DSTC: Direct Preference Learning with Only Self-Generated Tests and Code to Improve Code LMs

Direct preference learning offers a promising and computation-efficient beyond supervised fine-tuning (SFT) for improving code generation in coding large language models (LMs). However, the scarcity of reliable preference data is a bottleneck for the performance of direct preference learning to improve the coding accuracy of code LMs. In this paper, we introduce \underline{\textbf{D}}irect Preference Learning with Only \underline{\textbf{S}}elf-Generated \underline{\textbf{T}}ests and \underline{\textbf{C}}ode (DSTC), a framework that leverages only self-generated code snippets and tests to construct reliable preference pairs such that direct preference learning can improve LM coding accuracy without external annotations. DSTC combines a minimax selection process and test-code concatenation to improve preference pair quality, reducing the influence of incorrect self-generated tests and enhancing model performance without the need for costly reward models. When applied with direct preference learning methods such as Direct Preference Optimization (DPO) and Kahneman-Tversky Optimization (KTO), DSTC yields stable improvements in coding accuracy (pass@1 score) across diverse coding benchmarks, including HumanEval, MBPP, and BigCodeBench, demonstrating both its effectiveness and scalability for models of various sizes. This approach autonomously enhances code generation accuracy across LLMs of varying sizes, reducing reliance on expensive annotated coding datasets.

Language-Model-Assisted Bi-Level Programming for Reward Learning from Internet Videos

Learning from Demonstrations, particularly from biological experts like humans and animals, often encounters significant data acquisition challenges. While recent approaches leverage internet videos for learning, they require complex, task-specific pipelines to extract and retarget motion data for the agent. In this work, we introduce a language-model-assisted bi-level programming framework that enables a reinforcement learning agent to directly learn its reward from internet videos, bypassing dedicated data preparation. The framework includes two levels: an upper level where a vision-language model (VLM) provides feedback by comparing the learner's behavior with expert videos, and a lower level where a large language model (LLM) translates this feedback into reward updates. The VLM and LLM collaborate within this bi-level framework, using a "chain rule" approach to derive a valid search direction for reward learning. We validate the method for reward learning from YouTube videos, and the results have shown that the proposed method enables efficient reward design from expert videos of biological agents for complex behavior synthesis.

Reward-Augmented Data Enhances Direct Preference Alignment of LLMs

Preference alignment in Large Language Models (LLMs) has significantly improved their ability to adhere to human instructions and intentions. However, existing direct alignment algorithms primarily focus on relative preferences and often overlook the qualitative aspects of responses. Striving to maximize the implicit reward gap between the chosen and the slightly inferior rejected responses can cause overfitting and unnecessary unlearning of the high-quality rejected responses. The unawareness of the reward scores also drives the LLM to indiscriminately favor the low-quality chosen responses and fail to generalize to responses with the highest rewards, which are sparse in data. To overcome these shortcomings, our study introduces reward-conditioned LLM policies that discern and learn from the entire spectrum of response quality within the dataset, helping extrapolate to more optimal regions. We propose an effective yet simple data relabeling method that conditions the preference pairs on quality scores to construct a reward-augmented dataset. This dataset is easily integrated with existing direct alignment algorithms and is applicable to any preference dataset. The experimental results across instruction-following benchmarks including AlpacaEval, MT-Bench, and Arena-Hard-Auto demonstrate that our approach consistently boosts the performance of DPO by a considerable margin across diverse models. Additionally, our method improves the average accuracy on various academic benchmarks. When applying our method to on-policy data, the resulting DPO model achieves SOTA results on AlpacaEval. Through ablation studies, we demonstrate that our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere dataset expansion. Our code is available at https://github.com/shenao-zhang/reward-augmented-preference.

Just say what you want: only-prompting self-rewarding online preference optimization

We address the challenge of online Reinforcement Learning from Human Feedback (RLHF) with a focus on self-rewarding alignment methods. In online RLHF, obtaining feedback requires interaction with the environment, which can be costly when using additional reward models or the GPT-4 API. Current self-rewarding approaches rely heavily on the discriminator's judgment capabilities, which are effective for large-scale models but challenging to transfer to smaller ones. To address these limitations, we propose a novel, only-prompting self-rewarding online algorithm that generates preference datasets without relying on judgment capabilities. Additionally, we employ fine-grained arithmetic control over the optimality gap between positive and negative examples, generating more hard negatives in the later stages of training to help the model better capture subtle human preferences. Finally, we conduct extensive experiments on two base models, Mistral-7B and Mistral-Instruct-7B, which significantly bootstrap the performance of the reference model, achieving 34.5% in the Length-controlled Win Rates of AlpacaEval 2.0.

Safe MPC Alignment with Human Directional Feedback

In safety-critical robot planning or control, manually specifying safety constraints or learning them from demonstrations can be challenging. In this paper, we propose a certifiable alignment method for a robot to learn a safety constraint in its model predictive control (MPC) policy with human online directional feedback. To our knowledge, it is the first method to learn safety constraints from human feedback. The proposed method is based on an empirical observation: human directional feedback, when available, tends to guide the robot toward safer regions. The method only requires the direction of human feedback to update the learning hypothesis space. It is certifiable, providing an upper bound on the total number of human feedback in the case of successful learning of safety constraints, or declaring the misspecification of the hypothesis space, i.e., the true implicit safety constraint cannot be found within the specified hypothesis space. We evaluated the proposed method using numerical examples and user studies in two developed simulation games. Additionally, we implemented and tested the proposed method on a real-world Franka robot arm performing mobile water-pouring tasks in a user study. The simulation and experimental results demonstrate the efficacy and efficiency of our method, showing that it enables a robot to successfully learn safety constraints with a small handful (tens) of human directional corrections.

Toward Optimal LLM Alignments Using Two-Player Games

The standard Reinforcement Learning from Human Feedback (RLHF) framework primarily focuses on optimizing the performance of large language models using pre-collected prompts. However, collecting prompts that provide comprehensive coverage is both tedious and challenging, and often fails to include scenarios that LLMs need to improve on the most. In this paper, we investigate alignment through the lens of two-agent games, involving iterative interactions between an adversarial and a defensive agent. The adversarial agent's task at each step is to generate prompts that expose the weakness of the defensive agent. In return, the defensive agent seeks to improve its responses to these newly identified prompts it struggled with, based on feedback from the reward model. We theoretically demonstrate that this iterative reinforcement learning optimization converges to a Nash Equilibrium for the game induced by the agents. Experimental results in safety scenarios demonstrate that learning in such a competitive environment not only fully trains agents but also leads to policies with enhanced generalization capabilities for both adversarial and defensive agents.

Self-Exploring Language Models: Active Preference Elicitation for Online Alignment

Preference optimization, particularly through Reinforcement Learning from Human Feedback (RLHF), has achieved significant success in aligning Large Language Models (LLMs) to adhere to human intentions. Unlike offline alignment with a fixed dataset, online feedback collection from humans or AI on model generations typically leads to more capable reward models and better-aligned LLMs through an iterative process. However, achieving a globally accurate reward model requires systematic exploration to generate diverse responses that span the vast space of natural language. Random sampling from standard reward-maximizing LLMs alone is insufficient to fulfill this requirement. To address this issue, we propose a bilevel objective optimistically biased towards potentially high-reward responses to actively explore out-of-distribution regions. By solving the inner-level problem with the reparameterized reward function, the resulting algorithm, named Self-Exploring Language Models (SELM), eliminates the need for a separate RM and iteratively updates the LLM with a straightforward objective. Compared to Direct Preference Optimization (DPO), the SELM objective reduces indiscriminate favor of unseen extrapolations and enhances exploration efficiency. Our experimental results demonstrate that when finetuned on Zephyr-7B-SFT and Llama-3-8B-Instruct models, SELM significantly boosts the performance on instruction-following benchmarks such as MT-Bench and AlpacaEval 2.0, as well as various standard academic benchmarks in different settings. Our code and models are available at https://github.com/shenao-zhang/SELM.

Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer

Aligning generative models with human preference via RLHF typically suffers from overoptimization, where an imperfectly learned reward model can misguide the generative model to output undesired responses. We investigate this problem in a principled manner by identifying the source of the misalignment as a form of distributional shift and uncertainty in learning human preferences. To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model; one that simultaneously minimizes the maximum likelihood estimation of the loss and a reward penalty term. Here, the reward penalty term is introduced to prevent the policy from choosing actions with spurious high proxy rewards, resulting in provable sample efficiency of the algorithm under a partial coverage style condition. Moving from theory to practice, the proposed algorithm further enjoys an equivalent but surprisingly easy-to-implement reformulation. Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines: (i) a preference optimization loss that directly aligns the policy with human preference, and (ii) a supervised learning loss that explicitly imitates the policy with a (suitable) baseline distribution. In the context of aligning large language models (LLM), this objective fuses the direct preference optimization (DPO) loss with the supervised fune-tuning (SFT) loss to help mitigate the overoptimization towards undesired responses, for which we name the algorithm Regularized Preference Optimization (RPO). Experiments of aligning LLMs demonstrate the improved performance of RPO compared with DPO baselines. Our work sheds light on the interplay between preference optimization and SFT in tuning LLMs with both theoretical guarantees and empirical evidence.

A Mean-Field Analysis of Neural Gradient Descent-Ascent: Applications to Functional Conditional Moment Equations

We study minimax optimization problems defined over infinite-dimensional function classes. In particular, we restrict the functions to the class of overparameterized two-layer neural networks and study (i) the convergence of the gradient descent-ascent algorithm and (ii) the representation learning of the neural network. As an initial step, we consider the minimax optimization problem stemming from estimating a functional equation defined by conditional expectations via adversarial estimation, where the objective function is quadratic in the functional space. For this problem, we establish convergence under the mean-field regime by considering the continuous-time and infinite-width limit of the optimization dynamics. Under this regime, gradient descent-ascent corresponds to a Wasserstein gradient flow over the space of probability measures defined over the space of neural network parameters. We prove that the Wasserstein gradient flow converges globally to a stationary point of the minimax objective at a $\mathcal{O}(T^{-1} + \alpha^{-1} ) $ sublinear rate, and additionally finds the solution to the functional equation when the regularizer of the minimax objective is strongly convex. Here $T$ denotes the time and $\alpha$ is a scaling parameter of the neural network. In terms of representation learning, our results show that the feature representation induced by the neural networks is allowed to deviate from the initial one by the magnitude of $\mathcal{O}(\alpha^{-1})$, measured in terms of the Wasserstein distance. Finally, we apply our general results to concrete examples including policy evaluation, nonparametric instrumental variable regression, and asset pricing.

Advancing Object Goal Navigation Through LLM-enhanced Object Affinities Transfer

In object goal navigation, agents navigate towards objects identified by category labels using visual and spatial information. Previously, solely network-based methods typically rely on historical data for object affinities estimation, lacking adaptability to new environments and unseen targets. Simultaneously, employing Large Language Models (LLMs) for navigation as either planners or agents, though offering a broad knowledge base, is cost-inefficient and lacks targeted historical experience. Addressing these challenges, we present the LLM-enhanced Object Affinities Transfer (LOAT) framework, integrating LLM-derived object semantics with network-based approaches to leverage experiential object affinities, thus improving adaptability in unfamiliar settings. LOAT employs a dual-module strategy: a generalized affinities module for accessing LLMs' vast knowledge and an experiential affinities module for applying learned object semantic relationships, complemented by a dynamic fusion module harmonizing these information sources based on temporal context. The resulting scores activate semantic maps before feeding into downstream policies, enhancing navigation systems with context-aware inputs. Our evaluations in AI2-THOR and Habitat simulators demonstrate improvements in both navigation success rates and efficiency, validating the LOAT's efficacy in integrating LLM insights for improved object goal navigation.