Large language models (LLMs) have demonstrated capabilities that are highly satisfactory to a wide range of users by adapting to their culture and wisdom. Yet, this can translate into a propensity to produce responses that align with users’ viewpoints, even when the latter are wrong. This behaviour is known as sycophancy, the tendency of LLMs to generate misleading responses as long as they align with the user’s, inducing bias and reducing reliability. To make interactions consistent, reliable and safe, we introduce X-Agent, an Oversight Reasoning framework that audits human–LLM dialogues, reasons about them, captures sycophancy and corrects the final outputs. Concretely, X-Agent extends debate-based frameworks by (i) auditing human-LLM conversations, (ii) applying a defence layer that steers model behaviour and goes beyond user beliefs, and (iii) extracting reasoning traces from evaluations that serve as training signals for mitigating sycophancy. We evaluate X-Agent across diverse scenarios and languages, showing that it consistently detects sycophancy, reduces unwarranted agreement, and improves cross-turn consistency, advancing a reasoning-as-overview paradigm for safer LLM interaction. Our approach introduces a novel paradigm in which reasoning is not merely a means to solve problems, but as a mechanism for overseeing the problem-solving processes of other models.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
DeepResonance: Enhancing Multimodal Music Understanding via Music-centric Multi-way Instruction Tuning
PDF ↗Recent advancements in music large language models (LLMs) have significantly improved music understanding tasks, which involve the model’s ability to analyze and interpret various musical elements. These improvements primarily focused on integrating both music and text inputs. However, the potential of incorporating additional modalities such as images, videos and textual music features to enhance music understanding remains unexplored. To bridge this gap, we propose DeepResonance, a multimodal music understanding LLM fine-tuned via multi-way instruction tuning with multi-way aligned music, text, image, and video data. To this end, we construct Music4way-MI2T, Music4way-MV2T, and Music4way-Any2T, three 4-way training and evaluation datasets designed to enable DeepResonance to integrate both visual and textual music feature content. We also introduce multi-sampled ImageBind embeddings and a pre-LLM fusion Transformer to enhance modality fusion prior to input into text LLMs, tailoring for multi-way instruction tuning. Our model achieves state-of-the-art performances across six music understanding tasks, highlighting the benefits of the auxiliary modalities and the structural superiority of DeepResonance. We open-source the codes, models and datasets we constructed: https://github.com/sony/DeepResonance.
With the rapid expansion of large language model (LLM) applications, there is an emerging shift in the role of LLM-based AI chatbots from serving merely as general inquiry tools to acting as professional service agents. However, current studies often overlook a critical aspect of professional service agents: the act of communicating in a manner consistent with their professional identities. This is of particular importance in the healthcare sector, where effective communication with patients is essential for achieving professional goals, such as promoting patient well-being by encouraging healthy behaviors. To bridge this gap, we propose LAPI (LLM-based Agent with a Professional Identity), a novel framework for designing professional service agent tailored for medical question-and-answer (Q&A) services, ensuring alignment with a specific professional identity. Our method includes a theory-guided task planning process that decomposes complex professional tasks into manageable subtasks aligned with professional objectives and a pragmatic entropy method designed to generate professional and ethical responses with low uncertainty. Experiments on various LLMs show that the proposed approach outperforms baseline methods, including few-shot prompting, chain-of-thought prompting, across key metrics such as fluency, naturalness, empathy, patient-centricity, and ROUGE-L scores. Additionally, the ablation study underscores the contribution of each component to the overall effectiveness of the approach.
This paper investigates the enhancement of reasoning capabilities in language models through token-level multi-model collaboration. Our approach selects the optimal tokens from the next token distributions provided by multiple models to perform autoregressive reasoning. Contrary to the assumption that more models yield better results, we introduce a distribution distance-based dynamic selection strategy (DDS) to optimize the multi-model collaboration process. To address the critical challenge of vocabulary misalignment in multi-model collaboration, we propose the concept of minimal complete semantic units (MCSU), which is simple yet enables multiple language models to achieve natural alignment within the linguistic space. Experimental results across various benchmarks demonstrate the superiority of our method. The codes will be released soon.
Personalizing large language models (LLMs) is important for aligning outputs with diverse user preferences, yet existing methods struggle with flexibility and generalization. We propose CoPL (Collaborative Preference Learning), a graph-based collaborative filtering framework that models user-response relationships to enhance preference estimation, particularly in sparse annotation settings. By integrating a mixture of LoRA experts, CoPL efficiently fine-tunes LLMs while dynamically balancing shared and user-specific preferences. Additionally, an optimization-free adaptation strategy enables generalization to unseen users without fine-tuning. Experiments on TL;DR, UltraFeedback-P, and PersonalLLM datasets demonstrate that CoPL outperforms existing personalized reward models, effectively capturing both common and controversial preferences, making it a scalable solution for personalized LLM alignment.
Panic attacks are acute episodes of fear and distress, in which timely, appropriate intervention can significantly help individuals regain stability. However, suitable datasets for training such models remain scarce due to ethical and logistical issues. To address this, we introduce Pace, which is a dataset that includes high-distress episodes constructed from first-person narratives, and structured around the principles of Psychological First Aid (PFA). Using this data, we train Pacer, a counseling model designed to provide both empathetic and directive support, which is optimized through supervised learning and simulated preference alignment. To assess its effectiveness, we propose PanicEval, a multi-dimensional framework covering general counseling quality and crisis-specific strategies. Experimental results show that Pacer outperforms strong baselines in both counselor-side metrics and client affect improvement. Human evaluations further confirm its practical value, with Pacer consistently preferred over general, CBT-based, and GPT-4-powered models in panic scenarios.
Despite extensive efforts to align large language models (LLMs) with human values and safety rules, jailbreak attacks that exploit certain vulnerabilities continuously emerge, highlighting the need to strengthen existing LLMs with additional safety properties to defend against these attacks. However, tuning large models has become increasingly resource-intensive and may have difficulty ensuring consistent performance. We introduce Speculative Safety-Aware Decoding (SSD), a lightweight decoding-time approach that equips LLMs with the desired safety property while accelerating inference. We assume that there exists a small language model that possesses the desired safety property. SSD integrates speculative sampling during decoding and leverages the match ratio between the small and composite models to quantify jailbreak risks. This enables SSD to dynamically switch between decoding schemes to prioritize utility or safety, to handle the challenge of different model capacities. The output token is then sampled from a new distribution that combines the distributions of both models. Experimental results show that SSD successfully equips the large model with the desired safety property, and also allows the model to remain helpful to benign queries. Furthermore, SSD accelerates the inference time, thanks to the speculative sampling design.
Data to Defense: The Role of Curation in Aligning Large Language Models Against Safety Compromise
PDF ↗Large language models (LLMs) are widely adapted for downstream applications through fine-tuning, a process named customization. However, recent studies have identified a vulnerability during this process, where malicious samples can compromise the robustness of LLMs and amplify harmful behaviors. To address this challenge, we propose an adaptive data curation approach allowing any text to be curated to enhance its effectiveness in counteracting harmful samples during customization. To avoid the need for additional defensive modules, we further introduce a comprehensive mitigation framework spanning the lifecycle of the customization process: before customization to immunize LLMs against future compromise attempts, during customization to neutralize risks, and after customization to restore compromised models. Experimental results demonstrate a significant reduction in compromising effects, achieving up to a 100% success rate in generating safe responses. By combining adaptive data curation with lifecycle-based mitigation strategies, this work represents a solid step forward in mitigating compromising risks and ensuring the secure adaptation of LLMs.
DeepWell-Adol: A Scalable Expert-Based Dialogue Corpus for Adolescent Positive Mental Health and Wellbeing Promotion
PDF ↗Promoting positive mental health and well-being, especially in adolescents, is a critical yet underexplored area in natural language processing (NLP). Most existing NLP research focuses on clinical therapy or psychological counseling for the general population, which does not adequately address the preventative and growth-oriented needs of adolescents. In this paper, we introduce DeepWell-Adol, a domain-specific Chinese dialogue corpus grounded in positive psychology and coaching, designed to foster adolescents’ positive mental health and well-being. To balance the trade-offs between data quality, quantity, and scenario diversity, the corpus comprises two main components: human expert-written seed data (ensuring professional quality) and its mirrored expansion (automatically generated using a two-stage scenario-based augmentation framework). This approach enables large-scale data creation while maintaining domain relevance and reliability. Comprehensive evaluations demonstrate that the corpus meets general standards for psychological dialogue and emotional support, while also showing superior performance across multiple models in promoting positive psychological processes, character strengths, interpersonal relationships, and healthy behaviors. Moreover, the framework proposed for building and evaluating DeepWell-Adol offers a flexible and scalable method for developing domain-specific datasets. It significantly enhances automation and reduces development costs without compromising professional standards—an essential consideration in sensitive areas like adolescent and elderly mental health. We make our dataset publicly available.
Expanding before Inferring: Enhancing Factuality in Large Language Models through Premature Layers Interpolation
PDF ↗Large Language Models (LLMs) demonstrate remarkable capabilities in text understanding and generation. However, their tendency to produce factually inconsistent outputs—commonly referred to as “hallucinations”—remains a critical challenge. Existing approaches, such as retrieval-based and inference-time correction methods, primarily address this issue at the input or output level, often overlooking the intrinsic information refinement process and the role of premature layers. Meanwhile, alignment- and fine-tuning-based methods are resource-intensive. In this paper, we propose **PLI** (**P**remature **L**ayers **I**nterpolation), a novel, training-free, and plug-and-play intervention designed to enhance factuality. PLI mitigates hallucinations by inserting premature layers formed through mathematical interpolation with adjacent layers. Inspired by stable diffusion and sampling steps, PLI extends the depth of information processing and transmission in LLMs, improving factual coherence. Experiments on four publicly available datasets demonstrate that PLI effectively reduces hallucinations while outperforming existing baselines in most cases. Further analysis suggests that the success of layer interpolation is closely linked to LLMs’ internal mechanisms. To promote reproducibility, we will release our code and data upon acceptance.
Instruction tuning has empowered large language models (LLMs) to achieve remarkable performance, yet its success heavily depends on the availability of large-scale, high-quality instruction-response pairs. To meet this demand, various methods have been developed to synthesize data at scale. However, current methods for scaling up data generation often overlook a crucial aspect: the alignment between instructions and responses. We hypothesize that the quality of instruction-response pairs is determined not by the individual quality of each component, but by the degree of mutual alignment. To address this, we propose a Mutual Alignment Framework (MAIN) which enforces coherence between instructions and responses through mutual constraints. We demonstrate that MAIN generalizes well across model architectures and sizes, achieving state-of-the-art performance on LLaMA, Mistral, and Qwen models across diverse benchmarks. This work underscores the critical role of instruction-response alignment in enabling generalizable and high-quality instruction tuning for LLMs. All code is available from our repository.
Large pre-trained language models have demonstrated impressive capabilities, but there is still much to learn about how they operate. In this study, we conduct an investigation of the autoregressive transformer’s ability to perform basic addition operations. Specifically, by using causal analysis we found that a few different attention heads in the middle layers control the addition carry, with each head processing carries of different lengths. Due to the lack of global focus on the sequence within these attention heads, the model struggles to handle long-sequence addition tasks. By performing inference intervention on mistral-7B, partial task performance can be restored, with the accuracy on 20-digit long-sequence additions from 2% to 38%. Through fine-tuning, a new mechanism branches out for handling complex cases, yet it still faces challenges with length generalization. Our research reveals how the models perform basic arithmetic task, and further provides insights into the debate on whether these models are merely statistical.
VCSearch: Bridging the Gap Between Well-Defined and Ill-Defined Problems in Mathematical Reasoning
PDF ↗Large language models (LLMs) have demonstrated impressive performance on reasoning tasks, including mathematical reasoning. However, the current evaluation mostly focuses on carefully constructed benchmarks and neglects the consideration of real-world reasoning problems that present missing or contradictory conditions, known as ill-defined problems. To further study this problem, we develop a large-scale benchmark called Problems with Missing and Contradictory conditions (PMC) containing over 5,000 validated ill-defined mathematical problems. Our preliminary experiments through PMC reveal two challenges about existing methods: (1) traditional methods exhibit a trade-off between solving accuracy and rejection capabilities, and (2) formal methods struggle with modeling complex problems. To address these challenges, We develop Variable-Constraint Search (VCSearch), a training-free framework that leverages formal language to detect ill-defined problems, where a variable-constraint pair search strategy is incorporated to improve the modeling capability of formal language. Extensive experiments demonstrate that VCSearch improves the accuracy of identifying unsolvable problems by at least 12% across different LLMs, thus achieving stronger robust mathematical reasoning ability.
Large language models (LLMs) are now ubiquitous in user-facing applications, yet they still generate undesirable toxic outputs, including profanity, vulgarity, and derogatory remarks. Although numerous detoxification methods exist, most apply broad, surface-level fixes and can therefore easily be circumvented by jailbreak attacks. In this paper we leverage sparse autoencoders (SAEs) to identify toxicity-related directions in the residual stream of models and perform targeted activation steering using the corresponding decoder vectors. We introduce three tiers of steering aggressiveness and evaluate them on GPT-2 Small and Gemma-2-2B, revealing trade-offs between toxicity reduction and language fluency. At stronger steering strengths, these causal interventions surpass competitive baselines in reducing toxicity by up to 20%, though fluency can degrade noticeably on GPT-2 Small depending on the aggressiveness. Crucially, standard NLP benchmark scores upon steering remain stable, indicating that the model’s knowledge and general abilities are preserved. We further show that feature-splitting in wider SAEs hampers safety interventions, underscoring the importance of disentangled feature learning. Our findings highlight both the promise and the current limitations of SAE-based causal interventions for LLM detoxification, further suggesting practical guidelines for safer language-model deployment.
Diagrams play a central role in research papers for conveying ideas, yet they are often notoriously complex and labor-intensive to create. Although diagrams are presented as images, standard image generative models struggle to produce clear diagrams with well-defined structure. We argue that a promising direction is to generate demonstration diagrams directly in textual form as SVGs, which can leverage recent advances in large language models (LLMs). However, due to the complexity of components and the multimodal nature of diagrams, sufficiently discriminative and explainable metrics for evaluating the quality of LLM-generated diagrams remain lacking. In this paper, we propose DiagramEval, a novel evaluation metric designed to assess demonstration diagrams generated by LLMs. Specifically, DiagramEval conceptualizes diagrams as graphs, treating text elements as nodes and their connections as directed edges, and evaluates diagram quality using two new groups of metrics: node alignment and path alignment. For the first time, we effectively evaluate diagrams produced by state-of-the-art LLMs on recent research literature, quantitatively demonstrating the validity of our metrics. Furthermore, we show how the enhanced explainability of our proposed metrics offers valuable insights into the characteristics of LLM-generated diagrams.
A new trend uses LLMs as dense text encoders via contrastive learning. However, since LLM embeddings predict the probability distribution of the next token, they are inherently generative and distributive, conflicting with contrastive learning, which requires embeddings to capture full-text semantics and align via cosine similarity. This discrepancy hinders the full utilization of LLMs’ pre-training capabilities, resulting in inefficient learning. In response to this issue, we propose AutoRegEmbed, a new contrastive learning method built on embedding conditional probability distributions, which integrates two core tasks: information compression and conditional distribution alignment. The information compression task encodes text into the embedding space, ensuring that the embedding vectors capture global semantics. The conditional distribution alignment task focuses on aligning text embeddings with positive samples embeddings by leveraging the conditional distribution of embeddings while simultaneously reducing the likelihood of generating negative samples from text embeddings, thereby achieving embedding alignment and uniformity. Experimental results demonstrate that our method significantly outperforms traditional contrastive learning approaches and achieves performance comparable to state-of-the-art models when using the same amount of data.
Large language models (LLMs) have shown great potential in flagging harmful content in online communities. Yet, existing approaches for moderation require a separate model for every community and are opaque in their decision-making, limiting real-world adoption. We introduce Mixture of Moderation Experts (MoMoE), a modular, cross-community framework that adds post-hoc explanations to enable scalable content moderation. MoMoE orchestrates four operators—Allocate, Predict, Aggregate, Explain—and is instantiated as seven community-specialized experts (MoMoE-Community) and five norm-violation experts (MoMoE-NormVio). On 30 unseen subreddits, the best variants obtain Micro-F1 scores of 0.72 and 0.67, respectively, matching or surpassing strong fine-tuned baselines while consistently producing concise and reliable explanations. Although community-specialized experts deliver the highest peak accuracy, norm-violation experts provide steadier performance across domains. These findings show that MoMoE yields scalable, transparent moderation without needing per-community fine-tuning. More broadly, they suggest that lightweight, explainable expert ensembles can guide future NLP and HCI research on trustworthy human-AI governance of online communities.
Existing table understanding methods face challenges due to complex table structures and intricate logical reasoning. While supervised finetuning (SFT) dominates existing research, reinforcement learning (RL), such as Group Relative Policy Optimization (GRPO), has shown promise but struggled with low initial policy accuracy and coarse rewards in tabular contexts. In this paper, we introduce Table-R1, a three-stage RL framework that enhances multimodal table understanding through: (1) Warm-up that prompts initial perception and reasoning capabilities, (2) Perception Alignment GRPO (PA-GRPO), which employs continuous Tree-Edit-Distance Similarity (TEDS) rewards for recognizing table structures and contents, and (3) Hint-Completion GRPO (HC-GRPO), which utilizes fine-grained rewards of residual steps based on the hint-guided question. Extensive experiments demonstrate that Table-R1 can boost the model’s table reasoning performance obviously on both held-in and held-out datasets, outperforming SFT and GRPO largely. Notably, Qwen2-VL-7B with Table-R1 surpasses larger specific table understanding models (e.g., Table-LLaVA 13B), even achieving comparable performance to the closed-source model GPT-4o on held-in datasets, demonstrating the efficacy of each stage of Table-R1 in overcoming initialization bottlenecks and reward sparsity, thereby advancing robust multimodal table understanding.
Reasoning is a fundamental capability often required in real-world text-to-image (T2I) generation, e.g., generating “a bitten apple that has been left in the air for more than a week” necessitates understanding temporal decay and commonsense concepts. While recent T2I models have made impressive progress in producing photorealistic images, their reasoning capability remains underdeveloped and insufficiently evaluated. To bridge this gap, we introduce R2I-Bench, a comprehensive benchmark specifically designed to rigorously assess reasoning-driven T2I generation. R2I-Bench comprises 3068 meticulously curated data instances, spanning 7 core reasoning categories, including commonsense, mathematical, logical, compositional, numerical, causal, and concept mixing. To facilitate fine-grained evaluation, we design R2IScore, a QA-style metric based on instance-specific, reasoning-oriented evaluation questions that assess three critical dimensions: text-image alignment, reasoning accuracy, and image quality. Extensive experiments with 16 representative T2I models, including a strong pipeline-based framework that decouples reasoning and generation using the state-of-the-art language and image generation models, demonstrate consistently limited reasoning performance, highlighting the need for more robust, reasoning-aware architectures in the next generation of T2I systems.
We introduce ProcWorld, a large-scale benchmark for partially observable embodied spatial reasoning and long-term planning with large language models (LLM) and vision language models (VLM). ProcWorld features a wide range of challenging embodied navigation and object manipulation tasks, covering 16 task types, 5,000 rooms, and over 10 million evaluation trajectories with diverse data distribution. ProcWorld supports configurable observation modes, ranging from text-only descriptions to vision-only observations. It enables text-based actions to control the agent following language instructions. ProcWorld has presented significant challenges for LLMs and VLMs: (1) active information gathering given partial observations for disambiguation; (2) simultaneous localization and decision-making by tracking the spatio-temporal state-action distribution; (3) constrained reasoning with dynamic states subject to physical reachability. Our extensive evaluation of 15 foundation models and 5 reasoning algorithms (with over 1 million rollouts) indicates larger models perform better. However, ProcWorld remains highly challenging for existing state-of-the-art models and in-context learning methods due to constrained reachability and the need for combinatorial spatial reasoning.