NeurIPS 2024 论文列表

When answering questions, large language models (LLMs) can convey not only an answer to the question, but a level of confidence about the answer being correct. This includes explicit markers of confidence (e.g. giving a numeric confidence score) as well as implicit markers, like using an authoritative tone or elaborating with additional knowledge of a subject. For LLMs to be trustworthy sources of knowledge, the confidence they convey should match their actual expertise on a topic; however, this is currently not the case, with most models tending towards overconfidence. To calibrate both implicit and explicit confidence markers, we introduce a pragmatic, listener-aware finetuning method (LACIE) that directly models the listener, considering not only whether an answer is right, but whether it will be accepted by a listener. Specifically, we cast calibration as a preference optimization problem, creating data via a two-agent speaker-listener game, where a speaker model’s outputs are judged by a simulated listener. We then finetune three different LLMs (Mistral-7B, Llama3-8B, Llama3-70B) with LACIE, and show that the models resulting from this multi-agent optimization are better calibrated on TriviaQA with respect to a simulated listener. Crucially, these trends transfer to human listeners, helping them correctly predict model correctness: we conduct a human evaluation where annotators accept or reject an LLM’s answers to trivia questions, finding that training with LACIE results in 47% fewer incorrect answers being accepted while maintaining the same level of acceptance for correct answers. Furthermore, LACIE generalizes to another dataset, resulting in a large increase in truthfulness on TruthfulQA when trained on TriviaQA. Our analysis indicates that LACIE leads to a better separation in confidence between correct and incorrect examples. Qualitatively, we find that a LACIE-trained model hedges more when uncertain and adopts implicit cues to signal certainty when it is correct, such as using an authoritative tone or including details. Finally, finetuning with our listener- aware method leads to an emergent increase in model abstention (e.g. saying “I don’t know”) for answers that are likely to be wrong, trading recall for precision.

Many studies have revealed that large language models (LLMs) exhibit uneven awareness of different contextual positions. Their limited context awareness can lead to overlooking critical information and subsequent task failures. While several approaches have been proposed to enhance LLMs' context awareness, achieving both effectiveness and efficiency remains challenging. In this paper, for LLMs utilizing RoPE as position embeddings, we introduce a novel method called "Mixture of In-Context Experts" (MoICE) to address this challenge. MoICE comprises two key components: a router integrated into each attention head within LLMs and a lightweight router-only training optimization strategy:(1) MoICE views each RoPE angle as an 'in-context' expert, demonstrated to be capable of directing the attention of a head to specific contextual positions. Consequently, each attention head flexibly processes tokens using multiple RoPE angles dynamically selected by the router to attend to the needed positions. This approach mitigates the risk of overlooking essential contextual information. (2) The router-only training strategy entails freezing LLM parameters and exclusively updating routers for only a few steps. When applied to open-source LLMs including Llama and Mistral, MoICE surpasses prior methods across multiple tasks on long context understanding and generation, all while maintaining commendable inference efficiency.

Efficient fine-tuning of large language models for task-specific applications is imperative, yet the vast number of parameters in these models makes their training increasingly challenging. Despite numerous proposals for effective methods, a substantial memory overhead remains for gradient computations during updates. \thm{Can we fine-tune a series of task-specific small models and transfer their knowledge directly to a much larger model without additional training?} In this paper, we explore weak-to-strong specialization using logit arithmetic, facilitating a direct answer to this question. Existing weak-to-strong methods often employ a static knowledge transfer ratio and a single small model for transferring complex knowledge, which leads to suboptimal performance. To surmount these limitations, we propose a dynamic logit fusion approach that works with a series of task-specific small models, each specialized in a different task. This method adaptively allocates weights among these models at each decoding step, learning the weights through Kullback-Leibler divergence constrained optimization problems. We conduct extensive experiments across various benchmarks in both single-task and multi-task settings, achieving leading results. By transferring expertise from the 7B model to the 13B model, our method closes the performance gap by 96.4\% in single-task scenarios and by 86.3\% in multi-task scenarios compared to full fine-tuning of the 13B model. Notably, we achieve surpassing performance on unseen tasks. Moreover, we further demonstrate that our method can effortlessly integrate in-context learning for single tasks and task arithmetic for multi-task scenarios.

Recently, rapid advancements in Multi-Modal In-Context Learning (MM-ICL) have achieved notable success, which is capable of achieving superior performance across various tasks without requiring additional parameter tuning. However, the underlying rules for the effectiveness of MM-ICL remain under-explored. To fill this gap, this work aims to investigate the research question: "_What factors affect the performance of MM-ICL?_" To this end, we investigate extensive experiments on the three core steps of MM-ICL including demonstration retrieval, demonstration ordering, and prompt construction using 6 vision large language models and 20 strategies. Our findings highlight (1) the necessity of a multi-modal retriever for demonstration retrieval, (2) the importance of intra-demonstration ordering over inter-demonstration ordering, and (3) the enhancement of task comprehension through introductory instructions in prompts. We hope this study can serve as a foundational guide for optimizing MM-ICL strategies in future research.

When using large language models (LLMs) in high-stakes applications, we need to know when we can trust their predictions. Some works argue that prompting high-performance LLMs is sufficient to produce calibrated uncertainties, while others introduce sampling methods that can be prohibitively expensive. In this work, we first argue that prompting on its own is insufficient to achieve good calibration and then show that fine-tuning on a small dataset of correct and incorrect answers can create an uncertainty estimate with good generalization and small computational overhead. We show that a thousand graded examples are sufficient to outperform baseline methods and that training through the features of a model is necessary for good performance and tractable for large open-source models when using LoRA. We also investigate the mechanisms that enable reliable LLM uncertainty estimation, finding that many models can be used as general-purpose uncertainty estimators, applicable not just to their own uncertainties but also the uncertainty of other models. Lastly, we show that uncertainty estimates inform human use of LLMs in human-AI collaborative settings through a user study.

As language models continue to scale, Large Language Models (LLMs) have exhibited emerging capabilities in In-Context Learning (ICL), enabling them to solve language tasks by prefixing a few in-context demonstrations (ICDs) as context. Inspired by these advancements, researchers have extended these techniques to develop Large Multimodal Models (LMMs) with ICL capabilities. However, applying ICL usually faces two major challenges: 1) using more ICDs will largely increase the inference time and 2) the performance is sensitive to the selection of ICDs. These challenges are further exacerbated in LMMs due to the integration of multiple data types and the combinational complexity of multimodal ICDs. Recently, to address these challenges, some NLP studies introduce non-learnable In-Context Vectors (ICVs) which extract useful task information from ICDs into a single vector and then insert it into the LLM to help solve the corresponding task. However, although useful in simple NLP tasks, these non-learnable methods fail to handle complex multimodal tasks like Visual Question Answering (VQA). In this study, we propose \underline{\textbf{L}}earnable \underline{\textbf{I}}n-Context \underline{\textbf{Ve}}ctor (LIVE) to distill essential task information from demonstrations, improving ICL performance in LMMs. Experiments show that LIVE can significantly reduce computational costs while enhancing accuracy in VQA tasks compared to traditional ICL and other non-learnable ICV methods.

Reinforcement learning based fine-tuning of large language models (LLMs) on human preferences has been shown to enhance both their capabilities and safety behavior. However, in cases related to safety, without precise instructions to human annotators, the data collected may cause the model to become overly cautious, or to respond in an undesirable style, such as being judgmental. Additionally, as model capabilities and usage patterns evolve, there may be a costly need to add or relabel data to modify safety behavior. We propose a novel preference modeling approach that utilizes AI feedback and only requires a small amount of human data. Our method, Rule Based Rewards (RBR), uses a collection of rules for desired or undesired behaviors (e.g. refusals should not be judgmental) along with a LLM grader. In contrast to prior methods using AI feedback, our method uses fine-grained, composable, LLM-graded few-shot prompts as reward directly in RL training, resulting in greater control, accuracy and ease of updating. We show that RBRs are an effective training method, achieving an F1 score of 97.1, compared to a human-feedback baseline of 91.7, resulting in much higher safety-behavior accuracy through better balancing usefulness and safety.

Instruction tuning (IT) is crucial to tailoring large language models (LLMs) towards human-centric interactions. Recent advancements have shown that the careful selection of a small, high-quality subset of IT data can significantly enhance the performance of LLMs. Despite this, common approaches often rely on additional models or data, which increases costs and limits widespread adoption. In this work, we propose a novel approach, termed $\textit{SelectIT}$, that capitalizes on the foundational capabilities of the LLM itself. Specifically, we exploit the intrinsic uncertainty present in LLMs to more effectively select high-quality IT data, without the need for extra resources. Furthermore, we introduce a curated IT dataset, the $\textit{Selective Alpaca}$, created by applying SelectIT to the Alpaca-GPT4 dataset. Empirical results demonstrate that IT using Selective Alpaca leads to substantial model ability enhancement. The robustness of SelectIT has also been corroborated in various foundation models and domain-specific tasks. Our findings suggest that longer and more computationally intensive IT data may serve as superior sources of IT, offering valuable insights for future research in this area. Data, code, and scripts are freely available at https://github.com/Blue-Raincoat/SelectIT.

Large language models (LLMs) have rapidly advanced and demonstrated impressive capabilities. In-Context Learning (ICL) and Parameter-Efficient Fine-Tuning (PEFT) are currently two mainstream methods for augmenting LLMs to downstream tasks. ICL typically constructs a few-shot learning scenario, either manually or by setting up a Retrieval-Augmented Generation (RAG) system, helping models quickly grasp domain knowledge or question-answering patterns without changing model parameters. However, this approach involves trade-offs, such as slower inference speed and increased space occupancy. PEFT assists the model in adapting to tasks through minimal parameter modifications, but the training process still demands high hardware requirements, even with a small number of parameters involved. To address these challenges, we propose Reference Trustable Decoding (RTD), a paradigm that allows models to quickly adapt to new tasks without fine-tuning, maintaining low inference costs. RTD constructs a reference datastore from the provided training examples and optimizes the LLM's final vocabulary distribution by flexibly selecting suitable references based on the input, resulting in more trustable responses and enabling the model to adapt to downstream tasks at a low cost. Experimental evaluations on various LLMs using different benchmarks demonstrate that RTD establishes a new paradigm for augmenting models to downstream tasks. Furthermore, our method exhibits strong orthogonality with traditional methods, allowing for concurrent usage. Our code can be found at https://github.com/ShiLuohe/ReferenceTrustableDecoding.

InfoNCE loss is commonly used to train dense retriever in information retrieval tasks. It is well known that a large batch is essential to stable and effective training with InfoNCE loss, which requires significant hardware resources. Due to the dependency of large batch, dense retriever has bottleneck of application and research. Recently, memory reduction methods have been broadly adopted to resolve the hardware bottleneck by decomposing forward and backward or using a memory bank. However, current methods still suffer from slow and unstable train. To address these issues, we propose Contrastive Accumulation (ContAccum), a stable and efficient memory reduction method for dense retriever trains that uses a dual memory bank structure to leverage previously generated query and passage representations. Experiments on widely used five information retrieval datasets indicate that ContAccum can surpass not only existing memory reduction methods but also high-resource scenarios. Moreover, theoretical analysis and experimental results confirm that ContAccum provides more stable dual-encoder training than current memory bank utilization methods.

Large Language Models (LLMs) have demonstrated remarkable capabilities in many real-world applications. Nonetheless, LLMs are often criticized for their tendency to produce hallucinations, wherein the models fabricate incorrect statements on tasks beyond their knowledge and perception. To alleviate this issue, graph retrieval-augmented generation (GraphRAG) has been extensively explored which leverages the factual knowledge in knowledge graphs (KGs) to ground the LLM's responses in established facts and principles. However, most state-of-the-art LLMs are closed-source, making it challenging to develop a prompting framework that can efficiently and effectively integrate KGs into LLMs with hard prompts only. Generally, existing KG-enhanced LLMs usually suffer from three critical issues, including huge search space, high API costs, and laborious prompt engineering, that impede their widespread application in practice. To this end, we introduce a novel **Know**ledge **Gr**aph based **P**romp**T**ing framework, namely **KnowGPT**, to enhance LLMs with domain knowledge. KnowGPT contains a knowledge extraction module to extract the most informative knowledge from KGs, and a context-aware prompt construction module to automatically convert extracted knowledge into effective prompts. Experiments on three benchmarks demonstrate that KnowGPT significantly outperforms all competitors. Notably, KnowGPT achieves a 92.6% accuracy on OpenbookQA leaderboard, comparable to human-level performance.

Reasoning capabilities are crucial for Large Language Models~(LLMs), yet a notable gap exists between English and non-English languages. To bridge this disparity, some works fine-tune LLMs to relearn reasoning capabilities in non-English languages, while others replace non-English inputs with an external model's outputs such as English translation text to circumvent the challenge of LLM understanding non-English. Unfortunately, these methods often underutilize the built-in skilled reasoning and useful language understanding capabilities of LLMs. In order to better utilize the minds of reasoning and language understanding in LLMs, we propose a new method, namely MergeMinds, which merges LLMs with the external language understanding capabilities from multilingual models to boost the multilingual reasoning performance. Furthermore, a two-step training scheme is introduced to first train to embeded the external capabilities into LLMs and then train the collaborative utilization of the external capabilities and the built-in capabilities in LLMs. Experiments on three multilingual reasoning datasets and a language understanding dataset demonstrate that MergeMinds consistently outperforms all baselines, especially in low-resource languages. Without updating the parameters of LLMs, the average accuracy improved by 6.7 and 8.0 across all languages and low-resource languages on the MGSM dataset, respectively.

Reinforcement learning (RL) has emerged as a pivotal technique for fine-tuning large language models (LLMs) on specific tasks. However, prevailing RL fine-tuning methods predominantly rely on PPO and its variants. Though these algorithms are effective in general RL settings, they often exhibit suboptimal performance and vulnerability to distribution collapse when applied to the fine-tuning of LLMs. In this paper, we propose CORY, extending the RL fine-tuning of LLMs to a sequential cooperative multi-agent reinforcement learning framework, to leverage the inherent coevolution and emergent capabilities of multi-agent systems. In CORY, the LLM to be fine-tuned is initially duplicated into two autonomous agents: a pioneer and an observer. The pioneer generates responses based on queries, while the observer generates responses using both the queries and the pioneer’s responses. The two agents are trained together. During training, the agents exchange roles periodically, fostering cooperation and coevolution between them. Experiments evaluate CORY's performance by fine-tuning GPT-2 and Llama-2 under subjective and objective reward functions on the IMDB Review and GSM8K datasets, respectively. Results show that CORY outperforms PPO in terms of policy optimality, resistance to distribution collapse, and training robustness, thereby underscoring its potential as a superior methodology for refining LLMs in real-world applications.

Surveys have recently gained popularity as a tool to study large language models. By comparing models’ survey responses to those of different human reference populations, researchers aim to infer the demographics, political opinions, or values best represented by current language models. In this work, we critically examine language models' survey responses on the basis of the well-established American Community Survey by the U.S. Census Bureau. Evaluating 43 different language models using de-facto standard prompting methodologies, we establish two dominant patterns. First, models' responses are governed by ordering and labeling biases, for example, towards survey responses labeled with the letter “A”. Second, when adjusting for these systematic biases through randomized answer ordering, models across the board trend towards uniformly random survey responses, irrespective of model size or training data. As a result, models consistently appear to better represent subgroups whose aggregate statistics are closest to uniform for the survey under consideration, leading to potentially misguided conclusions about model alignment.

Understanding how language model performance varies with scale is critical to benchmark and algorithm development. Scaling laws are one approach to building this understanding, but the requirement of training models across many different scales has limited their use. We propose an alternative, observational approach that bypasses model training and instead builds scaling laws from ~100 publically available models. Building a single scaling law from multiple model families is challenging due to large variations in their training compute efficiencies and capabilities. However, we show that these variations are consistent with a simple, generalized scaling law where language model performance is a function of a low-dimensional capability space, and model families only vary in their efficiency in converting training compute to capabilities. Using this approach, we show the surprising predictability of complex scaling phenomena: we show that several emergent phenomena follow a smooth, sigmoidal behavior and are predictable from small models; we show that the agent performance of models such as GPT-4 can be precisely predicted from simpler non-agentic benchmarks; and we show how to predict the impact of post-training interventions like Chain-of-Thought and Self-Consistency as language model capabilities continue to improve.

Citation networks are critical infrastructures of modern science, serving as intricate webs of past literature and enabling researchers to navigate the knowledge production system. To mine information hiding in the link space of such networks, predicting which previous papers (candidates) will a new paper (query) cite is a critical problem that has long been studied. However, an important gap remains unaddressed: the roles of a paper's citations vary significantly, ranging from foundational knowledge basis to superficial contexts. Distinguishing these roles requires a deeper understanding of the logical relationships among papers, beyond simple edges in citation networks. The emergence of large language models (LLMs) with textual reasoning capabilities offers new possibilities for discerning these relationships, but there are two major challenges. First, in practice, a new paper may select its citations from gigantic existing papers, where the combined texts far exceed the context length of LLMs. Second, logical relationships between papers are often implicit, and directly prompting an LLM to predict citations may lead to results based primarily on surface-level textual similarities, rather than the deeper logical reasoning required. In this paper, we introduce the novel concept of core citation, which identifies the critical references that go beyond superficial mentions. Thereby, we elevate the citation prediction task from a simple binary classification to a more nuanced problem: distinguishing core citations from both superficial citations and non-citations. To address this, we propose $\textbf{HLM-Cite}$, a $\textbf{H}$ybrid $\textbf{L}$anguage $\textbf{M}$odel workflow for citation prediction, which combines embedding and generative LMs. We design a curriculum finetune procedure to adapt a pretrained text embedding model to coarsely retrieve high-likelihood core citations from vast candidate sets and then design an LLM agentic workflow to rank the retrieved papers through one-shot reasoning, revealing the implicit relationships among papers. With the two-stage pipeline, we can scale the candidate sets to 100K papers, vastly exceeding the size handled by existing methods. We evaluate HLM-Cite on a dataset across 19 scientific fields, demonstrating a 17.6\% performance improvement comparing SOTA methods. Our code is open-source at https://github.com/tsinghua-fib-lab/H-LM for reproducibility.

Mobile device operation tasks are increasingly becoming a popular multi-modal AI application scenario. Current Multi-modal Large Language Models (MLLMs), constrained by their training data, lack the capability to function effectively as operation assistants. Instead, MLLM-based agents, which enhance capabilities through tool invocation, are gradually being applied to this scenario. However, the two major navigation challenges in mobile device operation tasks — task progress navigation and focus content navigation — are difficult to effectively solve under the single-agent architecture of existing work. This is due to the overly long token sequences and the interleaved text-image data format, which limit performance. To address these navigation challenges effectively, we propose Mobile-Agent-v2, a multi-agent architecture for mobile device operation assistance. The architecture comprises three agents: planning agent, decision agent, and reflection agent. The planning agent condenses lengthy, interleaved image-text history operations and screens summaries into a pure-text task progress, which is then passed on to the decision agent. This reduction in context length makes it easier for decision agent to navigate the task progress. To retain focus content, we design a memory unit that updates with task progress by decision agent. Additionally, to correct erroneous operations, the reflection agent observes the outcomes of each operation and handles any mistake accordingly. Experimental results indicate that Mobile-Agent-v2 achieves over a 30% improvement in task completion compared to the single-agent architecture of Mobile-Agent. The code is open-sourced at https://github.com/X-PLUG/MobileAgent.

Urban knowledge graph has recently worked as an emerging building block to distill critical knowledge from multi-sourced urban data for diverse urban application scenarios. Despite its promising benefits, urban knowledge graph construction (UrbanKGC) still heavily relies on manual effort, hindering its potential advancement. This paper presents UrbanKGent, a unified large language model agent framework, for urban knowledge graph construction. Specifically, we first construct the knowledgeable instruction set for UrbanKGC tasks (such as relational triplet extraction and knowledge graph completion) via heterogeneity-aware and geospatial-infused instruction generation. Moreover, we propose a tool-augmented iterative trajectory refinement module to enhance and refine the trajectories distilled from GPT-4. Through hybrid instruction fine-tuning with augmented trajectories on Llama 2 and Llama 3 family, we obtain UrbanKGC agent family, consisting of UrbanKGent-7/8/13B version. We perform a comprehensive evaluation on two real-world datasets using both human and GPT-4 self-evaluation. The experimental results demonstrate that UrbanKGent family can not only significantly outperform 31 baselines in UrbanKGC tasks, but also surpass the state-of-the-art LLM, GPT-4, by more than 10% with approximately 20 times lower cost. Compared with the existing benchmark, the UrbanKGent family could help construct an UrbanKG with hundreds of times richer relationships using only one-fifth of the data. Our data and code are available at https://github.com/usail-hkust/UrbanKGent.

Large language models (LLMs) exhibit hallucinations in long-form question-answering tasks across various domains and wide applications. Current hallucination detection and mitigation datasets are limited in domain and size, which struggle to scale due to prohibitive labor costs and insufficient reliability of existing hallucination annotators. To facilitate the scalable oversight of LLM hallucinations, this paper introduces an iterative self-training framework that simultaneously and progressively scales up the annotation dataset and improves the accuracy of the annotator. Based on the Expectation Maximization algorithm, in each iteration, the framework first applies an automatic hallucination annotation pipeline for a scaled dataset and then trains a more accurate annotator on the dataset. This new annotator is adopted in the annotation pipeline for the next iteration. Extensive experimental results demonstrate that the finally obtained hallucination annotator with only 7B parameters surpasses GPT-4 and obtains new state-of-the-art hallucination detection results on HaluEval and HalluQA by zero-shot inference. Such an annotator can not only evaluate the hallucination levels of various LLMs on the large-scale dataset but also help to mitigate the hallucination of LLMs generations, with the Natural Language Inference metric increasing from 25% to 37% on HaluEval.