Large Language Models (LLMs) exhibit impressive capabilities in In-Context Learning (ICL) but are prone to label bias—an undesirable tendency to favor certain answers. Existing calibration methods mitigate bias by leveraging in-domain data, yet such data is often unavailable in real-world scenarios. To address this limitation, we propose SDC (Synthetic Data Calibration), a simple-yet-effective approach that generates synthetic in-domain data from a few in-context demonstrations and utilizes it for calibration. By approximating the benefits of real in-domain data, SDC effectively reduces label bias without requiring access to actual domain-specific inputs. Experimental evaluations on 279 classification and multiple-choice tasks from the Super-NaturalInstructions benchmark. The results show that SDC significantly reduces label bias, achieving an average Bias Score reduction of 57.5%, and outperforming all competitive baselines. Moreover, when combined with Leave-One-Out Calibration (LOOC), further improves performance, underscoring its effectiveness and generalizability in enhancing the reliability of LLMs.
论文检索
输入标题、作者或关键词，从 3,086 篇学术成果中精准定位
Language models today are widely used in education, yet their ability to tailor responses for learners with varied informational needs and knowledge backgrounds remains under-explored. To this end, we introduce ELI-Why, a benchmark of 13.4K “Why” questions to evaluate the pedagogical capabilities of language models. We then conduct two extensive human studies to assess the utility of language model-generated explanatory answers (explanations) on our benchmark, tailored to three distinct educational grades: elementary, high-school and graduate school. In our first study, human raters assume the role of an “educator” to assess model explanations’ fit to different educational grades. We find that GPT-4-generated explanations match their intended educational background only 50% of the time, compared to 79% for lay human-curated explanations. In our second study, human raters assume the role of a learner to assess if an explanation fits their own informational needs. Across all educational backgrounds, users deemed GPT-4-generated explanations 20% less suited on average to their informational needs, when compared to explanations curated by lay people. Additionally, automated evaluation metrics reveal that explanations generated across different language model families for different informational needs remain indistinguishable in their grade-level, limiting their pedagogical effectiveness.
Summary Factual Inconsistency Detection Based on LLMs Enhanced by Universal Information Extraction
PDF ↗Automatic text summarization has a potential flaw that affects the factuality of summaries. Recently, Large Language Models (LLMs) have been introduced as detectors for factual inconsistencies in summaries. However, LLM-based methods rely on reasoning capabilities and face challenges in terms of efficiency and explainability. We focus on decoupling LLMs’ information extraction and reasoning capabilities to address prominent challenges, and propose a novel framework, UIEFID (Universal Information Extraction-enhanced Factual Inconsistency Detection). Our idea is to define a self-adaptive structured schema to guide fine-tuned LLMs in extracting unified structured information from documents and summaries, ultimately detecting the origins of inconsistencies in extraction information. The evaluation on 5 open-source models shows that UIEFID not only enhances the detection accuracy on the AGGREFACT benchmark but also significantly reduces redundant reasoning.
The Reader is the Metric: How Textual Features and Reader Profiles Explain Conflicting Evaluations of AI Creative Writing
PDF ↗Recent studies comparing AI-generated and human-authored literary texts have produced conflicting results: some suggest AI already surpasses human quality, while others argue it still falls short. We start from the hypothesis that such divergences can be largely explained by genuine differences in how readers interpret and value literature, rather than by an intrinsic quality of the texts evaluated. Using five public datasets (1,471 stories, 101 annotators including critics, students, and lay readers), we (i) extract 17 reference-less textual features (e.g., coherence, emotional variance, average sentence length...); (ii) model individual reader preferences, deriving feature importance vectors that reflect their textual priorities; and (iii) analyze these vectors in a shared “preference space”. Reader vectors cluster into two profiles: _surface-focused readers_ (mainly non-experts), who prioritize readability and textual richness; and _holistic readers_ (mainly experts), who value thematic development, rhetorical variety, and sentiment dynamics. Our results quantitatively explain how measurements of literary quality are a function of how text features align with each reader’s preferences. These findings advocate for reader-sensitive evaluation frameworks in the field of creative text generation.
The growing use of large language models (LLMs) for text generation has led to widespread concerns about AI-generated content detection. However, an overlooked challenge is AI-polished text, where human-written content undergoes subtle refinements using AI tools. This raises a critical question: should minimally polished text be classified as AI-generated? Such classification can lead to false plagiarism accusations and misleading claims about AI prevalence in online content. In this study, we systematically evaluate *twelve* state-of-the-art AI-text detectors using our **AI-Polished-Text Evaluation (APT-Eval)** dataset, which contains 15K samples refined at varying AI-involvement levels. Our findings reveal that detectors frequently flag even minimally polished text as AI-generated, struggle to differentiate between degrees of AI involvement, and exhibit biases against older and smaller models. These limitations highlight the urgent need for more nuanced detection methodologies.
Recent advancements in state-of-the-art (SOTA) Large Language Model (LLM) agents, especially in multi-turn dialogue tasks, have been primarily driven by supervised fine-tuning and high-quality human feedback. However, as base LLM models continue to improve, acquiring meaningful human feedback has become increasingly challenging and costly. In certain domains, base LLM agents may eventually exceed human capabilities, making traditional feedback-driven methods impractical. In this paper, we introduce a novel self-improvement paradigm that empowers LLM agents to autonomously enhance their performance without external human feedback. Our method, Juxtaposed Outcomes for Simulation Harvesting (JOSH), is a self-alignment algorithm that leverages a sparse reward simulation environment to extract ideal behaviors and further train the LLM on its own outputs. We present ToolWOZ, a sparse reward tool-calling simulation environment derived from MultiWOZ. We demonstrate that models trained with JOSH, both small and frontier, significantly improve tool-based interactions while preserving general model capabilities across diverse benchmarks. Our code and data are publicly available on GitHub.
Large language models (LLMs) excel in complex reasoning tasks, and distilling their reasoning capabilities into smaller models has shown promise. However, we uncover an interesting phenomenon, which we term the Small Model Learnability Gap: small models (3B parameters) do not consistently benefit from long chain-of-thought (CoT) reasoning or distillation from larger models. Instead, they perform better when fine-tuned on shorter, simpler reasoning chains that better align with their intrinsic learning capacity. To address this, we propose Mix Distillation, a simple yet effective strategy that balances reasoning complexity by combining long and short CoT examples or reasoning from both larger and smaller models. Our experiments demonstrate that Mix Distillation significantly improves small model reasoning performance compared to training on either data alone. These findings highlight the limitations of direct strong model distillation and underscore the importance of adapting reasoning complexity for effective reasoning capability transfer.
Controlling the syntactic structure of text generated by language models is valuable for applications requiring clarity, stylistic consistency, or interpretability, yet it remains a challenging task. In this paper, we argue that sampling algorithms based on the posterior inference can effectively enforce a target constituency structure during generation. Our approach combines sequential Monte Carlo, which estimates the posterior distribution by sampling from a proposal distribution, with a syntactic tagger that ensures that each generated token aligns with the desired syntactic structure. Our experiments with GPT2 and Llama3-8B models show that with an appropriate proposal distribution, we can improve syntactic accuracy, increasing the F1 score from 12.31 (GPT2-large) and 35.33 (Llama3-8B) to about 93 in both cases without compromising the language model’s fluency. These results underscore both the complexity of syntactic control and the effectiveness of sampling algorithms, offering a promising approach for applications where precise control over syntax is essential.
Large Language Model (LLM) agents have demonstrated remarkable generalization capabilities across multi-domain tasks. Existing agent tuning approaches typically employ supervised finetuning on entire expert trajectories. However, behavior-cloning of full trajectories can introduce expert bias and weaken generalization to states not covered by the expert data. Additionally, critical steps—such as planning, complex reasoning for intermediate subtasks, and strategic decision-making—are essential to success in agent tasks, so learning these steps is the key to improving LLM agents. For more effective and efficient agent tuning, we propose ATLAS that identifies the critical steps in expert trajectories and finetunes LLMs solely on these steps with reduced costs. By steering the training’s focus to a few critical steps, our method mitigates the risk of overfitting entire trajectories and promotes generalization across different environments and tasks. In extensive experiments, an LLM finetuned on only 30% critical steps selected by ATLAS outperforms the LLM finetuned on all steps and recent open-source LLM agents. ATLAS maintains and improves base LLM skills as generalist agents interacting with diverse environments.
MAM: Modular Multi-Agent Framework for Multi-Modal Medical Diagnosis via Role-Specialized Collaboration
PDF ↗Recent advancements in medical Large Language Models (LLMs) have showcased their powerful reasoning and diagnostic capabilities. Despite their success, current unified multimodal medical LLMs face limitations in knowledge update costs, comprehensiveness, and flexibility. To address these challenges, we introduce the Modular Multi-Agent Framework for Multi-Modal Medical Diagnosis (MAM). Inspired by our empirical findings highlighting the benefits of role assignment and diagnostic discernment in LLMs, MAM decomposes the medical diagnostic process into specialized roles: a General Practitioner, Specialist Team, Radiologist, Medical Assistant, and Director, each embodied by an LLM-based agent. This modular and collaborative framework enables efficient knowledge updates and leverages existing medical LLMs and knowledge bases. Extensive experimental evaluations conducted on a wide range of publicly accessible multimodal medical datasets, incorporating text, image, audio, and video modalities, demonstrate that MAM consistently surpasses the performance of modality-specific LLMs. Notably, MAM achieves significant performance improvements ranging from 18% to 365% compared to baseline models. Our code, data, and prompts are released at https://github.com/yczhou001/MAM.
Finetuning large language models with a variety of instruction-response pairs has enhanced their capability to understand and follow instructions. Current instruction tuning primarily relies on teacher models or human intervention to generate and refine the instructions and responses for training, which are costly, non-sustainable, and may lack diversity. In this paper, we introduce Mosaic Instruction Tuning (Mosaic-IT), a human/model-free compositional data synthesis method that can efficiently create rich and diverse augmentations from existing instruction tuning data to enhance the LLMs. Mosaic-IT randomly concatenates multiple instruction data into one and trains the model to produce the corresponding responses with predefined higher-level meta-instructions to strengthen its multi-step instruction-following and format-following skills. Our extensive evaluations demonstrate a superior performance and training efficiency of Mosaic-IT, which achieves consistent performance improvements over various benchmarks and an 80% reduction in training costs compared with original instruction tuning.
Lemmas Matter, But Not Like That: Predictors of Lemma-Based Generalization in Morphological Inflection
PDF ↗Recent work has suggested that overlap –whether a given lemma or feature set is attested independently in train – drives model performance on morphological inflection tasks. The impact of lemma overlap, however, is debated, with recent work reporting accuracy drops from 0% to 30% between seen and unseen test lemmas. In this paper, we introduce a novel splitting algorithm designed to investigate predictors of accuracy on seen and unseen lemmas. We find only an 11% average drop from seen to unseen test lemmas, but show that the number of lemmas in train has a much stronger effect on accuracy on unseen than seen lemmas. We also show that the previously reported 30% drop is inflated due to the introduction of a near-30% drop in the number of training lemmas from the original splits to their novel splits.
Accurately grounding visual and textual elements within mobile user interfaces (UIs) remains a significant challenge for Vision-Language Models (VLMs). Visual grounding, a critical task in this domain, involves identifying the most relevant UI element or region based on a natural language query—a process that requires both precise perception and context-aware reasoning. In this work, we present - **MoUI**, a light-weight mobile UI understanding model trained on **MoIT**, an instruction-tuning dataset specifically tailored for mobile screen understanding and grounding, designed to bridge the gap between user intent and visual semantics. Complementing this dataset, we also present a human-annotated reasoning benchmark **MoIQ** that rigorously evaluates complex inference capabilities over mobile UIs. To harness these resources effectively, we propose a two-stage training approach that separately addresses perception and reasoning tasks, leading to stronger perception capabilities and improvement in reasoning abilities. Through extensive experiments, we demonstrate that our MoUI models achieve significant gains in accuracy across all perception tasks and _state-of-the-art_ results on public reasoning benchmark **ComplexQA (78%) and our MoIQ (49%)**. We will be open-sourcing our dataset, code, and models to foster further research and innovation in the field.
Ensuring the safety alignment of Large Language Models (LLMs) is critical for generating responses consistent with human values. However, LLMs remain vulnerable to jailbreaking attacks, where carefully crafted prompts manipulate them into producing toxic content. One category of such attacks reformulates the task as an optimization problem, aiming to elicit affirmative responses from the LLM. However, these methods heavily rely on predefined objectionable behaviors, limiting their effectiveness and adaptability to diverse harmful queries. In this study, we first identify why the vanilla target loss is suboptimal and then propose enhancements to the loss objective. We introduce \textit{DSN} (Don’t Say No) attack, which combines a cosine decay schedule method with refusal suppression to achieve higher success rates. Extensive experiments demonstrate that \textit{DSN} outperforms baseline attacks and achieves state-of-the-art attack success rates (ASR). \textit{DSN} also shows strong universality and transferability to unseen datasets and black-box models.
Revisiting Self-Consistency from Dynamic Distributional Alignment Perspective on Answer Aggregation
PDF ↗Self-consistency improves reasoning by aggregating diverse stochastic samples, yet the dynamics behind its efficacy remain underexplored. We reframe self-consistency as a dynamic distributional alignment problem, revealing that decoding temperature not only governs sampling randomness but also actively shapes the latent answer distribution. Given that high temperatures require prohibitively large sample sizes to stabilize, while low temperatures risk amplifying biases, we propose a confidence-driven mechanism that dynamically calibrates temperature: sharpening the sampling distribution under uncertainty to align with high-probability modes, and promoting exploration when confidence is high. Experiments on mathematical reasoning tasks show this approach outperforms fixed-diversity baselines under limited samples, improving both average and best-case performance across varying initial temperatures without additional data or modules. This establishes self-consistency as a synchronization challenge between sampling dynamics and evolving answer distributions.
Problem-Solving Therapy (PST) is a structured psychological approach that helps individuals manage stress and resolve personal issues by guiding them through problem identification, solution brainstorming, decision-making, and outcome evaluation. As mental health care increasingly adopts technologies like chatbots and large language models (LLMs), it is important to thoroughly understand how each session of PST is conducted before attempting to automate it. We developed a comprehensive framework for PST annotation using established PST Core Strategies and a set of novel Facilitative Strategies to analyze a corpus of real-world therapy transcripts to determine which strategies are most prevalent. Using various LLMs and transformer-based models, we found that GPT-4o outperformed all models, achieving the highest accuracy (0.76) in identifying all strategies. To gain deeper insights, we examined how strategies are applied by analyzing Therapeutic Dynamics (autonomy, self-disclosure, and metaphor), and linguistic patterns within our labeled data. Our research highlights LLMs’ potential to automate therapy dialogue analysis, offering a scalable tool for mental health interventions. Our framework enhances PST by improving accessibility, effectiveness, and personalized support for therapists.
LLMs as Planning Formalizers: A Survey for Leveraging Large Language Models to Construct Automated Planning Models
PDF ↗Large Language Models (LLMs) excel in various natural language tasks but often struggle with long-horizon planning problems requiring structured reasoning. This limitation has drawn interest in integrating neuro-symbolic approaches within the Automated Planning (AP) and Natural Language Processing (NLP) communities. However, identifying optimal AP deployment frameworks can be daunting and introduces new challenges. This paper aims to provide a timely survey of the current research with an in-depth analysis, positioning LLMs as tools for formalizing and refining planning specifications to support reliable off-the-shelf AP planners. By systematically reviewing the current state of research, we highlight methodologies, and identify critical challenges and future directions, hoping to contribute to the joint research on NLP and Automated Planning.
It has been demonstrated that carefully designed reasoning paradigms, like Chain-of-Thought(CoT) and Tree-of-Thought(ToT), can enhance the reasoning capabilities of small language models by detailed thinking and extensive thought searching, unbounded branching factors in the searching space create prohibitive reasoning consumption. However these methods fell into the trap of local optimum reasoning, which means the model lacks a global perspective while solving problems. We propose a novel reasoning paradigm called Reason from Future(RFF), which generates reasoning paths by bidirectional reasoning that combines top-down planning with bottom-up reasoning accumulation. The essence of RFF lies in its reverse reasoning mechanism, which prioritizes core logical relationships and imposes goal-oriented constraints on intermediate steps, thereby reducing the searching space and mitigating error accumulation inherent in sequential forward reasoning. Empirical evaluations across diverse experiments demonstrate that RFF outperforms conventional paradigms with higher accuracy and less searching space to solve complex tasks.
Chain-of-thought (CoT) reasoning has enabled large language models (LLMs) to utilize additional computation through intermediate tokens to solve complex tasks. However, we posit that typical reasoning traces contain many redundant tokens, incurring extraneous inference costs. Upon examination of the output distribution of current LLMs, we find evidence on their latent ability to reason more concisely, relative to their default behavior. To elicit this capability, we propose simple fine-tuning methods which leverage self-generated concise reasoning paths obtained by best-of-N sampling and few-shot conditioning, in task-specific settings. Our combined method achieves a 30% reduction in output tokens on average, across five model families on GSM8K and MATH, while maintaining average accuracy. By exploiting the fundamental stochasticity and in-context learning capabilities of LLMs, our self-training approach robustly elicits concise reasoning on a wide range of models, including those with extensive post-training.
We propose novel attention architectures, Multi-matrix Factorization Attention (MFA) and MFA-Key-Reuse (MFA-KR). Existing variants for standard Multi-Head Attention (MHA), including SOTA methods like MLA, fail to maintain as strong performance under stringent Key-Value cache (KV cache) constraints. MFA enhances model capacity by efficiently scaling up both the number and dimension of attention heads through low-rank matrix factorization in the Query-Key (QK) circuit. Extending MFA, MFA-KR further reduces memory requirements by repurposing the key cache as value through value projection re-parameterization. MFA’s design enables strong model capacity when working under tight KV cache budget, while MFA-KR is suitable for even harsher KV cache limits with minor performance trade-off. Notably, in our extensive and large-scale experiments, the proposed architecture outperforms MLA and performs comparably to MHA, while reducing KV cache usage by up to 56% and 93.7%, respectively.