Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. However, traditional RM training, which relies on response pairs tied to specific prompts, struggles to disentangle prompt-driven preferences from prompt-independent artifacts, such as response length and format. In this work, we expose a fundamental limitation of current RM training methods, where RMs fail to effectively distinguish between contextual signals and irrelevant artifacts when determining preferences. To address this, we introduce a causal framework that learns preferences independent of these artifacts and propose a novel data augmentation technique designed to eliminate them. Extensive experiments show that our approach successfully filters out undesirable artifacts, yielding a more robust reward model (RRM). Our RRM improves the performance of a pairwise reward model trained on Gemma-2-9b-it, on Reward-Bench, increasing accuracy from 80.61% to 84.15%. Additionally, we train two DPO policies using both the RM and RRM, demonstrating that the RRM significantly enhances DPO-aligned policies, improving MT-Bench scores from 7.27 to 8.31 and length-controlled win-rates in AlpacaEval-2 from 33.46% to 52.49%.

Recent studies have shown that large language models' (LLMs) mathematical problem-solving capabilities can be enhanced by integrating external tools, such as code interpreters, and employing multi-turn Chain-of-Thought (CoT) reasoning. While current methods focus on synthetic data generation and Supervised Fine-Tuning (SFT), this paper studies the complementary direct preference learning approach to further improve model performance. However, existing direct preference learning algorithms are originally designed for the single-turn chat task, and do not fully address the complexities of multi-turn reasoning and external tool integration required for tool-integrated mathematical reasoning tasks. To fill in this gap, we introduce a multi-turn direct preference learning framework, tailored for this context, that leverages feedback from code interpreters and optimizes trajectory-level preferences. This framework includes multi-turn DPO and multi-turn KTO as specific implementations. The effectiveness of our framework is validated through training of various language models using an augmented prompt set from the GSM8K and MATH datasets. Our results demonstrate substantial improvements: a supervised fine-tuned Gemma-1.1-it-7B model's performance increased from 77.5% to 83.9% on GSM8K and from 46.1% to 51.2% on MATH. Similarly, a Gemma-2-it-9B model improved from 84.1% to 86.3% on GSM8K and from 51.0% to 54.5% on MATH.

Existing preference optimization methods often assume scenarios where paired preference feedback (preferred/positive vs. dis-preferred/negative examples) is available. This requirement limits their applicability in scenarios where only unpaired feedback—for example, either positive or negative— is available. To address this, we introduce a novel approach that decouples learning from positive and negative feedback. This decoupling enables control over the influence of each feedback type and, importantly, allows learning even when only one feedback type is present. A key contribution is demonstrating stable learning from negative feedback alone, a capability not well-addressed by current methods. Our approach builds upon the probabilistic framework introduced in (Dayan and Hinton, 1997), which uses expectation-maximization (EM) to directly optimize the probability of positive outcomes (as opposed to classic expected reward maximization). We address a key limitation in current EM-based methods: they solely maximize the likelihood of positive examples, while neglecting negative ones. We show how to extend EM algorithms to explicitly incorporate negative examples, leading to a theoretically grounded algorithm that offers an intuitive and versatile way to learn from both positive and negative feedback. We evaluate our approach for training language models based on human feedback as well as training policies for sequential decision-making problems, where learned value functions are available.

Existing reinforcement learning (RL) methods for large language models (LLMs) rely on static prompt sets, where prompts are curated a priori, and sampled in a fixed schedule for training, regardless of their usefulness to the RL process. We design eva, the first method that allows LLMs to prioritize and adaptively create useful prompts during RL training by reward signals. In principle, eva (Evolving via A symmetric Self-Play) casts language model training as a game between: (1) a creator, who samples and generates training prompts, and (2) a solver, who generates responses to the prompts. eva is simple, suits both offline and online RL for LLMs, and sets a new state-of-the-art on challenging benchmarks without extra human prompts: it improves gemma-2-9b-it’s win-rate on Arena-Hard from 51.6% to 60.1% by DPO and 52.6% to 62.4% by RLOO, surpassing claude-3-opus and nearing gemini-1.5-pro, both are orders of magnitude larger. Further ablation studies show eva can induce meaningful learning curriculum, and effectively scale RL for LLMs beyond static human prompts.

Tianqi Liu, Zhen Qin, Junru Wu, Jiaming Shen, Misha Khalman, Rishabh Joshi, Yao Zhao, Mohammad Saleh, Simon Baumgartner, Jialu Liu, Peter J Liu, Xuanhui Wang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Improving the alignment of language models with human preferences remains an active research challenge. Previous approaches have primarily utilized online Reinforcement Learning from Human Feedback (RLHF). Recently, offline methods such as Sequence Likelihood Calibration (SLiC) and Direct Preference Optimization (DPO) have emerged as attractive alternatives, offering improvements in stability and scalability while maintaining competitive performance. SLiC refines its loss function using sequence pairs sampled from a supervised fine-tuned (SFT) policy, while DPO directly optimizes language models based on preference data, foregoing the need for a separate reward model. However, the maximum likelihood estimator (MLE) of the target optimal policy requires labeled preference pairs sampled from that policy. The absence of a reward model in DPO constrains its ability to sample preference pairs from the optimal policy. Meanwhile, SLiC can only sample preference pairs from the SFT policy. To address these limitations, we introduce a novel approach called Statistical Rejection Sampling Optimization (RSO) designed to source preference data from the target optimal policy using rejection sampling, enabling a more accurate estimation of the optimal policy. We also propose a unified framework that enhances the loss functions used in both SLiC and DPO from a preference modeling standpoint. Through extensive experiments across diverse tasks, we demonstrate that RSO consistently outperforms both SLiC and DPO as evaluated by both Large Language Models (LLMs) and human raters.

Ensuring alignment of language model’s outputs with human preferences is critical to guarantee a useful, safe, and pleasant user experience. Thus, human alignment has been extensively studied recently and several methods such as Reinforcement Learning from Human Feedback (RLHF), Direct Policy Optimisation (DPO) and Sequence Likelihood Calibration (SLiC) have emerged. In this paper, our contribution is two-fold. First, we show the equivalence between two recent alignment methods, namely Identity Policy Optimisation (IPO) and Nash Mirror Descent (Nash-MD). Second, we introduce a generalisation of IPO, named IPO-MD, that leverages the regularised sampling approach proposed by Nash-MD. This equivalence may seem surprising at first sight, since IPO is an offline method whereas Nash-MD is an online method using a preference model. However, this equivalence can be proven when we consider the online version of IPO, that is when both generations are sampled by the online policy and annotated by a trained preference model. Optimising the IPO loss with such a stream of data becomes then equivalent to finding the Nash equilibrium of the preference model through self-play. Building on this equivalence, we introduce the IPO-MD algorithm that generates data with a mixture policy (between the online and reference policy) similarly as the general Nash-MD algorithm. We compare online-IPO and IPO-MD to different online versions of existing losses on preference data such as DPO and SLiC on a summarisation task.

Ritam Dutt, Sayan Sinha, Rishabh Joshi, Surya Shekhar Chakraborty, Meredith Riggs, Xinru Yan, Haogang Bao, Carolyn Rose. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

OBJECTIVES: Biomedical natural language processing tools are increasingly being applied for broad-coverage information extraction-extracting medical information of all types in a scientific document or a clinical note. In such broad-coverage settings, linking mentions of medical concepts to standardized vocabularies requires choosing the best candidate concepts from large inventories covering dozens of types. This study presents a novel semantic type prediction module for biomedical NLP pipelines and two automatically-constructed, large-scale datasets with broad coverage of semantic types. METHODS: We experiment with five off-the-shelf biomedical NLP toolkits on four benchmark datasets for medical information extraction from scientific literature and clinical notes. All toolkits adopt a staged approach of mention detection followed by two stages of medical entity linking: (1) generating a list of candidate concepts, and (2) picking the best concept among them. We introduce a semantic type prediction module to alleviate the problem of overgeneration of candidate concepts by filtering out irrelevant candidate concepts based on the predicted semantic type of a mention. We present MedType, a fully modular semantic type prediction model which we integrate into the existing NLP toolkits. To address the dearth of broad-coverage training data for medical information extraction, we further present WikiMed and PubMedDS, two large-scale datasets for medical entity linking. RESULTS: Semantic type filtering improves medical entity linking performance across all toolkits and datasets, often by several percentage points of F-1. Further, pretraining MedType on our novel datasets achieves state-of-the-art performance for semantic type prediction in biomedical text. CONCLUSIONS: Semantic type prediction is a key part of building accurate NLP pipelines for broad-coverage information extraction from biomedical text. We make our source code and novel datasets publicly available to foster reproducible research.

The notion of face refers to the public selfimage of an individual that emerges both from the individual's own actions as well as from the interaction with others.Modeling face and understanding its state changes throughout a conversation is critical to the study of maintenance of basic human needs in and through interaction.Grounded in the politeness theory of Brown and Levinson (1978), we propose a generalized framework for modeling face acts in persuasion conversations, resulting in a reliable coding manual, an annotated corpus, and computational models.The framework reveals insights about differences in face act utilization between asymmetric roles in persuasion conversations.Using computational models, we are able to successfully identify face acts as well as predict a key conversational outcome (e.g.donation success).Finally, we model a latent representation of the conversational state to analyze the impact of predicted face acts on the probability of a positive conversational outcome and observe several correlations that corroborate previous findings.

The problem of building a coherent and non-monotonous conversational agent with proper discourse and coverage is still an area of open research. Current architectures only take care of semantic and contextual information for a given query and fail to completely account for syntactic and external knowledge which are crucial for generating responses in a chit-chat system. To overcome this problem, we propose an end to end multi-stream deep learning architecture that learns unified embeddings for query-response pairs by leveraging contextual information from memory networks and syntactic information by incorporating Graph Convolution Networks (GCN) over their dependency parse. A stream of this network also utilizes transfer learning by pre-training a bidirectional transformer to extract semantic representation for each input sentence and incorporates external knowledge through the neighborhood of the entities from a Knowledge Base (KB). We benchmark these embeddings on the next sentence prediction task and significantly improve upon the existing techniques. Furthermore, we use AMUSED to represent query and responses along with its context to develop a retrieval based conversational agent which has been validated by expert linguists to have comprehensive engagement with humans.

Distantly-supervised Relation Extraction (RE) methods train an extractor by automatically aligning relation instances in a Knowledge Base (KB) with unstructured text.In addition to relation instances, KBs often contain other relevant side information, such as aliases of relations (e.g., founded and co-founded are aliases for the relation founderOfCompany).RE models usually ignore such readily available side information.In this paper, we propose RESIDE, a distantly-supervised neural relation extraction method which utilizes additional side information from KBs for improved relation extraction.It uses entity type and relation alias information for imposing soft constraints while predicting relations.RE-SIDE employs Graph Convolution Networks (GCN) to encode syntactic information from text and improves performance even when limited side information is available.Through extensive experiments on benchmark datasets, we demonstrate RESIDE's effectiveness.We have made RESIDE's source code available to encourage reproducible research.