Classical reinforcement learning (RL) typically seeks a deterministic policy that maximizes the expected sum of a scalar reward. Yet, modern applications such as language model fine-tuning or scientific discovery demand diversity. Existing remedies such as entropy regularization or diversity bonuses often require fragile trade-offs that sacrifice performance for stochasticity or rely on heuristic metrics that can misalign policy rankings. We argue that diversity is more naturally understood as the rational response to uncertainty in the reward. When the reward function is not perfectly known–as is the case with ambiguous preferences or imperfect reward models–committing to a single action can be sub-optimal. Building on this, we propose a fundamental reformulation of the RL objective by replacing the scalar reward with a distribution over reward functions, and applying a non-linear objective over sets of actions. The result is a framework in which calibrated behavioural diversity emerges naturally, remains controllable through the reward function distribution, and is obtained without sacrificing expected reward. Focusing on the contextual bandit setting, we derive a principled gradient estimator for this objective and prove that our formulation naturally generalizes both vanilla policy gradient and more recently developed action-set approaches. Our empirical results demonstrate that this framework offers a robust and theoretically grounded alternative for complex RL tasks where the traditional formulation of the problem fails to induce the desired breadth of agent behaviour.

Supervised fine-tuning (SFT) improves in-domain performance but can degrade out-of-domain (OOD) generalization. Prior work suggests that this degradation is related to changes in dominant singular subspaces of pretrained weight matrices. However, directly identifying loss-sensitive directions with Hessian or Fisher information is computationally expensive at LLM scale. In this work, we propose preserving projected rotations in pretrained singular subspaces as an efficient proxy for Fisher-sensitive directions, which we call Rotation-Preserving Supervised Fine-Tuning (RPSFT). RPSFT penalizes changes in the projected top-$k$ singular-vector block of each pretrained weight matrix, limiting unnecessary rotation while preserving task adaptation. Across model families and sizes trained on math reasoning data, RPSFT improves the in-domain/OOD trade-off over standard SFT and strong SFT baselines, better preserves pretrained representations, and provides stronger initializations for downstream RL fine-tuning. Code is available at \href{https://github.com/jinhangzhan/RPSFT.git}{https://github.com/jinhangzhan/RPSFT}.

The Continual Learning (CL) literature has long been driven by the goal of mitigating catastrophic forgetting. This objective rests on a pervasive, often unstated assumption: that a lifelong learner should approximate the Joint-Task Learning (JTL) solution and retain all previously acquired knowledge. We challenge this retention-centered premise, arguing that in non-stationary environments prioritizing retention can impede real-time adaptation. Shifting the focus to the Average Lifelong Error (ALE), we formalize CL as an online optimization problem governed by the interaction between environmental and learning dynamics. We introduce Transfer Efficiency as a quantitative measure of the tension between Instability, the bias inherited from conflicting past experience, and Transient Error, the optimization cost of learning new tasks from scratch. Under mild convergence conditions, holding across linear and neural network models, this decomposition yields a Critical Task Duration: a closed-form threshold beyond which historical knowledge transitions from a warm-start advantage to an optimization liability whenever retention induces a positive stationary bias. We validate these theoretical predictions on continual image classification and reinforcement learning benchmarks. Finally, by connecting continual learning to the online learning framework of predictable sequences, we show that JTL is only one instance of a broader family of objectives, and we propose a new general class of continual learning algorithms, which we call Predictive Continual Learning. Predictive CL algorithms optimize expected future performance under an explicit, dynamically updated model of future tasks. As a proof of concept, we analyze a Window algorithm that interpolates between JTL and Independent-Task Learning (ITL), outperforming both under controlled distributional drift.

A long-standing finding in the causal learning literature is that adults struggle to identify conjunctive causal rules, where an effect requires the simultaneous presence of multiple causes, while performing better in disjunctive settings. However, most demonstrations of this “conjunctive handicap” rely on passive observation paradigms with limited evidence, where learners have no control over evidence generation. This paper asks whether this bias persists when adults are granted agency through active exploration. Using a modified “blicket detector” task, adult participants freely intervened to identify causal objects under conjunctive or disjunctive rule structures. We show that active exploration substantially improves adults' conjunctive causal reasoning, although conjunctive rules still require more tests to infer than disjunctive rules. We further compare human performance to a range of large language models in the same setting. While some state-of-the-art models approach human-level performance on hypothesis inference accuracy, they often exhibit less efficient exploration strategies and similar conjunctive-disjunctive performance gaps.

Effective model-based reinforcement learning in stochastic environments requires planning that accounts for predictive uncertainty. Propagating full state distributions analytically offers a principled way to do this, but has traditionally required restrictive policy or reward structures to remain tractable. Consequently, modern deep reinforcement learning has largely retreated to either stochastic sampling, which introduces significant target variance, or deterministic point estimates that ignore predictive covariance entirely. We investigate whether distribution-aware planning is possible without these constraints. Using a quadratic action-value parameterization, we first reduce the Bellman backup to an expectation over the state-value function alone; the key idea is then a compatibility principle between the predictive transition distribution and the value function class, under which this expectation is analytic in the distribution's moments. We instantiate this principle with a Gaussian transition model paired with a radial-basis value function, yielding a closed-form backup that propagates both predictive mean and covariance. Empirically, our approach reduces target variance and yields well-calibrated predictive uncertainty under stochastic observations in continuous control, providing a principled framework for planning with learned distribution models.

Personalized insulin therapy for individuals with Type 1 Diabetes via closed‑loop artificial pancreas systems requires rapid adaptation of dosing strategies to each patient's unique insulin response. However, learning patient‑specific policies from scratch demands extensive exploration, which is often impractical. In this work, we study a framework that integrates insulin-response-informed transfer learning with model-based reinforcement learning for insulin dosing. We first train an LSTM‑based insulin responsiveness predictor on virtual patients, using their glucose, insulin, and meal history to forecast future glucose levels. Analysis of insulin responsiveness of in-silico patients uncovers natural insulin‑response groups characterized by similar sensitivity and dynamics profiles. For a new patient, we identify a representative model from their response group and use it to generate synthetic trajectories. These trajectories are integrated into an enhanced H-step Deep Dyna-Q algorithm, enabling accelerated policy optimization through model-based planning. The dynamics model trained entirely in simulation achieves 91.31% accuracy in predicting blood glucose ranges on the Ohio Type 1 Diabetes dataset, indicating strong zero-shot generalization. Additionally, we find that bootstrapping a new patient with a physiologically-matched reference model accelerates convergence of effective dosing policies across in-silico cohorts of children, adolescents, and adults. These findings suggest that leveraging response-group-specific synthetic experience can expedite personalized insulin therapy, offering a promising pathway towards clinical validation.

Agents that can beat or tie any other under a model of pairwise preference have strong guarantees for both user satisfaction and overall social welfare. However, searching for these agents in long-term decision problems is not computationally tractable with current approaches, which require the size of an agent's policy to increase with the problem length. We introduce the \textit{Markov decision contest}, a model of learning from general preferences in long-term (infinite-horizon) decision problems. Within this model, we prove that agents only need a stationary Markov policy in order to be optimal (that is, to beat or tie any agent with a history-dependent policy); that the problem of finding an optimal policy is in P; and that a simple iterative algorithm (which we call Hedged Policy Iteration) converges to an optimal policy at a sublinear rate. In a suite of high-dimensional experiments, we demonstrate that Hedged Policy Iteration scales well to function approximation. Lastly, we present a near approximation of Hedged Policy Iteration, called HPI-Clip, which both matches the performance of Proximal Policy Optimization on reward-based tasks while also outperforming it on tasks with non-transitive preferences. These results show that learning from pairwise preferences in long-term decision problems can be far more tractable than what is known from prior work.

Background and Objectives: This study aimed at developing an AI-based predictive model for live birth based on a combination of a support vector machine (SVM) using clinical and embryological features, together with a convolutional neural network (CNN) using embryo time-lapse videos. Materials and Methods: This was a retrospective cohort analysis. Two hundred fifty-nine infertile couples treated between January 2012 and December 2019, with a total of 2330 embryos, were included in this study, and clinical data and images from 355 transferred embryos were used to build a predictive model. The main outcome was accuracy of live birth prediction. The secondary outcomes included accuracy in the prediction of biochemical pregnancy, clinical pregnancy and transferrable embryos. Results: The model was able to predict the transferrable embryo (i.e., embryos suitable for transfer or cryopreservation) with an accuracy of 0.98 in an internal set. The accuracy for predicting live birth, clinical pregnancy, and biochemical pregnancy exclusively using clinical data as input for an SVM model was 0.67, 0.68, and 0.67, respectively. With six frames from time-lapse embryo development, the CNN produced an accuracy of 0.57, 0.67, and 0.72. The predictive model performed best when combining input from clinical data and images from multiple embryo developmental frames, obtaining 0.71, 0.73, and 0.77 for predicting live birth, clinical pregnancy, and biochemical pregnancy. Conclusions: This study highlights the potential of combining clinical data and embryo development images to enhance predictive models in reproductive medicine.

Full models of the world require complex knowledge of immense detail. While pre-trained large models have been hypothesized to contain similar knowledge due to extensive pre-training on vast amounts of internet scale data, using them directly in a search procedure is inefficient and inaccurate. Conversely, partial models focus on making high quality predictions for a subset of state and actions: those linked through affordances that achieve user intents . Can we posit large models as partial world models? We provide a formal answer to this question, proving that agents achieving task-agnostic, language-conditioned intents necessarily possess predictive partial-world models informed by affordances. In the multi-task setting, we introduce distribution-robust affordances and show that partial models can be extracted to significantly improve search efficiency. Empirical evaluations in tabletop robotics tasks demonstrate that our affordance-aware partial models reduce the search branching factor and achieve higher rewards compared to full world models.

Reinforcement learning problems typically define the goal as maximizing the expected value of a scalar reward function. But, pairwise preferences are often easier to specify than scalar rewards, and they express certain goals that scalar rewards cannot. Methods for reinforcement learning with pairwise preferences have thus received growing interest. Unfortunately, these methods are inefficient in problems with long time horizons, and they lack guarantees on the performance of Markov policies relative to history-dependent policies, which bridge the theory and practice of reinforcement learning. We therefore propose the Markov decision contest as a new problem model for reinforcement learning with pairwise preferences. We prove that stationary Markov policies are optimal among all history-dependent policies, that solving a Markov decision contest exactly is in P, and that a simple iterative algorithm converges to an optimal policy at a sublinear rate. Lastly, in a set of high-dimensional decision problems with long time horizons, we show that our approximate algorithm is significantly more learning-efficient than prior work.

Pediatric serious illness communication (SIC) is critically important, yet scalable communication training for clinicians remains limited. Compared with other dialogue simulation settings, pediatric SIC poses additional challenges, including multi-party interactions, response to parental distress and strong dependence on feedback dynamics. Existing LLM-based simulators optimize generic dialogue quality rather than curriculum-contingent behavior required for effective SIC training. In collaboration with educators and pediatric clinicians, we introduce the first benchmark suite and simulation framework tailored to pediatric SIC training. Our benchmarks, PitfallBench and DialogueBench, evaluate simulators both at the turn-level and across full dialogues. We further propose SIC-Agents, a self-improving framework that generates a clinician-editable skill document to guide simulator behavior. Our experiments show that SIC-Agents outperforms static expert prompting. To support future research, we release our benchmarks for parent simulation in pediatric SIC at https://github.com/Beikewzh/sic-benchmarks

Reward models (RMs) are fundamental to aligning Large Language Models (LLMs) via human feedback, yet they often suffer from reward hacking. They tend to latch on to superficial or spurious attributes, such as response length or formatting, mistaking these cues learned from correlations in training data for the true causal drivers of quality (e.g., factuality, relevance). This occurs because standard training objectives struggle to disentangle these factors, leading to brittle RMs and misaligned policies. We introduce CROME (Causally Robust Reward Modeling), a novel framework inspired by an explicit causal model designed to mitigate reward hacking. CROME queries an oracle LLM for rubrics that are (or the oracle deems to be) causally relevant to answering a specific prompt. Then, it employs the following synthetic targeted augmentations during training: (1) Causal Augmentations, which are pairs that differ along specific causal attributes (subset of the Oracle identified rubrics), to enforce sensitivity along each causal attribute individually, and (2) Neutral Augmentations, which are tie-label pairs varying primarily in spurious attributes, to enforce invariance along spurious attributes. Notably, our neutral augmentations are produced without any knowledge of unknown spurious factors, via question swapping and response interventions only along causal rubrics. We show that the CROME augmentation strategy using rubrics from popular LLM APIs significantly outperforms standard baselines on RewardBench, improving average accuracy by up to 5.3% and achieving gains of up to 7.1% and 12.4% in reasoning and safety. The robustness of CROME is further testified by significant gains in DPO-aligned policies and Best-of-N alignment across various benchmarks, including AlpacaEval 2.0, RewardBench, safety-focused WildGuardTest, and the reasoning-specific GSM8k.

Effective decision-making in complex and changing environments requires balancing short-term and long-term consequences. In reinforcement learning (RL), this trade-off is typically controlled through a fixed discount factor, which imposes a single exponentially discounted temporal horizon. However, biological agents exhibit flexible and adaptive temporal discounting, suggesting that effective planning requires multiple timescales. Here, we propose a multi-horizon approach that adaptively selects and combines temporal horizons, enabling robust adaptation to changes in reward structure without manual discount-factor tuning. This flexibility makes the method particularly suitable for continual learning scenarios involving task switches and varying environmental configurations. Empirically, we demonstrate that our approach identifies effective discount factors across a range of MiniGrid environments, including continual settings composed of three sequentially changing tasks. These results suggest that adaptive temporal discounting can improve parameter efficiency and enhance adaptability in both artificial and biologically inspired learning systems.

Large language models (LLMs) trained on webscale data can produce toxic outputs, raising concerns for safe deployment. Prior defenses, based on applications of DPO, NPO, and similar algorithms, reduce the likelihood of harmful continuations, but not robustly so: they are vulnerable to adversarial prompting and easily undone by fine-tuning-based relearning attacks. Indeed, research has shown that these edits to the model are superficial: linear probing reveals that harmful "directions" remain present in representations. To address this, we propose Representation Erasure-based Preference Optimization (REPO), reformulating detoxification as a token-level preference problem. Using a novel objective with preference data, we force the representations of toxic continuations to converge toward their benign counterparts. Our mechanistic analysis reveals that this granular approach is critical: unlike baselines, REPO induces deep, localized edits to toxicity-encoding neurons while preserving general model utility. Exhaustive evaluations show that REPO achieves state-of-the-art robustness, stopping sophisticated threats-including relearning attacks and enhanced GCG jailbreaks-where existing representation- and output-based methods fail.

The primary focus of multi-agent reinforcement learning (MARL) has been to study interactions among a fixed number of agents embedded in an environment. However, in the real world, the number of agents is neither fixed nor known a priori. Moreover, an agent can decide to create other agents (for example, a cell may divide, or a company may spin off a division). In this paper, we propose a framework that allows agents to create other agents; we call this a fluid-agent environment. We present game-theoretic solution concepts for fluid-agent games and empirically evaluate the performance of several MARL algorithms within this framework. Our experiments include fluid variants of established benchmarks such as Predator-Prey and Level-Based Foraging, where agents can dynamically spawn, as well as a new environment we introduce that highlights how fluidity can unlock novel solution strategies beyond those observed in fixed-population settings. We demonstrate that this framework yields agent teams that adjust their size dynamically to match environmental demands.

A hallmark of intelligence is the ability to adapt in non-stationary environments, yet deep Reinforcement Learning (RL) agents often struggle in such settings. Most prior studies introduce non-stationarity through abrupt shifts in features or dynamics, whereas real-world changes might be more gradual, reflecting naturalistic continual drift in the underlying dynamics. This may have important implications for studies on the "stability versus plasticity dilemma" in RL, since abrupt changes in the task may necessitate more plasticity than real-world situations actually would demand. To address these concerns, we modify existing 3D Miniworld and MuJoCo environments to incorporate naturalistic, continual non-stationary changes, and use them to identify whether poor performance in RL systems arises from a loss of plasticity or stability. We find that in these settings, methods that preserve stability, such as synaptic consolidation, achieve better performance than those focused on plasticity, such as resetting a subset of the parameters. Motivated by this finding, and prior evidence that successor features (SFs) reduce interference in non-stationary settings, we investigate whether SFs provide a better target than Q-values for consolidation. Across both environments, we find that applying a neuro-inspired synaptic consolidation mechanism to SFs rather than Q-values yields superior performance on the naturalistic, continual changing MuJoCo tasks. Furthermore, we find that consolidation is most effective when SFs are stabilized across multiple timescales, as different timescales capture complementary aspects of the gradually changing environment. Together, these results show that stability may be more important in continual learning settings when abrupt changes in tasks do not occur. Moreover, to enhance stability, multi-timescale consolidation of predictive representations is an effective approach.

Reinforcement learning from human feedback usually models preferences using a reward function that does not distinguish between people. We argue that this is unlikely to be a good design choice in contexts with high potential for disagreement, like in the training of large language models. We formalise and analyse the problem of learning a reward model that can be specialised to a user. Using the principle of empirical risk minimisation, we derive a probably approximately correct (PAC) bound showing the dependency of the approximation error on the number of training examples, as usual, and also on the number of human raters who provided feedback on them. Based on our theoretical findings, we discuss how to best collect pairwise preference data and argue that adaptive reward models should be beneficial when there is considerable disagreement among users. We also propose a concrete architecture for an adaptive reward model. Our approach leverages the observation that individual preferences can be captured as a linear combination of a set of general reward features. We show how to learn such features and subsequently use them to quickly adapt the reward model to a specific individual, even if their preferences are not reflected in the training data. We present experiments with large language models illustrating our theoretical results and comparing the proposed architecture with a non-adaptive baseline. Consistent with our analysis, the benefits provided by our model increase with the number of raters and the heterogeneity of their preferences. We also show that our model compares favourably to adaptive counterparts, including those performing in-context personalisation.

Any agents we can possibly build are subject to capacity constraints, as memory and compute resources are inherently finite. However, comparatively little attention has been dedicated to understanding how agents with limited capacity should allocate their resources for optimal performance. The goal of this paper is to shed some light on this question by studying a simple yet relevant continual learning problem: the capacity-constrained linear-quadratic-Gaussian (LQG) sequential prediction problem. We derive a solution to this problem under appropriate technical conditions. Moreover, for problems that can be decomposed into a set of sub-problems, we also demonstrate how to optimally allocate capacity across these sub-problems in the steady state. We view the results of this paper as a first step in the systematic theoretical study of learning under capacity constraints.

Reinforcement Learning from Human Feedback (RLHF) is a widely used technique for aligning Large Language Models (LLMs) with human preferences, yet it often suffers from sparse reward signals, making effective credit assignment challenging. In typical setups, the reward model provides a single scalar score for an entire generated sequence, offering little insight into which token or span-level decisions were responsible for the outcome. To address this, we propose Shapley Credit Assignment Rewards (SCAR), a novel method that leverages Shapley values in cooperative game theory. SCAR distributes the total sequence-level reward among constituent tokens or text spans based on their principled marginal contributions. This creates dense reward signals, crucially, without necessitating the training of auxiliary critique models or recourse to fine-grained human annotations at intermediate generation stages. Unlike prior dense reward methods, SCAR offers a game-theoretic foundation for fair credit attribution. Theoretically, we demonstrate that SCAR preserves the original optimal policy, and empirically, across diverse tasks including sentiment control, text summarization, and instruction tuning, we show that SCAR converges significantly faster and achieves higher final reward scores compared to standard RLHF and attention-based dense reward baselines. Our findings suggest that SCAR provides a more effective and theoretically sound method for credit assignment in RLHF, leading to more efficient alignment of LLMs.

Recent progress in generative modeling has highlighted the importance of Reinforcement Learning (RL) for fine-tuning, with KL-regularized methods in particular proving to be highly effective for both autoregressive and diffusion models. Complementing this line of work, the Relative Trajectory Balance (RTB) objective was recently introduced in the context of Generative Flow Networks (GFlowNets) to serve the same role of improving fine-tuning in sequential generative models. Building on prior work linking GFlowNets and maximum-entropy RL, we establish in this paper an equivalence between RTB and Trust-PCL, an off-policy RL method with KL regularization. This equivalence situates RTB within the broader theoretical landscape of KL-regularized RL, and clarifies its relationship to earlier methods. Leveraging this insight, we revisit an illustrative example from the RTB paper and show that KL-regularized RL methods achieve comparable performance, offering an alternative perspective to what was previously reported.