Reinforcement Learning (RL) has been used to finetune Large Language Models (LLMs) using a reward model trained from preference data, to better align with human judgment. The recently introduced direct alignment methods, which are often simpler, more stable, and computationally lighter, can more directly achieve this. However, these approaches cannot optimize arbitrary rewards, and the preference-based ones are not the only rewards of interest for LLMs (eg, unit tests for code generation or textual entailment for summarization, among others). RL-finetuning is usually done with a variation of policy gradient, which calls for on-policy or near-on-policy samples, requiring costly generations. We introduce *Contrastive Policy Gradient*, or CoPG, a simple and mathematically principled new RL algorithm that can estimate the optimal policy even from off-policy data. It can be seen as an off-policy policy gradient approach that does not rely on important sampling techniques and highlights the importance of using (the right) state baseline. We show this approach to generalize the direct alignment method IPO (identity preference optimization) and classic policy gradient. We experiment with the proposed CoPGon a toy bandit problem to illustrate its properties, as well as for finetuning LLMs on a summarization task, using a learned reward function considered as ground truth for the purpose of the experiments.

While reinforcement learning (RL) has been proven essential for tuning large language models (LLMs), it can lead to reward over-optimization (ROO). Existing approaches address ROO by adding KL regularization, requiring computationally expensive hyperparameter tuning. Additionally, KL regularization focuses solely on regularizing the language policy, neglecting a potential source of regularization: the reward function itself. Inspired by demonstration-guided RL, we here introduce the Reward Calibration from Demonstration (RCfD), which leverages human demonstrations and a reward model to recalibrate the reward objective. Formally, given a prompt, the RCfD objective minimizes the distance between the demonstrations’ and LLM’s rewards rather than directly maximizing the reward function. This objective shift avoids incentivizing the LLM to exploit the reward model and promotes more natural and diverse language generation.We show the effectiveness of RCfD in three RL language tasks, where it achieves comparable performance to carefully tuned baselines while mitigating ROO.

This paper introduces TRUncated ReinForcement Learning for Language (TrufLL), an original approach to train conditional languagemodels without a supervised learning phase, by only using reinforcement learning (RL). As RL methods unsuccessfully scale to large action spaces, we dynamically truncate the vocabulary space using a generic language model. TrufLL thus enables to train a language agent by solely interacting with its environment without any task-specific prior knowledge; it is only guided with a task-agnostic language model. Interestingly, this approach avoids the dependency to labelled datasets and inherently reduces pretrained policy flaws such as language or exposure biases. We evaluate TrufLL on two visual question generation tasks, for which we report positive results over performance and language metrics, which we then corroborate with a human evaluation. To our knowledge, it is the first approach that successfully learns a language generation policy without pre-training, using only reinforcement learning.

Language drift has been one of the major obstacles to train language models through interaction. When word-based conversational agents are trained towards completing a task, they tend to invent their language rather than leveraging natural language. In recent literature, two general methods partially counter this phenomenon: Supervised Selfplay (S2P) and Seeded Iterated Learning (SIL). While S2P jointly trains interactive and supervised losses to counter the drift, SIL changes the training dynamics to prevent language drift from occurring. In this paper, we first highlight their respective weaknesses, i.e., late-stage training collapses and higher negative likelihood when evaluated on human corpus. Given these observations, we introduce Supervised Seeded Iterated Learning (SSIL) to combine both methods to minimize their respective weaknesses. We then show the effectiveness of in the language-drift translation game.