Large reasoning models (LRMs) achieve strong performance by externalizing explicit reasoning traces before producing the answer, yet suffer from overthinking challenge that allocates uniformly heavy computation to queries of varying difficulty.While proprietary models mitigate this via opaque routing, open-source LRMs still lack an efficient mechanism to internalize adaptive reasoning due to both expensive training cost and limited disclosure of training recipes.In response, we introduce RPO (Roottoken Policy Optimization), a framework that enables LRMs to self-determine when to reason by training only the initial root token (e.g., whether to invoke the tag) via group relative reward and group-wise advantages.By focusing on this pivotal branching point, RPO drastically reduces training overhead and VRAM usage.Across multiple model families and scales, RPO learns difficulty-aware adaptive thinking at just ∼2% of the training compute of prior adaptive-reasoning methods.

Offline-to-online reinforcement learning (RL) aims to integrate the complementary strengths of offline and online RL by pre-training an agent offline and subsequently fine-tuning it through online interactions. However, recent studies reveal that offline pre-trained agents often underperform during online fine-tuning due to inaccurate value estimation caused by distribution shift, with random initialization proving more effective in certain cases. In this work, we propose a novel method, Online Pre-Training for Offline-to-Online RL (OPT), explicitly designed to address the issue of inaccurate value estimation in offline pre-trained agents. OPT introduces a new learning phase, Online Pre-Training, which allows the training of a new value function tailored specifically for effective online fine-tuning. Implementation of OPT on TD3 and SPOT demonstrates an average 30% improvement in performance across a wide range of D4RL environments, including MuJoCo, Antmaze, and Adroit.

As the pre-training objectives (e.g., next token prediction) of language models (LMs) are inherently not aligned with task scores, optimizing LMs to achieve higher downstream task scores is essential. One of the promising approaches is to fine-tune LMs through reinforcement learning (RL). However, conventional RL methods based on PPO and a penalty of KL divergence are vulnerable to text degeneration where LMs do not generate natural texts anymore after RL fine-tuning. To address this problem, we provide Degeneration-free Policy Optimization (DfPO) that can fine-tune LMs to generate texts that achieve improved downstream task scores, while preserving the ability to generate natural texts. To achieve this, we introduce KL-masking which masks out the actions that potentially cause deviation from the reference policy when its likelihood is increased or decreased. Then, we devise truncated advantage functions for separately performing likelihood maximization and minimization to improve the task performance. In the experiments, we provide the results of DfPO and baseline algorithms on various generative NLP tasks including text continuation, text detoxification, and commonsense generation. Our experiments demonstrate that DfPO successfully improves the downstream task scores while preserving the ability to generate natural texts, without requiring additional hyperparameter search.

Multi-View Reinforcement Learning (MVRL) seeks to find an optimal control for an agent given multi-view observations from various sources. Despite recent advances in multi-view learning that aim to extract the latent representation from multi-view data, it is not straightforward to apply them to control tasks, especially when the observations are temporally dependent on one another. The problem can be even more challenging if the observations are intermittently missing for a subset of views. In this paper, we introduce Fuse2Control (F2C), an information-theoretic approach to capturing the underlying state space model from the sequences of multi-view observations. We conduct an extensive set of experiments in various control tasks showing that our method is highly effective in aggregating task-relevant information across many views, that scales linearly with the number of views while retaining robustness to arbitrary missing view scenarios.

We consider offline safe imitation learning (IL), where the agent aims to learn the safe policy that mimics preferred behavior while avoiding non-preferred behavior from non-preferred demonstrations and unlabeled demonstrations. This problem setting corresponds to various real-world scenarios, where satisfying safety constraints is more important than maximizing the expected return. However, it is very challenging to learn the policy to avoid constraint-violating (i.e. non-preferred) behavior, as opposed to standard imitation learning which learns the policy to mimic given demonstrations. In this paper, we present a hyperparameter-free offline safe IL algorithm, SafeDICE, that learns safe policy by leveraging the non-preferred demonstrations in the space of stationary distributions. Our algorithm directly estimates the stationary distribution corrections of the policy that imitate the demonstrations excluding the non-preferred behavior. In the experiments, we demonstrate that our algorithm learns a more safe policy that satisfies the cost constraint without degrading the reward performance, compared to baseline algorithms.

We consider the problem of learning from observation (LfO), in which the agent aims to mimic the expert's behavior from the state-only demonstrations by experts. We additionally assume that the agent cannot interact with the environment but has access to the action-labeled transition data collected by some agents with unknown qualities. This offline setting for LfO is appealing in many real-world scenarios where the ground-truth expert actions are inaccessible and the arbitrary environment interactions are costly or risky. In this paper, we present LobsDICE, an offline LfO algorithm that learns to imitate the expert policy via optimization in the space of stationary distributions. Our algorithm solves a single convex minimization problem, which minimizes the divergence between the two state-transition distributions induced by the expert and the agent policy. Through an extensive set of offline LfO tasks, we show that LobsDICE outperforms strong baseline methods.

We consider a strategic dialogue task, where the ability to infer the other agent's goal is critical to the success of the conversational agent. While this problem can be naturally formulated as Bayesian planning, it is known to be a very difficult problem due to its enormous search space consisting of all possible utterances. In this paper, we introduce an efficient Bayes-adaptive planning algorithm for goal-oriented dialogues, which combines RNN-based dialogue generation and MCTS-based Bayesian planning in a novel way, leading to robust decision-making under the uncertainty of the other agent's goal. We then introduce reinforcement learning for the dialogue agent that uses MCTS as a strong policy improvement operator, casting reinforcement learning as iterative alternation of planning and supervised-learning of self-generated dialogues. In the experiments, we demonstrate that our Bayes-adaptive dialogue planning agent significantly outperforms the state-of-the-art in a negotiation dialogue domain. We also show that reinforcement learning via MCTS further improves end-task performance without diverging from human language.

The goal-oriented dialogue system needs to be optimized for tracking the dialogue flow and carrying out an effective conversation under various situations to meet the user goal.The traditional approach to building such a dialogue system is to take a pipelined modular architecture, where its modules are optimized individually.However, such an optimization scheme does not necessarily yield an overall performance improvement of the whole system.On the other hand, end-to-end dialogue systems with monolithic neural architecture are often trained only with input-output utterances, without taking into account the entire annotations available in the corpus.This scheme makes it difficult for goal-oriented dialogues where the system needs to be integrated with external systems or to provide interpretable information about why the system generated a particular response.In this paper, we present an end-to-end neural architecture for dialogue systems that addresses both challenges above.Our dialogue system achieved the success rate of 68.32%, the language understanding score of 4.149, and the response appropriateness score of 4.287 in human evaluations, which ranked the system at the top position in the end-to-end multi-domain dialogue system task in the 8th dialogue systems technology challenge (DSTC8).

The recent development of flexible and scalable variational inference algorithms has popularized the use of deep probabilistic models in a wide range of applications. However, learning and reasoning about high-dimensional models with nondifferentiable densities are still a challenge. For such a model, inference algorithms struggle to estimate the gradients of variational objectives accurately, due to high variance in their estimates. To tackle this challenge, we present a novel variational inference algorithm for sequential data, which performs well even when the density from the model is not differentiable, for instance, due to the use of discrete random variables. The key feature of our algorithm is that it estimates future likelihoods at all time steps. The estimated future likelihoods form the core of our new low-variance gradient estimator. We formally analyze our gradient estimator from the perspective of variational objective, and show the effectiveness of our algorithm with synthetic and real datasets.

Stochastic variational inference has emerged as an effective method for performing inference on or learning complex models for data. Yet, one of the challenges in stochastic variational inference is handling high-dimensional data, such as sequential data, and models with non-differentiable densities caused by, for instance, the use of discrete latent variables. In such cases, it is challenging to control the variance of the gradient estimator used in stochastic variational inference, while low variance is often one of the key properties needed for successful inference. In this work, we present a new algorithm for stochastic variational inference of sequential models which trades off bias for variance to tackle this challenge effectively. Our algorithm is inspired by variance reduction techniques in reinforcement learning, yet it uniquely adopts their key ideas in the context of stochastic variational inference. We demonstrate the effectiveness of our approach through formal analysis and experiments on synthetic and real-world datasets.

Dialog state tracking, which refers to identifying the user intent from utterances, is one of the most important tasks in dialog management. In this paper, we present our dialog state tracker developed for the fifth dialog state tracking challenge, which focused on cross-language adaptation using a very scarce machine-translated training data when compared to the size of the ontology. Our dialog state tracker is based on the bi-directional long short-term memory network with a hierarchical attention mechanism in order to spot important words in user utterances. The user intent is predicted by finding the closest keyword in the ontology to the attention-weighted word vector. With the suggested methodology, our tracker can overcome various difficulties due to the scarce training data that existing machine learning-based trackers had, such as predicting user intents they have not seen before. We show that our tracker outperforms other trackers submitted to the challenge with respect to most of the performance measures.

In this paper, we consider the safe learning scenario where we need to restrict the exploratory behavior of a reinforcement learning agent. Specifically, we treat the problem as a form of Bayesian reinforcement learning in an environment that is modeled as a constrained MDP (CMDP) where the cost function penalizes undesirable situations. We propose a model-based Bayesian reinforcement learning (BRL) algorithm for such an environment, eliciting risk-sensitive exploration in a principled way. Our algorithm efficiently solves the constrained BRL problem by approximate linear programming, and generates a finite state controller in an off-line manner. We provide theoretical guarantees and demonstrate empirically that our approach outperforms the state of the art.

This paper presents a dialog state tracker submitted to Dialog State Tracking Challenge 5 (DSTC 5) with details. To tackle the challenging cross-language human-human dialog state tracking task with limited training data, we propose a tracker that focuses on words with meaningful context based on attention mechanism and bi-directional long short term memory (LSTM). The vocabulary including a plenty of proper nouns is vectorized with a sufficient amount of related texts crawled from web to learn a good embedding for words not existent in training dialogs. Despite its simplicity, our proposed tracker succeeded to achieve high accuracy without sophisticated pre- and post-processing.