Recent works have shown increasing interest in using natural language-based user profiles for recommender systems, as they offer greater transparency and interpretability compared to traditional embedding-based methods. Most existing approaches rely on zero-shot inference with large language models (LLMs) to generate these profiles, but the resulting quality remains insufficient, leading to suboptimal recommendation performance. In this paper, we present LangPTune, the first end-to-end training framework designed to directly optimize LLM-generated user profiles for recommendation tasks. By explicitly training the LLM for the recommendation objective, our approach significantly outperforms zero-shot baselines. Evaluations across training setups and benchmarks show that LangPTune not only exceeds the performance of zero-shot methods but also matches the performance of state-of-the-art embedding-based baselines. Additionally, we assess whether our training framework maintains the interpretability of user profiles, using both GPT-4 simulations and crowdworker studies.

Large Language Models (LLMs) have achieved remarkable success at tasks like summarization that involve a single turn of interaction. However, they can still struggle with multi-turn tasks like dialogue that require long-term planning. Previous works on multi-turn dialogue extend single-turn reinforcement learning from human feedback (RLHF) methods to the multi-turn setting by treating all prior dialogue turns as a long context. Such approaches suffer from covariate shift: the conversations in the training set have previous turns generated by some reference policy, which means that low training error may not necessarily correspond to good performance when the learner is actually in the conversation loop. In response, we introduce REgressing the RELative FUture (REFUEL), an efficient policy optimization approach designed to address multi-turn RLHF in LLMs. REFUEL employs a single model to estimate $Q$-values and trains on self-generated data, addressing the covariate shift issue. REFUEL frames the multi-turn RLHF problem as a sequence of regression tasks on iteratively collected datasets, enabling ease of implementation. Theoretically, we prove that REFUEL can match the performance of any policy covered by the training set. Empirically, we evaluate our algorithm by using Llama-3.1-70B-it to simulate a user in conversation with our model. REFUEL consistently outperforms state-of-the-art methods such as DPO and REBEL across various settings. Furthermore, despite having only 8 billion parameters, Llama-3-8B-it fine-tuned with REFUEL outperforms Llama-3.1-70B-it on long multi-turn dialogues. Implementation of REFUEL can be found at https://github.com/ZhaolinGao/REFUEL/, and models trained by REFUEL can be found at https://huggingface.co/Cornell-AGI.

Reinforcement learning (RL) has emerged as a powerful tool for fine-tuning large language models (LLMs) to improve complex reasoning abilities. However, state-of-the-art policy optimization methods often suffer from high computational overhead and memory consumption, primarily due to the need for multiple generations per prompt and the reliance on critic networks or advantage estimates of the current policy. In this paper, we propose $A^\star$-PO, a novel two-stage policy optimization framework that directly approximates the optimal advantage function and enables efficient training of LLMs for reasoning tasks. In the first stage, we leverage offline sampling from a reference policy to estimate the optimal value function $V^\star$, eliminating the need for costly online value estimation. In the second stage, we perform on-policy updates using a simple least-squares regression loss with only a single generation per prompt. Theoretically, we establish performance guarantees and prove that the KL-regularized RL objective can be optimized without requiring complex exploration strategies. Empirically, $A^\star$-PO achieves competitive performance across a wide range of mathematical reasoning benchmarks, while reducing training time by up to 2$\times$ and peak memory usage by over 30\% compared to PPO, GRPO, and REBEL. Implementation of $A^\star$-PO can be found at https://github.com/ZhaolinGao/A-PO.

Hidden Markov Models (HMMs) are fundamental tools for modeling sequential data with latent states that follow Markovian dynamics. However, they present significant challenges in model fitting and computational efficiency on real-world datasets. In this work, we demonstrate that pre-trained large language models (LLMs) can effectively model data generated by HMMs through in-context learning (ICL) — their ability to learn patterns from examples within the input context. We evaluate LLMs' performance on diverse synthetic HMMs, showing that their prediction accuracy converges to the theoretical optimum. We discover novel scaling trends influenced by HMM properties and provide theoretical conjectures for these empirical observations. Furthermore, we present practical guidelines for scientists on using ICL as a diagnostic tool for complex data. Applied to real-world animal decision-making tasks, ICL achieves competitive performance with models designed by human experts. Our results demonstrate potential for advancing understanding of LLMs' capabilities while opening new avenues for scientific discovery of biological mechanisms and hidden structures in real-world phenomena.

In this paper, we propose a simple and efficient method for value model training on long-context reasoning traces. Compared to existing process reward models (PRMs), our method does not require a fine-grained notion of ``step,'' which is difficult to define for long-context reasoning models. By collecting a dataset of 2.5 million reasoning traces, we train a 1.5B token-level value model and apply it to DeepSeek models for improved performance with test-time compute scaling. We find that block-wise value-guided search (\texttt{VGS}) with a final weighted majority vote achieves better test-time scaling than standard methods such as majority voting or best-of-$n$. Moreover, \texttt{VGS} significantly reduces the inference FLOPs required to achieve the same performance of majority voting. Our dataset, model and codebase are open-sourced at \codeurl.

While originally developed for continuous control problems, Proximal Policy Optimization (PPO) has emerged as the work-horse of a variety of reinforcement learning (RL) applications, including the fine-tuning of generative models. Unfortunately, PPO requires multiple heuristics to enable stable convergence (e.g. value networks, clipping), and is notorious for its sensitivity to the precise implementation of these components. In response, we take a step back and ask what a *minimalist* RL algorithm for the era of generative models would look like. We propose REBEL, an algorithm that cleanly reduces the problem of policy optimization to regressing the *relative reward* between two completions to a prompt in terms of the policy, enabling strikingly lightweight implementation. In theory, we prove that fundamental RL algorithms like Natural Policy Gradient can be seen as variants of REBEL, which allows us to match the strongest known theoretical guarantees in terms of convergence and sample complexity in the RL literature. REBEL can also cleanly incorporate offline data and be extended to handle the intransitive preferences we frequently see in practice. Empirically, we find that REBEL provides a unified approach to language modeling and image generation with stronger or similar performance as PPO and DPO, all while being simpler to implement and more computationally efficient than PPO. When fine-tuning Llama-3-8B-Instruct, REBEL achieves strong performance in AlpacaEval 2.0, MT-Bench, and Open LLM Leaderboard. Implementation of REBEL can be found at <https://github.com/ZhaolinGao/REBEL>, and models trained by REBEL can be found at <https://huggingface.co/Cornell-AGI>.

Online recommendation systems are prone to create filter bubbles, whereby users are only recommended content narrowly aligned with their historical interests. In the case of media recommendation, this can reinforce political polarization by recommending topical content (e.g., on the economy) at one extreme end of the political spectrum even though this topic has broad coverage from multiple political viewpoints that would provide a more balanced and informed perspective for the user. Historically, Maximal Marginal Relevance (MMR) has been used to diversify result lists and even mitigate filter bubbles, but suffers from three key drawbacks: (1)~MMR directly sacrifices relevance for diversity, (2)~MMR typically diversifies across all content and not just targeted dimensions (e.g., political polarization), and (3)~MMR is inefficient in practice due to the need to compute pairwise similarities between recommended items. To simultaneously address these limitations, we propose a novel methodology that trains Concept Activation Vectors (CAVs) for targeted topical dimensions (e.g., political polarization). We then modulate the latent embeddings of user preferences in a state-of-the-art VAE-based recommender system to diversify along the targeted dimension while preserving topical relevance across orthogonal dimensions. Our experiments show that our Targeted Diversification VAE-based Collaborative Filtering (TD-VAE-CF) methodology better preserves relevance of content to user preferences across a range of diversification levels in comparison to both untargeted and targeted variations of Maximum Marginal Relevance (MMR); TD-VAE-CF is also much more computationally efficient than the post-hoc re-ranking approach of MMR.

The majority of recent work in latent Collaborative Filtering (CF) has focused on developing new model architectures to learn accurate user and item representations. Typically, a standard pairwise loss function (BPR, Triplet, etc.) is used in these models, and little exploration is done on how to optimally extract signals from the available preference information. In the implicit setting, negative examples are sampled, and these losses allocate weights that solely depend on the difference in user distance between observed (positive) and negative item pairs. This can ignore valuable global information from other users and items, and lead to sub-optimal results. Motivated by this problem, we propose a novel loss which first leverages mining to select the most informative pairs, followed by a weighing process to allocate more weight to harder examples. Our weighting process consists of four different components, and incorporates distance information from other users, enabling the model to better position the learned representations. We conduct extensive experiments and demonstrate that our loss can be applied to different types of CF models leading to significant gains with each type. In particular, by applying our loss to the graph convolutional architecture, we achieve new state-of-the-art results on four different datasets. Further analysis shows that through our loss the model is able to learn better user-item representation space compared to other losses. Full code for this work is available here: https://github.com/layer6ai-labs/MCL.

Graph-based semi-supervised learning has been shown to be one of the most effective classification approaches, as it can exploit connectivity patterns between labeled and unlabeled samples to improve learning performance. However, we show that existing techniques perform poorly when labeled data are severely limited. To address the problem of semi-supervised learning in the presence of severely limited labeled samples, we propose a new framework, called Shoestring, that incorporates metric learning into the paradigm of graph-based semi-supervised learning. In particular, our base model consists of a graph embedding network, followed by a metric learning network that learns a semantic metric space to represent the semantic similarity between the sparsely labeled and large numbers of unlabeled samples. Then the classification can be performed by clustering the unlabeled samples according to the learned semantic space. We empirically demonstrate Shoestring's superiority over many baselines, including graph convolutional networks, label propagation and their recent label-efficient variations (IGCN and GLP). We show that our framework achieves state-of-the-art performance for node classification in the low-data regime. In addition, we demonstrate the effectiveness of our framework on image classification tasks in the few-shot learning regime, with significant gains on miniImageNet (2.57% ~ 3.59%) and tieredImageNet (1.05% ~ 2.70%).

In modern online social networks, each user is typically able to provide a value to indicate how trustworthy their direct friends are. Inferring such a value of social trust between any pair of nodes in online social networks is useful in a wide variety of applications, such as online marketing and recommendation systems. However, it is challenging to accurately and efficiently evaluate social trust between a pair of users in online social networks. Existing works either designed handcrafted rules that rely on specialized domain knowledge, or required a significant amount of computation resources, which affected their scalability.In recent years, graph convolutional neural networks (GCNs) have been shown to be powerful in learning on graph data. Their advantages provide great potential to trust evaluation as social trust can be represented as graph data. In this paper, we propose Guardian, a new end-to-end framework that learns latent factors in social trust with GCNs. Guardian is designed to incorporate social network structures and trust relationships to estimate social trust between any two users. Extensive experimental results demonstrated that Guardian can speedup trust evaluation by up to 2, 827 × with comparable accuracy, as compared to the stateof-the-art in the literature.