Neural Information Processing Systems http://nips.cc/

Direct alignment methods typically train large language models (LLMs) by contrasting the likelihoods of preferred and dispreferred responses. While effective at capturing relative preferences, these methods are widely observed to suppress the absolute likelihoods of example responses. As a result, aligned models can deviate from expected patterns, exhibiting rewar-hacking effect even without an explicit reward model. This fundamental limitation of contrastive alignment, which we term likelihood underdetermination, motivates us to revisit direct preference optimization (DPO) -- the seminal direct alignment method. Interestingly, we show that the DPO loss admits a principled decomposition. The reformulated loss not only extends naturally to a broader range of feedback types, but also unveils the root cause of likelihood underdetermination. Specifically, we identify that standard DPO implicitly oversimplifies a regularizer in the reformulated loss; restoring this full term effectively resolves the underdetermination. Building on these insights, we introduce PRoximalized PReference Optimization (PRO), a unified alignment method that accommodates diverse feedback types while eliminating likelihood underdetermination through an efficient approximation of the full regularizer. Empirical evaluations demonstrate the consistent superiority of PRO over existing methods across pairwise, binary and scalar feedback.

We investigate the problem of learning a Lipschitz function from binary feedback.

Although potential attacks against OL TR algorithms may cause serious losses in real-world applications, there is limited knowledge about adversarial attacks on OL TR. This paper studies attack strategies against multiple variants of OL TR.

Learning to plan in grounded environments typically requires carefully designed reward functions or high-quality annotated demonstrations. Recent works show that pretrained foundation models, such as large language models (LLMs) and vision language models (VLMs), capture background knowledge helpful for planning, which reduces the amount of reward design and demonstrations needed for policy learning. We evaluate how well LLMs and VLMs provide feedback across symbolic, language, and continuous control environments. We consider prominent types of feedback for planning including binary feedback, preference feedback, action advising, goal advising, and delta action feedback. We also consider inference methods that impact feedback performance, including in-context learning, chain-of-thought, and access to environment dynamics. We find that foundation models can provide diverse high-quality feedback across domains. Moreover, larger and reasoning models consistently provide more accurate feedback, exhibit less bias, and benefit more from enhanced inference methods. Finally, feedback quality degrades for environments with complex dynamics or continuous state spaces and action spaces.

State-of-the-art T2I models are capable of generating high-resolution images given textual prompts. However, they still struggle with accurately depicting compositional scenes that specify multiple objects, attributes, and spatial relations. We present CompAlign, a challenging benchmark with an emphasis on assessing the depiction of 3D-spatial relationships, for evaluating and improving models on compositional image generation. CompAlign consists of 900 complex multi-subject image generation prompts that combine numerical and 3D-spatial relationships with varied attribute bindings. Our benchmark is remarkably challenging, incorporating generation tasks with 3+ generation subjects with complex 3D-spatial relationships. Additionally, we propose CompQuest, an interpretable and accurate evaluation framework that decomposes complex prompts into atomic sub-questions, then utilizes a MLLM to provide fine-grained binary feedback on the correctness of each aspect of generation elements in model-generated images. This enables precise quantification of alignment between generated images and compositional prompts. Furthermore, we propose an alignment framework that uses CompQuest's feedback as preference signals to improve diffusion models' compositional image generation abilities. Using adjustable per-image preferences, our method is easily scalable and flexible for different tasks. Evaluation of 9 T2I models reveals that: (1) models remarkable struggle more with compositional tasks with more complex 3D-spatial configurations, and (2) a noticeable performance gap exists between open-source accessible models and closed-source commercial models. Further empirical study on using CompAlign for model alignment yield promising results: post-alignment diffusion models achieve remarkable improvements in compositional accuracy, especially on complex generation tasks, outperforming previous approaches.

Standard reinforcement learning (RL) assumes that an agent can observe a reward for each state-action pair. However, in practical applications, it is often difficult and costly to collect a reward for each state-action pair. While there have been several works considering RL with trajectory feedback, it is unclear if trajectory feedback is inefficient for learning when trajectories are long. In this work, we consider a model named RL with segment feedback, which offers a general paradigm filling the gap between per-state-action feedback and trajectory feedback. In this model, we consider an episodic Markov decision process (MDP), where each episode is divided into $m$ segments, and the agent observes reward feedback only at the end of each segment. Under this model, we study two popular feedback settings: binary feedback and sum feedback, where the agent observes a binary outcome and a reward sum according to the underlying reward function, respectively. To investigate the impact of the number of segments $m$ on learning performance, we design efficient algorithms and establish regret upper and lower bounds for both feedback settings. Our theoretical and experimental results show that: under binary feedback, increasing the number of segments $m$ decreases the regret at an exponential rate; in contrast, surprisingly, under sum feedback, increasing $m$ does not reduce the regret significantly.

By pretraining on trillions of tokens, an LLM gains the capability of text generation. However, to enhance its utility and reduce potential harm, SFT and alignment are applied sequentially to the pretrained model. Due to the differing nature and objective functions of SFT and alignment, catastrophic forgetting has become a significant issue. To address this, we introduce Unified Fine-Tuning (UFT), which integrates SFT and alignment into a single training stage using the same objective and loss functions through an implicit reward function. Our experimental results demonstrate that UFT outperforms SFT on instruction-tuning data alone. Moreover, when combining instruction-tuning data with alignment data, UFT effectively prevents catastrophic forgetting across these two stages and shows a clear advantage over sequentially applying SFT and alignment. This is evident in the significant improvements observed in the \textbf{ifeval} task for instruction-following and the \textbf{truthful-qa} task for factuality. The proposed general fine-tuning framework UFT establishes an effective and efficient pretraining-UFT paradigm for LLM training.

Contextual dueling bandit is used to model the bandit problems, where a learner's goal is to find the best arm for a given context using observed noisy preference feedback over the selected arms for the past contexts. However, existing algorithms assume the reward function is linear, which can be complex and non-linear in many real-life applications like online recommendations or ranking web search results. To overcome this challenge, we use a neural network to estimate the reward function using preference feedback for the previously selected arms. We propose upper confidence bound- and Thompson sampling-based algorithms with sub-linear regret guarantees that efficiently select arms in each round. We then extend our theoretical results to contextual bandit problems with binary feedback, which is in itself a non-trivial contribution. Experimental results on the problem instances derived from synthetic datasets corroborate our theoretical results.