arXiv subjects

Dale Schuurmans

Publications and source records attributed to Dale Schuurmans.

At least 19 recordsLinked to original sources

arXiv2026

Learning with a Single Rollout via Monte Carlo Pass@k Critic

Estimating token-level advantages in reinforcement learning (RL) for language models remains challenging because scaling up episodic experience collection is expensive. The difficulty intensifies for baseline advantage estimation methods, where repeated sampling causes trajectories to diverge into substantially different reasoning prefixes. In this context, RL algorithms such as GRPO prove limited: an outcome reward is too sparse to be attributed to specific actions like intermediate steps, and comparisons across sampled traces are non-trivial because they are heterogeneous. To mitigate both the computational cost of repeated sampling and the difficulty of credit assignment, we study single-rollout proximal policy optimization (SR-PPO) featuring token-level credit assignment in RL for language models. Instead of estimating advantages by normalizing episodic returns within the candidate group, we train a calibrated token-level credit critic using Monte Carlo outcomes from one rollout per prompt. Specifically, we use the critic to predict the Pass@k success probability at the prompt prefix, which is derived from a Pass@1 attempt. This choice yields a more selective learning signal than Pass@1: it discounts easily solved prefixes while prioritizing hard ones whose success probability remains marginal. We show that as $k$ increases, Pass@k converges to a reachability indicator, reflecting whether a prefix can lead to at least one successful continuation. In an explicit state graph, the limit ($k \rightarrow \infty$) can be computed in $O(|V|+|E|)$ time, offering a promising surrogate for direct credit assignment without the need to sample contrastive traces. As an initial validation, SR-PPO exhibits stable learning dynamics, along with consistent gains in Pass@128 success rates on mathematical reasoning benchmarks such as HMMT26 and AIME24.

cs.LG↗

arXiv2026

Spectral Ghost in Representation Learning: from Component Analysis to Self-Supervised Learning

Self-supervised learning (SSL) has improved empirical performance by unleashing the power of unlabeled data for practical applications. Specifically, SSL extracts the representation from massive unlabeled data, which will be transferred to a plenty of down streaming tasks with limited data. The significant improvement on diverse applications of representation learning has attracted increasing attention, resulting in a variety of dramatically different self-supervised learning objectives for representation extraction, with an assortment of learning procedures, but the lack of a clear and unified understanding. Such an absence hampers the ongoing development of representation learning, leaving a theoretical understanding missing, principles for efficient algorithm design unclear, and the use of representation learning methods in practice unjustified. The urgency for a unified framework is further motivated by the rapid growth in representation learning methods. In this paper, we are therefore compelled to develop a principled foundation of representation learning. We first theoretically investigate the sufficiency of the representation from a spectral representation view, which reveals the spectral essence of the existing successful SSL algorithms and paves the path to a unified framework for understanding and analysis. Such a framework work also inspires the development of more efficient and easy-to-use representation learning algorithms with principled way in real-world applications.

cs.LG↗

arXiv2026

Universal computation is intrinsic to language model decoding

Language models now provide an interface to express and often solve general problems in natural language, yet their ultimate computational capabilities remain a major topic of scientific debate. Unlike a formal computer, a language model is trained to autoregressively predict successive elements in human-generated text. We prove that chaining a language model's autoregressive output is sufficient to perform universal computation. That is, a language model can simulate the execution of any algorithm on any input. The challenge of eliciting desired computational behaviour can thus be reframed in terms of programmability: the ease of finding a suitable prompt. Strikingly, we demonstrate that even randomly initialized language models are capable of universal computation before training. This implies that training does not give rise to computational expressiveness -- rather, it improves programmability, enabling a natural language interface for accessing these intrinsic capabilities.

cs.CL↗

arXiv2025

The World Is Bigger! A Computationally-Embedded Perspective on the Big World Hypothesis

Continual learning is often motivated by the idea, known as the big world hypothesis, that "the world is bigger" than the agent. Recent problem formulations capture this idea by explicitly constraining an agent relative to the environment. These constraints lead to solutions in which the agent continually adapts to best use its limited capacity, rather than converging to a fixed solution. However, explicit constraints can be ad hoc, difficult to incorporate, and may limit the effectiveness of scaling up the agent's capacity. In this paper, we characterize a problem setting in which an agent, regardless of its capacity, is constrained by being embedded in the environment. In particular, we introduce a computationally-embedded perspective that represents an embedded agent as an automaton simulated within a universal (formal) computer. Such an automaton is always constrained; we prove that it is equivalent to an agent that interacts with a partially observable Markov decision process over a countably infinite state-space. We propose an objective for this setting, which we call interactivity, that measures an agent's ability to continually adapt its behaviour by learning new predictions. We then develop a model-based reinforcement learning algorithm for interactivity-seeking, and use it to construct a synthetic problem to evaluate continual learning capability. Our results show that deep nonlinear networks struggle to sustain interactivity, whereas deep linear networks sustain higher interactivity as capacity increases.

cs.AI↗

arXiv2025

Spectral Representation-based Reinforcement Learning

In real-world applications with large state and action spaces, reinforcement learning (RL) typically employs function approximations to represent core components like the policies, value functions, and dynamics models. Although powerful approximations such as neural networks offer great expressiveness, they often present theoretical ambiguities, suffer from optimization instability and exploration difficulty, and incur substantial computational costs in practice. In this paper, we introduce the perspective of spectral representations as a solution to address these difficulties in RL. Stemming from the spectral decomposition of the transition operator, this framework yields an effective abstraction of the system dynamics for subsequent policy optimization while also providing a clear theoretical characterization. We reveal how to construct spectral representations for transition operators that possess latent variable structures or energy-based structures, which implies different learning methods to extract spectral representations from data. Notably, each of these learning methods realizes an effective RL algorithm under this framework. We also provably extend this spectral view to partially observable MDPs. Finally, we validate these algorithms on over 20 challenging tasks from the DeepMind Control Suite, where they achieve performances comparable or superior to current state-of-the-art model-free and model-based baselines.

cs.LG↗

arXiv2025

Learning to Reason Efficiently with Discounted Reinforcement Learning

Large reasoning models (LRMs) often consume excessive tokens, inflating computational cost and latency. More broadly, in goal reaching sequential decision problems we often want to reach the goal quickly, and LRM reasoning can be viewed through this lens. We challenge the assumption that longer responses improve accuracy. By penalizing reasoning tokens using a discounted reinforcement learning setup (interpretable as a small token cost) and analyzing Blackwell optimality in restricted policy classes, we encourage concise yet accurate reasoning, analogous to preferring shorter successful trajectories in a stochastic shortest path problem. Experiments confirm our theoretical results that this approach shortens chains of thought while preserving accuracy.

cs.LG↗

arXiv2025

Rectifying Regression in Reinforcement Learning

This paper investigates the impact of the loss function in value-based methods for reinforcement learning through an analysis of underlying prediction objectives. We theoretically show that mean absolute error is a better prediction objective than the traditional mean squared error for controlling the learned policy's suboptimality gap. Furthermore, we present results that different loss functions are better aligned with these different regression objectives: binary and categorical cross-entropy losses with the mean absolute error and squared loss with the mean squared error. We then provide empirical evidence that algorithms minimizing these cross-entropy losses can outperform those based on the squared loss in linear reinforcement learning.

cs.LG↗

arXiv2025

Judging with Confidence: Calibrating Autoraters to Preference Distributions

The alignment of large language models (LLMs) with human values increasingly relies on using other LLMs as automated judges, or ``autoraters''. However, their reliability is limited by a foundational issue: they are trained on discrete preference labels, forcing a single ground truth onto tasks that are often subjective, ambiguous, or nuanced. We argue that a reliable autorater must learn to model the full distribution of preferences defined by a target population. In this paper, we propose a general framework for calibrating probabilistic autoraters to any given preference distribution. We formalize the problem and present two learning methods tailored to different data conditions: 1) a direct supervised fine-tuning for dense, probabilistic labels, and 2) a reinforcement learning approach for sparse, binary labels. Our empirical results show that finetuning autoraters with a distribution-matching objective leads to verbalized probability predictions that are better aligned with the target preference distribution, with improved calibration and significantly lower positional bias, all while preserving performance on objective tasks.

cs.CL↗

arXiv2025

Rethinking the Global Convergence of Softmax Policy Gradient with Linear Function Approximation

Policy gradient (PG) methods have played an essential role in the empirical successes of reinforcement learning. In order to handle large state-action spaces, PG methods are typically used with function approximation. In this setting, the approximation error in modeling problem-dependent quantities is a key notion for characterizing the global convergence of PG methods. We focus on Softmax PG with linear function approximation (referred to as $\texttt{Lin-SPG}$) and demonstrate that the approximation error is irrelevant to the algorithm's global convergence even for the stochastic bandit setting. Consequently, we first identify the necessary and sufficient conditions on the feature representation that can guarantee the asymptotic global convergence of $\texttt{Lin-SPG}$. Under these feature conditions, we prove that $T$ iterations of $\texttt{Lin-SPG}$ with a problem-specific learning rate result in an $O(1/T)$ convergence to the optimal policy. Furthermore, we prove that $\texttt{Lin-SPG}$ with any arbitrary constant learning rate can ensure asymptotic global convergence to the optimal policy.

cs.LG↗

arXiv2025

Improving Large Language Model Planning with Action Sequence Similarity

Planning is essential for artificial intelligence systems to look ahead and proactively determine a course of actions to reach objectives in the virtual and real world. Recent work on large language models (LLMs) sheds light on their planning capability in various tasks. However, it remains unclear what signals in the context influence the model performance. In this work, we explore how to improve the model planning capability through in-context learning (ICL), specifically, what signals can help select the exemplars. Through extensive experiments, we observe that commonly used problem similarity may result in false positives with drastically different plans, which can mislead the model. In response, we propose to sample and filter exemplars leveraging plan side action sequence similarity (AS). We propose GRASE-DC: a two-stage pipeline that first re-samples high AS exemplars and then curates the selected exemplars with dynamic clustering on AS to achieve a balance of relevance and diversity. Our experimental result confirms that GRASE-DC achieves significant performance improvement on various planning tasks (up to ~11-40 point absolute accuracy improvement with 27.3% fewer exemplars needed on average). With GRASE-DC* + VAL, where we iteratively apply GRASE-DC with a validator, we are able to even boost the performance by 18.9% more. Extensive analysis validates the consistent performance improvement of GRASE-DC with various backbone LLMs and on both classical planning and natural language planning benchmarks. GRASE-DC can further boost the planning accuracy by ~24 absolute points on harder problems using simpler problems as exemplars over a random baseline. This demonstrates its ability to generalize to out-of-distribution problems.

cs.AI↗

arXiv2025

Representation Learning via Non-Contrastive Mutual Information

Labeling data is often very time consuming and expensive, leaving us with a majority of unlabeled data. Self-supervised representation learning methods such as SimCLR (Chen et al., 2020) or BYOL (Grill et al., 2020) have been very successful at learning meaningful latent representations from unlabeled image data, resulting in much more general and transferable representations for downstream tasks. Broadly, self-supervised methods fall into two types: 1) Contrastive methods, such as SimCLR; and 2) Non-Contrastive methods, such as BYOL. Contrastive methods are generally trying to maximize mutual information between related data points, so they need to compare every data point to every other data point, resulting in high variance, and thus requiring large batch sizes to work well. Non-contrastive methods like BYOL have much lower variance as they do not need to make pairwise comparisons, but are much trickier to implement as they have the possibility of collapsing to a constant vector. In this paper, we aim to develop a self-supervised objective that combines the strength of both types. We start with a particular contrastive method called the Spectral Contrastive Loss (HaoChen et al., 2021; Lu et al., 2024), and we convert it into a more general non-contrastive form; this removes the pairwise comparisons resulting in lower variance, but keeps the mutual information formulation of the contrastive method preventing collapse. We call our new objective the Mutual Information Non-Contrastive (MINC) loss. We test MINC by learning image representations on ImageNet (similar to SimCLR and BYOL) and show that it consistently improves upon the Spectral Contrastive loss baseline.

cs.LG↗

arXiv2025

Ordering-based Conditions for Global Convergence of Policy Gradient Methods

We prove that, for finite-arm bandits with linear function approximation, the global convergence of policy gradient (PG) methods depends on inter-related properties between the policy update and the representation. textcolor{blue}{First}, we establish a few key observations that frame the study: \textbf{(i)} Global convergence can be achieved under linear function approximation without policy or reward realizability, both for the standard Softmax PG and natural policy gradient (NPG). \textbf{(ii)} Approximation error is not a key quantity for characterizing global convergence in either algorithm. \textbf{(iii)} The conditions on the representation that imply global convergence are different between these two algorithms. Overall, these observations call into question approximation error as an appropriate quantity for characterizing the global convergence of PG methods under linear function approximation. \textcolor{blue}{Second}, motivated by these observations, we establish new general results: \textbf{(i)} NPG with linear function approximation achieves global convergence \emph{if and only if} the projection of the reward onto the representable space preserves the optimal action's rank, a quantity that is not strongly related to approximation error. \textbf{(ii)} The global convergence of Softmax PG occurs if the representation satisfies a non-domination condition and can preserve the ranking of rewards, which goes well beyond policy or reward realizability. We provide experimental results to support these theoretical findings.

cs.LG↗

arXiv2025

Small steps no more: Global convergence of stochastic gradient bandits for arbitrary learning rates

We provide a new understanding of the stochastic gradient bandit algorithm by showing that it converges to a globally optimal policy almost surely using \emph{any} constant learning rate. This result demonstrates that the stochastic gradient algorithm continues to balance exploration and exploitation appropriately even in scenarios where standard smoothness and noise control assumptions break down. The proofs are based on novel findings about action sampling rates and the relationship between cumulative progress and noise, and extend the current understanding of how simple stochastic gradient methods behave in bandit settings.

cs.LG↗

arXiv2025

SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training

Supervised fine-tuning (SFT) and reinforcement learning (RL) are widely used post-training techniques for foundation models. However, their roles in enhancing model generalization capabilities remain unclear. This paper studies the difference between SFT and RL on generalization and memorization, focusing on text-based rule variants and visual variants. We introduce GeneralPoints, an arithmetic reasoning card game, and adopt V-IRL, a real-world navigation environment, to assess how models trained with SFT and RL generalize to unseen variants in both textual and visual domains. We show that RL, especially when trained with an outcome-based reward, generalizes across both rule-based textual and visual variants. SFT, in contrast, tends to memorize training data and struggles to generalize out-of-distribution scenarios. Further analysis reveals that RL improves the model's underlying visual recognition capabilities, contributing to its enhanced generalization in the visual domain. Despite RL's superior generalization, we show that SFT remains essential for effective RL training; SFT stabilizes the model's output format, enabling subsequent RL to achieve its performance gains. These findings demonstrates the capability of RL for acquiring generalizable knowledge in complex, multi-modal tasks.

cs.AI↗

arXiv2025

Evolving Deeper LLM Thinking

We explore an evolutionary search strategy for scaling inference time compute in Large Language Models. The proposed approach, Mind Evolution, uses a language model to generate, recombine and refine candidate responses. The proposed approach avoids the need to formalize the underlying inference problem whenever a solution evaluator is available. Controlling for inference cost, we find that Mind Evolution significantly outperforms other inference strategies such as Best-of-N and Sequential Revision in natural language planning tasks. In the TravelPlanner and Natural Plan benchmarks, Mind Evolution solves more than 98% of the problem instances using Gemini 1.5 Pro without the use of a formal solver.

cs.AI↗

arXiv2024

Improving Dynamic Object Interactions in Text-to-Video Generation with AI Feedback

Large text-to-video models hold immense potential for a wide range of downstream applications. However, they struggle to accurately depict dynamic object interactions, often resulting in unrealistic movements and frequent violations of real-world physics. One solution inspired by large language models is to align generated outputs with desired outcomes using external feedback. In this work, we investigate the use of feedback to enhance the quality of object dynamics in text-to-video models. We aim to answer a critical question: what types of feedback, paired with which specific self-improvement algorithms, can most effectively overcome movement misalignment and realistic object interactions? We first point out that offline RL-finetuning algorithms for text-to-video models can be equivalent as derived from a unified probabilistic objective. This perspective highlights that there is no algorithmically dominant method in principle; rather, we should care about the property of reward and data. While human feedback is less scalable, vision-language models could notice the video scenes as humans do. We then propose leveraging vision-language models to provide perceptual feedback specifically tailored to object dynamics in videos. Compared to popular video quality metrics measuring alignment or dynamics, the experiments demonstrate that our approach with binary AI feedback drives the most significant improvements in the quality of interaction scenes in video, as confirmed by AI, human, and quality metric evaluations. Notably, we observe substantial gains when using signals from vision language models, particularly in scenarios involving complex interactions between multiple objects and realistic depictions of objects falling.

cs.LG↗

arXiv2024

Toward Understanding In-context vs. In-weight Learning

It has recently been demonstrated empirically that in-context learning emerges in transformers when certain distributional properties are present in the training data, but this ability can also diminish upon further training. We provide a new theoretical understanding of these phenomena by identifying simplified distributional properties that give rise to the emergence and eventual disappearance of in-context learning. We do so by first analyzing a simplified model that uses a gating mechanism to choose between an in-weight and an in-context predictor. Through a combination of a generalization error and regret analysis we identify conditions where in-context and in-weight learning emerge. These theoretical findings are then corroborated experimentally by comparing the behaviour of a full transformer on the simplified distributions to that of the stylized model, demonstrating aligned results. We then extend the study to a full large language model, showing how fine-tuning on various collections of natural language prompts can elicit similar in-context and in-weight learning behaviour.

cs.LG↗

arXiv2024

Faster WIND: Accelerating Iterative Best-of-$N$ Distillation for LLM Alignment

Recent advances in aligning large language models with human preferences have corroborated the growing importance of best-of-N distillation (BOND). However, the iterative BOND algorithm is prohibitively expensive in practice due to the sample and computation inefficiency. This paper addresses the problem by revealing a unified game-theoretic connection between iterative BOND and self-play alignment, which unifies seemingly disparate algorithmic paradigms. Based on the connection, we establish a novel framework, WIN rate Dominance (WIND), with a series of efficient algorithms for regularized win rate dominance optimization that approximates iterative BOND in the parameter space. We provides provable sample efficiency guarantee for one of the WIND variant with the square loss objective. The experimental results confirm that our algorithm not only accelerates the computation, but also achieves superior sample efficiency compared to existing methods.

cs.LG↗