AI alignment in the shape of Reinforcement Learning from Human Feedback (RLHF) is increasingly treated as a crucial ingredient for high performance large language models. Proximal Policy Optimization (PPO) has been installed by the seminal literature as the standard method for the RL part of RLHF. However, it involves both high computational cost and sensitive hyperparameter tuning. We posit that most of the motivational principles that led to the development of PPO are less of a practical concern in RLHF and advocate for a less computationally expensive method that preserves and even increases performance. We revisit how alignment from human preferences is formulated in the context of RL. Keeping simplicity as a guiding principle, we show that many components of PPO are unnecessary in an RLHF context and that far simpler REINFORCE-style optimization variants outperform both PPO and newly proposed “RL-free” methods such as DPO and RAFT. Our work suggests that careful adaptation to LLMs alignment characteristics allows benefiting from online RL optimization at low cost.

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.

Despite the seeming success of contemporary grounded text generation systems, they often tend to generate factually inconsistent text with respect to their input. This phenomenon is emphasized in tasks like summarization, in which the generated summaries should be corroborated by their source article. In this work we leverage recent progress on textual entailment models to directly address this problem for abstractive summarization systems. We use reinforcement learning with reference-free, textual-entailment rewards to optimize for factual consistency and explore the ensuing trade-offs, as improved consistency may come at the cost of less informative or more extractive summaries. Our results, according to both automatic metrics and human evaluation, show that our method considerably improves the faithfulness, salience and conciseness of the generated summaries.

We introduce SPEAR-TTS, a multi-speaker text-to-speech (TTS) system that can be trained with minimal supervision. By combining two types of discrete speech representations, we cast TTS as a composition of two sequence-to-sequence tasks: from text to high-level semantic tokens (akin to “reading”) and from semantic tokens to low-level acoustic tokens (“speaking”). Decoupling these two tasks enables training of the “speaking” module using abundant audio-only data, and unlocks the highly efficient combination of pretraining and backtranslation to reduce the need for parallel data when training the “reading” component. To control the speaker identity, we adopt example prompting, which allows SPEAR-TTS to generalize to unseen speakers using only a short sample of 3 seconds, without any explicit speaker representation or speaker labels. Our experiments demonstrate that SPEAR-TTS achieves a character error rate that is competitive with state-of-the-art methods using only 15 minutes of parallel data, while matching ground-truth speech in naturalness and acoustic quality.

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.

We present MultiVec, a new toolkit for computing continuous representations for text at different granularity levels (word-level or sequences of words). MultiVec includes word2vec’s features, paragraph vector (batch and online) and bivec for bilingual distributed representations. MultiVec also includes different distance measures between words and sequences of words. The toolkit is written in C++ and is aimed at being fast (in the same order of magnitude as word2vec), easy to use, and easy to extend. It has been evaluated on several NLP tasks: the analogical reasoning task, sentiment analysis, and crosslingual document classification.

This paper describes a French Spoken Dialogue System (SDS) named NASTIA (Negotiating Appointment SeTting InterfAce). Appointment scheduling is a hybrid task halfway between slot-filling and negotiation. NASTIA implements three different negotiation strategies. These strategies were tested on 1734 dialogues with 385 users who interacted at most 5 times with the SDS and gave a rating on a scale of 1 to 10 for each dialogue. Previous appointment scheduling systems were evaluated with the same experimental protocol. NASTIA is different from these systems in that it can adapt its strategy during the dialogue. The highest system task completion rate with these systems was 81% whereas NASTIA had an 88% average and its best performing strategy even reached 92%. This strategy also significantly outperformed previous systems in terms of overall user rating with an average of 8.28 against 7.40. The experiment also enabled highlighting global recommendations for building spoken dialogue systems.

This paper describes the DINASTI (DIalogues with a Negotiating Appointment SeTting Interface) corpus, which is composed of 1734 dialogues with the French spoken dialogue system NASTIA (Negotiating Appointment SeTting InterfAce). NASTIA is a reinforcement learning-based system. The DINASTI corpus was collected while the system was following a uniform policy. Each entry of the corpus is a system-user exchange annotated with 120 automatically computable features. The corpus contains a total of 21587 entries, with 385 testers. Each tester performed at most five scenario-based interactions with NASTIA. The dialogues last an average of 10.82 dialogue turns, with 4.45 reinforcement learning decisions. The testers filled an evaluation questionnaire after each dialogue. The questionnaire includes three questions to measure task completion. In addition, it comprises 7 Likert-scaled items evaluating several aspects of the interaction, a numerical overall evaluation on a scale of 1 to 10, and a free text entry. Answers to this questionnaire are provided with DINASTI. This corpus is meant for research on reinforcement learning modelling for dialogue management.

Dans cet article, un environnement modulaire pour la simulation automatique de dialogues homme-machine est proposé. Cet environnement comprend notamment un modèle d’utilisateur consistant dirigé par le but et un module de simulation de compréhension de parole. Un réseau bayésien est à la base de ces deux modèles et selon les paramètres utilisés, il peut générer un comportement d’utilisateur cohérent ou servir de classificateur de concepts. L’environnement a été utilisé dans le contexte de l’optimisation de stratégies de dialogue sur une tâche simple de remplissage de formulaire et les résultats montrent qu’il est alors possible d’identifier certains dialogues problématiques du point de vue de la compréhension.