Retrieval augmented generation (RAG) is frequently used to mitigate hallucinations and provide up-to-date knowledge for large language models (LLMs). However, given that document retrieval is an imprecise task and sometimes results in erroneous or even harmful content being presented in context, this raises the question of how LLMs handle retrieved information: If the provided content is incorrect, does the model know to ignore it, or does it recapitulate the error? Conversely, when the model's initial response is incorrect, does it always know to use the retrieved information to correct itself, or does it insist on its wrong prior response? To answer this, we curate a dataset of over 1200 questions across six domains (e.g., drug dosages, Olympic records, locations) along with content relevant to answering each question. We further apply precise perturbations to the answers in the content that range from subtle to blatant errors.We benchmark six top-performing LLMs, including GPT-4o, on this dataset and find that LLMs are susceptible to adopting incorrect retrieved content, overriding their own correct prior knowledge over 60\% of the time. However, the more unrealistic the retrieved content is (i.e. more deviated from truth), the less likely the model is to adopt it. Also, the less confident a model is in its initial response (via measuring token probabilities), the more likely it is to adopt the information in the retrieved content. We exploit this finding and demonstrate simple methods for improving model accuracy where there is conflicting retrieved content. Our results highlight a difficult task and benchmark for LLMs -- namely, their ability to correctly discern when it is wrong in light of correct retrieved content and to reject cases when the provided content is incorrect.

When language model (LM) users aim to improve the quality of its generations, it is crucial to specify concrete behavioral attributes that the model should strive to reflect. However, curating such principles across many domains, even non-exhaustively, requires a labor-intensive annotation process. To automate this process, we propose eliciting these latent attributes that guide model reasoning toward human-preferred responses by explicitly modeling them in a self-correction setting. Our approach mines new principles from the LM itself and compresses the discovered elements to an interpretable set via clustering. Specifically, we employ a form of posterior-regularized Monte Carlo Expectation-Maximization to both identify a condensed set of the most effective latent principles and teach the LM to strategically invoke them in order to intrinsically refine its responses. We demonstrate that bootstrapping our algorithm over multiple iterations enables smaller language models (7-8B parameters) to self-improve, achieving +8-10% in AlpacaEval win-rate, an average of +0.3 on MT-Bench, and +19-23% in principle-following win-rate on IFEval. We also show that clustering the principles yields interpretable and diverse model-generated constitutions while retaining model performance. The gains that our method achieves highlight the potential of automated, principle-driven post-training recipes toward continual self-improvement.

Large Language Models (LLMs) have demonstrated remarkable success in conversational systems by generating human-like responses. However, they can fall short, especially when required to account for personalization or specific knowledge. In real-life settings, it is impractical to rely on users to detect these errors and request a new response. One way to address this problem is to refine the response before returning it to the user. While existing approaches focus on refining responses within a single LLM, this method struggles to consider diverse aspects needed for effective conversations. In this work, we propose refining responses through a multi-agent framework, where each agent is assigned a specific role for each aspect. We focus on three key aspects crucial to conversational quality: factuality, personalization, and coherence. Each agent is responsible for reviewing and refining one of these aspects, and their feedback is then merged to improve the overall response. To enhance collaboration among them, we introduce a dynamic communication strategy. Instead of following a fixed sequence of agents, our approach adaptively selects and coordinates the most relevant agents based on the specific requirements of each query. We validate our framework on challenging conversational datasets, demonstrating that ours significantly outperforms relevant baselines, particularly in tasks involving knowledge or user's persona, or both.

In many practical applications of AI, an AI model is used as a decision aid for human users. The AI provides advice that a human (sometimes) incorporates into their decision-making process. The AI advice is often presented with some measure of "confidence" that the human can use to calibrate how much they

Large language models (LLMs) have achieved impressive results in natural language understanding, yet their reasoning capabilities remain limited when operating as single agents. Multi-Agent Debate (MAD) has been proposed to address this limitation by enabling collaborative reasoning among multiple models in a round-table debate manner. While effective, MAD introduces substantial computational overhead due to the number of agents involved and the frequent communication required. In this paper, we propose MARS (Multi-Agent Review System), a role-based collaboration framework inspired by the review process. In MARS, an author agent generates an initial solution, reviewer agents provide decisions and comments independently, and a meta-reviewer integrates the feedback to make the final decision and guide further revision. This design enhances reasoning quality while avoiding costly reviewer-to-reviewer interactions, thereby controlling token consumption and inference time. We compared MARS with both MAD and other state-of-the-art reasoning strategies across multiple benchmarks. Extensive experiments with different LLMs show that MARS matches the accuracy of MAD while reducing both token usage and inference time by approximately 50\%. Code is available at https://github.com/xwang97/MARS.

Self-correction has demonstrated potential in code generation by allowing language models to revise and improve their outputs through successive refinement. Recent studies have explored prompting-based strategies that incorporate verification or feedback loops using proprietary models, as well as training-based methods that leverage their strong reasoning capabilities. However, whether smaller models possess the capacity to effectively guide their outputs through self-reflection remains unexplored. Our findings reveal that smaller models struggle to exhibit reflective revision behavior across both self-correction paradigms. In response, we introduce CoCoS, an approach designed to enhance the ability of small language models for multi-turn code correction. Specifically, we propose an online reinforcement learning objective that trains the model to confidently maintain correct outputs while progressively correcting incorrect outputs as turns proceed. Our approach features an accumulated reward function that aggregates rewards across the entire trajectory and a fine-grained reward better suited to multi-turn correction scenarios. This facilitates the model in enhancing initial response quality while achieving substantial improvements through self-correction. With 1B-scale models, CoCoS achieves improvements of 35.8% on the MBPP and 27.7% on HumanEval compared to the baselines.

There have been numerous studies evaluating bias of LLMs towards political topics. However, how positions towards these topics in model outputs are highly sensitive to the prompt. What happens when the prompt itself is suggestive of certain arguments towards those positions remains underexplored. This is crucial for understanding how robust these bias evaluations are and for understanding model behaviour, as these models frequently interact with opinionated text. To that end, we conduct experiments for political bias evaluation in presence of supporting and refuting arguments. Our experiments show that such arguments substantially alter model responses towards the direction of the provided argument in both single-turn and multi-turn settings. Moreover, we find that the strength of these arguments influences the directional agreement rate of model responses. These effects point to a sycophantic tendency in LLMs adapting their stance to align with the presented arguments which has downstream implications for measuring political bias and developing effective mitigation strategies.

The distillation of knowledge from Large Language Models (LLMs) into Smaller Language Models (SLMs), preserving the capabilities and performance of LLMs while reducing model size, has played a key role in the proliferation of LLMs. Because SLMs are considerably smaller than LLMs, they are often utilized in domains where human interaction is frequent but resources are highly constrained, e.g., smart phones. Therefore, it is crucial to ensure that empathy, a fundamental aspect of positive human interactions, already instilled into LLMs, is retained by SLMs after distillation. In this paper, we develop a comprehensive approach for effective empathy distillation from LLMs into SLMs. Our approach features a two-step fine-tuning process that fully leverages datasets of empathetic dialogue responses distilled from LLMs. We explore several distillation methods beyond basic direct prompting and propose four unique sets of prompts for targeted empathy improvement to significantly enhance the empathy distillation process. Our evaluations demonstrate that SLMs fine-tuned through the two-step fine-tuning process with distillation datasets enhanced by the targeted empathy improvement prompts significantly outperform the base SLM at generating empathetic responses with a win rate of 90%. Our targeted empathy improvement prompts substantially outperform the basic direct prompting with a 10% improvement in win rate.

Large language models (LLMs) have demonstrated remarkable evaluation and critique capabilities, providing insightful feedback and identifying flaws in various tasks. However, limited research has explored which types of critiques are most effective for improving model responses or how to generate such critiques. To address this gap, we introduce \textbf{R}efinement-oriented \textbf{C}ritique \textbf{O}ptimization (RCO), a novel framework designed to train critic models using refinement signals. RCO uses a feedback loop where critiques, generated by the critic model, guide the actor model in refining its responses. The critique utility (CU) quantifies the effectiveness of these refinements, serving as the reward signal for training the critic model. By focusing on critiques that lead to better refinements, RCO eliminates the need for direct critique preference assessment, ensuring that critiques driving meaningful improvements are rewarded. We evaluate RCO across five tasks, i.e., dialog generation, summarization, question answering, mathematical reasoning, and code generation, and show that it significantly outperforms traditional methods and open-source models in terms of critique quality and refinement outcomes. Our contributions include the introduction of RCO, a novel supervision scheme based on refined response preferences, and comprehensive experimental results that highlight the method's effectiveness in enhancing LLM critique-refinement loops.

Retrieval augmented generation (RAG) is frequently used to mitigate hallucinations and provide up-to-date knowledge for large language models (LLMs). However, given that document retrieval is an imprecise task and sometimes results in erroneous or even harmful content being presented in context, this raises the question of how LLMs handle retrieved information: If the provided content is incorrect, does the model know to ignore it, or does it recapitulate the error? Conversely, when the model's initial response is incorrect, does it always know to use the retrieved information to correct itself, or does it insist on its wrong prior response? To answer this, we curate a dataset of over 1200 questions across six domains (e.g., drug dosages, Olympic records, locations) along with content relevant to answering each question. We further apply precise perturbations to the answers in the content that range from subtle to blatant errors.We benchmark six top-performing LLMs, including GPT-4o, on this dataset and find that LLMs are susceptible to adopting incorrect retrieved content, overriding their own correct prior knowledge over 60\% of the time.