Automated evaluation is crucial for assessing the quality of natural language text, especially in open-ended generation tasks, given the costly and time-consuming nature of human evaluation. Existing automatic evaluation metrics like ROUGE and BLEU often show low correlation with human judgments. As large language models (LLMs) continue to evolve, researchers have explored their use as alternatives to human evaluators. Although single-agent approaches have shown potential, results indicate that further progress is required to close the gap between their performance and the quality of human assessments. Acknowledging that human evaluations involve multiple annotators, the multi-agent approach allows LLMs to collaborate, enhancing efficiency and effectiveness in handling complex tasks. In this paper, we present CourtEval, a novel Multi-Agent Evaluation Framework modeled after courtroom dynamics. Each agent takes on a distinct role: the Grader, similar to a judge, assigns an initial score; the Critic, like a prosecutor, challenges this score; and the Defender, akin to a defense attorney, defends it. Based on the input from both the Critic and Defender, the Grader re-evaluates the score, leading to a more balanced and fair final decision through this adversarial process. CourtEval substantially outperforms the previous state-of-the-art methods in two meta-evaluation benchmarks in NLG evaluation, SummEval and TopicalChat.
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
Large language models (LLMs) encode vast amounts of pre-trained knowledge in their parameters, but updating them as real-world information evolves remains a challenge. Existing methodologies and benchmarks primarily target entity substitutions, failing to capture the full breadth of complex real-world dynamics. In this paper, we introduce Knowledge Update Playground (KUP), an automatic pipeline for simulating realistic knowledge updates reflected in an evidence corpus. KUP’s evaluation framework includes direct and indirect probes to both test memorization of updated facts and reasoning over them, for any update learning methods. Next, we present a lightweight method called memory conditioned training (MCT), which conditions tokens in the update corpus on self-generated ”memory” tokens during training. Our strategy encourages LLMs to surface and reason over newly memorized knowledge at inference. Our results on two LLM families show that (1) KUP benchmark is highly challenging, with the best CPT models achieving <2% in indirect probing setting (reasoning) and (2) MCT training significantly outperforms prior continued pre-training (CPT) baselines, improving direct probing (memorization) results by up to 25.4%.
This paper presents LEMONADE, a large-scale conflict event dataset comprising 39,786 events across 20 languages and 171 countries, with extensive coverage of region-specific entities. LEMONADE is based on a partially reannotated subset of the Armed Conflict Location & Event Data (ACLED), which has documented global conflict events for over a decade.To address the challenge of aggregating multilingual sources for global event analysis, we introduce abstractive event extraction (AEE) and its subtask, abstractive entity linking (AEL). Unlike conventional span-based event extraction, our approach detects event arguments and entities through holistic document understanding and normalizes them across the multilingual dataset. We evaluate various large language models (LLMs) on these tasks, adapt existing zero-shot event extraction systems, and benchmark supervised models. Additionally, we introduce ZEST, a novel zero-shot retrieval-based system for AEL.Our best zero-shot system achieves an end-to-end F1 score of 58.3%, with LLMs outperforming specialized event extraction models such as GoLLIE. For entity linking, ZEST achieves an F1 score of 45.7%, significantly surpassing OneNet, a state-of-the-art zero-shot baseline that achieves only 23.7%. However, these zero-shot results lag behind the best supervised systems by 20.1% and 37.0% in the end-to-end and AEL tasks, respectively, highlighting the need for further research.
Despite the widespread use of ‘artificial intelligence’ (AI) framing in Natural Language Processing (NLP) research, it is not clear what researchers mean by ”intelligence”. To that end, we present the results of a survey on the notion of ”intelligence” among researchers and its role in the research agenda. The survey elicited complete responses from 303 researchers from a variety of fields including NLP, Machine Learning (ML), Cognitive Science, Linguistics, and Neuroscience.We identify 3 criteria of intelligence that the community agrees on the most: generalization, adaptability, & reasoning.Our results suggests that the perception of the current NLP systems as ”intelligent” is a minority position (29%).Furthermore, only 16.2% of the respondents see developing intelligent systems as a research goal, and these respondents are more likely to consider the current systems intelligent.
LLM use in annotation is becoming widespread, and given LLMs’ overall promising performance and speed, putting humans in the loop to simply “review” LLM annotations can be tempting. In subjective tasks with multiple plausible answers, this can impact both evaluation of LLM performance, and analysis using these labels in a social science task downstream. In a pre-registered experiment with 350 unique annotators and 7,000 annotations across 4 conditions, 2 models, and 2 datasets, we find that presenting crowdworkers with LLM-generated annotation suggestions did not make them faster annotators, but did improve their self-reported confidence in the task. More importantly, annotators strongly took the LLM suggestions, significantly changing the label distribution compared to the baseline. We show that when these labels created with LLM assistance are used to evaluate LLM performance, reported model performance significantly increases. We show how changes in label distributions as a result of LLM assistance can affect conclusions drawn by analyzing even “human-approved” LLM-annotated datasets. We believe our work underlines the importance of understanding the impact of LLM-assisted annotation on subjective, qualitative tasks, on the creation of gold data for training and testing, and on the evaluation of NLP systems on subjective tasks.
Low-Resource Grammatical Error Correction: Selective Data Augmentation with Round-Trip Machine Translation
PDF ↗Supervised state-of-the-art methods for grammatical error correction require large amounts of parallel data for training. Due to lack of gold-labeled data, techniques that create synthetic training data have become popular. We show that models trained on synthetic data tend tocorrect a limited range of grammar and spelling mistakes that involve character-level changes, but perform poorly on (more complex) phenomena that require word-level changes. We propose to address the performance gap on such errors by generating synthetic data through selective data augmentation via round-trip machine translation. We show that the proposed technique, SeLex-RT, is capable of generating mistakes that are similar to those observed with language learners. Using the approach with two types of state-of-the-art learning frameworks and two low-resource languages (Russian and Ukrainian), we achieve substantial improvements, compared to training on synthetic data produced with standard techniques. Analysis of the output reveals that models trained on data noisified with the SeLex-RT approach are capable of making word-level changes and correct lexical errors common with language learners.
Artificial Text Detection (ATD) is becoming increasingly important with the rise of advanced Large Language Models (LLMs). Despite numerous efforts, no single algorithm performs consistently well across different types of unseen text or guarantees effective generalization to new LLMs. Interpretability plays a crucial role in achieving this goal. In this study, we enhance ATD interpretability by using Sparse Autoencoders (SAE) to extract features from Gemma-2-2B’s residual stream. We identify both interpretable and efficient features, analyzing their semantics and relevance through domain- and model-specific statistics, a steering approach, and manual or LLM-based interpretation of obtained features. Our methods offer valuable insights into how texts from various models differ from human-written content. We show that modern LLMs have a distinct writing style, especially in information-dense domains, even though they can produce human-like outputs with personalized prompts. The code for this paper is available at https://github.com/pyashy/SAE_ATD.
Enhancing Medical Dialogue Generation through Knowledge Refinement and Dynamic Prompt Adjustment
PDF ↗Medical dialogue systems (MDS) have emerged as crucial online platforms for enabling multi-turn, context-aware conversations with patients. However, existing MDS often struggle to (1) identify relevant medical knowledge and (2) generate personalized, medically accurate responses. To address these challenges, we propose MedRef, a novel MDS that incorporates knowledge refining and dynamic prompt adjustment. First, we employ a knowledge refining mechanism to filter out irrelevant medical data, improving predictions of critical medical entities in responses. Additionally, we design a comprehensive prompt structure that incorporates historical details and evident details. To enable real-time adaptability to diverse patient conditions, we implement two key modules, Triplet Filter and Demo Selector, providing appropriate knowledge and demonstrations equipped in the system prompt.Extensive experiments on MedDG and KaMed benchmarks show that MedRef outperforms state-of-the-art baselines in both generation quality and medical entity accuracy, underscoring its effectiveness and reliability for real-world healthcare applications.
Large language models (LLMs) have demonstrated remarkable success across a wide range of tasks; however, they still encounter challenges in reasoning tasks that require understanding and inferring relationships between distinct pieces of information within text sequences. This challenge is particularly pronounced in tasks involving multi-step processes, such as logical reasoning and multi-hop question answering, where understanding implicit relationships between entities and leveraging multi-hop connections in the given context are crucial. Graphs, as fundamental data structures, explicitly represent pairwise relationships between entities, thereby offering the potential to enhance LLMs’ reasoning capabilities. External graphs have proven effective in supporting LLMs across multiple tasks. However, in many reasoning tasks, no pre-existing graph structure is provided. Can we structure implicit knowledge derived from context into graphs to assist LLMs in reasoning? In this paper, we propose Reasoning with Graphs (RwG) by first constructing explicit graphs from the context and then leveraging these graphs to enhance LLM reasoning performance on reasoning tasks. Extensive experiments demonstrate the effectiveness of the proposed method in improving both logical reasoning and multi-hop question answering tasks.
Next token prediction paradigm has been prevailing for autoregressive models in the era of LLMs. The current default sampling choice for popular LLMs is temperature scaling together with nucleus sampling to balance diversity and coherence. Nevertheless, such approach leads to inferior performance in various NLP tasks when the model is not certain about testing questions. To this end, we propose a brand new training-free decoding strategy, dubbed as Cautious Next Token Prediction (CNTP). In the decoding process, if the model has comparatively high prediction entropy at a certain step, we sample multiple trials starting from the step independently and stop when encountering any punctuation. Then we select the trial with the lowest perplexity score viewed as the most probable and reliable trial path given the model’s capacity. The trial number is negatively correlated with the prediction confidence, i.e., the less confident the model is, the more trials it should sample. This is consistent with human beings’ behaviour: when feeling uncertain or unconfident, one tends to think more creatively, exploring multiple thinking paths, to cautiously select the path one feels most confident about. Extensive experiments on both LLMs and MLLMs show that our proposed CNTP approach outperforms existing standard decoding strategies consistently by a clear margin. Moreover, the integration of CNTP with self consistency can further improve over vanilla self consistency. We believe our proposed CNTP has the potential to become one of the default choices for LLM decoding. Code is available at https://github.com/wyzjack/CNTP.
Why Multi-Interest Fairness Matters: Hypergraph Contrastive Multi-Interest Learning for Fair Conversational Recommender System
PDF ↗Unfairness is a well-known challenge in Recommender Systems (RSs), often resulting in biased outcomes that disadvantage users or items based on attributes such as gender, race, age, or popularity. Although some approaches have started to improve fairness recommendation in offline or static contexts, the issue of unfairness often exacerbates over time, leading to significant problems like the Matthew effect, filter bubbles, and echo chambers. To address these challenges, we proposed a novel framework, Hypergraph Contrastive Multi-Interest Learning for Fair Conversational Recommender System (HyFairCRS), aiming to promote multi-interest diversity fairness in dynamic and interactive Conversational Recommender Systems (CRSs). HyFairCRS first captures a wide range of user interests by establishing diverse hypergraphs through contrastive learning. These interests are then utilized in conversations to generate informative responses and ensure fair item predictions within the dynamic user-system feedback loop. Experiments on two CRS-based datasets show that HyFairCRS achieves a new state-of-the-art performance while effectively alleviating unfairness.
Do Language Models Mirror Human Confidence? Exploring Psychological Insights to Address Overconfidence in LLMs
PDF ↗Psychology research has shown that humans are poor at estimating their performance on tasks, tending towards underconfidence on easy tasks and overconfidence on difficult tasks. We examine three LLMs, Llama-3-70B-instruct, Claude-3-Sonnet, and GPT-4o, on a range of QA tasks of varying difficulty, and show that models exhibit subtle differences from human patterns of overconfidence: less sensitive to task difficulty, and when prompted to answer based on different personas—e.g., expert vs layman, or different race, gender, and ages—the models will respond with stereotypically biased confidence estimations even though their underlying answer accuracy remains the same. Based on these observations, we propose Answer-Free Confidence Estimation (AFCE) to improve confidence calibration and LLM interpretability in these settings. AFCE is a self-assessment method that employs two stages of prompting, first eliciting only confidence scores on questions, then asking separately for the answer. Experiments on the MMLU and GPQA datasets spanning subjects and difficulty show that this separation of tasks significantly reduces overconfidence and delivers more human-like sensitivity to task difficulty.
IMPARA-GED: Grammatical Error Detection is Boosting Reference-free Grammatical Error Quality Estimator
PDF ↗We propose IMPARA-GED, a novel reference-free automatic grammatical error correction (GEC) evaluation method with grammatical error detection (GED) capabilities. We focus on the quality estimator of IMPARA, an existing automatic GEC evaluation method, and construct that of IMPARA-GED using a pre-trained language model with enhanced GED capabilities. Experimental results on SEEDA, a meta-evaluation dataset for automatic GEC evaluation methods, demonstrate that IMPARA-GED achieves the highest correlation with human sentence-level evaluations.
Large Vision-Language Models often generate hallucinated content that is not grounded in its visual inputs. While prior work focuses on mitigating hallucinations, we instead explore leveraging hallucination correction as a training objective to improve video-language alignment. We introduce HACA, a self-training framework learning to correct hallucinations in descriptions that do not align with the video content. By identifying and correcting inconsistencies, HACA enhances the model’s ability to align video and textual representations for spatio-temporal reasoning. Our experimental results show consistent gains in video-caption binding and text-to-video retrieval tasks, demonstrating that hallucination correction-inspired tasks serve as an effective strategy for improving vision and language alignment.
DiaLLMs: EHR-Enhanced Clinical Conversational System for Clinical Test Recommendation and Diagnosis Prediction
PDF ↗Recent advances in Large Language Models (LLMs) have led to remarkable progresses in medical consultation.However, existing medical LLMs overlook the essential role of Electronic Health Records (EHR) and focus primarily on diagnosis recommendation, limiting their clinical applicability. We propose DiaLLM, the first medical LLM that integrates heterogeneous EHR data into clinically grounded dialogues, enabling clinical test recommendation, result interpretation, and diagnosis prediction to better align with real-world medical practice. To construct clinically grounded dialogues from EHR, we design a Clinical Test Reference (CTR) strategy that maps each clinical code to its corresponding description and classifies test results as “normal” or “abnormal”. Additionally, DiaLLM employs a reinforcement learning framework for evidence acquisition and automated diagnosis. To handle the large action space, we introduce a reject sampling strategy to reduce redundancy and improve exploration efficiency. Furthermore, a confirmation reward and a class-sensitive diagnosis reward are designed to guide accurate diagnosis prediction.Extensive experimental results demonstrate that DiaLLM outperforms baselines in clinical test recommendation and diagnosis prediction. Our code is available at Github.
Natural language processing evaluation has made significant progress, largely driven by the proliferation of powerful large language mod-els (LLMs). New evaluation benchmarks are of increasing priority as the reasoning capabilities of LLMs are expanding at a rapid pace. In particular, while multi-document (MD) reasoning is an area of extreme relevance given LLM capabilities in handling longer-context inputs, few benchmarks exist to rigorously examine model behavior in this setting. Moreover, the multi-document setting is historically challenging for benchmark creation due to the expensive cost of annotating long inputs. In this work, we introduce MDBench, a new dataset for evaluating LLMs on the task of multi-document reasoning. Notably, MDBench is created through a novel synthetic generation process, allowing us to controllably and efficiently generate challenging document sets and the corresponding question-answer (QA) examples. Our novel technique operates on condensed structured seed knowledge, modifying it through LLM-assisted edits to induce MD-specific reasoning challenges. We then convert this structured knowledge into a natural text surface form, generating a document set and corresponding QA example. We analyze the behavior of popular LLMs and prompting techniques, finding that MDBench poses significant challenges for all methods, even with relatively short document sets. We also see our knowledge-guided generation technique (1) allows us to readily perform targeted analysis of MD-specific reasoning capabilities and (2) can be adapted quickly to account for new challenges and future modeling improvements.
The rapid advancement of large language models (LLMs) has unlocked diverse opportunities across domains and applications but has also raised concerns about their tendency to generate harmful responses under jailbreak attacks. However, most existing jailbreak strategies are single-turn with explicit malicious intent, failing to reflect the real-world scenario where interactions can be multi-turn and users can conceal their intents. Recent studies on Theory of Mind (ToM) reveal that LLMs often struggle to infer users’ latent intent in such scenarios. Building on these limitations, we propose a novel jailbreak attack, RED QUEEN ATTACK, which constructs a multi-turn scenario, concealing the malicious intent under the guise of preventing harm. We generate 56k multi-turn concealment data points across 40 scenarios and 14 harmful categories, evaluating four LLM families of different sizes. Results show all models are vulnerable to RED QUEEN ATTACK, reaching 87.6% attack success rate (ASR) on GPT-4o and 77.1% on Llama3-70B. Compared to prior jailbreak attacks, the RED QUEEN ATTACK achieves superior performance on nine out of ten models, with ASR improvements ranging from 2% to 64%. Further analysis reveals that larger models exhibit greater vulnerability to our attack, primarily due to the combination of multi-turn structures and concealment strategies. To enhance safety, we propose RED QUEEN GUARD, a mitigation strategy reducing ASR to below 1% while maintaining model performance on standard benchmarks. Full implementation and dataset are publicly accessible at https://github.com/kriti-hippo/red_queen.
Large Language Models encode behaviors like refusal within their activation space, but identifying these behaviors remains challenging. Existing methods depend on predefined refusal templates detectable in output tokens or manual review. We introduce **COSMIC** (Cosine Similarity Metrics for Inversion of Concepts), an automated framework for direction selection that optimally identifies steering directions and target layers using cosine similarity, entirely independent of output text. COSMIC achieves steering effectiveness comparable to prior work without any prior knowledge or assumptions of a model’s refusal behavior such as the use of certain refusal tokens. Additionally, COSMIC successfully identifies refusal directions in adversarial scenarios and models with weak safety alignment, demonstrating its robustness across diverse settings.
Aspect Sentiment Triplet Extraction (ASTE) is a subtask of Aspect-Based Sentiment Analysis (ABSA) that aims to extract aspect terms, corresponding opinion terms, and their associated sentiment polarities from text. Current end-to-end approaches, whether employing Large Language Models (LLMs) or complex neural network structures, struggle to effectively model the intricate latent relationships between aspects and opinions. Therefore, in this work, we propose Polarity-Aware Sentiment Triplet Extraction with LLM-as-a-judge (PASTEL), a novel pipeline that decomposes the ASTE task into structured subtasks. We employ finetuned LLMs to separately extract the aspect and opinion terms, incorporating a polarity-aware mechanism to enhance opinion extraction. After generating a candidate set through the Cartesian product of the extracted aspect and opinion-sentiment sets, we leverage an LLM-as-a-Judge to validate and prune these candidates. Experimental evaluations demonstrate that PASTEL outperforms existing baselines. Our findings highlight the necessity of modular decomposition in complex sentiment analysis tasks to fully exploit the capabilities of current LLMs.
CogSteer: Cognition-Inspired Selective Layer Intervention for Efficiently Steering Large Language Models
PDF ↗Large Language Models (LLMs) achieve remarkable performance through pretraining on extensive data. This enables efficient adaptation to diverse downstream tasks. However, the lack of interpretability in their underlying mechanisms limits the ability to effectively steer LLMs for specific applications. In this work, we investigate the intrinsic mechanisms of LLMs from a cognitive perspective using eye movement measures. Specifically, we analyze the layer-wise correlation between human cognitive indicators and LLM representations. Building on these insights, we propose a heuristic approach for selecting the optimal steering layer to modulate LLM semantics. To this end, we introduce an efficient selective layer intervention based on prominent parameter-efficient fine-tuning methods, which conventionally adjust either all layers or only the final layer. Additionally, we present an implicit layer contrastive intervention during inference to steer LLMs away from toxic outputs. Extensive experiments on natural language understanding, reasoning, and generation tasks, conducted on GPT-2, LLaMa2-7B, and Mixtral-7B, demonstrate the effectiveness and efficiency of our approach. As a model-agnostic framework, it enhances the interpretability of LLMs while improving efficiency for safe deployment.