In Reinforcement Learning, agents learn policies by exploring and interacting with the environment. Due to the curse of dimensionality, learning policies that map high-dimensional sensory input to motor output is particularly challenging. During training, state of the art methods (SAC, PPO, etc.) explore the environment by perturbing the actuation with independent Gaussian noise. While this unstructured exploration has proven successful in numerous tasks, it can be suboptimal for overactuated systems. When multiple actuators, such as motors or muscles, drive behavior, uncorrelated perturbations risk diminishing each other's effect, or modifying the behavior in a task-irrelevant way.

Finetuning language models with reinforcement learning (RL), e.g. from human feedback (HF), is a prominent method for alignment. But optimizing against a reward model can improve on reward while degrading performance in other areas, a phenomenon known as reward hacking, alignment tax, or language drift. First, we argue that commonly-used test metrics are insufficient and instead measure how different algorithms tradeoff between reward and drift. The standard method modified the reward with a Kullback-Lieber (KL) penalty between the online and initial model. We propose Elastic Reset, a new algorithm that achieves higher reward with less drift without explicitly modifying the training objective.

Direct Preference Optimization (DPO) optimizes language models to align with human preferences. Utilizing on-policy samples, generated directly by the policy model, typically results in better performance due to its distribution consistency with the model compared to off-policy samples. This paper identifies the quality of candidate preference samples as another critical factor. While the quality of on-policy data is inherently constrained by the capabilities of the policy model, off-policy data, which can be derived from diverse sources, offers greater potential for quality despite experiencing distribution shifts. However, current research mostly relies on on-policy data and neglects the value of off-policy data in terms of data quality, due to the challenge posed by distribution shift. In this paper, we propose InCo-DPO, an efficient method for synthesizing preference data by integrating on-policy and off-policy data, allowing dynamic adjustments to balance distribution shifts and data quality, thus finding an optimal trade-off. Consequently, InCo-DPO overcomes the limitations of distribution shifts in off-policy data and the quality constraints of on-policy data. We evaluated InCo-DPO with the Alpaca-Eval 2.0 and Arena-Hard benchmarks. Experimental results demonstrate that our approach not only outperforms both on-policy and off-policy data but also achieves a state-of-the-art win rate of 60.8 on Arena-Hard with the vanilla DPO using Gemma-2 model.

Finetuning language models with reinforcement learning (RL), e.g. from human feedback (HF), is a prominent method for alignment. But optimizing against a reward model can improve on reward while degrading performance in other areas, a phenomenon known as reward hacking, alignment tax, or language drift. First, we argue that commonly-used test metrics are insufficient and instead measure how different algorithms tradeoff between reward and drift. The standard method modified the reward with a Kullback-Lieber (KL) penalty between the online and initial model. We propose Elastic Reset, a new algorithm that achieves higher reward with less drift without explicitly modifying the training objective.

In Reinforcement Learning, agents learn policies by exploring and interacting with the environment. Due to the curse of dimensionality, learning policies that map high-dimensional sensory input to motor output is particularly challenging. During training, state of the art methods (SAC, PPO, etc.) explore the environment by perturbing the actuation with independent Gaussian noise. While this unstructured exploration has proven successful in numerous tasks, it can be suboptimal for overactuated systems. When multiple actuators, such as motors or muscles, drive behavior, uncorrelated perturbations risk diminishing each other's effect, or modifying the behavior in a task-irrelevant way.

Offline safe reinforcement learning (OSRL) involves learning a decision-making policy to maximize rewards from a fixed batch of training data to satisfy pre-defined safety constraints. However, adapting to varying safety constraints during deployment without retraining remains an under-explored challenge. To address this challenge, we introduce constraint-adaptive policy switching (CAPS), a wrapper framework around existing offline RL algorithms. During training, CAPS uses offline data to learn multiple policies with a shared representation that optimize different reward and cost trade-offs. During testing, CAPS switches between those policies by selecting at each state the policy that maximizes future rewards among those that satisfy the current cost constraint. Our experiments on 38 tasks from the DSRL benchmark demonstrate that CAPS consistently outperforms existing methods, establishing a strong wrapper-based baseline for OSRL. The code is publicly available at https://github.com/yassineCh/CAPS.

Large language models (LLMs) are aligned with human preferences by reinforcement learning from human feedback (RLHF). Effective data sampling is crucial for RLHF, as it determines the efficiency of model training, ensuring that models learn from the informative samples. To achieve better data generation, we propose a new sampling method called Preference-Guided Reflective Sampling (PRS). PRS frames the response generation as an optimization process to the explicitly specified user preference described in natural language. It employs a tree-based generation framework to enable an efficient sampling process, which guides the direction of generation through preference and better explores the sampling space with adaptive self-refinement. Notably, PRS can align LLMs to diverse preferences. We study preference-controlled text generation for instruction following and keyword-focused document summarization. Our findings indicate that PRS, across different LLM policies, generates training data with much higher rewards than strong baselines. PRS also excels in post-RL training.

Reinforcement Learning (RL) has achieved impressive results on complex tasks but struggles in multi-task settings with different embodiments. World models offer scalability by learning a simulation of the environment, yet they often rely on inefficient gradient-free optimization methods. We introduce Policy learning with large World Models (PWM), a novel model-based RL algorithm that learns continuous control policies from large multi-task world models. By pre-training the world model on offline data and using it for first-order gradient policy learning, PWM effectively solves tasks with up to 152 action dimensions and outperforms methods using ground-truth dynamics. Additionally, PWM scales to an 80-task setting, achieving up to 27% higher rewards than existing baselines without the need for expensive online planning. Visualizations and code available at https://www.imgeorgiev.com/pwm

This paper introduces two learning schemes for distributed agents in Reinforcement Learning (RL) environments, namely Reward-Weighted (R-Weighted) and Loss-Weighted (L-Weighted) gradient merger. The R/L weighted methods replace standard practices for training multiple agents, such as summing or averaging the gradients. The core of our methods is to scale the gradient of each actor based on how high the reward (for R-Weighted) or the loss (for L-Weighted) is compared to the other actors. During training, each agent operates in differently initialized versions of the same environment, which gives different gradients from different actors. In essence, the R-Weights and L-Weights of each agent inform the other agents of its potential, which again reports which environment should be prioritized for learning. This approach of distributed learning is possible because environments that yield higher rewards, or low losses, have more critical information than environments that yield lower rewards or higher losses. We empirically demonstrate that the R-Weighted methods work superior to the state-of-the-art in multiple RL environments.