COLM 2024 论文列表

Large language models (LLMs) have recently attracted considerable interest for their ability to perform complex reasoning tasks, such as chain-of-thought (CoT) reasoning. However, most of the existing approaches to enhance this ability rely heavily on data-driven methods, while neglecting the structural aspects of the model's reasoning capacity. To encourage a more structural generation of CoT steps, we propose a hierarchical generation scheme: we let the LM generate a planning token at the start of each reasoning step, intuitively serving as a high-level plan of the current step, and add their embeddings to the model parameters. Our approach requires a negligible increase in trainable parameters (0.001%) and can be applied through either full fine-tuning or a more parameter-efficient scheme. We demonstrate our method's effectiveness by applying it to three different LLMs, showing notable accuracy improvements across three math word problem datasets and one multihop QA dataset with respect to standard fine-tuning baselines.

Large Language Models (LLMs) are often paired with a reported cutoff date, the time at which training data was gathered. Such information is crucial for applications where the LLM must provide up-to-date information. However, a reported cutoff only scratches the surface. Do all sub-resources in the training data share the same cutoff? Does the model's demonstrated knowledge for these sub-resources closely align to their cutoff? We define the notion of an effective cutoff, which is distinct from the LLM's reported cutoff and differs between sub-resources. We propose a simple approach to estimate effective cutoffs of an LLM on the resource-level by probing across versions of the data. Crucially, our method does not require access to a model's pre-training data. Through our analysis, we find that effective cutoffs often drastically differ from reported cutoffs. To understand the root cause of this observation, we conduct a large-scale analysis on open pre-training datasets. Our analysis reveals two reasons for these inconsistencies: (1) temporal misalignments of CommonCrawl data due to non-trivial amounts of old data in new dumps; and (2) complications in LLM deduplication schemes involving semantic duplicates and lexical near-duplicates. Overall, our results show that cutoffs are not as simple as they have seemed and that care must be taken both by LLM dataset curators as well as practitioners who seek to use these models.

We present a comprehensive evaluation of large language models' capability to reason compositional relations through a benchmark encompassing 1,800 test cases in both English and Chinese, covering six distinct categories of composition relations: Positional, Comparative, Personal, Mathematical, Identity, and Other. We expand our assessment to the multilingual realm by including translations of the benchmark suite into Japanese, French, and Korean. Our Multilingual Composition Relation (MCR) benchmark aims at investigating the robustness and adaptability of LLMs in handling compositional relation reasoning across diverse linguistic contexts.

Large language models (LLMs) are initially pretrained for broad capabilities and then finetuned with instruction-following datasets to improve their performance in interacting with humans. Despite advances in finetuning, a standardized guideline for selecting high-quality datasets to optimize this process remains elusive. In this paper, we first propose InstructMining, an innovative method designed for automatically selecting premium instruction-following data for finetuning LLMs. Specifically, InstructMining utilizes natural language indicators as a measure of data quality, applying them to evaluate unseen datasets. During experimentation, we discover that double descent phenomenon exists in large language model finetuning. Based on this observation, we further leverage BlendSearch to help find the best subset among the entire datase. Experiment results show that InstructMining-7B achieves state-of-the-art performance on two of the most popular benchmarks: LLM-as-a-judge and OpenLLM benchmark.

Pretraining datasets for large language models (LLMs) have grown to trillions of tokens composed of large amounts of CommonCrawl (CC) web scrape along with smaller, domain-specific datasets. It is expensive to understand the impact of these domain-specific datasets on model capabilities as training at large FLOP scales is required to reveal significant changes to difficult and emergent benchmarks. Given the increasing cost of experimenting with pretraining data, how does one determine the optimal balance between the diversity in general web scrapes and the information density of domain specific data? In this work, we show how to leverage the smaller domain specific datasets by upsampling them relative to CC at the end of training to drive performance improvements on difficult benchmarks. This simple technique allows us to improve up to 6.90 pp on MMLU, 8.26 pp on GSM8K, and 6.17 pp on HumanEval relative to the base data mix for a 7B model trained for 1 trillion (T) tokens, thus rivaling Llama-2 (7B)—a model trained for twice as long. We experiment with ablating the duration of domain upsampling from 5% to 30% of training and find that 10% to 20% percent is optimal for navigating the tradeoff between general language modeling capabilities and targeted benchmarks. We also use domain upsampling to characterize at scale the utility of individual datasets for improving various benchmarks by removing them during this final phase of training. This tool opens up the ability to experiment with the impact of different pretraining datasets at scale, but at an order of magnitude lower cost compared to full pretraining runs.

A fundamental open challenge in modern LLM scaling is the lack of understanding around emergent capabilities. In particular, language model pretraining loss is known to be highly predictable as a function of compute. However, downstream capabilities are far less predictable---sometimes even exhibiting emergent jumps---which makes it challenging to anticipate the capabilities of future models. In this work, we first pose the task of emergence prediction: given access to current LLMs that have random few-shot accuracy on a task, can we predict whether future models (GPT-N+1) will have non-trivial accuracy on that task? We then discover a simple insight for this problem: directly finetuning LLMs on a given task can shift the point in scaling at which emergence occurs towards less capable models. To operationalize this insight, we can finetune LLMs with varying amounts of data and fit a parametric function that predicts when emergence will occur (i.e., ``emergence laws''). To validate this approach, we use four standard NLP benchmarks where large-scale open-source LLMs already demonstrate emergence (MMLU, GSM8K, CommonsenseQA, and CoLA). Using only small-scale LLMs, we find that, in some cases, we are able to accurately predict whether models trained with up to 4x more compute have emerged.

Reinforcement Learning from Human Feedback (RLHF) is a popular method for aligning Language Models (LM) with human values and preferences. RLHF requires a large number of preference pairs as training data, which are often used in both the Supervised Fine-Tuning and Reward Model training and therefore publicly available datasets are commonly used. In this work, we study to what extent a malicious actor can manipulate the LMs generations by poisoning the preferences, i.e., injecting poisonous preference pairs into these datasets and the RLHF training process. We propose strategies to build poisonous preference pairs and test their performance by poisoning two widely used preference datasets. Our results show that preference poisoning is highly effective: injecting a small amount of poisonous data (1-5\% of the original dataset), we can effectively manipulate the LM to generate a target entity in a target sentiment (positive or negative). The findings from our experiments also shed light on strategies to defend against the preference poisoning attack.

The dramatic improvements in Large Language Models (LLMs) come at the cost of increased computational resources for inference. Recent studies ameliorate the computational costs of LLMs by increasing their activation sparsity but suffer from significant performance degradation on downstream tasks. In this work, we introduce a new framework for sparsifying the activations of LLMs and reducing inference costs, dubbed $\underline{C}$ontextually $\underline{A}$ware $\underline{T}$hresholding for $\underline{S}$parsity (CATS). CATS is a relatively simple algorithm that is easy to implement and highly effective. At the heart of our framework is a new non-linear activation function. We demonstrate that CATS can be applied to various models, including Mistral-7B and Llama2-7B \& 13B, and outperforms existing sparsification techniques across multiple tasks. More precisely, CATS-based models achieve downstream task performance within $\sim$ 99\% of their base models at activation sparsity levels of 50\%, even without any fine-tuning. Moreover, with fine-tuning that targets only 1\% of the parameters, CATS-based models not only converge faster but also achieve better task performance than competing techniques. Finally, we develop a custom GPU kernel for the efficient implementation of CATS that translates the activation sparsity of CATS to real wall-clock time speedups. Our custom kernel implementation of CATS results in a $\sim$15\% improvement in wall-clock inference latency of token generation. We release our code, experiments, and datasets at https://github.com/ScalingIntelligence/CATS.

Transformer models have achieved superior performance in various natural language processing tasks. However, the quadratic computational cost of the attention mechanism limits its practicality for long sequences. There are existing attention variants that improve the computational efficiency, but they have limited ability to effectively compute global information. In parallel to Transformer models, state space models (SSMs) are tailored for long sequences, but they are not flexible enough to capture complicated local information. We propose SPADE, short for State Space Augmented Transformer. Specifically, we augment a SSM into the bottom layer of SPADE, and we employ efficient local attention methods for the other layers. The SSM augments global information, which complements the lack of long-range dependency issue in local attention methods. Experimental results on the Long Range Arena benchmark and language modeling tasks demonstrate the effectiveness of the proposed method. To further demonstrate the scalability of SPADE, we pre-train large encoder-decoder models and present fine-tuning results on natural language understanding and natural language generation tasks.

The capabilities of pretrained language models (LMs) have opened opportunities to explore new application areas, but applications involving human-human interaction are limited by the fact that most data is protected from public release for privacy reasons. Problem-solving human-human dialogues in real applications can be much more complex than existing Wizard-of-Oz collections, preventing successful domain transfer. To support information extraction (IE) for a private call center dataset (AIC), we introduce a human-in-the-loop dialogue generation framework capable of synthesizing realistic dialogues. In IE experiments with AIC dialogues, we observe 25% relative improvement in F1 after augmenting a small set of real human-human conversations with synthetic data. In controlled experiments, we compare training with our human-in-the-loop-synthesized data vs. fully automatically LM-generated data and find that collaborating humans adds value both in the generation and annotation stages. We release code and our synthetic dataset to illustrate the complexity of call center conversations and encourage development of complex dialogue datasets that are more representative of natural data.

Are $n$-gram language models still relevant in this era of neural large language models (LLMs)? Our answer is *yes*, and we showcase their values in both text analysis and improving neural LLMs. This was done by modernizing $n$-gram LMs in two aspects. First, we train them at the same data scale as neural LLMs -- **5 trillion tokens**. This is one of the largest $n$-gram LMs ever built. Second, existing $n$-gram LMs use small $n$ which hinders their performance; we instead allow $n$ to be arbitrarily large, by introducing a new **$\infty$-gram LM** with backoff. Instead of pre-computing $n$-gram count tables (which would be very expensive), we develop an engine named infini-gram -- powered by suffix arrays -- that can compute $\infty$-gram (as well as $n$-gram with arbitrary $n$) probabilities with **millisecond-level latency**. The $\infty$-gram framework and infini-gram engine enable us to conduct many novel and interesting analyses of human-written and machine-generated text: we find that the $\infty$-gram LM has fairly high accuracy for next-token prediction (47%), and can complement neural LLMs to greatly reduce their perplexity. When analyzing machine-generated text, we also observe irregularities in the machine: $\infty$-gram agreement level with respect to the suffix length, which indicates deficiencies in neural LLM pretraining and the positional embeddings of Transformers.

Large Language Models (LLMs) exhibit remarkable capabilities but are prone to generating inaccurate or hallucinatory responses. This limitation stems from their reliance on vast pretraining datasets, making them susceptible to errors in unseen scenarios. To tackle these challenges, Retrieval Augmented Generation (RAG) addresses this by incorporating external, relevant documents into the response generation process, thus leveraging non-parametric knowledge alongside LLMs’ in-context learning abilities. However, existing RAG implementations primarily focus on initial input for context retrieval, overlooking the nuances of ambiguous or complex queries that necessitate further clarification or decomposition for accurate responses. To this end, we propose learning to Refine Queries for Retrieval Augmented Generation (RQ-RAG) in this paper, endeavoring to enhance the model by equipping it with capabilities for explicit rewriting, decomposition, and disambiguation. Our experimental results indicate that our method, when applied to a 7B Llama2 model, surpasses the previous state-of-the-art (SOTA) by an average of 1.9% across three single-hop QA datasets, and when applied to a 8B Llama3 model, it also demonstrates enhanced performance in handling complex, multi-hop QA datasets.

Selecting influential data for fine-tuning on downstream tasks is a key factor for both performance and computation efficiency. Recent works have shown that training with only limited data can show a superior performance on general tasks. However, the feasibility on mathematical reasoning tasks has not been validated. To go further, there exist two open questions for mathematical reasoning: how to select influential data and what is an influential data composition. For the former one, we propose a Quality-aware Diverse Selection (QaDS) strategy adaptable for mathematical reasoning. A comparison with other selection strategies validates the superiority of QaDS. For the latter one, we first enlarge our setting and explore the influential data composition. We conduct a series of experiments and highlight: scaling up reasoning data, and training with general data selected by QaDS is helpful. Then, we define our optimal mixture as OpenMathMix, an influential data mixture with open-source data selected by QaDS. With OpenMathMix, we achieve a state-of-the-art 48.8% accuracy on MATH with 7B base model. Additionally, we showcase the use of QaDS in creating efficient fine-tuning mixtures with various selection ratios, and analyze the quality of a wide range of open-source datasets, which can perform as a reference for future works on mathematical reasoning tasks.

This paper introduces LalaEval, a holistic framework designed for the human evaluation of domain-specific large language models (LLMs). LalaEval proposes a comprehensive suite of end-to-end protocols that cover five main components including domain specification, criteria establishment, benchmark dataset creation, construction of evaluation rubrics, and thorough analysis and interpretation of evaluation outcomes. This initiative aims to fill a crucial research gap by providing a systematic methodology for conducting standardized human evaluations within specific domains, a practice that, despite its widespread application, lacks substantial coverage in the literature and human evaluation are often criticized to be less reliable due to subjective factors, so standardized procedures adapted to the nuanced requirements of specific domains or even individual organizations are in great need. Furthermore, the paper demonstrates the framework's application within the logistics industry and a comparative analysis of LLMs for the logistics domain use, highlighting the framework's capacity to elucidate performance differences and guide model selection and development for domain-specific LLMs. Through real-world deployment, the paper underscores the framework's effectiveness in advancing the field of domain-specific LLM evaluation, thereby contributing significantly to the ongoing discussion on LLMs' practical utility and performance in domain-specific applications.

Measuring personal disclosures made in human-chatbot interactions can provide a better understanding of users’ AI literacy and facilitate privacy research for large language models (LLMs). We run an extensive, fine-grained analysis on the personal disclosures made by real users to commercial GPT models, investigating the leakage of personally identifiable and sensitive information. To understand the contexts in which users disclose to chatbots, we develop a taxonomy of tasks and sensitive topics, based on qualitative and quantitative analysis of naturally occurring conversations. We discuss these potential privacy harms and observe that: (1) personally identifiable information (PII) appears in unexpected contexts such as in translation or code editing (48% and 16% of the time, respectively) and (2) PII detection alone is insufficient to capture the sensitive topics that are common in human-chatbot interactions, such as detailed sexual preferences or specific drug use habits. We believe that these high disclosure rates are of significant importance for researchers and data curators, and we call for the design of appropriate nudging mechanisms to help users moderate their interactions.

Foundation models, now powering most of the exciting applications in deep learning, are almost universally based on the Transformer architecture and its core attention module. Many subquadratic-time architectures such as linear attention, gated convolution and recurrent models, and structured state space models (SSMs) have been developed to address Transformers' computational inefficiency on long sequences, but they have not performed as well as attention on important modalities such as language. We identify that a key weakness of such models is their inability to perform content-based reasoning, and make several improvements. First, simply letting the SSM parameters be functions of the input addresses their weakness with discrete modalities, allowing the model to selectively propagate or forget information along the sequence length dimension depending on the current token. Second, even though this change prevents the use of efficient convolutions, we design a hardware-aware parallel algorithm in recurrent mode. We integrate these selective SSMs into a simplified end-to-end neural network architecture without attention or even MLP blocks (Mamba). Mamba enjoys fast inference (5x higher throughput than Transformers) and linear scaling in sequence length, and its performance improves on real data up to million-length sequences. As a general sequence model backbone, Mamba achieves state-of-the-art performance across several modalities such as language, audio, and genomics. On language modeling, our Mamba-3B model outperforms Transformers of the same size and matches Transformers twice its size, both in pretraining and downstream evaluation.

Retrieval-augmented generation (RAG) augments large language models (LLM) by retrieving relevant knowledge, showing promising potential in mitigating LLM hallucinations and enhancing response quality, thereby facilitating the great adoption of LLMs in practice. However, we find that existing RAG systems are inadequate in answering multi-hop queries, which require retrieving and reasoning over multiple pieces of supporting evidence. Furthermore, to our knowledge, no existing RAG benchmarking dataset focuses on multi-hop queries. In this paper, we develop a novel dataset, MultiHop-RAG, which consists of a knowledge base, a large collection of multi-hop queries, their ground-truth answers, and the associated supporting evidence. We detail the procedure of building the dataset, utilizing an English news article dataset as the underlying RAG knowledge base. We demonstrate the benchmarking utility of MultiHop-RAG in two experiments. The first experiment compares different embedding models for retrieving evidence for multi-hop queries. In the second experiment, we examine the capabilities of various state-of-the-art LLMs, including GPT-4, PaLM, and Llama2-70B, in reasoning and answering multi-hop queries given the evidence. Both experiments reveal that existing RAG methods perform unsatisfactorily in retrieving and answering multi-hop queries. We hope MultiHop-RAG will be a valuable resource for the community in developing effective RAG systems, thereby facilitating greater adoption of LLMs in practice. We make the dataset and benchmarking code publicly available via GitHub.

Model collapse, as introduced in (Shumailov et al., 2023), refers to the phenomenon where training models on synthetic data generated from previously trained models leads to a deterioration in performance. This recursive training loop makes the tails of the original distribution disappear, thereby making future-generation models forget about the initial (real) distribution. With the aim of rigorously understanding model collapse in language models, we consider in this paper a statistical model that allows us to characterize the impact of various recursive training scenarios. Specifically, we demonstrate that model collapse cannot be avoided when training solely on synthetic data. However, when mixing both real and synthetic data, we provide an estimate of a maximal amount of synthetic data below which model collapse can eventually be avoided. Our theoretical conclusions are further supported by empirical validations.

Common self-improvement approaches for large language models (LLMs), such as STaR (Zelikman et al., 2022), iteratively fine-tune LLMs on self-generated solutions to improve their problem-solving ability. However, these approaches discard the large amounts of incorrect solutions generated during this process, potentially neglecting valuable information in such solutions. To address this shortcoming, we propose V-STaR that utilizes both the correct and incorrect solutions generated during the self-improvement process to train a verifier using DPO that judges correctness of model-generated solutions. This verifier is used at inference time to select one solution among many candidate solutions. Running V-STaR for multiple iterations results in progressively better reasoners and verifiers, delivering a 4% to 17% test accuracy improvement over existing self-improvement and verification approaches on common code generation and math reasoning benchmarks with LLaMA2 models.