New large language models (LLMs) are being released every day. Some perform significantly better or worse than expected given their parameter count. Therefore, there is a need for a method to independently evaluate models. The current best way to evaluate a model is to measure its Elo score by comparing it to other models in a series of contests—an expensive operation since humans are ideally required to compare LLM outputs. We observe that when an LLM is asked to judge such contests, the consistency with which it selects a model as the best in a matchup produces a metric that is 91% correlated with its own human-produced Elo score. This provides a simple proxy for Elo scores that can be computed cheaply, without any human data or prior knowledge.
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
We introduce a zero-shot merging framework for large language models (LLMs) that consolidates specialized domain experts into a single model without any further training. Our core contribution lies in leveraging relative task vectors—difference representations encoding each expert’s unique traits with respect to a shared base model—to guide a principled and efficient merging process. By dissecting parameters into common dimensions (averaged across experts) and complementary dimensions (unique to each expert), we strike an optimal balance between generalization and specialization. We further devise a compression mechanism for the complementary parameters, retaining only principal components and scalar multipliers per expert, thereby minimizing overhead. A dynamic router then selects the most relevant domain at inference, ensuring that domain-specific precision is preserved. Experiments on code generation, mathematical reasoning, medical question answering, and instruction-following benchmarks confirm the versatility and effectiveness of our approach. Altogether, this framework enables truly adaptive and scalable LLMs that seamlessly integrate specialized knowledge for improved zero-shot performance.
Expectation Preference Optimization: Reliable Preference Estimation for Improving the Reasoning Capability of Large Language Models
PDF ↗Pairwise preference optimization, such as Direct Preference Optimization (DPO), was originally designed to align large language models (LLMs) with human values. It has recently been used to improve the supervised fine-tuning (SFT) performance of LLMs. Using pairs of single samples, DPO estimates the probability distribution of the preferences of picking one response over another. However, in tasks that involve more complicated preferences (e.g., reasoning tasks) than those in the human value alignment task, this sampling method is likely to bring deviations from the ground-truth distribution. To solve the problem, extra efforts (e.g., external annotations or amendment of the loss function) are often required. In this paper, we hypothesise that the preferences can be better estimated through a multi-sampling process. Accordingly, we propose an Expectation Preference Optimization (EPO) algorithm that takes pairs of sample groups, instead of pairs of single samples as in DPO, for preference learning. Compared to pairwise DPO, the proposed EPO tends to produce more reliable preference estimations. Applying different preference optimization methods in a self-training paradigm, we have conducted extensive experiments on various reasoning benchmarks. The results show that our EPO approach outperforms a range of baseline approaches in terms of zero-shot accuracy on all benchmarks.
In sensitive domains such as medical and legal, protecting sensitive information is critical, with protective laws strictly prohibiting the disclosure of personal data. This poses challenges for sharing valuable data such as medical reports and legal cases summaries. While language models (LMs) have shown strong performance in text summarization, it is still an open question to what extent they can provide privacy-preserving summaries from non-private source documents. In this paper, we perform a comprehensive study of privacy risks in LM-based summarization across two closed- and four open-weight models of different sizes and families. We experiment with both prompting and fine-tuning strategies for privacy-preservation across a range of summarization datasets including medical and legal domains. Our quantitative and qualitative analysis, including human evaluation, shows that LMs frequently leak personally identifiable information in their summaries, in contrast to human-generated privacy-preserving summaries, which demonstrate significantly higher privacy protection levels. These findings highlight a substantial gap between current LM capabilities and expert human expert performance in privacy-sensitive summarization tasks.
Large language models (LLMs) offer a powerful opportunity to simulate the results of social science experiments. In this work, we demonstrate that finetuning LLMs directly on individual-level responses from past experiments meaningfully improves the accuracy of such simulations. We construct SocSci210 via an automatic pipeline, a dataset comprising 2.9 million responses from 400,491 participants in 210 open-source social science experiments. Through finetuning, we achieve multiple levels of generalization. In completely unseen studies, our strongest model, Socrates-Qwen-14B, produces predictions that are 36% more aligned with distributions of human responses to diverse outcome questions under varying conditions relative to its base model (Qwen2.5-14B), outperforming GPT-4o by 15%. By finetuning on a subset of conditions in a study, generalization to new unseen conditions is particularly robust, improving by 71%. Since SocSci210 contains rich demographic information, we reduce demographic parity difference, a measure of bias, by 10.6% through finetuning. Because social sciences routinely generate rich, topic-specific datasets, our findings indicate that finetuning on such data could enable more accurate simulations for experimental hypothesis screening. We release our data, models and finetuning code.
Large Language Models (LLMs) excel across diverse tasks, with post-training methods like Supervised Fine-Tuning (SFT), Preference Learning, and Model Merging enabling effective domain and task adaptation. While outcomes can vary with data orderings or component combinations, yet manual pipeline optimization is costly and labor-intensive. Existing approaches typically rely on manual design or focus narrowly on optimizing individual components, such as data ordering or merging parameters. We propose LaMDAgent, an LLM Agent-driven framework that autonomously constructs and optimizes end-to-end post-training pipelines by exploring various model improving methods, objects, and their applied orderings based on task-based feedback. LaMDAgent achieves a 9.0-point gain in tool-use accuracy without degrading instruction-following, and identifies high-performing strategies overlooked by manual design.We further analyze the impact of data and model scaling to reduce computational costs on the exploration, finding that model size scalings introduces new challenges, whereas scaling data size enables cost-effective pipeline discovery.
GRAID: Synthetic Data Generation with Geometric Constraints and Multi-Agentic Reflection for Harmful Content Detection
PDF ↗We address the problem of data scarcity in harmful text classification for guardrailing applications and introduce GRAID (Geometric and Reflective AI-Driven Data Augmentation), a novel pipeline that leverages Large Language Models (LLMs) for dataset augmentation. GRAID consists of two stages: (i) generation of geometrically controlled examples using a constrained LLM, and (ii) augmentation through a multi-agentic reflective process that promotes stylistic diversity and uncovers edge cases. This combination enables both reliable coverage of the input space and nuanced exploration of harmful content. Using two benchmark data sets, we demonstrate that augmenting a harmful text classification dataset with GRAID leads to significant improvements in downstream guardrail model performance.
Correct-Detect: Balancing Performance and Ambiguity Through the Lens of Coreference Resolution in LLMs
PDF ↗Large Language Models (LLMs) are intended to reflect human linguistic competencies. But humans have access to a broad and embodied context, which is key in detecting and resolving linguistic ambiguities, even in isolated text spans. A foundational case of semantic ambiguity is found in the task of coreference resolution: how is a pronoun related to an earlier person mention? This capability is implicit in nearly every downstream task, and the presence of ambiguity at this level can alter performance significantly. We show that LLMs can achieve good performance with minimal prompting in both coreference disambiguation and the detection of ambiguity in coreference, however, they cannot do both at the same time. We present the CORRECT-DETECT trade-off: though models have both capabilities and deploy them implicitly, successful performance balancing these two abilities remains elusive.
Societal stereotypes are at the center of a myriad of responsible AI interventions targeted at reducing the generation and propagation of potentially harmful outcomes. While these efforts are much needed, they tend to be fragmented and often address different parts of the issue without adopting a unified or holistic approach to social stereotypes and how they impact various parts of the machine learning pipeline. As a result, current interventions fail to capitalize on the underlying mechanisms that are common across different types of stereotypes, and to anchor on particular aspects that are relevant in certain cases. In this paper, we draw on social psychological research and build on NLP data and methods, to propose a unified framework to operationalize stereotypes in generative AI evaluations. Our framework identifies key components of stereotypes that are crucial in AI evaluation, including the target group, associated attribute, relationship characteristics, perceiving group, and context. We also provide considerations and recommendations for its responsible use.
Lemmatization is crucial for NLP tasks in morphologically rich languages with ambiguous orthography like Arabic, but existing tools face challenges due to inconsistent standards and limited genre coverage. This paper introduces two novel approaches that frame lemmatization as classification into a Lemma-POS-Gloss (LPG) tagset, leveraging machine translation and semantic clustering. We also present a new Arabic lemmatization test set covering diverse genres, standardized alongside existing datasets. We evaluate character-level sequence-to-sequence models, which perform competitively and offer complementary value, but are limited to lemma prediction (not LPG) and prone to hallucinating implausible forms. Our results show that classification and clustering yield more robust, interpretable outputs, setting new benchmarks for Arabic lemmatization.
Arabic dialects form a diverse continuum, yet NLP models often treat them as discrete categories. Recent work addresses this issue by modeling dialectness as a continuous variable, notably through the Arabic Level of Dialectness (ALDi). However, ALDi reduces complex variation to a single dimension. We propose a complementary measure: the Arabic Generality Score (AGS), which quantifies how widely a word is used across dialects. We introduce a pipeline that combines word alignment, etymology-aware edit distance, and smoothing to annotate a parallel corpus with word-level AGS. A regression model is then trained to predict AGS in context. Our approach outperforms strong baselines, including state-of-the-art dialect ID systems, on a multi-dialect benchmark. AGS offers a scalable, linguistically grounded way to model lexical generality, enriching representations of Arabic dialectness. Code is publicly available at https://github.com/CAMeL-Lab/arabic-generality-score.
Recent advances in large language models (LLMs) have enabled zero-shot automated essay scoring (AES), providing a promising way to reduce the cost and effort of essay scoring in comparison with manual grading. However, most existing zero-shot approaches rely on LLMs to directly generate absolute scores, which often diverge from human evaluations owing to model biases and inconsistent scoring. To address these limitations, we propose LLM-based Comparative Essay Scoring (LCES), a method that formulates AES as a pairwise comparison task. Specifically, we instruct LLMs to judge which of two essays is better, collect many such comparisons, and convert them into continuous scores. Considering that the number of possible comparisons grows quadratically with the number of essays, we improve scalability by employing RankNet to efficiently transform LLM preferences into scalar scores. Experiments using AES benchmark datasets show that LCES outperforms conventional zero-shot methods in accuracy while maintaining computational efficiency. Moreover, LCES is robust across different LLM backbones, highlighting its applicability to real-world zero-shot AES.
KGE Calibrator: An Efficient Probability Calibration Method of Knowledge Graph Embedding Models for Trustworthy Link Prediction
PDF ↗Knowledge graph embedding (KGE) models are designed for the task of link prediction, which aims to infer missing triples by learning representations for entities and relations. While KGE models excel at ranking-based link prediction, the critical issue of probability calibration has been largely overlooked, resulting in uncalibrated estimates that limit their adoption in high-stakes domains where trustworthy predictions are essential. Addressing this is challenging, as we demonstrate that existing calibration methods are ill-suited to KGEs, often significantly degrading the essential ranking performance they are meant to support. To overcome this, we introduce the KGE Calibrator (KGEC), the first probability calibration method tailored for KGE models to enhance the trustworthiness of their predictions. KGEC integrates three key techniques: a Jump Selection Strategy that improves efficiency by selecting the most informative instances while filtering out less significant ones; Multi-Binning Scaling, which models different confidence levels separately to increase capacity and flexibility; and a Wasserstein distance-based calibration loss that further boosts calibration performance. Extensive experiments across multiple datasets demonstrate that KGEC consistently outperforms existing calibration methods in terms of both effectiveness and efficiency, making it a promising solution for calibration in KGE models.
Sparse Autoencoders (SAEs) provide potential for uncovering structured, human-interpretable representations in Large Language Models (LLMs), making them a crucial tool for transparent and controllable AI systems. We systematically analyze SAE for interpretable feature extraction from LLMs in safety-critical classification tasks. Our framework evaluates (1) model-layer selection and scaling properties, (2) SAE architectural configurations, including width and pooling strategies, and (3) the effect of binarizing continuous SAE activations. SAE-derived features achieve macro F1 > 0.8, outperforming hidden-state and BoW baselines while demonstrating cross-model transfer from Gemma 2 2B to 9B-IT models. These features generalize in a zero-shot manner to cross-lingual toxicity detection and visual classification tasks. Our analysis highlights the significant impact of pooling strategies and binarization thresholds, showing that binarization offers an efficient alternative to traditional feature selection while maintaining or improving performance. These findings establish new best practices for SAE-based interpretability and enable scalable, transparent deployment of LLMs in real-world applications.
It has been demonstrated that incorporating external information as textual modality can effectively improve time series forecasting accuracy. However, current multi-modal models ignore the dynamic and different relations between time series patterns and textual features, which leads to poor performance in temporal-textual feature fusion. In this paper, we propose a lightweight and model-agnostic temporal-textual fusion framework named Cross-MoE. It replaces Cross Attention with Cross-Ranker to reduce computational complexity, and enhances modality-aware correlation memorization with Mixture-of-Experts (MoE) networks to tolerate the distributional shifts in time series. The experimental results demonstrate a 8.78% average reduction in Mean Squared Error (MSE) compared to the SOTA multi-modal time series framework. Notably, our method requires only 75% of computational overhead and 12.5% of activated parameters compared with Cross Attention mechanism. Our codes are available at https://github.com/Kilosigh/Cross-MoE.git
Reasoning over temporal and numerical data, such as time series, is a crucial aspect of fact-checking. While many systems have recently been developed to handle this form of evidence, their evaluation remains limited by existing datasets, which often lack structured evidence, provide insufficient justifications for verdicts, or rely on synthetic claims. In this paper, we introduce TSVer, a new benchmark dataset for fact verification focusing on temporal and numerical reasoning with time-series evidence. TSVer contains 304 real-world claims sourced from 41 fact-checking organizations and a curated database of 400 time series covering diverse domains. Each claim is annotated with time frames across all pertinent time series, along with a verdict and justifications reflecting how the evidence is used to reach the verdict. Using an LLM-assisted multi-step annotation process, we improve the quality of our annotations and achieve an inter-annotator agreement of \kappa = 0.77 on verdicts. We also develop a baseline for verifying claims against time-series evidence and show that even the state-of-the-art reasoning models like Gemini-2.5-Pro are challenged by time series, achieving a 63.57 accuracy score on verdicts and an \mathrm{Ev}^{2}\mathrm{R} score of 48.63 on verdict justifications.
Schema linking is widely recognized as a key factor in improving text-to-SQL performance. Supervised fine-tuning approaches enhance SQL generation quality by explicitly fine-tuning schema linking as an extraction task. However, they suffer from two major limitations: (i) The training corpus of small language models restricts their cross-domain generalization ability. (ii) The extraction-based fine-tuning process struggles to capture complex linking patterns. To address these issues, we propose GenLink, a generation-driven schema-linking framework based on multi-model learning. Instead of explicitly extracting schema elements, GenLink enhances linking through a generation-based learning process, effectively capturing implicit schema relationships. By integrating multiple small language models, GenLink improves schema-linking recall rate and ensures robust cross-domain adaptability. Experimental results on the BIRD and Spider benchmarks validate the effectiveness of GenLink, achieving execution accuracies of 67.34% (BIRD), 89.7% (Spider development set), and 87.8% (Spider test set), demonstrating its superiority in handling diverse and complex database schemas.
Interaction between learner and feedback-giver has come into focus recently for post-training of Large Language Models (LLMs), through the use of reward models that judge the appropriateness of a model’s response. In this paper, we investigate whether Dialogue Games—goal-directed and rule-governed activities driven predominantly by verbal actions—can also serve as a source of feedback signals for learning.We introduce Playpen, an environment for off- and online learning through Dialogue Game self-play, and investigate a representative set of post-training methods: supervised fine-tuning; direct alignment (DPO); and reinforcement learning with Group Relative Policy Optimization (GRPO). We experiment with post-training a small LLM (Llama-3.1-8B-Instruct), evaluating performance on unseen instances of training games as well as unseen games, and on standard benchmarks. We find that imitation learning through SFT improves performance on unseen instances, but negatively impacts other skills, while interactive learning with GRPO shows balanced improvements without loss of skills. We release the framework and the baseline training setups to foster research in this promising new direction of “learning in (synthetic) interaction”.
Recent work has demonstrated that using chain of thought (CoT), on soft-reasoning problems such as analytical and commonsense reasoning, often yields limited or even negative performance gains. CoT can also be unfaithful to the model’s actual reasoning. This paper investigates dynamics and unfaithfulness of CoT in soft-reasoning tasks across instruction-tuned, reasoning and reasoning-distilled models. Our findings show that distilled‐reasoning models rely heavily on CoT for these tasks, while instruction‐tuned and reasoning models often use it post‐hoc. Additionally, we find that CoT can steer model predictions without faithfully reflecting reasoning, indicating a disconnect between CoT influence and faithfulness.
Hallucination remains a critical failure mode of large language models (LLMs), undermining their trustworthiness in real-world applications. In this work, we focus on confabulation, a foundational aspect of hallucination where the model fabricates facts about unknown entities. We introduce a targeted dataset designed to isolate and analyze this behavior across diverse prompt types. Using this dataset, and building on recent progress in interpreting LLM internals, we extract latent directions associated with confabulation using sparse projections. A simple vector-based steering method demonstrates that these directions can modulate model behavior with minimal disruption, shedding light on the inner representations that drive factual and non-factual output. Our findings contribute to a deeper mechanistic understanding of LLMs and pave the way toward more trustworthy and controllable generation. We release the code and dataset at https://github.com/Thibaud-Ardoin/where-confabulation-lives.