Reward models (RMs) are central to the alignment of language models (LMs). An RM often serves as a proxy for human preferences to guide downstream LM behavior. However, our understanding of RM behavior is limited. Our work (i) formalizes a framework for measuring the alignment of opinions captured by RMs, (ii) investigates the extent to which RMs demonstrate sociodemographic biases, and (iii) explores the effects of prompting to steer rewards towards the preferences of a target group. We study the subjective and diverse perspectives on controversial topics, which allows us to quantify \textit{RM perspectives} in terms of their opinions, attitudes, and values. We show that RMs are poorly aligned with several demographic groups and can systematically reward harmful stereotypes, and steering alone is not enough to overcome these limitations. Our findings underscore the need for more careful consideration of RM behavior in model alignment during preference learning to prevent the propagation of unwanted social biases in the language technologies that we use.
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
Unraveling Interwoven Roles of Large Language Models in Authorship Privacy: Obfuscation, Mimicking, and Verification
PDF ↗Recent advancements in large language models (LLMs) have been fueled by large-scale training corpora drawn from diverse sources such as websites, news articles, and books. These datasets often contain explicit user information, such as person names, addresses, that LLMs may unintentionally reproduce in their generated outputs. Beyond such explicit content, LLMs can also leak identity-revealing cues through implicit signals such as distinctive writing styles, raising significant concerns about authorship privacy. There are three major automated tasks in authorship privacy, namely authorship obfuscation (AO), authorship mimicking (AM), and authorship verification (AV). Prior research has studied AO, AM, and AV independently. However, their interplays remain under-explored, which leaves a major research gap, especially in the era of LLMs, where they are profoundly shaping how we curate and share user-generated content, and the distinction between machine‐generated and human‐authored text is also increasingly blurred. This work then presents the first unified framework for analyzing the dynamic relationships among LLM-enabled AO, AM, and AV in the context of authorship privacy. We quantify how they interact with each other to transform human‐authored text, examining effects at a single point in time and iteratively over time. We also examine the role of demographic metadata, such as gender, academic background, in modulating their performances, inter-task dynamics, and privacy risks. The code is available at https://github.com/nguyentuc/authorship_privacy.
Although large language models have enhanced automated travel planning abilities, current systems remain misaligned with real-world scenarios. First, they assume users provide explicit queries, while in reality requirements are often implicit. Second, existing solutions ignore diverse environmental factors and user preferences, limiting the feasibility of plans. Third, systems can only generate plans with basic POI arrangements, failing to provide all-in-one plans with rich details. To mitigate these challenges, we construct a novel dataset RETAIL, which supports decision-making for implicit queries while covering explicit queries, both with and without revision needs. It also enables environmental awareness to ensure plan feasibility under real-world scenarios, while incorporating detailed POI information for all-in-one travel plans. Furthermore, we propose a topic-guided multi-agent framework, termed TGMA. Our experiments reveal that even the strongest existing model achieves merely a 1.0% pass rate, indicating real-world travel planning remains extremely challenging. In contrast, TGMA demonstrates substantially improved performance 2.72%, offering promising directions for real-world travel planning.
Recent studies have explored the use of large language models (LLMs) in psychotherapy; however, text-based cognitive behavioral therapy (CBT) models often struggle with client resistance, which can weaken therapeutic alliance. To address this, we propose a multimodal approach that incorporates nonverbal cues, which allows the AI therapist to better align its responses with the client’s negative emotional state.Specifically, we introduce a new synthetic dataset, Mirror (Multimodal Interactive Rolling with Resistance), which is a novel synthetic dataset that pairs each client’s statements with corresponding facial images. Using this dataset, we train baseline vision language models (VLMs) so that they can analyze facial cues, infer emotions, and generate empathetic responses to effectively manage client resistance.These models are then evaluated in terms of both their counseling skills as a therapist, and the strength of therapeutic alliance in the presence of client resistance. Our results demonstrate that Mirror significantly enhances the AI therapist’s ability to handle resistance, which outperforms existing text-based CBT approaches.Human expert evaluations further confirm the effectiveness of our approach in managing client resistance and fostering therapeutic alliance.
Evaluating Taxonomy Free Character Role Labeling (TF-CRL) in News Stories using Large Language Models
PDF ↗We introduce Taxonomy-Free Character Role Labeling (TF-CRL); a novel task that assigns open-ended narrative role labels to characters in news stories based on their functional role in the narrative. Unlike fixed taxonomies, TF-CRL enables more nuanced and comparative analysis by generating compositional labels (e.g., Resilient Leader, Scapegoated Visionary). We evaluate several large language models (LLMs) on this task using human preference rankings and ratings across four criteria: faithfulness, relevance, informativeness, and generalizability. LLMs almost uniformly outperform human annotators across all dimensions. We further show how TF-CRL supports rich narrative analysis by revealing novel latent taxonomies and enabling cross-domain narrative comparisons. Our approach offers new tools for studying media portrayals, character framing, and the socio-political impacts of narrative roles at-scale.
Power doesn’t reside in size: A Low Parameter Hybrid Language Model (HLM) for Sentiment Analysis in Code-mixed data
PDF ↗Code-mixed text—where multiple languages are used within the same utterance—is increasingly common in both spoken and written communication. However, it presents significant challenges for machine learning models due to the interplay of distinct grammatical structures, effectively forming a hybrid language. While fine-tuning large language models (LLMs) such as GPT-3, or Llama-3 on code-mixed data has led to performance improvements, these models still lag behind their monolingual counterparts and incur high computational costs due to the large number of trainable parameters.In this paper, we focus on the task of sentiment detection in code-mixed text and propose a Hybrid Language Model (HLM) that combines a multilingual encoder (e.g., mBERT) with a lightweight decoder (e.g., Sarvam-1) (3B parameters). Despite having significantly fewer trainable parameters, HLM achieves sentiment classification performance comparable to that of fine-tuned Large Language Models (LLMs) (> 7B parameters). Furthermore, our results demonstrate that HLM significantly outperforms models trained individually, underscoring its effectiveness for low-resource, code-mixed sentiment analysis.
This work presents a novel trie (prefix-tree)-based parallel decoding method that addresses the memory inefficiency of batch-based beam search. By sharing a single KV cache across beams with common prefixes, our approach dramatically reduces memory usage and enables efficient decoding. We evaluated our method across three attention architectures, Multi-Head Attention (Phi-3.5-mini-instruct), Grouped Query Attention (Llama-3.1-8B-Instruct), and Sliding Window Attention (Mistral-Small-24B-Instruct-2501), using CNN/DailyMail for abstractive summarization and HumanEval for code generation. Our experiments demonstrate substantial memory savings (4–8\times) and up to 2.4\times faster decoding, without compromising generation quality. These results highlight our method’s suitability for memory-constrained environments and large-scale deployments.
PPC-GPT: Federated Task-Specific Compression of Large Language Models via Pruning and Chain-of-Thought Distillation
PDF ↗Compressing Large Language Models (LLMs) into task-specific Small Language Models (SLMs) encounters two significant challenges: safeguarding domain-specific knowledge privacy and managing limited resources. To tackle these challenges, we propose PPC-GPT, a novel unified framework that systematically addresses both privacy preservation and model compression in federated settings. PPC-GPT works on a server-client federated architecture, where the client sends differentially private (DP) perturbed task-specific data to the server’s LLM. The LLM then generates synthetic data along with their corresponding rationales. This synthetic data is subsequently used for both LLM pruning and retraining processes. Our framework’s key innovation lies in its holistic integration of privacy-preserving mechanisms, synthetic data generation, and task-specific compression techniques, creating unique benefits through component interaction. Our experiments across diverse text generation tasks demonstrate that PPC-GPT successfully achieves dual objectives: maintaining competitive performance comparable to full-sized LLMs while ensuring robust privacy protection through its federated architecture. Our code has been contributed to the FATE open-source project and is now publicly accessible at https://github.com/FederatedAI/FATE-LLM/tree/main/python/fate_llm/algo/ppc-gpt
In-Context Learning (ICL) has become a powerful paradigm that enables LLMs to perform a wide range of tasks without task-specific fine-tuning. However, the effectiveness of ICL heavily depends on the quality of exemplar selection. In particular, for structured prediction tasks such as semantic parsing, existing ICL selection strategies often overlook structural alignment, leading to suboptimal performance and poor generalization. To address this issue, we propose a novel two-stage exemplar selection strategy that achieves a strong balance between efficiency, generalizability, and performance. First, we fine-tune a BERT-based retriever using structure-aware supervision, guiding it to select exemplars that are both semantically relevant and structurally aligned. Then, we enhance the retriever with a plug-in module, which amplifies syntactically meaningful information in the hidden representations. This plug-in is model-agnostic, requires minimal overhead, and can be seamlessly integrated into existing pipelines. Experiments on four benchmarks spanning three semantic parsing tasks demonstrate that our method consistently outperforms existing baselines with multiple recent LLMs as inference-time models.
In recent years, developing compact and efficient large language models (LLMs) has emerged as a thriving area of research. However, traditional Supervised Fine-Tuning (SFT), which relies on singular ground truth labels, often fails to capture token-level dependencies and linguistic diversity. To address these limitations, we propose a logits-based fine-tuning framework that integrates the strengths of supervised learning and knowledge distillation. Our approach constructs enriched training targets by combining teacher logits with ground truth labels, preserving both correctness and linguistic diversity. This ensures more reliable and effective training. To validate our approach, we constructed a large-scale 1.2M logits dataset and trained a series of science-focused models. Experimental results demonstrate that our method achieves significant improvements over current SOTA, with accuracy gains of 18% on Mawps and 22.7% on TabMWP. Across nine widely used mathematical benchmarks, our method consistently outperforms prior SFT models, achieving an average improvement of 7.28%. All code and datasets will be open-sourced.
Benchmark data contamination (BDC) silently inflate the evaluation performance of large language models (LLMs), yet current work on BDC has centered on direct token overlap (data/label level), leaving the subtler and equally harmful semantic level BDC largely unexplored. This gap is critical in fake news detection task, where prior exposure to semantic BDC lets a model “remember” the answer instead of reasoning. In this work, (1) we are the first to formally define semantic contamination for this task and (2) introduce the Semantic Sensitivity Amplifier (SSA), a lightweight, model-agnostic framework that detects BDC risks across semantic to label level via an entity shift perturbation and a comprehensive interpretable metric, the SSA Factor. Evaluating 45 variants of nine LLMs (0.5B–72B parameters) across four BDC levels, we find LIAR2 accuracy climbs monotonically with injected contamination, while the SSA Factor escalates in near-perfect lock-step (r\geq.97, for models \geq3B, p<.05; \rho \geq.9 overall, p<.05). These results show that SSA provides a sensitive and scalable audit of comprehensive BDC risk and paves the way for a more integrity evaluation of the LLM-driven fake news detection task.
Effective interactions between artificial intelligence (AI) and humans require an equitable and accurate representation of diverse cultures. It is known that current AI, particularly large language models (LLMs), possesses some degrees of cultural knowledge but not without limitations. We present a framework aimed at understanding the origin of these limitations. We hypothesize that there is a fundamental discordance between embedded ethics—how LLMs represent right versus wrong, and cultural inference—how LLMs infer cultural knowledge, specifically cultural norms. We demonstrate this by extracting low-dimensional subspaces that embed ethical principles of LLMs based on established benchmarks. We then show that how LLMs make errors in culturally distinctive scenarios significantly correlates with how they represent cultural norms with respect to these embedded ethics subspaces. Furthermore, we show that coercing cultural norms to be more aligned with the embedded ethics increases LLM performance in cultural inference. Our analyses of 12 language models, two large-scale cultural benchmarks spanning 75 countries and two ethical datasets indicate that 1) the ethics-culture discordance tends to be exacerbated in instruct-tuned models, and 2) how current LLMs represent ethics can impose limitations on their adaptation to diverse cultures particularly pertaining to non-Western and low-income regions.
Large Language Models (LLMs) have demonstrated inherent calibration capabilities, where predicted probabilities align well with correctness, despite prior findings that deep neural networks are often overconfident. Recent studies have linked this behavior to specific components in the final layer, such as entropy neurons and the unembedding matrix’s null space. In this work, we provide a complementary perspective by investigating how calibration evolves throughout the network’s depth. Analyzing multiple open-weight models on the MMLU benchmark, we uncover a distinct confidence correction phase in the upper/later layers, where model confidence is actively recalibrated after decision certainty has been reached. Furthermore, we identify a low-dimensional calibration direction in the residual stream whose perturbation significantly improves calibration metrics (ECE and MCE) without harming accuracy. Our findings suggest that calibration is a distributed phenomenon, shaped throughout the network’s forward pass, not just in its final projection, providing new insights into how confidence-regulating mechanisms operate within LLMs.
WISE: Weak-Supervision-Guided Step-by-Step Explanations for Multimodal LLMs in Image Classification
PDF ↗Multimodal Large Language Models (MLLMs) have shown promise in visual-textual reasoning, with Multimodal Chain-of-Thought (MCoT) prompting significantly enhancing interpretability. However, existing MCoT methods rely on rationale-rich datasets and largely focus on inter-object reasoning, overlooking the intra-object understanding crucial for image classification. To address this gap, we propose WISE, a Weak-supervision-guided Step-by-step Explanation method that augments any image classification dataset with MCoTs by reformulating the concept-based representations from Concept Bottleneck Models (CBMs) into concise, interpretable reasoning chains under weak supervision. Experiments across ten datasets show that our generated MCoTs not only improve interpretability by 37% but also lead to gains in classification accuracy when used to fine-tune MLLMs. Our work bridges concept-based interpretability and generative MCoT reasoning, providing a generalizable framework for enhancing MLLMs in fine-grained visual understanding.
Link Prediction (LP) aims to predict missing triple information within a Knowledge Graph (KG). Existing LP methods have sought to improve the performance by integrating structural and textual information. However, for lexico-semantic KGs designed to document fine-grained sense distinctions, these types of information may not be sufficient to support effective LP. From a linguistic perspective, word senses within lexico-semantic relations usually show systematic differences in their sememic components. In light of this, we are motivated to enhance LP with sememe knowledge. We first construct a Sememe Prediction (SP) dataset, SememeDef, for learning such knowledge, and two Chinese datasets, HN7 and CWN5, for LP evaluation; Then, we propose a method, SememeLP, to leverage this knowledge for LP fully. It consistently and significantly improves the LP performance in both English and Chinese, achieving SOTA MRR of 75.1%, 80.5%, and 77.1% on WN18RR, HN7, and CWN5, respectively; Finally, an in-depth analysis is conducted, making clear how sememic components can benefit LP for lexico-semantic KGs, which provides promising progress for the completion of them.
ESGenius: Benchmarking LLMs on Environmental, Social, and Governance (ESG) and Sustainability Knowledge
PDF ↗We introduce ESGenius, a comprehensive benchmark for evaluating and enhancing the proficiency of Large Language Models (LLMs) in Environmental, Social, and Governance (ESG) and sustainability-focused question answering. ESGenius comprises two key components: (i) ESGenius-QA, a collection of 1,136 Multiple-Choice Questions (MCQs) generated by LLMs and rigorously validated by domain experts, covering a broad range of ESG pillars and sustainability topics. Each question is systematically linked to its corresponding source text, enabling transparent evaluation and supporting Retrieval-Augmented Generation (RAG) methods; and (ii) ESGenius-Corpus, a meticulously curated repository of 231 foundational frameworks, standards, reports, and recommendation documents from 7 authoritative sources. Moreover, to fully assess the capabilities and adaptation potential of LLMs, we implement a rigorous two-stage evaluation protocol—Zero-Shot and RAG. Extensive experiments across 50 LLMs (0.5B to 671B) demonstrate that state-of-the-art models achieve only moderate performance in zero-shot settings, with accuracies around 55–70%, highlighting a significant knowledge gap for LLMs in this specialized, interdisciplinary domain. However, models employing RAG demonstrate significant performance improvements, particularly for smaller models. For example, DeepSeek-R1-Distill-Qwen-14B improves from 63.82% (zero-shot) to 80.46% with RAG. These results demonstrate the necessity of grounding responses in authoritative sources for enhanced ESG understanding. To the best of our knowledge, ESGenius is the first comprehensive QA benchmark designed to rigorously evaluate LLMs on ESG and sustainability knowledge, providing a critical tool to advance trustworthy AI in this vital domain.
Identifying/retrieving relevant statutes and prior cases/precedents for a given legal situation are common tasks exercised by law practitioners. Researchers till date have addressed the two tasks independently, thus developing completely different datasets and models for each task; however, both retrieval tasks are inherently related, e.g., similar cases tend to cite similar statutes (due to similar factual situation). In this paper, we address this gap. We propose IL-PCSR (Indian Legal corpus for Prior Case and Statute Retrieval), which is a unique corpus that provides a common testbed for developing models for both the tasks (Statute Retrieval and Precedent Retrieval) that can exploit the dependence between the two. We experiment extensively with several baseline models on the tasks, including lexical models, semantic models and ensemble based on GNNs. Further, to exploit the dependence between the two tasks, we develop an LLM based re-ranking approach that gives the best performance.
Tuning Less, Prompting More: In-Context Preference Learning Pipeline for Natural Language Transformation
PDF ↗Natural language transformation (NLT) tasks, such as machine translation (MT) and text style transfer (TST), require models to generate accurate and contextually appropriate outputs. However, existing approaches face significant challenges, including the computational costs of leveraging large pre-trained models and the limited generalization ability of fine-tuned smaller models. In this paper, we propose a novel framework that combines the flexibility of prompting with the cost-effectiveness of fine-tuning. Our method enhances smaller models by integrating In-Context Examples (ICE) from retrieval, enabling the model to better capture contextual information and align with user-level preferences. We further improve performance through hierarchical contrastive learning and dynamic preference inference mechanisms. Experimental results demonstrate that our approach outperforms existing methods, such as Supervised Fine Tuning (SFT), Direct Preference Optimization (DPO), and Contrastive Preference Optimization (CPO), across both MT and TST tasks, providing a more efficient solution for resource-constrained environments.
Preference mechanisms, such as human preference, LLM-as-a-Judge (LaaJ), and reward models, are central to aligning and evaluating large language models (LLMs). Yet, the underlying concepts that drive these preferences remain poorly understood. In this work, we propose a fully automated method for generating local and global concept-based explanations of preferences across multiple domains. Our method utilizes an LLM to identify concepts (rubrics) that distinguish between chosen and rejected responses, and to represent them with concept-based vectors. To model the relationships between concepts and preferences, we propose a white-box Hierarchical Multi-Domain Regression model that captures both domain-general and domain-specific effects. To evaluate our method, we curate a dataset spanning eight diverse domains and explain twelve mechanisms. Our method achieves strong preference prediction performance, outperforming baselines while also being explainable. Additionally, we assess explanations in two application-driven settings. First, guiding LLM outputs with concepts from LaaJ explanations yields responses that those judges consistently prefer. Second, prompting LaaJs with concepts explaining humans improves their preference predictions. Together, our work establishes a new paradigm for explainability in the era of LLMs.
Selecting high-quality data can improve the pretraining efficiency of large language models (LLMs). Existing methods generally rely on heuristic techniques or single quality signals, limiting their ability to evaluate data quality comprehensively. In this work, we propose FIRE, a flexible and scalable framework for integrating multiple data quality raters, which allows for a comprehensive assessment of data quality across various dimensions. FIRE aligns multiple quality signals into a unified space, and integrates diverse data quality raters to provide a comprehensive quality signal for each data point. Further, we introduce a progressive data selection scheme based on FIRE that iteratively refines the selection of high-quality data points. Extensive experiments show that FIRE outperforms other data selection methods and significantly boosts pretrained model performance across a wide range of downstream tasks, while requiring less than 37.5% tokens needed by the \textit{Random} baseline to reach the target performance.