Information flows by routes inside the network via mechanisms implemented in the model. These routes can be represented as graphs where nodes correspond to token representations and edges to computations. We automatically build these graphs in a top-down manner, for each prediction leaving only the most important nodes and edges. In contrast to the existing workflows relying on activation patching, we do this through attribution: this allows us to efficiently uncover existing circuits with just a single forward pass. Unlike with patching, we do not need a human to carefully design prediction templates, and we can extract information flow routes for any prediction (not just the ones among the allowed templates). As a result, we can analyze model behavior in general, for specific types of predictions, or different domains. We experiment with Llama 2 and show that some attention head roles are overall important, e.g. previous token heads and subword merging heads. Next, we find similarities in Llama 2 behavior when handling tokens of the same part of speech. Finally, we show that some model components can be specialized on domains such as coding or multilingual texts.
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
Aligning large language models (LLMs) with human expectations requires high-quality instructional dialogues, which can be achieved by raising diverse, in-depth, and insightful instructions that deepen interactions. Existing methods target instructions from real instruction dialogues as a learning goal and fine-tune a user simulator for posing instructions. However, the user simulator struggles to implicitly model complex dialogue flows and pose high-quality instructions. In this paper, we take inspiration from the cognitive abilities inherent in human learning and propose the explicit modeling of complex dialogue flows through instructional strategy reuse. Specifically, we first induce high-level strategies from various real instruction dialogues. These strategies are applied to new dialogue scenarios deductively, where the instructional strategies facilitate high-quality instructions. Experimental results show that our method can generate diverse, in-depth, and insightful instructions for a given dialogue history. The constructed multi-turn instructional dialogues can outperform competitive baselines on the downstream chat model.
Semantic similarity between two sentences is inherently subjective and can vary significantly based on the specific aspects emphasized. Consequently, traditional sentence encoders must be capable of generating conditioned sentence representations that account for diverse conditions or aspects. In this paper, we propose a novel yet efficient framework based on transformer-style language models that facilitates advanced conditioned sentence representation while maintaining model parameters and computational efficiency. Empirical evaluations on the Conditional Semantic Textual Similarity and Knowledge Graph Completion tasks demonstrate the superiority of our proposed framework.
Artificial intelligence has advanced in Medical Visual Question Answering (Med-VQA), but prevalent research tends to focus on the accuracy of the answers, often overlooking the reasoning paths and interpretability, which are crucial in clinical settings. Besides, current Med-VQA algorithms, typically reliant on singular models, lack the robustness needed for real-world medical diagnostics which usually require collaborative expert evaluation. To address these shortcomings, this paper presents MedCoT, a novel hierarchical expert verification reasoning chain method designed to enhance interpretability and accuracy in biomedical imaging inquiries. MedCoT is predicated on two principles: The necessity for explicit reasoning paths in Med-VQA and the requirement for multi-expert review to formulate accurate conclusions. The methodology involves an Initial Specialist proposing diagnostic rationales, followed by a Follow-up Specialist who validates these rationales, and finally, a consensus is reached through a vote among a sparse Mixture of Experts within the locally deployed Diagnostic Specialist, which then provides the definitive diagnosis. Experimental evaluations on four standard Med-VQA datasets demonstrate that MedCoT surpasses existing state-of-the-art approaches, providing significant improvements in performance and interpretability.
StorySparkQA: Expert-Annotated QA Pairs with Real-World Knowledge for Children’s Story-Based Learning
PDF ↗Interactive story reading is common in early childhood education, where teachers expect to teach both language skills and real-world knowledge beyond the story. While many story reading systems have been developed for this activity, they often fail to infuse real-world knowledge into the conversation. This limitation can be attributed to the existing question-answering (QA) datasets used for children’s education, upon which the systems are built, failing to capture the nuances of how education experts think when conducting interactive story reading activities. To bridge this gap, we design an annotation framework, empowered by existing knowledge graph to capture experts’ annotations and thinking process, and leverage this framework to construct StorySparkQA dataset, which comprises 5, 868 expert-annotated QA pairs with real-world knowledge. We conduct automated and human expert evaluations across various QA pair generation settings to demonstrate that our StorySparkQA can effectively support models in generating QA pairs that target real-world knowledge beyond story content. StorySparkQA is available at https://huggingface.co/datasets/NEU-HAI/StorySparkQA.
The complexity of text-embedded images presents a formidable challenge in machine learning given the need for multimodal understanding of multiple aspects of expression conveyed by them. While previous research in multimodal analysis has primarily focused on singular aspects such as hate speech and its subclasses, this study expands this focus to encompass multiple aspects of linguistics: hate, targets of hate, stance, and humor. We introduce a novel dataset PrideMM comprising 5,063 text-embedded images associated with the LGBTQ+ Pride movement, thereby addressing a serious gap in existing resources. We conduct extensive experimentation on PrideMM by using unimodal and multimodal baseline methods to establish benchmarks for each task. Additionally, we propose a novel framework MemeCLIP for efficient downstream learning while preserving the knowledge of the pre-trained CLIP model. The results of our experiments show that MemeCLIP achieves superior performance compared to previously proposed frameworks on two real-world datasets. We further compare the performance of MemeCLIP and zero-shot GPT-4 on the hate classification task. Finally, we discuss the shortcomings of our model by qualitatively analyzing misclassified samples. Our code and dataset are publicly available at: https://github.com/SiddhantBikram/MemeCLIP.
Medical texts are notoriously challenging to read. Properly measuring their readability is the first step towards making them more accessible. Here, we present the first systematic study on fine-grained readability measurements in the medical domain, at both sentence-level and span-level. We first introduce a new dataset MedReadMe, which consists of manually annotated readability ratings and fine-grained complex span annotation for 4,520 sentences, featuring two novel “Google-Easy” and “Google-Hard” categories. It supports our quantitative analysis, which covers 650 linguistic features and additional complex span features, to answer “why medical sentences are so hard.” Enabled by our high-quality annotation, we benchmark several state-of-the-art sentence-level readability metrics, including unsupervised, supervised, and prompting-based methods using recently developed large language models (LLMs). Informed by our fine-grained complex span annotation, we find that adding a single feature, capturing the number of jargon spans, into existing readability formulas can significantly improve their correlation with human judgments, and also make them more stable. We will publicly release data and code.
Vision Language Models (VLMs) are rapidly advancing in their capability to answer information-seeking questions. As these models are widely deployed in consumer applications, they could lead to new privacy risks due to emergent abilities to identify people in photos, geolocate images, etc. As we demonstrate, somewhat surprisingly, current open-source and proprietary VLMs are very capable image geolocators, making widespread geolocation with VLMs an immediate privacy risk, rather than merely a theoretical future concern. As a first step to address this challenge, we develop a new benchmark, GPTGeoChat, to test the capability of VLMs to moderate geolocation dialogues with users. We collect a set of 1,000 image geolocation conversations between in-house annotators and GPT-4v, which are annotated with the granularity of location information revealed at each turn. Using this new dataset we evaluate the ability of various VLMs to moderate GPT-4v geolocation conversations by determining when too much location information has been revealed. We find that custom fine-tuned models perform on par with prompted API-based models when identifying leaked location information at the country or city level, however fine-tuning on supervised data appears to be needed to accurately moderate finer granularities, such as the name of a restaurant or building.
Adaptive Question Answering: Enhancing Language Model Proficiency for Addressing Knowledge Conflicts with Source Citations
PDF ↗Resolving knowledge conflicts is a crucial challenge in Question Answering (QA) tasks, as the internet contains numerous conflicting facts and opinions. While some research has made progress in tackling ambiguous settings where multiple valid answers exist, these approaches often neglect to provide source citations, leaving users to evaluate the factuality of each answer. On the other hand, existing work on citation generation has focused on unambiguous settings with single answers, failing to address the complexity of real-world scenarios. Despite the importance of both aspects, no prior research has combined them, leaving a significant gap in the development of QA systems. In this work, we bridge this gap by proposing the novel task of QA with source citation in ambiguous settings, where multiple valid answers exist. To facilitate research in this area, we create a comprehensive framework consisting of: (1) five novel datasets, obtained by augmenting three existing reading comprehension datasets with citation meta-data across various ambiguous settings, such as distractors and paraphrasing; (2) the first ambiguous multi-hop QA dataset featuring real-world, naturally occurring contexts; (3) two new metrics to evaluate models’ performances; and (4) several strong baselines using rule-based, prompting, and finetuning approaches over five large language models. We hope that this new task, datasets, metrics, and baselines will inspire the community to push the boundaries of QA research and develop more trustworthy and interpretable systems.
As machine learning models continue to swiftly advance, calibrating their performance has become a major concern prior to practical and widespread implementation. Most existing calibration methods often negatively impact model accuracy due to the lack of diversity of validation data, resulting in reduced generalizability. To address this, we propose a calibration method that incorporates synthetic data without compromising accuracy. We derive the expected calibration error (ECE) bound using the Probably Approximately Correct (PAC) learning framework. Large language models (LLMs), known for their ability to mimic real data and generate text with mixed class labels, are utilized as a synthetic data generation strategy to lower the ECE bound and improve model accuracy on real test data. Additionally, we propose data generation mechanisms for efficient calibration. Testing our method on four different natural language processing tasks, we observed an average up to 34% increase in accuracy and 33% decrease in ECE.
Framing is used to present some selective aspects of an issue and making them more salient, which aims to promote certain values, interpretations, or solutions (Entman, 1993). This study investigates the nuances of media framing on public perception and understanding by examining how events are presented within news articles. Unlike previous research that primarily focused on word choice as a framing device, this work explores the comprehensive narrative construction through events and their relations. Our method integrates event extraction, cross-document event coreference, and causal relationship mapping among events to extract framing devices employed by media to assess their role in framing the narrative. We evaluate our approach with a media attitude detection task and show that the use of event mentions, event cluster descriptors, and their causal relations effectively captures the subtle nuances of framing, thereby providing deeper insights into the attitudes conveyed by news articles. The experimental results show the framing device models surpass the baseline models and offers a more detailed and explainable analysis of media framing effects. We make the source code and dataset publicly available.
Evaluations of creative stories generated by large language models (LLMs) often focus on objective properties of the text, such as its style, coherence, and diversity. While these metrics are indispensable, they do not speak to a story’s subjective, psychological impact from a reader’s perspective. We introduce the Psychological Depth Scale (PDS), a novel framework rooted in literary theory that measures an LLM’s ability to produce authentic and narratively complex stories that provoke emotion, empathy, and engagement. We empirically validate our framework by showing that humans can consistently evaluate stories based on PDS (0.72 Krippendorff’s alpha). We also explore techniques for automating the PDS to easily scale future analyses. GPT-4o, combined with a novel Mixture-of-Personas (MoP) prompting strategy, achieves an average Spearman correlation of 0.51 with human judgment while Llama-3-70B with constrained decoding scores as high as 0.68 for empathy. Finally, we compared the depth of stories authored by both humans and LLMs. Surprisingly, GPT-4 stories either surpassed or were statistically indistinguishable from highly-rated human-written stories sourced from Reddit. By shifting the focus from text to reader, the Psychological Depth Scale is a validated, automated, and systematic means of measuring the capacity of LLMs to connect with humans through the stories they tell.
Large Language Models (LLMs) possess the potential to exert substantial influence on public perceptions and interactions with information. This raises concerns about the societal impact that could arise if the ideologies within these models can be easily manipulated. In this work, we investigate how effectively LLMs can learn and generalize ideological biases from their instruction-tuning data. Our findings reveal a concerning vulnerability: exposure to only a small amount of ideologically driven samples significantly alters the ideology of LLMs. Notably, LLMs demonstrate a startling ability to absorb ideology from one topic and generalize it to even unrelated ones. The ease with which LLMs’ ideologies can be skewed underscores the risks associated with intentionally poisoned training data by malicious actors or inadvertently introduced biases by data annotators. It also emphasizes the imperative for robust safeguards to mitigate the influence of ideological manipulations on LLMs.
A critical component of the current generation of language models is preference alignment, which aims to precisely control the model’s behavior to meet human needs and values. The most notable among such methods is Reinforcement Learning with Human Feedback (RLHF) and its offline variant Direct Preference Optimization (DPO), both of which seek to maximize a reward model based on human preferences. In particular, DPO derives reward signals directly from the offline preference data, but in doing so overfits the reward signals and generates suboptimal responses that may contain human biases in the dataset. In this work, we propose a practical application of a diversity-seeking RL algorithm called GFlowNet-DPO (GDPO) in an offline preference alignment setting to curtail such challenges. Empirical results show GDPO can generate far more diverse responses than the baseline methods that are still relatively aligned with human values in dialog generation and summarization tasks.
Do large multilingual language models learn language universals? We consider a candidate universal much-discussed in the linguistics literature, the Final-over-Final Condition (Sheehan et al., 2017b). This Condition is syntactic in the sense that it can only be stated by reference to abstract sentence properties such as nested phrases and head direction. A study of typologically diverse “mixed head direction” languages confirms that the Condition holds in corpora. But in a targeted syntactic evaluation, Gemini Pro only seems to respect the Condition in German, Russian, Hungarian and Serbian. These relatively high-resource languages contrast with Basque, where Gemini Pro does not seem to have learned the Condition at all. This result suggests that modern language models may need additional sources of bias in order to become truly human-like, within a developmentally-realistic budget of training data.
As large language models (LLMs) evolve, evaluating their output reliably becomes increasingly difficult due to the high cost of human evaluation. To address this, we introduce FLAMe, a family of Foundational Large Autorater Models. FLAMe is trained on a diverse set of over 100 quality assessment tasks, incorporating 5M+ human judgments curated from publicly released human evaluations. FLAMe outperforms models like GPT-4 and Claude-3 on various held-out tasks, and serves as a powerful starting point for fine-tuning, as shown in our reward model evaluation case study (FLAMe-RM). On Reward-Bench, FLAMe-RM-24B achieves 87.8% accuracy, surpassing GPT-4-0125 (85.9%) and GPT-4o (84.7%). Additionally, we introduce FLAMe-Opt-RM, an efficient tail-patch fine-tuning approach that offers competitive RewardBench performance using 25×fewer training datapoints. Our FLAMe variants outperform popular proprietary LLM-as-a-Judge models on 8 of 12 autorater benchmarks, covering 53 quality assessment tasks, including RewardBench and LLM-AggreFact. Finally, our analysis shows that FLAMe is significantly less biased than other LLM-as-a-Judge models on the CoBBLEr autorater bias benchmark.
Synthetic long-context LLM benchmarks (e.g., “needle-in-the-haystack”) test only surface-level retrieval capabilities; but how well can long-context LLMs retrieve, synthesize, and reason over information across book-length inputs? We address this question by creating NoCha, a dataset of 1,001 minimally different pairs of true and false claims about 67 recently-published English fictional books, written by human readers of those books. In contrast to existing long-context benchmarks, our annotators confirm that the largest share of pairs in NoCha require global reasoning over the entire book to verify. Our experiments show that while human readers easily perform this task, it is enormously challenging for all ten long-context LLMs that we evaluate: no open-weight model performs above random chance (despite their strong performance on synthetic benchmarks), while GPT-4o achieves the highest pair accuracy at 55.8%. Further analysis reveals that (1) on average, models perform much better on pairs that require only sentence-level retrieval vs. global reasoning; (2) model-generated explanations for their decisions are often inaccurate even for correctly-labeled claims; and (3) models perform substantially worse on speculative fiction books that contain extensive world-building. The methodology proposed in NoCha allows for the evolution of the benchmark dataset and the easy analysis of future models.
Reasoning Paths with Reference Objects Elicit Quantitative Spatial Reasoning in Large Vision-Language Models
PDF ↗Despite recent advances demonstrating vision- language models’ (VLMs) abilities to describe complex relationships among objects in images using natural language, their capability to quantitatively reason about object sizes and distances remains underexplored. In this work, we introduce a manually annotated benchmark of 241 questions across five categories specifically designed for quantitative spatial reasoning, and systematically investigate the performance of SoTA VLMs on this task. Our analysis reveals that questions involving reasoning about distances between objects are particularly challenging for SoTA VLMs; however, some VLMs perform significantly better at this task than others, with an almost 40 points gap between the two best performing models. We also make the surprising observation that the success rate of the top-performing VLM increases by 19 points when a reasoning path using a reference object emerges naturally in the response. Inspired by this observation, we develop a zero-shot prompting technique, SpatialPrompt, that encourages VLMs to answer quantitative spatial questions using references objects as visual cues. Specifically, we demonstrate that instruct- ing VLMs to use reference objects in their reasoning paths significantly improves their quantitative spatial reasoning performance, bypassing the need for external data, architectural modifications, or fine-tuning. Remarkably, by solely using SpatialPrompt, Gemini 1.5 Pro, GPT-4V, and GPT-4o improve by 56.2, 28.5, and 6.7 points on average in Q-Spatial Bench without the need for more data, model architectural modifications, or fine-tuning.
We introduce Mathador-LM, a new benchmark for evaluating the mathematical reasoning on large language models (LLMs), combining ruleset interpretation, planning, and problem-solving. This benchmark is inspired by the Mathador game, where the objective is to reach a target number using basic arithmetic operations on a given set of base numbers, following a simple set of rules. We show that, across leading LLMs, we obtain stable average performance while generating benchmark instances dynamically, following a target difficulty level. Thus, our benchmark alleviates concerns about test-set leakage into training data, an issue that often undermines popular benchmarks. Additionally, we conduct a comprehensive evaluation of both open and closed-source state-of-the-art LLMs on Mathador-LM. Our findings reveal that contemporary models struggle with Mathador-LM, scoring significantly lower than average 3rd graders. This stands in stark contrast to their strong performance on popular mathematical reasoning benchmarks. The implementation of Mathador-LM benchmark is available at https://github.com/IST-DASLab/Mathador-LM.