PyTorchKR Papers

최신 논문

총 15,266편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2402.03469공식 코드 · Papers with Code

Rethinking the Role of Proxy Rewards in Language Model Alignment

Learning from human feedback via proxy reward modeling has been studied to align Large Language Models (LLMs) with human values. However, achieving reliable training through that proxy reward model (RM) is not a trivial problem, and its behavior remained as a black-box. In this paper, we study the role of proxy rewards in the LLM alignment via `reverse reward engineering' by composing interpretable features as a white-box reward function. We aim to replicate the ground truth (gold) reward signal by achieving a monotonic relationship between the proxy and gold reward signals after training the model using the proxy reward in reinforcement learning (RL). Our findings indicate that successfully emulating the gold reward requires generating responses that are relevant with enough length to open-ended questions, while also ensuring response consistency in closed-ended questions. Furthermore, resulting models optimizing our devised white-box reward show competitive performances with strong open-source RMs in alignment benchmarks. We highlight its potential usage as a simple but strong reward baseline for the LLM alignment, not requiring explicit human feedback dataset and RM training. Our code is available at https://github.com/naver-ai/rethinking-proxy-reward.

arXiv:2404.18870공식 코드 · Papers with Code

More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness

The trustworthiness of Large Language Models (LLMs) refers to the extent to which their outputs are reliable, safe, and ethically aligned, and it has become a crucial consideration alongside their cognitive performance. In practice, Reinforcement Learning From Human Feedback (RLHF) has been widely used to align LLMs with labeled human preferences, but its assumed effect on model trustworthiness hasn't been rigorously evaluated. To bridge this knowledge gap, this study investigates how models aligned with general-purpose preference data perform across five trustworthiness verticals: toxicity, stereotypical bias, machine ethics, truthfulness, and privacy. Our results demonstrate that RLHF on human preferences doesn't automatically guarantee trustworthiness, and reverse effects are often observed. Furthermore, we propose to adapt efficient influence function based data attribution methods to the RLHF setting to better understand the influence of fine-tuning data on individual trustworthiness benchmarks, and show its feasibility by providing our estimated attribution scores. Together, our results underscore the need for more nuanced approaches for model alignment from both the data and framework perspectives, and we hope this research will guide the community towards developing language models that are increasingly capable without sacrificing trustworthiness.

arXiv:2404.18085공식 코드 · Papers with Code

CRE-LLM: A Domain-Specific Chinese Relation Extraction Framework with Fine-tuned Large Language Model

Domain-Specific Chinese Relation Extraction (DSCRE) aims to extract relations between entities from domain-specific Chinese text. Despite the rapid development of PLMs in recent years, especially LLMs, DSCRE still faces three core challenges: complex network structure design, poor awareness, and high consumption of fine-tuning. Given the impressive performance of large language models (LLMs) in natural language processing, we propose a new framework called CRE-LLM. This framework is based on fine-tuning open-source LLMs, such as Llama-2, ChatGLM2, and Baichuan2. CRE-LLM enhances the logic-awareness and generative capabilities of the model by constructing an appropriate prompt and utilizing open-source LLMs for instruction-supervised fine-tuning. And then it directly extracts the relations of the given entities in the input textual data, which improving the CRE approach. To demonstrate the effectiveness of the proposed framework, we conducted extensive experiments on two domain-specific CRE datasets, FinRE and SanWen. The experimental results show that CRE-LLM is significantly superior and robust, achieving state-of-the-art (SOTA) performance on the FinRE dataset. This paper introduces a novel approach to domain-specific relation extraction (DSCRE) tasks that are semantically more complex by combining LLMs with triples. Our code is publicly available.

arXiv:2404.17460

Ruffle&Riley: Insights from Designing and Evaluating a Large Language Model-Based Conversational Tutoring System

Conversational tutoring systems (CTSs) offer learning experiences through interactions based on natural language. They are recognized for promoting cognitive engagement and improving learning outcomes, especially in reasoning tasks. Nonetheless, the cost associated with authoring CTS content is a major obstacle to widespread adoption and to research on effective instructional design. In this paper, we discuss and evaluate a novel type of CTS that leverages recent advances in large language models (LLMs) in two ways: First, the system enables AI-assisted content authoring by inducing an easily editable tutoring script automatically from a lesson text. Second, the system automates the script orchestration in a learning-by-teaching format via two LLM-based agents (Ruffle&Riley) acting as a student and a professor. The system allows for free-form conversations that follow the ITS-typical inner and outer loop structure. We evaluate Ruffle&Riley's ability to support biology lessons in two between-subject online user studies (N = 200) comparing the system to simpler QA chatbots and reading activity. Analyzing system usage patterns, pre/post-test scores and user experience surveys, we find that Ruffle&Riley users report high levels of engagement, understanding and perceive the offered support as helpful. Even though Ruffle&Riley users require more time to complete the activity, we did not find significant differences in short-term learning gains over the reading activity. Our system architecture and user study provide various insights for designers of future CTSs. We further open-source our system to support ongoing research on effective instructional design of LLM-based learning technologies.

arXiv:2404.17546공식 코드 · Papers with Code

Probabilistic Inference in Language Models via Twisted Sequential Monte Carlo

Numerous capability and safety techniques of Large Language Models (LLMs), including RLHF, automated red-teaming, prompt engineering, and infilling, can be cast as sampling from an unnormalized target distribution defined by a given reward or potential function over the full sequence. In this work, we leverage the rich toolkit of Sequential Monte Carlo (SMC) for these probabilistic inference problems. In particular, we use learned twist functions to estimate the expected future value of the potential at each timestep, which enables us to focus inference-time computation on promising partial sequences. We propose a novel contrastive method for learning the twist functions, and establish connections with the rich literature of soft reinforcement learning. As a complementary application of our twisted SMC framework, we present methods for evaluating the accuracy of language model inference techniques using novel bidirectional SMC bounds on the log partition function. These bounds can be used to estimate the KL divergence between the inference and target distributions in both directions. We apply our inference evaluation techniques to show that twisted SMC is effective for sampling undesirable outputs from a pretrained model (a useful component of harmlessness training and automated red-teaming), generating reviews with varied sentiment, and performing infilling tasks.

arXiv:2404.17475공식 코드 · Papers with Code

CEval: A Benchmark for Evaluating Counterfactual Text Generation

Counterfactual text generation aims to minimally change a text, such that it is classified differently. Judging advancements in method development for counterfactual text generation is hindered by a non-uniform usage of data sets and metrics in related work. We propose CEval, a benchmark for comparing counterfactual text generation methods. CEval unifies counterfactual and text quality metrics, includes common counterfactual datasets with human annotations, standard baselines (MICE, GDBA, CREST) and the open-source language model LLAMA-2. Our experiments found no perfect method for generating counterfactual text. Methods that excel at counterfactual metrics often produce lower-quality text while LLMs with simple prompts generate high-quality text but struggle with counterfactual criteria. By making CEval available as an open-source Python library, we encourage the community to contribute more methods and maintain consistent evaluation in future work.

arXiv:2404.17027

Player-Driven Emergence in LLM-Driven Game Narrative

We explore how interaction with large language models (LLMs) can give rise to emergent behaviors, empowering players to participate in the evolution of game narratives. Our testbed is a text-adventure game in which players attempt to solve a mystery under a fixed narrative premise, but can freely interact with non-player characters generated by GPT-4, a large language model. We recruit 28 gamers to play the game and use GPT-4 to automatically convert the game logs into a node-graph representing the narrative in the player's gameplay. We find that through their interactions with the non-deterministic behavior of the LLM, players are able to discover interesting new emergent nodes that were not a part of the original narrative but have potential for being fun and engaging. Players that created the most emergent nodes tended to be those that often enjoy games that facilitate discovery, exploration and experimentation.

arXiv:2404.17000공식 코드 · Papers with Code

Evaluating Class Membership Relations in Knowledge Graphs using Large Language Models

A backbone of knowledge graphs are their class membership relations, which assign entities to a given class. As part of the knowledge engineering process, we propose a new method for evaluating the quality of these relations by processing descriptions of a given entity and class using a zero-shot chain-of-thought classifier that uses a natural language intensional definition of a class. We evaluate the method using two publicly available knowledge graphs, Wikidata and CaLiGraph, and 7 large language models. Using the gpt-4-0125-preview large language model, the method's classification performance achieves a macro-averaged F1-score of 0.830 on data from Wikidata and 0.893 on data from CaLiGraph. Moreover, a manual analysis of the classification errors shows that 40.9% of errors were due to the knowledge graphs, with 16.0% due to missing relations and 24.9% due to incorrectly asserted relations. These results show how large language models can assist knowledge engineers in the process of knowledge graph refinement. The code and data are available on Github.

arXiv:2404.16891공식 코드 · Papers with Code

Attacks on Third-Party APIs of Large Language Models

Large language model (LLM) services have recently begun offering a plugin ecosystem to interact with third-party API services. This innovation enhances the capabilities of LLMs, but it also introduces risks, as these plugins developed by various third parties cannot be easily trusted. This paper proposes a new attacking framework to examine security and safety vulnerabilities within LLM platforms that incorporate third-party services. Applying our framework specifically to widely used LLMs, we identify real-world malicious attacks across various domains on third-party APIs that can imperceptibly modify LLM outputs. The paper discusses the unique challenges posed by third-party API integration and offers strategic possibilities to improve the security and safety of LLM ecosystems moving forward. Our code is released at https://github.com/vk0812/Third-Party-Attacks-on-LLMs.

arXiv:2404.17525

Large Language Model Agent as a Mechanical Designer

Conventional mechanical design follows an iterative process in which initial concepts are refined through cycles of expert assessment and resource-intensive Finite Element Method (FEM) analysis to meet performance goals. While machine learning models have been developed to assist in parts of this process, they typically require large datasets, extensive training, and are often tailored to specific tasks, limiting their generalizability. To address these limitations, we propose a framework that leverages a pretrained Large Language Model (LLM) in conjunction with an FEM module to autonomously generate, evaluate, and refine structural designs based on performance specifications and numerical feedback. The LLM operates without domain-specific fine-tuning, using general reasoning to propose design candidates, interpret FEM-derived performance metrics, and apply structurally sound modifications. Using 2D truss structures as a testbed, we show that the LLM can effectively navigate highly discrete and multi-faceted design spaces, balance competing objectives, and identify convergence when further optimization yields diminishing returns. Compared to Non-dominated Sorting Genetic Algorithm II (NSGA-II), our method achieves faster convergence and fewer FEM evaluations. Experiments with varying temperature settings (0.5, 1.0, 1.2) and model sizes (GPT-4.1 and GPT-4.1-mini) indicate that smaller models yield higher constraint satisfaction with fewer steps, while lower temperatures enhance design consistency. These results establish LLMs as a promising new class of reasoning-based, natural language-driven optimizers for autonomous design and iterative structural refinement.

arXiv:2404.16222공식 코드 · Papers with Code

Step Differences in Instructional Video

Comparing a user video to a reference how-to video is a key requirement for AR/VR technology delivering personalized assistance tailored to the user's progress. However, current approaches for language-based assistance can only answer questions about a single video. We propose an approach that first automatically generates large amounts of visual instruction tuning data involving pairs of videos from HowTo100M by leveraging existing step annotations and accompanying narrations, and then trains a video-conditioned language model to jointly reason across multiple raw videos. Our model achieves state-of-the-art performance at identifying differences between video pairs and ranking videos based on the severity of these differences, and shows promising ability to perform general reasoning over multiple videos. Project page: https://github.com/facebookresearch/stepdiff

arXiv:2404.15087공식 코드 · Papers with Code

Detection of circular permutations by Protein Language Models

Protein circular permutations are crucial for understanding protein evolution and functionality. Traditional detection methods, sequence-based or structure-based, struggle with accuracy and computational efficiency, the latter also limited by treating proteins as rigid bodies. The plmCP method, utilizing a protein language model, not only speeds up the detection process but also enhances the accuracy of identifying circular permutations, contributing significantly to protein research and engineering by acknowledging structural flexibility.

arXiv:2404.16053

Human Latency Conversational Turns for Spoken Avatar Systems

A problem with many current Large Language Model (LLM) driven spoken dialogues is the response time. Some efforts such as Groq address this issue by lightning fast processing of the LLM, but we know from the cognitive psychology literature that in human-to-human dialogue often responses occur prior to the speaker completing their utterance. No amount of delay for LLM processing is acceptable if we wish to maintain human dialogue latencies. In this paper, we discuss methods for understanding an utterance in close to real time and generating a response so that the system can comply with human-level conversational turn delays. This means that the information content of the final part of the speaker's utterance is lost to the LLM. Using the Google NaturalQuestions (NQ) database, our results show GPT-4 can effectively fill in missing context from a dropped word at the end of a question over 60% of the time. We also provide some examples of utterances and the impacts of this information loss on the quality of LLM response in the context of an avatar that is currently under development. These results indicate that a simple classifier could be used to determine whether a question is semantically complete, or requires a filler phrase to allow a response to be generated within human dialogue time constraints.

arXiv:2404.16048

GUIDE: Graphical User Interface Data for Execution

In this paper, we introduce GUIDE, a novel dataset tailored for the advancement of Multimodal Large Language Model (MLLM) applications, particularly focusing on Robotic Process Automation (RPA) use cases. Our dataset encompasses diverse data from various websites including Apollo(62.67\%), Gmail(3.43\%), Calendar(10.98\%) and Canva(22.92\%). Each data entry includes an image, a task description, the last action taken, CoT and the next action to be performed along with grounding information of where the action needs to be executed. The data is collected using our in-house advanced annotation tool NEXTAG (Next Action Grounding and Annotation Tool). The data is adapted for multiple OS, browsers and display types. It is collected by multiple annotators to capture the variation of design and the way person uses a website. Through this dataset, we aim to facilitate research and development in the realm of LLMs for graphical user interfaces, particularly in tasks related to RPA. The dataset's multi-platform nature and coverage of diverse websites enable the exploration of cross-interface capabilities in automation tasks. We believe that our dataset will serve as a valuable resource for advancing the capabilities of multi-platform LLMs in practical applications, fostering innovation in the field of automation and natural language understanding. Using GUIDE, we build V-Zen, the first RPA model to automate multiple websites using our in-House Automation tool AUTONODE

arXiv:2404.15675

Hi-Gen: Generative Retrieval For Large-Scale Personalized E-commerce Search

Leveraging generative retrieval (GR) techniques to enhance search systems is an emerging methodology that has shown promising results in recent years. In GR, a text-to-text model maps string queries directly to relevant document identifiers (docIDs), dramatically simplifying the retrieval process. However, when applying most GR models in large-scale E-commerce for personalized item search, we must face two key problems in encoding and decoding. (1) Existing docID generation methods ignore the encoding of efficiency information, which is critical in E-commerce. (2) The positional information is important in decoding docIDs, while prior studies have not adequately discriminated the significance of positional information or well exploited the inherent interrelation among these positions. To overcome these problems, we introduce an efficient Hierarchical encoding-decoding Generative retrieval method (Hi-Gen) for large-scale personalized E-commerce search systems. Specifically, we first design a representation learning model using metric learning to learn discriminative feature representations of items to capture semantic relevance and efficiency information. Then, we propose a category-guided hierarchical clustering scheme that makes full use of the semantic and efficiency information of items to facilitate docID generation. Finally, we design a position-aware loss to discriminate the importance of positions and mine the inherent interrelation between different tokens at the same position. This loss boosts the performance of the language model used in the decoding stage. Besides, we propose two variants of Hi-Gen (Hi-Gen-I2I and Hi-Gen-Cluster) to support online real-time large-scale recall in the online serving process. Hi-Gen gets 3.30% and 4.62% improvements over SOTA for Recall@1 on the public and industry datasets, respectively.

arXiv:2404.16198

Towards Efficient Patient Recruitment for Clinical Trials: Application of a Prompt-Based Learning Model

Objective: Clinical trials are essential for advancing pharmaceutical interventions, but they face a bottleneck in selecting eligible participants. Although leveraging electronic health records (EHR) for recruitment has gained popularity, the complex nature of unstructured medical texts presents challenges in efficiently identifying participants. Natural Language Processing (NLP) techniques have emerged as a solution with a recent focus on transformer models. In this study, we aimed to evaluate the performance of a prompt-based large language model for the cohort selection task from unstructured medical notes collected in the EHR. Methods: To process the medical records, we selected the most related sentences of the records to the eligibility criteria needed for the trial. The SNOMED CT concepts related to each eligibility criterion were collected. Medical records were also annotated with MedCAT based on the SNOMED CT ontology. Annotated sentences including concepts matched with the criteria-relevant terms were extracted. A prompt-based large language model (Generative Pre-trained Transformer (GPT) in this study) was then used with the extracted sentences as the training set. To assess its effectiveness, we evaluated the model's performance using the dataset from the 2018 n2c2 challenge, which aimed to classify medical records of 311 patients based on 13 eligibility criteria through NLP techniques. Results: Our proposed model showed the overall micro and macro F measures of 0.9061 and 0.8060 which were among the highest scores achieved by the experiments performed with this dataset. Conclusion: The application of a prompt-based large language model in this study to classify patients based on eligibility criteria received promising scores. Besides, we proposed a method of extractive summarization with the aid of SNOMED CT ontology that can be also applied to other medical texts.

arXiv:2404.16192

Fusion of Domain-Adapted Vision and Language Models for Medical Visual Question Answering

Vision-language models, while effective in general domains and showing strong performance in diverse multi-modal applications like visual question-answering (VQA), struggle to maintain the same level of effectiveness in more specialized domains, e.g., medical. We propose a medical vision-language model that integrates large vision and language models adapted for the medical domain. This model goes through three stages of parameter-efficient training using three separate biomedical and radiology multi-modal visual and text datasets. The proposed model achieves state-of-the-art performance on the SLAKE 1.0 medical VQA (MedVQA) dataset with an overall accuracy of 87.5% and demonstrates strong performance on another MedVQA dataset, VQA-RAD, achieving an overall accuracy of 73.2%.

arXiv:2404.16032공식 코드 · Papers with Code

Studying Large Language Model Behaviors Under Context-Memory Conflicts With Real Documents

Retrieval-augmented generation (RAG) mitigates many problems of fully parametric language models, such as temporal degradation, hallucinations, and lack of grounding. In RAG, the model's knowledge can be updated from documents provided in context. This leads to cases of conflict between the model's parametric knowledge and the contextual information, where the model may not always update its knowledge. Previous work studied context-memory knowledge conflicts by creating synthetic documents that contradict the model's correct parametric answers. We present a framework for studying such knowledge conflicts in a realistic setup. We update incorrect parametric knowledge using real conflicting documents. This reflects how knowledge conflicts arise in practice. In this realistic scenario, we find that knowledge updates fail less often than previously reported. In cases where the models still fail to update their answers, we find a parametric bias: the incorrect parametric answer appearing in context makes the knowledge update likelier to fail. These results suggest that the factual parametric knowledge of LLMs can negatively influence their reading abilities and behaviors. Our code is available at https://github.com/kortukov/realistic_knowledge_conflicts/ .

arXiv:2404.15928

Generalization Measures for Zero-Shot Cross-Lingual Transfer

A model's capacity to generalize its knowledge to interpret unseen inputs with different characteristics is crucial to build robust and reliable machine learning systems. Language model evaluation tasks lack information metrics about model generalization and their applicability in a new setting is measured using task and language-specific downstream performance, which is often lacking in many languages and tasks. In this paper, we explore a set of efficient and reliable measures that could aid in computing more information related to the generalization capability of language models in cross-lingual zero-shot settings. In addition to traditional measures such as variance in parameters after training and distance from initialization, we also measure the effectiveness of sharpness in loss landscape in capturing the success in cross-lingual transfer and propose a novel and stable algorithm to reliably compute the sharpness of a model optimum that correlates to generalization.

arXiv:2404.15848

Detecting Conceptual Abstraction in LLMs

We present a novel approach to detecting noun abstraction within a large language model (LLM). Starting from a psychologically motivated set of noun pairs in taxonomic relationships, we instantiate surface patterns indicating hypernymy and analyze the attention matrices produced by BERT. We compare the results to two sets of counterfactuals and show that we can detect hypernymy in the abstraction mechanism, which cannot solely be related to the distributional similarity of noun pairs. Our findings are a first step towards the explainability of conceptual abstraction in LLMs.