Counterfactual Rollouts

Learned world models are becoming essential to general-purpose agents: by predicting action consequences, they support planning and decision-making while reducing reliance on costly trial and error. This reliance creates a fundamental ambiguity: when a world-model-guided decision fails, the trajectory alone may not reveal whether the agent's decision rule or the world model caused the loss. We formalize this failure-attribution problem as a counterfactual decomposition of return loss and prove that its components are not identifiable from passive interaction, even for finite-horizon planners. This obstruction motivates Dual-Frontier, a learning principle that admits a world-model-guided decision only when its predicted advantage exceeds a certified bound on decision-relevant world-model error; otherwise, evidence is allocated to world-model verification. Action-conditioned value bounds and a closed-loop extension guarantee non-decreasing return for admitted decisions. Calibrated gates and simultaneous confidence sequences support adaptive evidence reuse, with sufficient and necessary verification bounds. Controlled learned-model experiments validate the predicted failure modes and certification behavior, while cross-backbone tool-use benchmarks instantiate the same verify-then-promote rule in realistic agent world-model pipelines, consistently improving decision quality and reliability.

Generative world models aim to predict future states conditioned on actions, where action controllability is fundamental for reliable dynamics modeling. While recent efforts leverage simulator-generated data to enhance this capability, existing training pipelines face two fundamental limitations. First, static offline data collection leads to a distribution misalignment between training sets and the model's evolving error patterns, failing to resolve critical long-tail scenarios where dynamics predictions remain unreliable. Second, the standard objective of minimizing observational discrepancy often encourages the model to exploit spurious correlations instead of capturing the underlying action-effect causality. To address these limitations, we propose OnlineWM, an online training framework that continuously improves world modeling through active simulator interaction and causality-aware optimization. OnlineWM introduces two key innovations: (1) Active Online Learning: Instead of using fixed datasets, OnlineWM adaptively queries the simulator for new interaction sequences that target the model's current predictive weaknesses, ensuring high-utility data acquisition. (2) Causality-Aware Fine-Tuning: We propose a counterfactual learning strategy that contrasts the outcomes of different actions from identical states, forcing the model to attribute state transitions to specific actions rather than ambient environmental evolution, thereby grounding its predictions in reliable causal mechanisms. By integrating active data acquisition with causal optimization, OnlineWM establishes a closed-loop refinement process that ensures the model is both robust to diverse scenarios and precise in its causal attribution. Extensive experiments demonstrate that OnlineWM significantly enhances action controllability and generalizes effectively to unseen domains.

We introduce loss-conditioned state execution, a model-agnostic method that decides whether to execute a world model's fixed feasible proposal or retain the current state. Predictive informativeness alone, however, does not establish whether an update will reduce downstream loss. Occurrence ranking can approach perfection while persistence remains the unique absolute-loss Bayes action. Two transition laws can also share occurrence information and conditional variance yet require opposite absolute-loss decisions. We formalize state movability as the existence of a loss-reducing feasible correction and distinguish it from the benefit of a particular proposal. Our method constructs a loss-specific feasible proposal from a predictive distribution and evaluates its groupwise bounded-loss gain over persistence on independent calibration units. The proposal is executed only in groups with a positive simultaneous lower confidence bound. For fixed proposals and groups with bounded unit losses, we prove that every accepted group has lower expected loss than persistence with high probability when calibration units are i.i.d. draws from the target population. Experiments on public forecasting and action-conditioned dynamics benchmarks show supported updates and a trade-off between certification and coverage. On 28,684 held-out M4 Monthly series, the method executes the proposal for 14.0% of series and achieves bounded loss 0.588, compared with 0.599 for persistence and 0.621 for always executing the proposal. The paired 95% bootstrap intervals for both comparisons lie below zero. In constrained forecasting of six unhealthy-inventory types from JD$\mbox{.}$com, a leading e-retailer in China, strong occurrence-ranking signal coexists with a loss-based preference for persistence, illustrating why event predictability and state execution must be evaluated separately.

Continual world models must decide whether new data justify changing the model. Fixed replay schedules and prediction-error triggers specify when to update, but neither reveals the value of an individual update: one deployment run cannot show how the same model would have performed at that moment had it held its parameters. We introduce the fork ledger, which branches a deployment stream at pre-registered decision points into matched update and hold continuations under common random numbers. It evaluates both continuations on the same episodes and records $\Delta R = R_{\mathrm{update}} - R_{\mathrm{hold}}$. Always applying one fixed update mechanism lowers return on all three simulated control tasks: CartPole ($-144.0$; checkpoint-bootstrap $95\%$ CI $[-185.4,-116.1]$, against a converged return near $650$), Walker ($-82.8$; $[-101.1,-61.7]$) and Cheetah ($-18.6$; $[-29.0,-6.6]$). Divergence is an outcome of applying the update, so the estimand counts every attempted fork; restricted to the $693$ of $720$ that did not collapse, CartPole and Walker are unchanged in sign ($-113.4$ and $-82.1$) and Cheetah becomes unresolved ($-3.9$; $[-17.5,+13.0]$). The task is the unit of inference: each contributes $240$ attempted forks over five pretrained checkpoints crossed with two drift directions. The ledger makes counterfactual utility observable for a fixed mechanism, allowing triggers to be judged by the updates they select rather than by surprise detection alone.

The performance of an LLM agent depends on the scaffold around a frozen model. A common way to improve that scaffold is to use a coding agent as an optimizer: it reads current scores and traces and iteratively edits the source, producing a new candidate each round. Each edit is chosen according to a belief about how the environment will respond: what went wrong, and which change should help. That belief is typically implicit. It lives in the coding agent's reasoning on the current call, or remains latent in its parameters, rather than as something written down. Later calls therefore see scores and traces, but they do not use that belief. We introduce Belief-Calibrated Optimization (BCO), a method that writes that belief down as a persistent in-context document and continually revises that document as new candidates are evaluated. The resulting document is a world model: the current account of how the environment responds to edits. Added to an otherwise standard loop, BCO reaches a higher train passrate than a matched control that lacks only the world model, on five benchmarks spanning memory QA, tool-use QA, code-as-action app agents, and terminal agents. The gap remains on every held-out split, which is not used to select the candidate. After a target-model swap, in which the frozen model is replaced and the scaffold is not, the selected BCO scaffold leads on the tasks we test, except where context-window overruns leave it unfinished. An offline ablation then asks whether that gap comes from what the world model says. A fresh predictor given the accumulated document forecasts how the environment will respond more accurately than predictors given either no document or a same-form copy whose content has been falsified. The comparison indicates that the document carries reusable information in its content, not only in its form.

We ask whether a small, directly addressable hidden-state intervention can place a learned world model on an intended counterfactual future and then let the model's own dynamics carry that future forward. In a controlled two-object collision environment, we study a 192-dimensional recurrent model trained on factual and locally edited counterfactual trajectories. Candidate carriers are learned from training-only counterfactual-minus-factual hidden differences, and an affine map predicts carrier coordinates from the factual state and requested edit without access to the native counterfactual hidden state at test time. For bounded single-component velocity edits, rank 4 is the smallest tested rank on the preregistered grid that satisfies the development criteria. A one-shot rank-4 patch launches a 12-transition autonomous rollout without future observations, teacher forcing, repeated hidden-state correction, or physical-state clamping. The frozen procedure satisfies the preregistered 2-of-3 fresh-checkpoint replication rule and remains reusable at nearby anchors. The same Single-derived carrier and Single-only affine map also support bounded same-object two-component requests. Across the matched training regimes, broader counterfactual support was associated mainly with better Joint rollout accuracy and more additive Joint hidden responses. Composition-related structure is enriched in the rank-4 subspace but is not confined to it, and local recurrent diagnostics show strong one-step coupling from the carrier to the rest of the hidden state. A position-edit stress test fails the required specificity controls. Together, these results support a compact dynamics-effective intervention-entry interface, not a closed four-dimensional state or an intrinsic state dimension.

Detecting or attributing a supply-chain disruption is not the same as selecting the intervention that maximizes recoverable net value. We present CriticalSCM-Bench v1, a controlled synthetic benchmark with causal ground truth, paired factual/counterfactual rollouts, and an explicit net-value objective. Relative to a full-information train-selected static benchmark, LambdaMART improves median normalized net value by 5.7--16.2%, with paired statistical support on the semiconductor and critical-material archetypes but not on digital infrastructure. On digital infrastructure, a domain-informed constant-buffer policy remains stronger, showing that greater model complexity is not uniformly justified. Across partial and delayed settings, LambdaMART retains 33--75% of full-clamp value. Stress tests further show that intervention fidelity, timing, cost, and held-out disruptions can alter policy ordering. Critical materials show the weakest out-of-distribution retention. Separately, a guarded explanation study over 540 generations preserves every fixed intervention decision after deterministic validation and template fallback, although exact wording remains unstable. Within this controlled setting, the results identify regimes in which adaptive ranking adds value and those in which simpler structural policies remain preferable.

Reinforcement learning (RL) seeks to optimize sequential decisions to maximize population-level benefits over time. However, when deployed in high-stakes settings such as healthcare, RL decisions might systematically restrict some subpopulation's access to valuable services in a manner contrary to the values and goals of stakeholders. Counterfactual fairness (CF) offers a promising framework to address this problem based on causal reasoning. This paper develops a data preprocessing algorithm that, when used in tandem with policy learning, enables CF in RL. Our algorithm relies on a novel quantile distribution mapping method for sequentially estimating the counterfactual states and rewards in the data preprocessing step, subsuming common additivity assumptions used for counterfactual prediction as a special case. We theoretically prove that the per-step level of counterfactual unfairness and infinite-horizon suboptimality gap can be bounded under mild regularity conditions. We also empirically test our algorithm in numerical experiments as well as in application to a real-world interventional digital health dataset.

A world model is only useful for physical AI if it changes what the agent does, and only safe if it declines to do so when it is wrong. We study both halves of that requirement with CausalNav, a controller built around a signed, action-conditioned transition graph over identified state coordinates. At deployment CausalNav simulates a small library of intervention sequences, converts their objective error into policy-logit advice, and admits that advice only when a scale-free predictive-reliability certificate, a policy-margin gate, and an argmax-agreement gate all pass; otherwise it falls back exactly to its own model-based base controller. We evaluate against nine controlled baselines (transformer, recurrent, split-latent, graph, causal-induction, and three recent model-based reasoning modules) on CartPole-v1 and discretized Pendulum-v1 with physical-parameter shifts, under one shared PPO trainer, one interaction budget, and ten held-out seeds (200 runs). CausalNav attains the best average rank (1.25 of ten). The diagnostic result is more informative than the ranking: the learned graph recovers structure well above chance (CartPole F1 = 0.59 +/- 0.09), yet per-seed structural fidelity is uncorrelated with per-seed control benefit (r = -0.15, p = 0.67), and the certificate abstains on 10/10 Pendulum seeds, where forcing the planner on costs return. Model fidelity did not predict downstream control utility in our setting; certified abstention, not better prediction, is what made the world model safe to deploy.

While many algorithms blend reinforcement learning (RL) with counterfactual regret (CFR) methods to leverage tradeoffs in computational speed and performance, there are fewer investigations into generative sampling frameworks in game theoretic applications in incomplete information games. We extend a generative flow network framework, Adversarial Flow Networks (AFlowNets), to incomplete information games, called Information Flow Networks (IFNs). We prove that previously established constraints for generative flow networks in complete information games are inadmissible for obtaining valid densities (corresponding to player strategies) and a valid training objective. We show that our proposed generalization, IFlowNets, alleviates this issue and strictly generalizes AFlowNets. In preliminary results for three standard game environments, IFlowNets perform comparably to or better than Outcome Sampling Monte Carlo Counterfactual Regret (OSMCCFR) and standard RL-based methods in performance and speed.

Counterfactual inference approaches for sequential decision-making typically assume deterministic causal models, where all randomness stems from latent variables. However, Markov Decision Processes (MDPs) are inherently stochastic. We address this by formalising counterfactual policy optimisation under probabilistic nondeterministic causal models, which properly separates latent confounding from irreducible stochasticity, and here propose a first practical optimisation problem for identifying robust counterfactual policies under a sensitivity analysis framework. We validate our approach on a sepsis treatment simulator, where diabetes status acts as a hidden global confounder.

Algorithmic recourse aims to provide individuals with actionable changes to improve their predicted outcomes in high-stakes classification settings, such as loan and mortgage applications. However, most existing approaches focus only on flipping a model's prediction, without accounting for whether the recommended changes lead to genuine improvement in an individual's true qualifications or merely enable strategic gaming of the classifier. Consequently, deployed recourse policies can induce behavioral responses that degrade predictive accuracy and become ineffective after model retraining. In this work, we formalize this failure mode through a causal performative framework for recourse. We model how recourse actions propagate through a structural causal model, capturing interactions among features as well as their effect on the true label. These causal responses induce a non-convex optimization problem, even under standard convex losses. We characterize conditions under which performatively stable solutions exist and can be efficiently computed via simple iterative dynamics. Our analysis reveals that recourse policies that ignore causal structure can induce large, misaligned behavioral responses, whereas causal recourse leads to stable equilibria that reduce incentives for gaming. Experiments on both semi-synthetic and real credit datasets demonstrate that our approach consistently outperforms standard empirical risk minimization while reducing the need for repeated model retraining to accommodate distribution shifts caused by strategic agent behavior.

Markov Decision Processes (MDPs) are widely used as decision-making models, commonly specified over factored state spaces through state variables and their valuations. The exponential blowup in the number of states renders many reasoning tasks in MDPs challenging. Abstractions are promising techniques to reduce MDPs and thus mitigate scalability issues. In this work, we introduce a notion of causality on factored MDPs and a novel property-driven causal abstraction technique that retains many characteristics of the original MDP model. For this, we rely on causal relations over state variable predicates and identify those states that share the same reasons for fulfilling or violating a given abstraction property. We theoretically and empirically compare various causal MDP abstractions using different model types such as MDPs, interval MDPs, or stochastic games. Our evaluation demonstrates the potential of our approach: For several standard benchmarks, we obtain small abstractions that allow us to compute near-optimal policies for the original MDP. Furthermore, our causal abstractions often generalize to related large-scale MDP models.

In model-based reinforcement learning, world models exist as internal simulators, but their training often conflates statistical correlations with causal mechanisms. This problem is exacerbated in multi-agent systems where physical transitions are intertwined with strategic agent intents, causing world models to fail under distribution shift. We introduce Implicit Causal World Models to recover environmental dynamics from offline demonstrations without requiring pre-defined causal graphs. By incorporating policy variance, we render world models discoverable via the sequential backdoor condition. Evaluations across coordination tasks (Two-Door, Navigation, and Giveway) demonstrate that these models provide interpretable causal representations under both full and partial observability, with model accuracy scaling directly with interventional strength.

The Credit Assignment Problem (CAP) is fundamental to developing efficient and explainable Reinforcement Learning (RL) agents. Existing frameworks, whether relying on temporal contiguity or hindsight-conditioned reward reweighting, frequently fail to attribute properly between an agent's policy (skill) and environmental stochasticity (luck). A principled approach to CAP must isolate the true causal drivers of observed outcomes from spurious correlations and environmental randomness. We introduce Counterfactual Shapley Credit Assignment, a novel framework grounded in causal theory that attributes credit and blame via the Counterfactual Shapley Value ($φ$-value). By redistributing environmental rewards, $φ$-values enhance temporal credit assignment across three critical dimensions: sparse causality, high stochasticity, and delayed rewards, all while preserving the optimal policy. We derive a consistent estimator that computes $φ$-values efficiently, enabling a new class of policy gradient methods, $φ$-PPO, combined with Prioritized Trajectory Replay (PTR). Empirical results demonstrate that $φ$-values align precisely to the ground truth causes of task rewards with superior sample efficiency in challenging environments where prior state-of-the-art methods fail to converge.

We introduce COAT (Counterfactual Optimal Action Tree), a framework for learning interpretable prescriptive policies from observational data. COAT combines counterfactual outcome estimation with large-scale mixed-integer optimization, using column generation to translate causal predictions into feasible, transparent decisions under business and regulatory constraints. We apply COAT to airline ancillary pricing, a setting characterized by complex business rules and limited experimental flexibility. In a 17-week field pilot with a major global airline, COAT increased upsell revenue per booking by 6.9%, with the airline projecting $50-$150 million in incremental annual premium seat revenue across eligible domestic markets. The success of the pilot led to scaled adoption and informed broader AI-driven decision initiatives within the organization.

Long-horizon failure in world models is conventionally attributed to compounding error, a generic framing that does not distinguish what kind of error compounds. We propose a kinematic-vs-dynamic reframing: world models tend to imagine kinematically rather than dynamically. We operationalize this as the imagined Kinematic-Consistency Error, a per-step diagnostic that measures how far a rollout departs from a closed-form kinematic null, paired with a perturbation protocol that tests whether iKCE responds when physical conditions cross a regime boundary. We instantiate the diagnostic on a released DreamerV3 checkpoint trained on DMC walker-walk, where imagined iKCE runs roughly two orders of magnitude above that of matched real-physics rollouts. Across a friction sweep that crosses the gait-collapse boundary, the model's iKCE stays statistically flat even as the trained policy's reward collapses through the same range, providing the kinematic-not-dynamic signature. The diagnostic distinguishes kinematic from dynamic imagination at horizons longer than the embodiment's gait period.

Recent advances in reinforcement learning have produced superhuman agents across a wide range of competitive games. As a byproduct, researchers have begun studying how these agents play, extracting behavioral representations, analyzing decision structure, and modeling the latent geometry of expert performance. However, this growing body of work has overwhelmingly focused on defeating human players rather than providing feedback, leaving a critical gap in creating model solutions to improve human players. Unlike chess and Go, where AI has become integral to player training, real-time strategy (RTS) games lack principled frameworks for translating expert knowledge into actionable feedback. We introduce Latent Maps of Performance, a framework for counterfactual path generation. We focus on StarCraft~II data to model player improvement as an algorithmic recourse within a learned representation space. As inspiration for our work, we have looked at the championship model used in sports science. We trained a Guided Variational Autoencoder model on 23,305 professional tournament replays, enabling counterfactual traversal between losing and winning gameplay profiles. To fulfill our goal, we have devised and verified four traversal strategies on out-of-distribution (OOD) data randomly sampled from a dataset of amateur replays, namely linear interpolation, iterative optimal transport, density-regularized gradient ascent, and neural flow matching, each designed to generate multi-step improvement trajectories that remain grounded in observed expert behavior while moving a player's profile toward winning configurations. Feedback is extracted at multiple granularities to support players at different stages of improvement. Finally, we conclude that there is a trade-off between the path-finding methods we employ and hope that future research will focus on developing model solutions for human improvement.

Causal inference provides a set of principles and tools that allow one to combine data and knowledge about an environment to reason with questions of counterfactual nature, i.e., what would have happened had reality been different, even when no data of this unrealized reality is currently available. Reinforcement learning provides methods to learn a policy that optimizes a specific measure (e.g., reward, regret) when the agent is deployed in an environment and pursues an exploratory, trial-and-error approach. These two disciplines have evolved independently and with virtually no interaction between them. We note that they operate over different aspects of the same building block, counterfactual relations, which makes them umbilically connected. Based on these observations, novel learning opportunities arise when this connection is explicitly acknowledged and mathematized. To realize this potential, we note that any environment where the RL agent is deployed can be decomposed as a collection of autonomous mechanisms with different causal invariances, parsimoniously modeled as a structural causal model; any standard RL setting implicitly encodes such a model. This formalization allows us to put under a unifying treatment different modes of learning, including online, off-policy, and causal calculus learning, which appear unrelated in the literature. However, these modalities are not exhaustive: we introduce several natural and pervasive classes of learning settings that entail novel dimensions of analysis. Specifically, we introduce and discuss through causal lenses generalized policy learning, where to intervene, imitation learning, and counterfactual learning. These tasks lead to a broader view of counterfactual learning and suggest great potential for studying causal inference and reinforcement learning side by side, which we call causal reinforcement learning (CRL).

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal reasoning. However, prevailing reinforcement learning (RL) paradigms lack explicit counterfactual enhancement and causal learning mechanisms. This fundamental deficiency results in severe grounding failures, manifesting as a tendency to ignore visual evidence in favor of language priors or exhibiting hallucination drift during long chain-of-thought reasoning. To address this root cause, we propose CounterFactual Policy Optimization (CFPO), a novel framework that enforces causal consistency between visual perception and textual reasoning. CFPO introduces a cross-modal counterfactual enhancement mechanism, which regularizes the policy by maximizing the discrepancy between the model's predictions and those from a counterfactual state where critical visual cues are suppressed. This approach seamlessly integrates with standard algorithms like GRPO and DAPO without requiring external reward models or additional supervision. Extensive experiments demonstrate that CFPO significantly improves reasoning fidelity, achieving consistent gains of 3.17%-6.25% over standard RL baselines and 1.32%-2.13% over the state-of-the-art perception-aware method (PAPO). Code is available at https://github.com/Raven-July/CFPO.

Although pitch sequencing is a central topic in baseball analytics, previous studies have primarily focused on optimizing the final pitch within a single plate appearance, leaving the role of preceding setup pitches and their impact on long-term season-level performance insufficiently examined. To address these issues, this study conducted counterfactual analyses using MLB Statcast data. A Transformer-based machine-learning model was trained to predict whether a target pitch would result in an in-play outcome or swing-out. Counterfactual pitch sequences were then generated by replacing either the final pitch or the preceding setup pitch with alternative pitch types and locations while keeping the surrounding contextual information fixed. Optimal counterfactual selections were defined as those that minimized the predicted in-play probability, and their expected effects on pitchers' seasonal statistics were estimated using regression models linking model outputs to season statistics. The results suggest that the optimization of both final and setup pitches may substantially influence season-level performance, including improvements of more than 1.0 in K/9. The analyses also provided several practical insights, including velocity-band-specific effective locations, the importance of pitch commands, and the expansion of pitch-selection options through middle-velocity pitches. These findings quantitatively support the strategic importance of pitch sequencing in baseball.

Reward models trained from pairwise preferences often exploit superficial shortcut cues rather than learning true response quality. We propose DynaCF, a dynamic reweighting framework for mitigating shortcut learning in reward model training. Unlike static shortcut heuristics, DynaCF measures shortcut sensitivity online during optimization by applying semantics-preserving counterfactual perturbations and tracking the resulting margin shifts and preference flips under the current model. Samples with higher shortcut sensitivity are dynamically downweighted in the Bradley-Terry objective, encouraging the model to rely less on superficial patterns and more on task-relevant preference signals. Extensive experiments show that DynaCF consistently improves robustness in preference modeling.

Many theories of decision making -- planning, reinforcement learning, causal intervention, online learning, and game-theoretic equilibrium -- turn local information into globally coherent behavior. This paper proposes a common categorical formulation: a Universal Decision Learner (UDL) extends a partially specified decision functor from observed contexts to new contexts by a pair of universal constructions. Left Kan extensions express rollout, aggregation, and candidate generation; right Kan extensions express consistency, constraint satisfaction, and fixed-point semantics. The central claim is not that every decision problem has the same algorithm, but that many decision formalisms instantiate the same universal problem: extend local behavioral data canonically, then characterize the globally coherent extensions. We give the abstract UDL construction, prove its universal comparison property, define Kan-invariant behavioral equivalence and minimal abstractions, and show how Bellman equations, planning recursions, causal interventions, online regret, and equilibria arise as special cases. The supplementary material develops the reinforcement-learning specialization in more detail.

Epidemic forecasting faces a fundamental challenge: human behavior dynamically responds to disease spread, creating feedback loops that induce distribution shifts at policy intervention points. This renders data-driven models unreliable under distribution shift. We propose \textbf{SL-BiLEM} (Structured Learnable Behavior-in-the-Loop Epidemic Model), leveraging physical constraints as regularization for robust extrapolation. The framework decomposes effective transmission as $β_{\text{eff}}(t,g) = β_0(g) \times m_{\text{policy}}(t) \times m_{\text{media}}(t) \times m_{\text{comp}}(t,g)$, where monotonicity, smoothness, and bounded-jump constraints on the learned compliance function maintain predictive validity under novel policy regimes. Beyond forecasting, SL-BiLEM enables counterfactual analysis for intervention decision support. We validate forecasting on three real-world datasets (cruise ship, school influenza, and school-district COVID-19 surveillance) and evaluate counterfactual recovery on synthetic benchmarks with known ground truth. SL-BiLEM demonstrates: (1) 76% improvement over neural-mechanistic baselines, with only 53% OOD degradation versus 1142% for neural baselines under policy-induced shift; (2) 100% bootstrap CI coverage across 27 synthetic counterfactual experiments; and (3) Treatment Effect Accuracy exceeding 0.85. These results establish SL-BiLEM as an interpretable tool for public health decision-makers seeking accurate prediction and principled intervention planning.

World models for partially observed environments must imagine multiple compatible hidden futures and steer between them under counterfactual actions. Joint Embedding Predictive Architectures (JEPAs) do this in latent space, but a vector-valued latent has no internal structure for carrying the belief over hidden continuations through blind rollout. We introduce the Unitary World Model JEPA (UWM-JEPA), a JEPA world model with a density-matrix latent on a joint system-environment space and a learned unitary predictor. The construction preserves the joint-state spectrum exactly during rollout, so the predictor itself cannot dissipate the represented uncertainty. On a hidden-velocity indicator task requiring five-step forward simulation under a given action sequence with the target observation masked, UWM-JEPA reaches 0.77 accuracy and degrades monotonically as actions are perturbed; a parameter-matched LSTM-JEPA trained under the same counterfactual-target objective and action head collapses to majority-class accuracy (0.53) under every action condition. Under blind rollout, UWM-JEPA loses fewer than ten points of probe R^2 at short horizons while vector-latent baselines lose forty-one and sixty-eight; both nevertheless tie on a held-out context probe, locating the separation in the predictor rather than the encoder. Action sensitivity itself requires training against counterfactual rather than teacher-forced targets, a finding that applies beyond the unitary parameterisation. For JEPA world models to imagine under partial observability, latent geometry and predictor dynamics matter, not frozen context-encoding capacity alone.

We give an attribution method for neural combinatorial-optimisation (CO) policies that (i) decomposes a decision by constraint families via LP-relaxation duals, (ii) certifies counterfactuals through a combinatorial feasibility model (implemented as a CSP feasibility-decision model), and (iii) bounds the size of a PAC-sufficient explanation with a Bonferroni-corrected Hoeffding sufficient-subset test along a greedy ordering. Across three CO problems and three seeds, our LP-anchored $Λ$-attribution matches the CF-derived signal at 96.5% on CVRPTW (n_cert=344) and 77.2% on the Orienteering Problem (n_cert=281) vs 75.0% and 35.2% for proxy gradient (paired diffs +0.215 and +0.420; McNemar exact $p \le 10^{-14}$). In the rank-aligned regime of the Flexible Job-Shop Scheduling Problem, both backends agree on every CSP-certified flip (n_cert=59), confirming the no-gain prediction. Bonferroni-PAC subsets average 5.0 nodes per step ($M=70$, $\varepsilon=δ=0.2$, $k_{\max}=25$). Reference implementation: https://github.com/sohaibafifi/neuro-co-cax

Reinforcement learning algorithms are generally designed to maximize the expected return across a population. However, a policy that is optimal on average may be suboptimal for certain individuals, leading to potential safety concerns. To address this, we first formalize the notion of individual harm from a counterfactual perspective and define harm as the event in which a chosen action results in a strictly worse outcome than a baseline alternative. We then propose a general two-stage procedure for learning policies that maximize the expected return while accounting for individual harm. We further establish the finite-sample properties of the learned policy, derive an upper bound on its sub-optimality gap, and show that the harm rate remains well-controlled. Numerical experiments on both simulated and real-world datasets demonstrate the effectiveness of the proposed approach.

Reasoning language models do not distinguish tokens used for computation from tokens that constitute persistent state: once generated, all hidden thoughts remain in context and influence future predictions. As a result, downstream reasoning may depend on failed attempts, dead ends, and private scratch work that should not be safely relied on later. We recast this phenomenon as a new training objective, state commitment learning: training models to explicitly distinguish information that should be committed as persistent state from temporary computation that can be discarded. We define a counterfactual criterion, persistent-state sufficiency, which makes it trainable and measurable whether an answer remains usable after hidden thoughts are erased. We then propose Counterfactual Erasure RL (CERL), which evaluates, under the same prefix, both a path that keeps hidden thoughts and a path that erases them, and gives reward only when the erasure path remains correct. We also introduce the Erasure Dependence Protocol and show across mathematics, long-chain logic, scientific QA, and multi-turn tool-use evaluation that CERL substantially reduces answer dependence on hidden thoughts without sacrificing accuracy, consistently outperforming correctness-only RL and long-answer SFT baselines.

Video large language models (Video LLMs) achieve strong benchmark accuracy, yet often answer video questions through shortcuts such as single-frame cues and language priors rather than by tracking spatiotemporal dynamics. This issue is exacerbated in RL post-training, where correctness-only rewards can further reinforce shortcut policies that obtain high reward without tracking video dynamics. We address this by asking a controlled counterfactual question: if the visual world changed while the question remained fixed, should the answer change or stay the same? Based on this view, we propose \textbf{Counterfactual Relational Policy Optimization (CRPO)}, a dual-branch RL framework for improving \emph{spatiotemporal sensitivity}. CRPO constructs counterfactual videos through horizontal flips and temporal reversals, trains on both original and counterfactual branches, and introduces a \textbf{Counterfactual Relation Reward (CRR)} between their answers. CRR encourages answers to change for dynamic questions and remain unchanged for static questions. This cross-branch constraint makes it difficult for shortcut policies to be consistently rewarded across both branches. To evaluate this property, we introduce \textbf{DyBench}, a paired counterfactual video benchmark with 3,014 videos covering reversible dynamics, moving direction, and event sequence, together with a strict pair-accuracy metric that prevents fixed-answer shortcuts from inflating scores. Experiments show that CRPO outperforms prior RL methods on spatiotemporal-sensitive evaluations while maintaining competitive general video performance. On Qwen3-VL-8B, CRPO improves DyBench P-Acc by +7.7 and TimeBlind I-Acc by +8.2 over the base model, indicating improved spatiotemporal sensitivity rather than stronger reliance on static shortcuts. The project website can be found at https://ddz16.github.io/crpo.github.io/ .

Vision-Language-Action (VLA) policies increasingly rely on asynchronous inference to hide large-model latency behind ongoing robot motion. While this avoids the stop-and-go behavior of synchronous action-chunk execution, it creates a prediction-execution mismatch: the next chunk is computed from a stale observation at inference start but executed only after the robot and scene have evolved. As a result, actions that fit the prediction-time state can become misaligned with the execution-time state. Existing runtime repair, behavior-cloning, and preference-alignment approaches do not directly teach the policy to resolve this stale-input mismatch. We propose DEFLECT, an offline post-training framework for delay-robust asynchronous VLAs. DEFLECT converts latency-induced mismatch into counterfactual preference supervision: a frozen reference VLA generates a preferred chunk from the future execution-time observation and a rejected chunk from the stale prediction-time observation. The trainable policy scores both chunks under the same deployment-time input, learning to favor execution-time-aligned actions while a supervised fine-tuning anchor preserves the expert action manifold. DEFLECT requires no human preference labels, reward models, online robot rollouts, architectural changes, or additional inference-time computation. Across Kinetix, LIBERO, and three real-robot tasks, DEFLECT improves delay robustness over strong asynchronous VLA baselines, raising high-latency success by up to 6.4 percentage points and achieving a 4.6 percentage-point gain at the longest delay on a real-scale VLA.

We formulate machine unlearning for online L-BFGS as a counterfactual state-alignment problem. Given an actual event stream and a deletion-edited counterfactual stream, the target of unlearning is the optimizer state that would have arisen had the deleted samples never been processed. We introduce state-aware metrics that separately measure parameter error, memory-operator error, combined state error, and update-direction error. The memory metric compares the inverse-Hessian actions induced by the o-L-BFGS memory, rather than treating curvature pairs as of finite influence. Under convexity assumptions, we derive a recursive bound on counterfactual state deviation. We then evaluate a state-aware benchmark of deletion interventions, including memory-only and parameter-only corrections, against an counterfactual oracle model. These results show that unlearning for online L-BFGS is not merely a parameter-correction problem: it requires alignment with a realizable counterfactual optimizer state.

Large language models (LLMs) have become a strong foundation for multi-agent systems, but their effectiveness depends heavily on orchestration design. Across different tasks, role design, capacity assignment, and dependency construction jointly affect both solution quality and execution efficiency. Existing approaches automate parts of this design process, yet they often optimize these decisions partially or sequentially, and rely on execution-level feedback that provides limited credit assignment for local orchestration decisions. We propose LEMON (\textbf{L}earning \textbf{E}xecutable \textbf{M}ulti-agent \textbf{O}rchestratio\textbf{N} via Counterfactual Reinforcement Learning), an LLM-based orchestrator that generates an executable orchestration specification. The specification integrates task-specific roles, customized duties, capacity levels, and dependency structure into a single deployable system. To train the orchestrator, we augment the orchestration-level GRPO objective with a localized counterfactual signal that edits role, capacity, or dependency fields and applies the resulting reward contrast only to the edited spans. Experiments on six reasoning and coding benchmarks, including MMLU, GSM8K, AQuA, MultiArith, SVAMP, and HumanEval, show that LEMON achieves state-of-the-art performance among the evaluated multi-agent orchestration methods. Our code is available at https://anonymous.4open.science/r/LEMON-B23C.

Parallelization has played an instrumental role in the field of artificial intelligence (AI), drastically reducing the time taken to train and evaluate large AI models. In contrast to its impact in the broader field of AI, applying parallelization to computational game solving is relatively unexplored, despite its great potential. In this paper, we parallelize the family of counterfactual regret minimization (CFR) algorithms, which were central to important breakthroughs for solving large imperfect-information games. We present a generalized parallelization framework, reframing CFR as a series of linear algebra operations. Then, existing techniques for parallelizing linear algebra operations can be applied to accelerate CFR. We also describe how our technique can be applied to other tabular members of the CFR family of algorithms, including the state-of-the-art, such as CFR+, discounted CFR, and predictive variants of CFR. Experimentally, we show that our CFR implementation on a GPU is up to four orders of magnitude faster than Google DeepMind OpenSpiel's CFR implementations on a CPU.

Adaptive test-time compute for LLM agents aims to invoke extra computation only when it improves performance. Existing methods typically use confidence-, uncertainty-, or difficulty-based gates, assuming a fixed direction from the gating signal through compute need to the value of computation. This makes gating a utility-calibration problem: gating signals should align with whether extra computation improves the final outcome over the base policy. We show that this alignment is unstable: the same signal predicts rollout benefit in one setting and rollout harm in another, with reversals across environments and backbones even when the task is fixed. Wrong-direction gates can therefore worsen performance by precisely selecting harmful states. This reversal reflects a deeper distinction between compute need and compute suitability: a high uncertainty signal may indicate decision-difficult states where rollouts help compare alternatives, or intervention-unsuitable states where the current context does not support useful rollout-based improvement. Under this two-source model, fixed-direction gates are unreliable across heterogeneous settings. To address this, we propose DIAL (Direction-Informed Adaptive Learning), a sparse gate trained from signal-agnostic counterfactual exploration to learn the utility direction of state features per (environment, backbone). Across six environments and three backbones, DIAL yields a stronger overall success-cost trade-off than fixed-direction baselines.

When a machine learning (ML) model forecasts an undesired event, one often seeks a decision to avoid it, known as the avoiding undesired future (AUF) problem. Many rehearsal learning methods have been proposed for AUF, but they rely on an underlying graph structure; learning such a graph from observational data is challenging and can incur substantial estimation error. In this work, we demonstrate that the order structure can be sufficient for AUF decision-making, and propose the first order-based rehearsal learning method. Although an order is less informative than a graph, it can be sufficient to identify the influence of decisions from observational data, suggesting that learning the entire graph is not always necessary. To learn the order, we develop an information-theoretic method that imposes no restrictions on the form of structural functions or the type of noise distributions. For AUF decision-making, we construct an order-based sampler to approximate the influence of decisions and, combined with a surrogate objective for maximizing the post-decision success probability, reduce the AUF task to a differentiable optimization problem. Experiments show that our order learning method outperforms existing methods, and that our AUF approach not only surpasses methods relying on learned graphs or learned orders, but also matches or even exceeds oracle baselines that are given the true graph.

Open-loop imitation learning has advanced modern autonomous driving policy architectures, but closed-loop deployment remains vulnerable to policy-induced distribution shift. Existing post-training paradigms exhibit fundamental trade-offs: closed-loop RL fine-tuning provides grounded feedback from executed actions but is constrained by the sparsity of informative events, whereas counterfactual fine-tuning provides dense supervision over candidate futures but inherits bias from imperfect future estimates. We introduce Counterfactual-to-Interactive Reinforcement Fine-Tuning (CRAFT), an on-policy framework that formulates closed-loop post-training as proxy-residual optimization. CRAFT uses group-normalized counterfactual advantages as a dense proxy for real closed-loop advantages and aligns this proxy with the closed-loop world through grounded residual correction from interaction-critical events. To stabilize adaptation, CRAFT regularizes the online policy toward an EMA teacher via asymmetric KL self-distillation. Theoretically, CRAFT decomposes the real closed-loop policy gradient into proxy and residual terms under the same visited-state distribution, reducing residual variance with an aligned proxy while mitigating proxy bias through grounded residual approximation. Empirically, CRAFT achieves the strongest closed-loop gains on Bench2Drive across hierarchical planning, vision-language-action, and vocabulary-scoring architectures. Ablations, scaling behavior, stability analyses, and transfer results further validate the complementary roles of dense counterfactual proxy and grounded residual correction. Project page: https://currychen77.github.io/CRAFT.

A world model matters to an agent only through the state it constructs. That state must preserve some information, discard other information, and support some future function: prediction, control, planning, memory, grounding, or counterfactual reasoning. This paper treats world-model research as latent state design under sufficiency constraints. We propose a functional taxonomy that groups methods by what their latent state is for, rather than by architecture or application domain: predictive embedding, recurrent belief state, object/causal structure, latent action interface, grounded planning interface, and memory substrate. These roles expose distinctions that architecture-based groupings hide, including the gap between predictive sufficiency and control sufficiency, and the gap between passive video prediction and counterfactual action modeling. The taxonomy supports an evaluation framework that judges a model by the sufficiency constraint its latent state was built to satisfy. We compare methods along seven axes: representation, prediction, planning, controllability, causal/counterfactual support, memory, and uncertainty. We use the resulting matrix as a diagnostic for what a latent state preserves, discards, and enables. The conclusion that follows is that an actionable world model is the one whose state construction matches the task, not the one that preserves the most information.

In cooperative teams where agents act in a fixed order and share a single team-level reward (multi-agent language systems, sequential robotic tasks), per-agent credit assignment is under-determined. Critic-based approaches scale poorly as the number of agents grows owing to the costly maintenance of joint/factored critic(s), whereas the existing critic-free alternatives have other issues: common credit across agents that couples every agent's signal to teammate noise, importance-sampling corrections for upstream-update staleness that incur variance exponential in team size, or per-agent counterfactual replay that isolates each agent's effect at the price of extra environment or reward calls. We propose COSAC, a critic-free per-agent policy gradient for sequential cooperative teams. COSAC fits an additive per-agent decomposition of the team reward by a single ridge regression on the rollout batch (giving each agent a learning signal decoupled from teammate noise), and computes each agent's counterfactual advantage from fictitious continuations of the current policy (policy forward passes that replace both importance-sampling reweighting and per-agent environment replay, at no extra environment or reward cost). The estimator instantiates the Sequential Aristocrat Utility (SeqAU), our extension of Wolpert and Tumer's (2001) aristocrat utility to sequential teams. We prove bias and variance bounds on SeqAU credits that stay controlled as the team grows. Our controlled study on sequential bandits demonstrates that COSAC attains the lowest advantage MSE and consistently low learning regret across team sizes up to $K = 16$. On the AI2 Reasoning Challenge (ARC) task, where four Qwen3-0.6B agents reason in turn about a grade-school science question, COSAC attains faster convergence than the other critic-free baselines.

Machine unlearning has gained increasing attention in recent years, as a promising technique to selectively remove unwanted privacy or copyrighted information from Large Language Models that are trained on a massive scale of human data. However, the emergence of Large Reasoning Models (LRMs), which emphasize long chain-of-thought (CoT) reasoning to address complex questions, presents a dilemma to unlearning: existing methods either struggle to completely eliminate undesired knowledge from the CoT traces or degrade the reasoning performances due to the interference with the reasoning process. To this end, we introduce Counterfactual Unlearning through iterative Preference Optimization (CiPO), a novel framework that redefines unlearning as the targeted intervention of the CoT reasoning in LRMs. More specifically, given a desired unlearning target answer, CiPO instructs LRMs to generate a logically valid counterfactual reasoning trace for preference tuning. As the LRM adjusts to the counterfactual trace, CiPO iteratively updates the preference learning data to increase the discrepancy from the original model. This iterative loop ensures both desirable unlearning and smooth optimization, effectively mitigating the dilemma. Experiments on challenging benchmarks demonstrate that CiPO excels at unlearning, completely removing knowledge from both the intermediate CoT steps and the final answer, while preserving the reasoning abilities of LRMs.

Reinforcement Fine-Tuning (RFT) has established itself as a critical paradigm for the alignment of Multi-modal Large Language Models (MLLMs) with complex human values and domain-specific requirements. Nevertheless, current research primarily focuses on mitigating exogenous distribution shifts arising from data-centric factors, the non-stationarity inherent in the endogenous reasoning remains largely unexplored. In this work, a critical vulnerability is revealed within MLLMs: they are highly susceptible to endogenous reasoning drift, across both thinking and perception perspectives. It manifests as unpredictable distribution changes that emerge spontaneously during the autoregressive generation process, independent of external environmental perturbations. To adapt it, we first theoretically define endogenous reasoning drift within the RFT of MLLMs as the multi-modal concept drift. In this context, this paper proposes Counterfactual Preference Optimization ++ (CPO++), a comprehensive and autonomous framework adapted to the multi-modal concept drift. It integrates counterfactual reasoning with domain knowledge to execute controlled perturbations across thinking and perception, employing preference optimization to disentangle spurious correlations. Extensive empirical evaluations across two highly dynamic and safety-critical domains: medical diagnosis and autonomous driving. They demonstrate that the proposed framework achieves superior performance in reasoning coherence, decision-making precision, and inherent robustness against extreme interference. The methodology also exhibits exceptional zero-shot cross-domain generalization, providing a principled foundation for reliable multi-modal reasoning in safety-critical applications.

Imperfect Information Games (IIGs) are used to model games under uncertainty or lack complete information. Counterfactual Regret Minimization (CFR) is one of the most successful families of algorithms for IIGs. The integration of skill-based strategy learning with CFR could potentially mirror more human-like decision-making and improve learning on complex IIGs. It enables the learning of a hierarchical strategy, wherein low-level components represent skills for solving subgames and the high-level component manages the transition between skills. In this paper, we introduce the first hierarchical version of Deep CFR (HDCFR), an innovative method that boosts learning efficiency in tasks involving extensively large state spaces and deep game trees. Notably, HDCFR enables learning with predefined (human) expertise and extracting skills transferable to similar tasks. We first present the algorithm and establish its theory in a tabular setting, including hierarchical CFR update rules and a variance-reduced Monte Carlo sampling extension for the model-free setting, where backtracking is infeasible. We then extend HDCFR to large-scale tasks via deep learning objectives that match the tabular targets under exact function fitting. Code: https://anonymous.4open.science/r/HDCFR_RUN-677B.