--AI collaboration remain less explored. LLM research places particular emphasis on state representation and generalization, whereas grounded explanation appears more frequently in hybrid approaches combining language models with engines, expert knowledge, or other external structures. Two distinctions emerge that the map aggregates rather than resolves: hybrid systems differ in where and when heterogeneous capabilities combine, and evaluations that show improved human performance do not thereby establish human--AI synergy. We propose both as extensions of the mapping framework. We argue that chess provides a useful bridge between cognitive and computational perspectives on strategic reasoning, and identify explicit planning, grounded and faithful explanation, metacognitive calibration, and human--AI complementarity as directions for future research.

Abstract: The agent-centric general value function (ACGVF) construction of \citet{tasse2026goal} lets the agent make two decisions that are normally imposed by the environment or agent designer: which goal to pursue and when to declare a goal as finished (in addition to choosing the action). This is a very general framework that subsumes almost all prior work on reinforcement learning, control and planning, as well as more general formalisms proposed in the cognitive sciences. However, it assumes the environment is fully observed, i.e., that the observation is a sufficient statistic. In \citet{murphy2025rl}, a general agent design was proposed where the policy is based on an internal belief state $z_t$ and an internal goal; however, the goals were assumed to be externally provided. In this note, we unify and extend these two approaches using the formalism of hierarchical hidden Markov models (HHMM) \citep{murphy2001hhmm}.

Abstract: Social-deduction games such as Werewolf are increasingly used to evaluate LLM agents, but existing evaluations often rely on final game outcomes. We propose a belief-shift evaluation benchmark in Werewolf for analyzing communication skills through belief updating. Using LLM-played games, we annotate suspicion and accusation messages and measure how an observing village-side model's beliefs change after each message. We evaluate 40 open-weight LLM configurations on 1,224 annotated messages. Our results show that larger models better distinguish true wolves from villagers based on game history, but accusations still strongly influence their beliefs. Models become more suspicious of the accused target and less suspicious of the accuser, especially when the accuser is trusted, even if the accuser is wolf-aligned. Larger models better resist accusations from accusers they already distrust. Overall, our findings suggest that current open-weight LLMs up to 120B parameters still struggle to integrate accusation content with source trust in strategic communication. Our benchmark and code are available at https://rlg.iis.sinica.edu.tw/papers/werewolf-accusation-benchmark.

Abstract: Lux AI Season 3 requires agents to act under partial observability, randomized episode level dynamics, and a best of five match structure that rewards both tactical execution and fast adaptation. We present HORIZON, a hierarchical agent that combines symmetry aware spatial perception, dual memory belief tracking, relic centric graph attention, information gain driven exploration, and an opponent conditioned policy mixture. HORIZON separates short horizon control from cross match meta reasoning, while auxiliary belief and world model objectives stabilize learning. Trained with PPO in a large scale JAX simulator, the resulting agent explicitly infers hidden game parameters and opponent style. Experiments show consistent gains in match win rate, episode win rate, adaptation gain, and league rating over strong recurrent and feed forward baselines.

Abstract: The Core is a unique competitive co-evolution algorithm that allows agents to evolve autonomous control without utilizing a traditional fitness function. The agents evolve via local interactions through tournament selection, crossover, and mutation, producing offspring by evolving better controllers. Previous works have shown The Core's ability to evolve agents capable of combat and navigation in the Xpilot video game. This research expands upon that premise by niching agents to specific subsets of the original environment The Core was tested in. Our results demonstrate the niched agents capacity for success over agents niched to the entire system and agents niched to different sub-environments.

Abstract: Cooperative AI agents are evaluated against other AIs, yet human cooperation relies on implicit conventions---shared protocols for reading meaning beyond the literal message---which AI-AI benchmarks may not capture. We propose the \emph{convention gap}, the difference between the failure probability predicted from the literal content of communication and the observed failure rate, as a metric of implicit communication. In the card game Hanabi, the finite deck and deterministic hint constraints make this posterior exactly computable. We replayed about 101,000 play actions from three public datasets of human-human (hanab.live), AI-AI (HOAD), and human-AI (HanabiData) games. The gap was +26.2 percentage points (pp) in human pairs, $-$0.7~pp in AI pairs, and +16.4~pp in human-AI pairs, and was concentrated on plays of cards that had received no hints (+46~pp in human pairs). Within human-AI play, the literal information available to humans was similar across the three AI partners (mean predicted failure 38--41\%), but human failure rates ranged from 14.4\% to 34.4\% and the gap from +24.1 to +6.2~pp; the partner eliciting the largest gap produced the fewest human failures. Game score carried different information: it depended on each corpus's roster composition, whereas the gap separated human from AI play at the agent level. As a known-answer check, Off-Belief Learning agents, whose convention content is controlled by construction, gave a gap of +1.6~pp at the convention-free level, rising monotonically to +21.7~pp. These results suggest that convention compatibility, rather than AI-AI performance, may predict an AI's effectiveness with human partners.

Abstract: The integration of graphs with Goal-Conditioned Hierarchical Reinforcement Learning (GCHRL) has received increasing attention, as graphs naturally encode task hierarchies for effective subgoal sampling. However, existing methods often overlook intrinsic connectivity information, failing to fully leverage the underlying topology for efficient learning. Most graph-based GCHRL methods use the graph as a stochastic sampling tool rather than as an environmental model that encodes connectivity and state-accessibility information. This limitation is particularly acute in quasimetric environments, where the inherent asymmetry of state transitions poses a fundamental challenge to stable policy learning and robust path planning. In this paper, we address these problems by introducing a state connectivity model designed to predict pairwise state connectivity strength in asymmetric environments. We transform these connectivity strengths into scalar auxiliary dense rewards, providing continuous guidance across multiple hierarchical levels. We demonstrate that our proposed framework, Graph-Guided Quasimetric Dense Reward (G2QDR), can theoretically be integrated into any existing GCHRL architecture, and the state connectivity model is efficiently implemented via a neural network trained on a directed state graph generated during exploration. Empirical results across a wide range of sparse reward environments indicate that, in general, G2QDR can enhance the performance of baseline GCHRL approaches with acceptable computational overhead.

Abstract: In continual reinforcement learning, carefully managing the stability-plasticity tradeoff remains a core challenge. Recent work by Abel et al. (2025) formalized this dilemma by defining plasticity as the generalized directed information from an agent's observations to its actions, and empowerment as the generalized directed information from its actions to its observations. This formulation successfully reframes the traditional stability-plasticity tradeoff as an empowerment-plasticity tradeoff. However, while extensive literature exists on optimizing for empowerment, there is currently no research addressing the optimization of plasticity under this new definition. This paper presents preliminary work toward optimizing plasticity within Markov decision processes. We show that there exists a Bellman optimality equation for optimizing plasticity similar to previous work for empowerment.

Abstract: Combining search with function approximation has driven major advances in game-playing programs, making self-play algorithms more competitive than ever. Still, the computational overhead of the most popular methods, based on Monte Carlo Tree Search (MCTS), can be substantial. In this work, we investigate whether simpler methods remain competitive in non-trivial, moderately sized games such as Connect Four, Hex(7x7) and synthetic games. We train a minimal self-play implementation of Approximate Value Iteration (AVI) and use ground-truth oracles for exact evaluation. Contrary to expectations, our results demonstrate the surprising effectiveness of AVI: it learns more accurate value functions than those learned by AlphaZero, while its one-step-lookahead greedy policies remain competitive with MCTS-based policies at substantially lower training and inference costs. Preliminary experiments on Othello and Go(9x9) show that AVI trains stably on larger games and learns effective value functions. These findings suggest that the success of MCTS-based methods may have eclipsed simpler approaches that have become increasingly practical with modern deep-learning tools.