Pretrained language models sometimes possess knowledge that we do not wish them to, including memorized personal information and knowledge that could be used to harm people. They can also output toxic or harmful text. To mitigate these safety and informational issues, we propose an attack-and-defense framework for studying the task of deleting sensitive information directly from model weights. We study direct edits to model weights because (1) this approach should guarantee that particular deleted information is never extracted by future prompt attacks, and (2) it should protect against whitebox attacks, which is necessary for making claims about safety/privacy in a setting where publicly available model weights could be used to elicit sensitive information. Our threat model assumes that an attack succeeds if the answer to a sensitive question is located among a set of B generated candidates, based on scenarios where the information would be insecure if the answer is among B candidates. Experimentally, we show that even state-of-the-art model editing methods such as ROME struggle to truly delete factual information from models like GPT-J, as our whitebox and blackbox attacks can recover "deleted" information from an edited model 38% of the time. These attacks leverage two key observations: (1) that traces of deleted information can be found in intermediate model hidden states, and (2) that applying an editing method for one question may not delete information across rephrased versions of the question. Finally, we provide new defense methods that protect against some extraction attacks, but we do not find a single universally effective defense method. Our results suggest that truly deleting sensitive information is a tractable but difficult problem, since even relatively low attack success rates have potentially severe societal implications for real-world deployment of language models.
相關內容
The rise of large language models (LLMs) has brought a critical need for high-quality human-labeled data, particularly for processes like human feedback and evaluation. A common practice is to label data via consensus annotation over crowdworker judgments. However, annotators' judgments for subjective tasks can differ in many ways: they may have different qualitative judgments about an example, and they may map those to a labeling scheme in different ways. We show that these nuances can be captured by natural language explanations, and propose a method to rescale ordinal annotations and explanations using LLMs. Specifically, we feed annotators' Likert ratings and corresponding explanations into an LLM and prompt it to produce a numeric score anchored in a scoring rubric. These scores should reflect the annotators' underlying assessments of the example. The rubric can be designed or modified after annotation, and include distinctions that may not have been known when the original error taxonomy was devised. We explore our technique in the context of rating system outputs for a document-grounded question answering task, where LLMs achieve near-human performance. Our method rescales the raw judgments without impacting agreement and brings the scores closer to human judgments grounded in the same scoring rubric.
While recent advancements in large language models (LLMs) bring us closer to achieving artificial general intelligence, the question persists: Do LLMs truly understand language, or do they merely mimic comprehension through pattern recognition? This study seeks to explore this question through the lens of syntax, a crucial component of sentence comprehension. Adopting a natural language question-answering (Q&A) scheme, we craft questions targeting nine syntactic knowledge points that are most closely related to sentence comprehension. Experiments conducted on 24 LLMs suggest that most have a limited grasp of syntactic knowledge, exhibiting notable discrepancies across different syntactic knowledge points. In particular, questions involving prepositional phrase attachment pose the greatest challenge, whereas those concerning adjectival modifier and indirect object are relatively easier for LLMs to handle. Furthermore, a case study on the training dynamics of the LLMs reveals that the majority of syntactic knowledge is learned during the initial stages of training, hinting that simply increasing the number of training tokens may not be the `silver bullet' for improving the comprehension ability of LLMs.
With the increasing amount of data available to scientists in disciplines as diverse as bioinformatics, physics, and remote sensing, scientific workflow systems are becoming increasingly important for composing and executing scalable data analysis pipelines. When writing such workflows, users need to specify the resources to be reserved for tasks so that sufficient resources are allocated on the target cluster infrastructure. Crucially, underestimating a task's memory requirements can result in task failures. Therefore, users often resort to overprovisioning, resulting in significant resource wastage and decreased throughput. In this paper, we propose a novel online method that uses monitoring time series data to predict task memory usage in order to reduce the memory wastage of scientific workflow tasks. Our method predicts a task's runtime, divides it into k equally-sized segments, and learns the peak memory value for each segment depending on the total file input size. We evaluate the prototype implementation of our method using workflows from the publicly available nf-core repository, showing an average memory wastage reduction of 29.48% compared to the best state-of-the-art approach
A hallmark property of explainable AI models is the ability to teach other agents, communicating knowledge of how to perform a task. While Large Language Models perform complex reasoning by generating explanations for their predictions, it is unclear whether they also make good teachers for weaker agents. To address this, we consider a student-teacher framework between two LLM agents and study if, when, and how the teacher should intervene with natural language explanations to improve the student's performance. Since communication is expensive, we define a budget such that the teacher only communicates explanations for a fraction of the data, after which the student should perform well on its own. We decompose the teaching problem along four axes: (1) if teacher's test time intervention improve student predictions, (2) when it is worth explaining a data point, (3) how the teacher should personalize explanations to better teach the student, and (4) if teacher explanations also improve students on future unexplained data. We first show that teacher LLMs can indeed intervene on student reasoning to improve their performance. Next, inspired by the Theory of Mind abilities of effective teachers, we propose building two few-shot mental models of the student. The first model defines an Intervention Function that simulates the utility of an intervention, allowing the teacher to intervene when this utility is the highest and improving student performance at lower budgets. The second model enables the teacher to personalize explanations for a particular student and outperform unpersonalized teachers. We also demonstrate that in multi-turn interactions, teacher explanations generalize and learning from explained data improves student performance on future unexplained data. Finally, we verify that misaligned teachers can lower student performance to random chance by intentionally misleading them.
Path reasoning methods over knowledge graphs have gained popularity for their potential to improve transparency in recommender systems. However, the resulting models still rely on pre-trained knowledge graph embeddings, fail to fully exploit the interdependence between entities and relations in the KG for recommendation, and may generate inaccurate explanations. In this paper, we introduce PEARLM, a novel approach that efficiently captures user behaviour and product-side knowledge through language modelling. With our approach, knowledge graph embeddings are directly learned from paths over the KG by the language model, which also unifies entities and relations in the same optimisation space. Constraints on the sequence decoding additionally guarantee path faithfulness with respect to the KG. Experiments on two datasets show the effectiveness of our approach compared to state-of-the-art baselines. Source code and datasets: AVAILABLE AFTER GETTING ACCEPTED.
Sparse matrix representations are ubiquitous in computational science and machine learning, leading to significant reductions in compute time, in comparison to dense representation, for problems that have local connectivity. The adoption of sparse representation in leading ML frameworks such as PyTorch is incomplete, however, with support for both automatic differentiation and GPU acceleration missing. In this work, we present an implementation of a CSR-based sparse matrix wrapper for PyTorch with CUDA acceleration for basic matrix operations, as well as automatic differentiability. We also present several applications of the resulting sparse kernels to optimization problems, demonstrating ease of implementation and performance measurements versus their dense counterparts.
Recent large language models (LLM) exhibit sub-optimal performance on low-resource languages, as the training data of these models is usually dominated by English and other high-resource languages. Furthermore, it is challenging to train models for low-resource languages, especially from scratch, due to a lack of high quality training data. Adapting pretrained LLMs reduces the need for data in the new language while also providing cross lingual transfer capabilities. However, naively adapting to new languages leads to catastrophic forgetting and poor tokenizer efficiency. In this work, we study how to efficiently adapt any existing pretrained LLM to a new language without running into these issues. In particular, we improve the encoding efficiency of the tokenizer by adding new tokens from the target language and study the data mixing recipe to mitigate forgetting. Our experiments on adapting an English LLM to Hungarian and Thai show that our recipe can reach better performance than open source models on the target language, with minimal regressions on English.
Recently, graph neural networks (GNNs) have been widely used for document classification. However, most existing methods are based on static word co-occurrence graphs without sentence-level information, which poses three challenges:(1) word ambiguity, (2) word synonymity, and (3) dynamic contextual dependency. To address these challenges, we propose a novel GNN-based sparse structure learning model for inductive document classification. Specifically, a document-level graph is initially generated by a disjoint union of sentence-level word co-occurrence graphs. Our model collects a set of trainable edges connecting disjoint words between sentences and employs structure learning to sparsely select edges with dynamic contextual dependencies. Graphs with sparse structures can jointly exploit local and global contextual information in documents through GNNs. For inductive learning, the refined document graph is further fed into a general readout function for graph-level classification and optimization in an end-to-end manner. Extensive experiments on several real-world datasets demonstrate that the proposed model outperforms most state-of-the-art results, and reveal the necessity to learn sparse structures for each document.
Recent contrastive representation learning methods rely on estimating mutual information (MI) between multiple views of an underlying context. E.g., we can derive multiple views of a given image by applying data augmentation, or we can split a sequence into views comprising the past and future of some step in the sequence. Contrastive lower bounds on MI are easy to optimize, but have a strong underestimation bias when estimating large amounts of MI. We propose decomposing the full MI estimation problem into a sum of smaller estimation problems by splitting one of the views into progressively more informed subviews and by applying the chain rule on MI between the decomposed views. This expression contains a sum of unconditional and conditional MI terms, each measuring modest chunks of the total MI, which facilitates approximation via contrastive bounds. To maximize the sum, we formulate a contrastive lower bound on the conditional MI which can be approximated efficiently. We refer to our general approach as Decomposed Estimation of Mutual Information (DEMI). We show that DEMI can capture a larger amount of MI than standard non-decomposed contrastive bounds in a synthetic setting, and learns better representations in a vision domain and for dialogue generation.
Knowledge graphs (KGs), which could provide essential relational information between entities, have been widely utilized in various knowledge-driven applications. Since the overall human knowledge is innumerable that still grows explosively and changes frequently, knowledge construction and update inevitably involve automatic mechanisms with less human supervision, which usually bring in plenty of noises and conflicts to KGs. However, most conventional knowledge representation learning methods assume that all triple facts in existing KGs share the same significance without any noises. To address this problem, we propose a novel confidence-aware knowledge representation learning framework (CKRL), which detects possible noises in KGs while learning knowledge representations with confidence simultaneously. Specifically, we introduce the triple confidence to conventional translation-based methods for knowledge representation learning. To make triple confidence more flexible and universal, we only utilize the internal structural information in KGs, and propose three kinds of triple confidences considering both local and global structural information. In experiments, We evaluate our models on knowledge graph noise detection, knowledge graph completion and triple classification. Experimental results demonstrate that our confidence-aware models achieve significant and consistent improvements on all tasks, which confirms the capability of CKRL modeling confidence with structural information in both KG noise detection and knowledge representation learning.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191