Do LLMs align with human perceptions of safety? We study this question via *annotation alignment*, the extent to which LLMs and humans agree when annotating the safety of user-chatbot conversations. We leverage the recent DICES dataset (Aroyo et al. 2023), in which 350 conversations are each rated for safety by 112 annotators spanning 10 race-gender groups. GPT-4 achieves a Pearson correlation of r=0.59 with the average annotator rating, higher than the median annotator’s correlation with the average (r=0.51). We show that larger datasets are needed to resolve whether GPT-4 exhibits disparities in how well it correlates with different demographic groups. Also, there is substantial idiosyncratic variation in correlation within groups, suggesting that race & gender do not fully capture differences in alignment. Finally, we find that GPT-4 cannot predict when one demographic group finds a conversation more unsafe than another.
论文检索
输入标题、作者或关键词，从 2,271 篇学术成果中精准定位
Static word embeddings are ubiquitous in computational social science applications and contribute to practical decision-making in a variety of fields including law and healthcare. However, assessing the statistical uncertainty in downstream conclusions drawn from word embedding statistics has remained challenging. When using only point estimates for embeddings, researchers have no streamlined way of assessing the degree to which their model selection criteria or scientific conclusions are subject to noise due to sparsity in the underlying data used to generate the embeddings. We introduce a method to obtain approximate, easy-to-use, and scalable reconstruction error variance estimates for GloVe, one of the most widely used word embedding models, using an analytical approximation to a multivariate normal model. To demonstrate the value of embeddings with variance (GloVe-V), we illustrate how our approach enables principled hypothesis testing in core word embedding tasks, such as comparing the similarity between different word pairs in vector space, assessing the performance of different models, and analyzing the relative degree of ethnic or gender bias in a corpus using different word lists.
Modular AI systems can be developed using LLM-prompts-based modules to minimize deployment time even for complex tasks. However, these systems do not always perform well and improving them using the data traces collected from a deployment remains an open challenge. The data traces contain LLM inputs and outputs, but the annotations from LLMs are noisy. We hypothesize that Active Label Correction (ALC) can be use on the collected data to train smaller task-specific improved models that can replace LLM-based modules. In this paper, we study the noise in three GPT-3.5-annotated datasets and their denoising with human feedback. We also propose a novel method ALC3 that iteratively applies three updates to the training dataset: auto-correction, correction using human feedback and filtering. Our results show that ALC3 can lead to oracle performance with feedback on 17-24% fewer examples than the number of noisy examples in the dataset across three different NLP tasks.
Language model alignment research often attempts to ensure that models are not only helpful and harmless, but also truthful and unbiased. However, optimizing these objectives simultaneously can obscure how improving one aspect might impact the others. In this work, we focus on analyzing the relationship between two concepts essential in both language model alignment and political science: truthfulness and political bias. We train reward models on various popular truthfulness datasets and subsequently evaluate their political bias. Our findings reveal that optimizing reward models for truthfulness on these datasets tends to result in a left-leaning political bias. We also find that existing open-source reward models (i.e., those trained on standard human preference datasets) already show a similar bias and that the bias is larger for larger models. These results raise important questions about the datasets used to represent truthfulness, potential limitations of aligning models to be both truthful and politically unbiased, and what language models capture about the relationship between truth and politics.
Retrieval-augmented generation has gained popularity as a framework to enhance large language models with external knowledge. However, its effectiveness hinges on the retrieval robustness of the model. If the model lacks retrieval robustness, its performance is constrained by the accuracy of the retriever, resulting in significant compromises when the retrieved context is irrelevant. In this paper, we evaluate the “implicit” retrieval robustness of various large language models, instructing them to directly output the final answer without explicitly judging the relevance of the retrieved context. Our findings reveal that fine-tuning on a mix of gold and distracting context significantly enhances the model’s robustness to retrieval inaccuracies, while still maintaining its ability to extract correct answers when retrieval is accurate. This suggests that large language models can implicitly handle relevant or irrelevant retrieved context by learning solely from the supervision of the final answer in an end-to-end manner. Introducing an additional process for explicit relevance judgment can be unnecessary and disrupts the end-to-end approach.
The most effective techniques to detect LLM-generated text rely on inserting a detectable signature—or watermark—during the model’s decoding process. Most existing watermarking methods require access to the underlying LLM’s logits, which LLM API providers are loath to share due to fears of model distillation. As such, these watermarks must be implemented independently by each LLM provider. In this paper, we develop PostMark, a modular post-hoc watermarking procedure in which an input-dependent set of words (determined via a semantic embedding) is inserted into the text after the decoding process has completed. Critically, PostMark does not require logit access, which means it can be implemented by a third party. We also show that PostMark is more robust to paraphrasing attacks than existing watermarking methods: our experiments cover eight baseline algorithms, five base LLMs, and three datasets. Finally, we evaluate the impact of PostMark on text quality using both automated and human assessments, highlighting the trade-off between quality and robustness to paraphrasing. We release our code, outputs, and annotations at https://github.com/lilakk/PostMark.
Language agents, built on top of language models (LMs), are systems that can interact with complex environments, such as the open web. In this work, we examine whether such agents can perform realistic and time-consuming tasks on the web, e.g., monitoring real-estate markets or locating relevant nearby businesses. We introduce AssistantBench, a challenging new benchmark consisting of 214 realistic tasks that can be automatically evaluated, covering different scenarios and domains. We find that AssistantBench exposes the limitations of current systems, including language models and retrieval-augmented language models, as no model reaches an accuracy of more than 25 points. While closed-book LMs perform well in terms of accuracy, they exhibit low precision and tend to hallucinate facts. State-of-the-art web agents reach a score of near zero. Additionally, we introduce SeePlanAct (SPA), a new web agent that significantly outperforms previous agents, and an ensemble of SPA and closed-book models reaches the best overall performance. Moreover, we analyze failures of current systems and highlight that open web navigation remains a major challenge.
An open challenge in recent machine learning is about how to improve the reasoning capability of large language models (LLMs) in a black-box setting, i.e., without access to detailed information such as output token probabilities. Existing approaches either rely on accessibility (which is often unrealistic) or involve significantly increased train- and inference-time costs. This paper addresses those limitations or shortcomings by proposing a novel approach, namely CoBB (Correct for improving QA reasoning of Black-Box LLMs). It uses a trained adaptation model to perform a seq2seq mapping from the often-imperfect reasonings of the original black-box LLM to the correct or improved reasonings. Specifically, the adaptation model is initialized with a relatively small open-source LLM and adapted over a collection of sub-sampled training pairs. To select the representative pairs of correct and incorrect reasonings, we formulated the dataset construction as an optimization problem that minimizes the statistical divergence between the sampled subset and the entire collection, and solved it via a genetic algorithm. We then train the adaptation model over the sampled pairs by contrasting the likelihoods of correct and incorrect reasonings. Our experimental results demonstrate that CoBB significantly improves reasoning accuracy across various QA benchmarks, compared to the best-performing adaptation baselines.
Synthetic data is widely used in speech recognition due to the availability of text-to-speech models, which facilitate adapting models to previously unseen text domains. However, existing methods suffer in performance when they fine-tune an automatic speech recognition (ASR) model on synthetic data as they suffer from the distributional shift commonly referred to as the synthetic-to-real gap. In this paper, we find that task arithmetic is effective at mitigating this gap. Our proposed method, SYN2REAL task vector, shows an average improvement of 10.03% improvement in word error rate over baselines on the SLURP dataset. Additionally, we show that an average of SYN2REAL task vectors, when we have real speeches from multiple different domains, can further adapt the original ASR model to perform better on the target text domain.
We present RICHES, a novel approach that interleaves retrieval with sequence generation tasks. RICHES offers an alternative to conventional RAG systems by eliminating the need for separate retriever and generator. It retrieves documents by directly decoding their contents, constrained on the corpus. Unifying retrieval with generation allows us to adapt to diverse new tasks via prompting alone. RICHES can work with any Instruction-tuned model, without additional training. It provides attributed evidence, supports multi-hop retrievals and interleaves thoughts to plan on what to retrieve next, all within a single decoding pass of the LLM. We demonstrate the strong performance of RICHES across ODQA tasks including attributed and multi-hop QA.
Rapid advancements over the years have helped machine learning models reach previously hard-to-achieve goals, sometimes even exceeding human capabilities. However, achieving desired accuracy comes at the cost of larger model sizes and increased computational demands. Thus, serving predictions from these models to meet any latency and cost requirements of applications remains a key challenge, despite recent work in building inference serving systems as well as algorithmic approaches that dynamically adapt models based on inputs. Our paper introduces a new form of dynamism, modality selection, where we adaptively choose modalities from inference inputs while maintaining the model quality. We introduce MOSEL, an automated inference serving system for multi-modal ML models that carefully picks input modalities per request based on user-defined performance and accuracy requirements. MOSEL exploits modality configurations extensively, improving system throughput by 3.6 \times with an accuracy guarantee. It also reduces job completion times by 11\times compared to modality-agnostic approaches.
Beyond Label Attention: Transparency in Language Models for Automated Medical Coding via Dictionary Learning
PDF ↗Medical coding, the translation of unstructured clinical text into standardized medical codes, is a crucial but time-consuming healthcare practice. Though large language models (LLM) could automate the coding process and improve the efficiency of such tasks, interpretability remains paramount for maintaining patient trust. Current efforts in interpretability of medical coding applications rely heavily on label attention mechanisms, which often leads to the highlighting of extraneous tokens irrelevant to the ICD code. To facilitate accurate interpretability in medical language models, this paper leverages dictionary learning that can efficiently extract sparsely activated representations from dense language model embeddings in superposition. Compared with common label attention mechanisms, our model goes beyond token-level representations by building an interpretable dictionary which enhances the mechanistic-based explanations for each ICD code prediction, even when the highlighted tokens are medically irrelevant. We show that dictionary features are human interpretable, can elucidate the hidden meanings of upwards of 90% of medically irrelevant tokens, and steer model behavior.
Recognizing if LLM output can be grounded in evidence is central to many tasks in NLP: retrieval-augmented generation, summarization, document-grounded dialogue, and more. Current approaches to this kind of fact-checking are based on verifying each piece of a model generation against potential evidence using an LLM. However, this process can be very computationally expensive, requiring many calls to a model to check a single response. In this work, we show how to build small fact-checking models that have GPT-4-level performance but for 400x lower cost. We do this by constructing synthetic training data with GPT-4, which involves creating realistic yet challenging instances of factual errors via a structured generation procedure. Training on this data teaches models to check each fact in the claim and recognize synthesis of information across sentences. For evaluation, we unify datasets from recent work on fact-checking and grounding LLM generations into a new benchmark, LLM-AggreFact. Our best system MiniCheck-FT5 (770M parameters) outperforms all systems of comparable size and reaches GPT-4 accuracy. We release LLM-AggreFact, code for data synthesis, and models.
A Comprehensive Survey of Scientific Large Language Models and Their Applications in Scientific Discovery
PDF ↗In many scientific fields, large language models (LLMs) have revolutionized the way text and other modalities of data (e.g., molecules and proteins) are handled, achieving superior performance in various applications and augmenting the scientific discovery process. Nevertheless, previous surveys on scientific LLMs often concentrate on one or two fields or a single modality. In this paper, we aim to provide a more holistic view of the research landscape by unveiling cross-field and cross-modal connections between scientific LLMs regarding their architectures and pre-training techniques. To this end, we comprehensively survey over 260 scientific LLMs, discuss their commonalities and differences, as well as summarize pre-training datasets and evaluation tasks for each field and modality. Moreover, we investigate how LLMs have been deployed to benefit scientific discovery. Resources related to this survey are available at https://github.com/yuzhimanhua/Awesome-Scientific-Language-Models.
This paper explores the utilization of LLMs for data preprocessing (DP), a crucial step in the data mining pipeline that transforms raw data into a clean format. We instruction-tune local LLMs as universal DP task solvers that operate on a local, single, and low-priced GPU, ensuring data security and enabling further customization. We select a collection of datasets across four representative DP tasks and construct instruction data using data configuration, knowledge injection, and reasoning data distillation techniques tailored to DP. By tuning Mistral-7B, Llama 3-8B, and OpenOrca-Platypus2-13B, our models, Jellyfish-7B/8B/13B, deliver competitiveness compared to GPT-3.5/4 models and strong generalizability to unseen tasks while barely compromising the base models’ abilities in NLP tasks. Meanwhile, Jellyfish offers enhanced reasoning capabilities compared to GPT-3.5. Our models are available at: https://huggingface.co/NECOUDBFM/JellyfishOur instruction dataset is available at: https://huggingface.co/datasets/NECOUDBFM/Jellyfish-Instruct
Alignment is a crucial step to enhance the instruction-following and conversational abilities of language models. Despite many recent works proposing new algorithms, datasets, and training pipelines, there is a lack of comprehensive studies measuring the impact of various design choices throughout the whole training process. We first conduct a rigorous analysis over a three-stage training pipeline consisting of supervised fine-tuning, offline preference learning, and online preference learning. We have found that using techniques like sequence packing, loss masking in SFT, increasing the preference dataset size in DPO, and online DPO training can significantly improve the performance of language models. We then train from Gemma-2b-base and LLama-3-8b-base, and find that our best models exceed the performance of the official instruct models tuned with closed-source data and algorithms. Our code and models can be found at https://github.com/Columbia-NLP-Lab/LionAlignment.
Revisiting Who’s Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective
PDF ↗This paper investigates Who’s Harry Potter (WHP), a pioneering yet insufficiently understood method for LLM unlearning. We explore it in two steps. First, we introduce a new task of LLM targeted unlearning, where given an unlearning target (e.g., a person) and some unlearning documents, we aim to unlearn only the information about the target, rather than everything in the unlearning documents. We further argue that a successful unlearning should satisfy criteria such as not outputting gibberish, not fabricating facts about the unlearning target, and not releasing factual information under jailbreak attacks. Second, we construct a causal intervention framework for targeted unlearning, where the knowledge of the unlearning target is modeled as a confounder between LLM input and output, and the unlearning process as a deconfounding process. This framework justifies and extends WHP, deriving a simple unlearning algorithm that includes WHP as a special case. Experiments on existing and new datasets show that our approach, without explicitly optimizing for the aforementioned criteria, achieves competitive performance in all of them.
Weight-based model editing methods update the parametric knowledge of language models post-training. However, these methods can unintentionally alter unrelated parametric knowledge representations, potentially increasing the risk of harm. In this work, we investigate how weight editing methods unexpectedly amplify model biases after edits. We introduce a novel benchmark dataset, Seesaw-CF, for measuring bias amplification of model editing methods for demographic traits such as race, geographic origin, and gender. We use Seesaw-CF to examine the impact of model editing on bias in five large language models. Our results demonstrate that edited models exhibit, to various degrees, more biased behavior for certain demographic groups than before they were edited, specifically becoming less confident in properties for Asian and African subjects. Additionally, editing facts about place of birth, country of citizenship, or gender has particularly negative effects on the model’s knowledge about unrelated properties, such as field of work, a pattern observed across multiple models.
The rapid scaling of large language models (LLMs) has raised concerns about the transparency and fair use of the data used in their pretraining. Detecting such content is challenging due to the scale of the data and limited exposure of each instance during training. We propose ReCaLL (Relative Conditional Log-Likelihood), a novel membership inference attack (MIA) to detect LLMs’ pretraining data by leveraging their conditional language modeling capabilities. ReCaLL examines the relative change in conditional log-likelihoods when prefixing target data points with non-member context. Our empirical findings show that conditioning member data on non-member prefixes induces a larger decrease in log-likelihood compared to non-member data. We conduct comprehensive experiments and show that ReCaLL achieves state-of-the-art performance on the WikiMIA dataset, even with random and synthetic prefixes, and can be further improved using an ensemble approach. Moreover, we conduct an in-depth analysis of LLMs’ behavior with different membership contexts, providing insights into how LLMs leverage membership information for effective inference at both the sequence and token level.
Exploring Nested Named Entity Recognition with Large Language Models: Methods, Challenges, and Insights
PDF ↗Nested Named Entity Recognition (NER) poses a significant challenge in Natural Language Processing (NLP), demanding sophisticated techniques to identify entities within entities. This research investigates the application of Large Language Models (LLMs) to nested NER, exploring methodologies from prior work and introducing specific reasoning techniques and instructions to improve LLM efficacy. Through experiments conducted on the ACE 2004, ACE 2005, and GENIA datasets, we evaluate the impact of these approaches on nested NER performance. Results indicate that output format critically influences nested NER performance, methodologies from previous works are less effective, and our nested NER-tailored instructions significantly enhance performance. Additionally, we find that label information and descriptions of nested cases are crucial in eliciting the capabilities of LLMs for nested NER, especially in specific domains (i.e., the GENIA dataset). However, these methods still do not outperform BERT-based models, highlighting the ongoing need for innovative approaches in nested NER with LLMs.