Transformer-Based RL

Reinforcement learning (RL) has become a standard tool for post-training language models on reasoning tasks, where the policy is updated by reward feedback while exploring the space of responses. Despite its empirical success, theoretical understanding of RL post-training remains limited, in particular of why on-policy exploration combined with a neural reward model is effective. In this paper, we address this question by modeling the reward as a hierarchical function on the response space: the reward consists of infinitely many local components, each of which becomes relevant only after the preceding ones have been resolved. We show that a natural Transformer-based actor--critic algorithm, which alternates between sampling from the current KL-regularized policy, fitting a Transformer critic to the observed rewards, and updating the policy, achieves the minimax optimal rates in the query budget and in the regularization strength up to logarithmic factors, and is minimax optimal for a fixed number of prompts. In contrast, we prove that sampling from the fixed reference distribution, as in offline reward modeling, can limit regret decay to a logarithmic rate. These results show that on-policy exploration progressively zooms in on the region where the reward is concentrated, and quantify its benefit for RL post-training.

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

This paper addresses the problem of safe offline reinforcement learning, which involves training a policy to satisfy safety constraints using an offline dataset. This problem is inherently challenging as it requires balancing three highly interconnected and competing objectives: satisfying safety constraints, maximizing rewards, and adhering to the behavior regularization imposed by the offline dataset. To tackle this trilogy challenge, we propose Q-learning Penalized Transformer policy (QPT), a \emph{training--inference consistent} framework that bridges conditional sequence modeling with constraint-aware value estimation. QPT trains a Transformer policy that generates actions conditioned on trajectory context and target return/cost, retaining strong behavior regularization. To inject explicit safety semantics during learning, we augment sequence-model training with a Q-shaped penalty using learned reward and cost Q-functions to favor high return under low constraint violation. At inference, the same Q-functions enforce the cost threshold and choose the highest-reward feasible action, closing the loop between training and deployment. We provide a principled analysis under stylized near-deterministic CMDPs, characterizing how Q-penalized conditional generation improve safety and performance. Empirically, QPT consistently outperforms strong safe offline RL baselines across 38 tasks on the DSRL benchmark, and exhibits robust zero-shot adaptation to different constraint thresholds.

Decision Transformer performance degrades on long rollouts because the conditioning context drifts out of the training distribution. We show that this drift is visible through the model's own next state prediction error, which rises during rollout and stays elevated, giving a direct signal of when context has become unreliable. We introduce Trust Guided Decision Transformer (TGDT), which selects context before applying value guidance. At each step, TGDT evaluates several recent context suffixes using rolling next state prediction error, calibrated against held out offline data via split conformal prediction. It keeps only suffixes whose error stays within the calibrated threshold, then uses a frozen critic to choose the highest value action among the trusted suffixes. This reverses the order used by value only elastic selection, where the critic may choose an action generated from a context the model itself has flagged as unreliable. Experiments on D4RL navigation and locomotion tasks show that state prediction, critic guidance, and hard context reset each solve only part of the problem. TGDT reduces persistent high error runs and improves return over vanilla Decision Transformer, reset based context control, and value only context selection.

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

This paper investigates collaborative mobile edge computing (MEC) servers for large language model (LLM) inference under soft deadline constraints. In this system, to improve the quality of service, computations are expected to be completed within their deadlines. However, due to dependencies among tasks or subtasks, any missed deadline can lead to catastrophic consequences for the entire request. In this context, this work proposes an extended deadline mechanism with constrained flexibility. The main challenges lie in handling large-scale computations under strict latency constraints while limiting the number of allowable deadline extensions, especially in the presence of task dependencies within each request. To tackle these challenges, we develop a transformer-enhanced proximal policy optimization (PPO) framework that enables efficient collaboration among MEC servers. The proposed approach aims to maximize the number of tasks completed within their deadlines while minimizing the use of deadline extensions. By capturing temporal dependencies and cross-server interactions, the transformer improves decision-making for task migration. Simulation results demonstrate that the proposed method significantly outperforms conventional PPO and heuristic-based approaches in terms of task completion rate and overall system efficiency.

Reinforcement learning (RL) has become a central component of post-training large language models (LLMs), yet little is understood about how RL adaptation is distributed across transformer layers. Existing approaches typically update all model parameters uniformly, implicitly assuming that every layer contributes similarly to the gains obtained during RL post-training. In this work, we challenge this assumption through a systematic layer-wise study of RL training. Surprisingly, we find that training a single transformer layer can recover most of the gains achieved by full-parameter RL training, and in some cases even surpass it. To quantify this phenomenon, we introduce the quantity layer contribution, which measures the fraction of full RL improvement recovered by training a layer in isolation. Across seven models spanning two model families (Qwen3, Qwen2.5), three RL algorithms (GRPO, GiGPO, Dr. GRPO), and multiple task domains including mathematical reasoning, code generation, and agentic decision-making, we observe a remarkably stable pattern: RL gains are highly concentrated in a small subset of, and in many cases even a single, transformer layers. More strikingly, the same structural pattern consistently emerges: high-contribution layers concentrate in the middle of the transformer stack, while layers near the input and output ends contribute substantially less. The resulting layer rankings remain strongly correlated across datasets, tasks, model families, and RL algorithms.

Unmanned Traffic Management (UTM) systems are cloud-based platforms designed to manage and coordinate multiple aerial vehicles remotely. UTM systems are safety-critical which cannot tolerate failures like crash or collision. To reveal latent vulnerabilities, there are neither optimal failure-exposing demonstrations nor clear reward signals. Additionally, UTM's self-healing capability introduces the ``long-tail effect'' of critical failures. We propose framing UTM vulnerability discovery as a sequence modeling problem amenable to transformer-based RL architectures. Our approach leverages attention mechanisms to directly model the relationship among system states, and predict optimal actions. Our framework introduces a Policy Model that generates targeted test scenarios and an Action Sampler that enforces domain constraints. We use a risk-based reward function to guide exploration. Through extensive evaluation on a 700-hour simulation study, we demonstrate an 8$\times$ improvement in vulnerability discovery efficiency compared to expert-guided testing. It also discovers critical edge cases that traditional methods have missed.

Recent progress in large-scale sequence modeling has shown that a single model can learn useful representations across highly diverse data distributions. Inspired by these advances, we investigate whether a unified transformer policy can be trained across large collections of heterogeneous reinforcement learning environments. We introduce LDM-v0, a Large Decision Model trained offline on trajectories collected from thousands of environments spanning multiple domains and modalities. LDM-v0 is a multi-task, multi-modal transformer policy conditioned on histories of observations, actions, rewards, and termination signals, and trained through supervised next-action prediction over offline trajectories. We describe the environment infrastructure, automated data generation pipeline, model architecture, and training methodology used to build LDM-v0, and evaluate its performance across diverse environments. We show that a single pretrained model matches the performance of independently trained task-specific reference policies on approximately 1,000 environments including robotics, autonomous driving, inventory management, cybersecurity, trading, and video games. These results demonstrate the feasibility of large-scale offline pretraining across heterogeneous reinforcement learning environments using a single transformer policy.

In this work we study offline reinforcement learning (RL) under temporally extended task constraints expressed in Linear Temporal Logic over finite traces (LTLf). Recently, transformer-based approaches such as Trajectory Transformers and Decision Transformers have been adopted to address RL as a sequence modeling problem. However, these methods optimize purely for reward and do not account for high-level temporal requirements. Here, we introduce a neurosymbolic framework that injects LTLf background knowledge into such transformer-based RL policies. Our approach compiles LTLf formulas into deterministic finite automata (DFAs) and integrates them into the learning process through a differentiable representation and a logic-based loss function. In particular, we derive differentiable satisfaction signals from DFA progression and use them as a regularization term during training. The resulting method is architecture-agnostic across different models. We evaluate the proposed framework on navigation environments with specification suites covering combinations of safety and reachability temporal properties. Experimental results show that incorporating background knowledge not only improves constraint satisfaction, but also maintains competitive return compared to vanilla baselines.

Tree search is a central abstraction behind many language-agent reasoning and decision-making tasks: agents must explore actions, remember failures, and backtrack toward promising alternatives. Yet, we lack a theoretical understanding of how transformer-based policies acquire such search capabilities from the training dynamics of reinforcement learning (RL). We study this question in a stochastic $k$-ary tree environment, where an agentic transformer observes only its trajectory history through interaction and receives a terminal reward for reaching a hidden leaf goal node. We first construct a two-head transformer that implements randomized depth-first search (DFS): one head tracks previous actions, while the other detects failure outcomes and triggers backtracking. We then analyze the training dynamics of policy gradient under a depth-wise curriculum, showing that this same DFS mechanism emerges in stages from sparse reinforcement feedback without expert demonstrations. The resulting policy exhibits depth generalization: after training only on depth-$1$ and depth-$2$ trees, it succeeds on deeper full trees. We further show that, under imbalanced goal distributions, discounting the return leads to a ranked DFS policy that prioritizes higher-probability branches. Overall, our results identify a mechanistic normal form for transformer-based search, in which attention heads specialize and cooperate to extract decision-relevant traces from context and convert them into agentic action selection via RL training.

When a large language model under reinforcement learning commits a wrong reasoning step early in a trajectory, standard algorithms force it to keep generating until the maximum horizon, spending compute on tokens that never receive positive reward and polluting advantage estimates with post-failure noise. We propose ESPO (Early-Stopping Proximal Policy Optimization), which detects trajectory failure on-the-fly and terminates rollouts early. At each generation step, ESPO computes a surrogate regret using only the logits already computed during sampling, and terminates when the smoothed cumulative regret significantly exceeds its estimated values. Truncated trajectories are treated as absorbing failure states with a terminal reward, concentrating negative temporal-difference (TD) errors near the detected failure step without any additional reward model or human annotation. On DeepSeek-R1-Distill-Qwen-7B trained for mathematical reasoning, ESPO surpasses PPO on AIME2024 (46.28% vs. 45.25%), AMC2023 (85.83% vs. 82.94%), and MATH-500 (87.42% vs. 85.43%), while saving more than 20% rollout tokens cumulatively.

Reinforcement Learning (RL) has long been a powerful solution to various problems in communication networks. However, traditional RL models still face with several limitations. Not only do they rely on large numbers of interactions with the environment, but they are also limited in terms of modeling long-term relationships and tackling partial observability. In recent years, the Transformer model has demonstrated the ability to enhance RL models, allowing them to overcome these issues. Particularly, the self-attention mechanism within the Transformer enables efficient modeling of long-range dependencies and global correlations, as well as accelerates training processes and handles heterogeneous data modalities. In this paper, we present a comprehensive survey of Transformer-based RL algorithms and their applications in communication networks. Specifically, the paper provides the mathematical background of RL and Transformer architectures, along with insights into key issues such as resource allocation, computation offloading, routing, and trajectory control, and network security. We conclude the paper by discussing challenges, open issues, and notable future research directions, including Transformer-enhanced DRL algorithms for semantic communication and network optimization.

In dynamic urban logistics, the stochastic emergence of time-sensitive tasks poses a significant optimality challenge for heterogeneous AAVs logistics task allocation. To address this problem, a reinforcement learning enhanced overlapping coalition formation game approach is proposed. A dynamic task allocation model is established, where global optimality is mathematically quantified by a generalized logistics cost coupling service quality and resource consumption. To deal with the time-varying task sets induced by stochastic order arrivals, a transformer-based soft actor-critic network is designed. By leveraging multi-head self-attention to encode variable-length logistics states and capture task-wise spatiotemporal dependencies, the learned policy adaptively guides coalition updates, replacing heuristic rules in the overlapping coalition formation game. On this basis, heterogeneous AAVs can form more efficient overlapping coalitions for dynamic logistics tasks. The resulting coalition formation process is proven to constitute an exact potential game, which guarantees convergence to a Nash-stable equilibrium within a finite number of iterations. Numerical simulations demonstrate that the proposed algorithm effectively improves the optimality of task allocation under the generalized logistics cost criterion. In a scenario with 32 AAVs and 80 tasks, our algorithm achieves a 39.76% cost reduction compared with the heuristic OCF baseline. Indoor flight experiments further validate its practicality.

Token-based transformer world models have shown strong performance in visual reinforcement learning, but often suffer from temporal inconsistency in long-horizon rollouts, including object duplication, disappearance, and transmutation. A key reason is that most existing approaches treat next-frame prediction purely as a token generation problem, without considering the persistence of tokens across time. We introduce Identifiable Token Correspondence (ITC), a decoding step for token-based transformer world models that formulates next-frame prediction as a structured assignment problem with latent token correspondence variables: each next-frame token is explained either by copying a token from the previous frame or by generating a new one. ITC leaves the transformer architecture and training procedure unchanged and can be added on top of existing backbones. Our experiments show state-of-the-art performance on 4 challenging benchmarks. The proposed method achieves a return of 72.5% and a score of 35.6% on the Craftax-classic benchmark, significantly surpassing the previous best of 67.4% and 27.9%. We release our source code on https://github.com/snu-mllab/Identifiable-Token-Correspondence.

A central challenge in reinforcement learning (RL) is to learn models that generalize beyond the tasks on which they are trained, a goal traditionally pursued through multi-task and meta RL. Recently, transformer architectures have emerged as a promising approach, enabling adaptation to new tasks via in-context learning without explicit parameter updates. From a functional perspective, a transformer can be viewed as a functional operator that maps a context to a task-specific function. It is thus fundamental to understand and design this operator to support stronger generalization in RL. In this work, we address this resulting question of generalization from a kernel-based perspective by establishing a connection between non-linear transformers and kernel-based temporal difference learning. By interpreting the transformer as performing regression in a Reproducing Kernel Hilbert Space (RKHS), we show that value functions from different domains can be represented using a shared set of weights, provided they lie within the same RKHS. Experiments on multiple MetaWorld domains support this interpretation, demonstrating convergence of the temporal-difference objective.

Long-horizon, sparse-reward tasks pose a fundamental challenge for reinforcement learning, since single-step TD learning suffers from bootstrapping error accumulation across successive Bellman updates. Actor-critic methods with action chunking address this by operating over temporally extended actions, which reduce the effective horizon, enable fast value backups, and support temporally consistent exploration. However, existing methods rely on a fixed chunk size and therefore cannot adaptively balance reactivity against temporal consistency. A large fixed chunk size reduces responsiveness to new observations, while a small one produces incoherent motions, forcing task-specific tuning of the chunk size. To address this limitation, we propose Adaptive Chunk Size Actor-Critic (ACSAC). ACSAC leverages a causal Transformer critic to evaluate expected returns for action chunks of different sizes. At each chunk boundary, it adaptively selects the chunk size that maximizes the expected return, supporting flexible, state-dependent chunk sizes without task-specific tuning. We prove that the ACSAC Bellman operator is a contraction whose unique fixed point is the action-value function of the adaptive policy. Experiments on OGBench demonstrate that ACSAC achieves state-of-the-art performance on long-horizon, sparse-reward manipulation tasks across both offline RL and offline-to-online RL settings.

In-context reinforcement learning (ICRL) studies agents that, after pretraining, adapt to new tasks by conditioning on additional context without parameter updates. Existing theoretical analyses of ICRL largely rely on linear attention, which replaces the softmax function in the standard attention with an identity mapping. This paper provides the first theoretical understanding of ICRL without making the unrealistic linear attention simplification. In particular, we consider the standard softmax attention used in practice. We show that, with certain parameters, the layerwise forward pass of a Transformer with such softmax attention is equivalent to iterative updates of a weighted softmax temporal difference (TD) learning algorithm. Here, weighted softmax TD is a new RL algorithm that performs policy evaluation in kernel space and adopts both linear TD and tabular TD as special cases. We also prove that under a certain contraction condition, the policy evaluation error decays as the number of layers grows, with the identified parameters above. Finally, we prove that those parameters are a global minimizer of a pretraining loss, explaining their emergence in our numerical experiments.

In-context reinforcement learning (ICRL) refers to the ability of RL agents to adapt to new tasks at inference time without parameter updates by conditioning on additional context. Recent empirical studies further demonstrate that Chain-of-Thought (CoT) generation can amplify this ICRL capability. This paper is the first to provide a theoretical understanding on how CoT interacts with ICRL. We conduct our analysis in a policy evaluation setup with linear Transformer. We prove that with specific Transformer parameters, the CoT generation process is equivalent to repeatedly executing temporal difference learning updates. Additionally, we provide finite sample convergence analysis showing that the policy evaluation error decreases geometrically with CoT length and eventually saturates at a statistical floor determined by the context length. We also prove that the desired Transformer parameters are a global minimizer of the pretraining loss, providing a theoretical understanding on the empirical emergence of those parameters.

Decision Transformer (DT) formulates offline reinforcement learning as autoregressive sequence modeling, achieving promising results by predicting actions from a sequence of Return-to-Go (RTG), state, and action tokens. However, RTG is a scalar that summarizes future rewards, containing far less information than typical state or action vectors, yet it consumes the same computational budget per token. Worse, the self-attention cost of Transformers grows quadratically with sequence length, so including RTG as a separate token adds unnecessary overhead. We propose SlimDT, which removes RTG from the autoregressive sequence. Instead, we inject RTG information into the state representations before the sequential modeling step, allowing the Transformer to process only a compact (state, action) sequence. This reduces the sequence length by one-third, directly improving inference efficiency. On the D4RL benchmark, SlimDT surpasses standard DT across various tasks and achieves performance comparable to existing state-of-the-art methods. Decoupling a sparse conditioning signal from an information-rich sequence thus yields both computational gains and higher task performance.

We investigate the ability of transformers to perform in-context reinforcement learning (ICRL), where a model must infer and execute learning algorithms from trajectory data without parameter updates. We show that a linear self-attention transformer block can provably implement policy-improvement methods, including semi-gradient SARSA and actor-critic, via explicit parameter constructions. Beyond existence, we design a teacher-mimicking training procedure, analyze its gradient-flow dynamics, and establish the first convergence guarantee in the ICRL literature: under suitable richness conditions on the training MDP distribution, gradient flow converges locally and exponentially to an optimal parameter manifold corresponding to the desired RL update. Empirically, training transformers on randomly generated tabular MDPs confirms these predictions: the learned models recover the parameter structure of our explicit constructions and, when deployed on unseen MDPs, deliver strong in-context control performance. Together, these results illuminate how transformer architectures internalize and execute classical reinforcement learning algorithms in context, bridging mechanistic understanding and training dynamics in ICRL.

We investigate how embedding dimension affects the emergence of an internal "world model" in a transformer trained with reinforcement learning to perform bubble-sort-style adjacent swaps. Models achieve high accuracy even with very small embedding dimensions, but larger dimensions yield more faithful, consistent, and robust internal representations. In particular, higher embedding dimensions strengthen the formation of structured internal representation and lead to better interpretability. After hundreds of experiments, we observe two consistent mechanisms: (1) the last row of the attention weight matrix monotonically encodes the global ordering of tokens; and (2) the selected transposition aligns with the largest adjacent difference of these encoded values. Our results provide quantitative evidence that transformers build structured internal world models and that model size improves representation quality in addition to end performance. We release our metrics and analyses, which can be used to probe similar algorithmic tasks.