COLM 2024 Papers

The goal of detoxifying language models is to reduce the chances of producing offensive or harmful output in pre-trained language models (PLMs), ensuring their safer use. A recently proposed detoxification method utilizes the task vector obtained by subtraction from the fine-tuned model on toxic datasets to the pre-trained model. This approach has shown effectiveness for detoxification but still suffers from degradation. This study focuses on further mitigating degradation while maintaining detoxification performance. To mitigate the degradation, we propose a method that detoxifies the PLMs by fine-tuning multiple models on split toxic datasets and by merging the subtracted task vectors. We conducted experiments on two toxic datasets (Civil Comments and Toxigen) with five PLMs (GPT2-small, GPT2-medium, GPT2-large, Phi-1.5, and Llama2-7b), demonstrating that our method consistently achieves a lower toxicity score while preventing the degradation compared to baseline methods. Especially, with the GPT2-small model on the Toxigen dataset, degradation was reduced by 38.9\% compared to that of an existing task vector method while maintaining a similar toxicity score. In addition, we found that merging multiple detoxified models tends to increase the number of parameters that remained almost unchanged from the pre-trained model. We assume that by merging multiple detoxified models, "decoupling noise and toxic parameters" is implicitly achieved. The accidental noise in the parameter shift unrelated to detoxification disappears by averaging noise, whereas the parameter shift associated with detoxification is maintained. We hope that the findings of this study will be applied not only to detoxification but also to many other research domains that seek to suppress undesirable outputs of language models.

Researchers and practitioners operating on a limited budget face the well-known cost-performance trade-off dilemma. The challenging decision often centers on whether to use a large LLM with better performance or a smaller one with reduced costs. This has motivated recent research in the optimisation of LLM calls. Either a cascading strategy is used, where a smaller LLM or both are called causally, or a routing strategy is used, where only one model is ever called. This is dependent on a decision criterion in both scenarios which is typically an auxiliary neural model. In this work, we propose a cost-effective solution; we use only the uncertainty of the generations of the small LLM as the decision criterion. We compare our approach with both cascading and routing strategies using three different pairs of pre-trained small and large LLMs, on nine different tasks and against approaches that require an additional neural model. Our experiments reveal this simple solution optimally balances cost and performance, outperforming existing methods on 25 out of 27 experimental setups.

Large language models (LLMs) has demonstrated superior performance on various downstream tasks. However, their practical applications are hindered by their immense memory and computation requirements. Although recent post-training quantization methods can effectively reduce memory usage and improve computational efficiency, they often overlook the varying sensitivity of different layer weights to bit precision. Additionally, the previous methods suffer from significant accuracy loss under low-bit quantization (2-3 bits). To address these limitations, we propose Adaptive Mixed Precision and Low-Rank Quantization Error Reconstruction for LLMs (AMLQ), which achieves state-of-the-art performance under the approximate average bit precision overall. Furthermore, we introduce the low-rank decomposition to reconstruct quantization error based on the output features. Experimental results demonstrate that this method can be effectively combined with various quantization techniques and bring considerable performance gains. Our approach comprehensively considers model performance and inference efficiency, offering more than 3$\times$ speedup over the FP16 execution.

The rapid advancement of large language models (LLMs) has significantly propelled the development of text-based chatbots, demonstrating their capability to engage in coherent and contextually relevant dialogues. However, extending these advancements to enable end-to-end speech-to-speech conversation bots remains a formidable challenge, primarily due to the extensive dataset and computational resources required. The conventional approach of cascading automatic speech recognition (ASR), LLM, and text-to-speech (TTS) models in a pipeline, while effective, suffers from unnatural prosody because it lacks direct interactions between the input audio and its transcribed text and the output audio. These systems are also limited by their inherent latency from the ASR process for real-time applications. This paper introduces Style-Talker, an innovative framework that fine-tunes an audio LLM alongside a style-based TTS model for fast spoken dialog generation. Style-Talker takes user input audio and uses transcribed chat history and speech styles to generate both the speaking style and text for the response. Subsequently, the TTS model synthesizes the speech, which is then played back to the user. While the response speech is being played, the input speech undergoes ASR processing to extract the transcription and speaking style, serving as the context for the ensuing dialogue turn. This novel pipeline accelerates the traditional cascade ASR-LLM-TTS systems while integrating rich paralinguistic information from input speech. Our experimental results show that Style-Talker significantly outperforms the conventional cascade and speech-to-speech baselines in terms of both dialogue naturalness and coherence while being more than 50\% faster. The demo and code are available at https://styletalker.github.io/.

With the digital imagery landscape rapidly evolving, image stocks and AI-generated image marketplaces have become central to visual media. Traditional stock images now exist alongside innovative platforms that trade in prompts for AI-generated visuals, driven by sophisticated APIs like DALL-E 3 and Midjourney. This paper studies the possibility of employing multi-modal models with enhanced visual understanding to mimic the outputs of these platforms, introducing an original attack strategy. Our method leverages fine-tuned CLIP models, a multi-label classifier, and the descriptive capabilities of GPT-4V to create prompts that generate images similar to those available in marketplaces and from premium stock image providers, yet at a markedly lower expense. In presenting this strategy, we aim to spotlight a new class of economic and security considerations within the realm of digital imagery. Our findings, supported by both automated metrics and human assessment, reveal that comparable visual content can be produced for a fraction of the prevailing market prices (\$0.23 - \$0.27 per image), emphasizing the need for awareness and strategic discussions about the integrity of digital media in an increasingly AI-integrated landscape. Additionally, this approach holds promise as a tool for data augmentation, potentially enhancing machine learning models by providing varied and cost-effective training data. Our work also contributes to the field by assembling a dataset consisting of approximately 19 million prompt-image pairs generated by the popular Midjourney platform, which we plan to release publicly.

There is a belief that learning to compress well will lead to intelligence. Recently, language modeling has been shown to be equivalent to compression, which offers a compelling rationale for the success of large language models (LLMs): development of more advanced language models is essentially enhancing compression which facilitates intelligence. Despite such appealing discussions, little empirical evidence is present for the interplay between compression and intelligence. In this work, we examine the relationship between compression and intelligence in the context of LLMs, treating LLMs as data compressors. Given the abstract concept of "intelligence", we adopt the average downstream benchmark scores as a surrogate, specifically targeting intelligence related to knowledge and commonsense, coding, and mathematical reasoning. Across 12 benchmarks, our study brings together 31 public LLMs that vary in size and originate from diverse organizations. Remarkably, we find that LLMs' intelligence -- reflected by benchmark scores -- almost **linearly** correlates with their ability to compress external text corpora. These results provide concrete evidence supporting the belief that superior compression indicates greater intelligence. Furthermore, our findings suggest that compression efficiency, as an unsupervised metric derived from raw text corpora, serves as a reliable evaluation measure that is linearly associated with the model capabilities. This work advocates for the adoption of compression performance as a stable, flexible, and reliable metric for evaluating LLMs. We open-source our compression datasets as well as our data collection pipelines to facilitate future researchers to assess compression properly.

While Transformers have enabled tremendous progress in various application settings, such architectures still struggle with solving planning and sequential decision-making tasks. In this work, we demonstrate how to train Transformers to solve complex planning tasks. This is accomplished by first designing a synthetic language that captures the computation performed by the $A^*$ search algorithm when solving a planning task. Then, an encoder-decoder Transformer model is trained to predict this language, resulting in a language model that can correctly solve novel planning tasks by generating $A^*$'s search dynamics. We fine tune this model to obtain a Searchformer, a Transformer model that optimally solves previously unseen Sokoban puzzles 93.7\% of the time, while using up to 26.8\% fewer search steps than our $A^*$ reference implementation. Searchformer significantly outperforms baselines that predict the optimal plan directly with a 5-10$\times$ smaller model size and a 10$\times$ smaller training dataset. Lastly, we demonstrate how Searchformer scales to larger and more complex decision making tasks with improved percentage of solved tasks and shortened search dynamics.

By leveraging the retrieval of information from external knowledge databases, Large Language Models (LLMs) exhibit enhanced capabilities for accomplishing many knowledge-intensive tasks. However, due to the inherent flaws of current retrieval systems, there might exist irrelevant information within those retrieving top-ranked passages. In this work, we present a comprehensive investigation into the robustness of LLMs to different types of irrelevant information under various conditions. We initially introduce a framework to construct high-quality irrelevant information that ranges from semantically unrelated, partially related, and related to questions. Furthermore, our analysis demonstrates that the constructed irrelevant information not only scores highly on similarity metrics, being highly retrieved by existing systems, but also bears semantic connections to the context. Our investigation reveals that current LLMs still face challenges in discriminating highly semantically related information and can be easily distracted by these irrelevant yet misleading content. Besides, we also find that current solutions for handling irrelevant information have limitations in improving the robustness of LLMs to such distractions. All the resources are available on [GitHub](https://github.com/Di-viner/LLM-Robustness-to-Irrelevant-Information).

The success of Large Language Models (LLMs) in multicultural environments hinges on their ability to understand users' diverse cultural backgrounds. We measure this capability by having an LLM simulate human profiles representing various nationalities within the scope of a questionnaire-style psychological experiment. Specifically, we employ GPT-3.5 to reproduce reactions to persuasive news articles of 7,286 participants from 15 countries; comparing the results with a dataset of real participants sharing the same demographic traits. Our analysis shows that specifying a person's country of residence improves GPT-3.5's alignment with their responses. In contrast, using native language prompting introduces shifts that significantly reduce overall alignment, with some languages particularly impairing performance. These findings suggest that while direct nationality information enhances the model's cultural adaptability, native language cues do not reliably improve simulation fidelity and can detract from the model's effectiveness.

Recent advancements have significantly augmented the reasoning capabilities of Large Language Models (LLMs) through various methodologies, especially chain-of-thought (CoT) reasoning. However, previous methods often struggle to address reasoning errors in intermediate steps, which can lead to accumulative errors. In this paper, we propose Deductive Beam Search (DBS), which seamlessly integrates CoT and deductive reasoning with step-wise beam search for LLMs. Our approach deploys a verifier, verifying the deducibility of a reasoning step and its premises, thus alleviating the error accumulation. Furthermore, we introduce a scalable and labor-free data construction method to amplify our model’s verification capabilities. Extensive experiments demonstrate that our approach significantly enhances the base performance of LLMs of various scales (7B, 13B, 70B, and ChatGPT) across 8 reasoning datasets from 3 diverse reasoning genres, including arithmetic, commonsense, and symbolic. Moreover, our analysis proves DBS’s capability of detecting diverse and subtle reasoning errors and robustness on different model scales.

Humans are not homo economicus (i.e., rational economic beings). As humans, we exhibit systematic behavioral biases such as loss aversion, anchoring, framing, etc., which lead us to make suboptimal economic decisions. Insofar as such biases may be embedded in text data on which large language models (LLMs) are trained, to what extent are LLMs prone to the same behavioral biases? Understanding these biases in LLMs is crucial for deploying LLMs to support human decision-making. We propose utility theory-a paradigm at the core of modern economic theory-as an approach to evaluate the economic biases of LLMs. Utility theory enables the quantification and comparison of economic behavior against benchmarks such as perfect rationality or human behavior. To demonstrate our approach, we quantify and compare the economic behavior of a variety of open- and closed-source LLMs. We find that the economic behavior of current LLMs is neither entirely human-like nor entirely economicus-like. We also find that most current LLMs struggle to maintain consistent economic behavior across settings. Finally, we illustrate how our approach can measure the effect of interventions such as prompting on economic biases.

As language models (LMs) become increasingly powerful and widely used, it is important to quantify them for sociodemographic bias with potential for harm. Prior measures of bias are sensitive to perturbations in the templates designed to compare performance across social groups, due to factors such as low diversity or limited number of templates. Also, most previous work considers only one NLP task. We introduce Comprehensive Assessment of Language Models (CALM) for robust measurement of social biases. We use sixteen datasets for question-answering, sentiment analysis and natural language inference and filter them to produce 224 templates with high diversity (e.g., length, vocabulary). This helps us create a novel dataset of 78,400 prompts covering the three NLP tasks. Our empirical evaluation shows that CALM bias scores are more robust and far less sensitive than previous bias measurements to perturbations in the templates, such as synonym substitution, or to random subset selection of templates. We apply CALM to 20 large language models, and find that for 2 LM series, larger parameter models tend to be more biased than smaller ones. The T0 series is the least biased model families, of the 20 LLMs investigated here.

In this study, we introduce $\textbf{CT-LLM}$, a groundbreaking 2B large language model (LLM) that illustrates a pivotal shift towards prioritizing the Chinese language in the development of LLMs. Uniquely initiated from scratch, CT-LLM diverges from the conventional methodology by primarily incorporating Chinese textual data, utilizing an extensive corpus of 1,200 billion tokens, including 800 billion Chinese tokens and 400 billion English tokens. This strategic composition facilitates the model's exceptional proficiency in understanding and processing Chinese, a capability further enhanced through alignment techniques including supervised fine-tuning (SFT) and direct preference optimization (DPO). Demonstrating remarkable performance on the ChineseHardCase Benchmark, CT-LLM not only excels in Chinese language tasks but also showcases its adeptness in English through SFT. This research challenges the prevailing paradigm of training LLMs predominantly on English corpora and then adapting them to other languages, broadening the horizons for LLM training methodologies. By open-sourcing full-process of CT-LLM, we aim to foster further exploration and innovation within both the academic and industrial spheres, paving the way for more inclusive and versatile language models in the future.

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 human judgments. However, annotators' judgments for subjective tasks can differ in many ways: they may reflect different qualitative judgments about an example, and they may be mapped 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.

The recent surge in open-source Large Language Models (LLMs), such as LLaMA, Falcon, and Mistral, provides diverse options for AI practitioners and researchers. However, most LLMs have only released partial artifacts, such as the final model weights or inference code, and technical reports increasingly limit their scope to high-level design choices and surface statistics. These choices hinder progress in the field by degrading transparency into the training of LLMs and forcing teams to rediscover many details in the training process. We present **LLM360**, an initiative to fully open-source LLMs, which advocates for all training code and data, model checkpoints, and intermediate results to be made available to the community. The goal of LLM360 is to support open and collaborative AI research by making the end-to-end LLM training process transparent and reproducible by everyone. As a first step of LLM360, we release two 7B parameter LLMs pre-trained from scratch, Amber and Crystal, including their training code, data, intermediate checkpoints, and analyses. We are committed to continually pushing the boundaries of LLMs through this open-source effort. More large-scale and stronger models are underway and will be released in the future.

In settings where most deaths occur outside the healthcare system, verbal autopsies (VAs) are a common tool to monitor trends in causes of death (COD). VAs are interviews with a surviving caregiver or relative that are used to predict the decedent’s COD. Turning VAs into actionable insights for researchers and policymakers requires two steps (i) predicting likely COD using the VA interview and (ii) performing inference with predicted CODs (e.g. modeling the breakdown of causes by demographic factors using a sample of deaths). In this paper, we develop a method for valid inference using outcomes (in our case COD) predicted from free-form text using state-of-the-art NLP techniques. This method, which we call multiPPI++, extends recent work in “prediction-powered inference” to multinomial classification. We leverage a suite of NLP techniques for COD prediction and, through empirical analysis of VA data, we demonstrate the effectiveness of our approach in handling transportability issues. multiPPI++ recovers ground truth estimates, regardless of which NLP model produced predictions and regardless of whether they were produced by a more accurate predictor like GPT-4-32k or a less accurate predictor like KNN. Our findings demonstrate the practical importance of inference correction for public health decision-making and suggests that if inference tasks are the end goal, having a small amount of contextually relevant, high quality labeled data is essential regardless of the NLP algorithm.

It is a common belief that large language models (LLMs) are better than smaller-sized ones. However, larger models also require significantly more time and compute during inference. This begs the question: what happens when both models operate under the same budget? (e.g., compute, run-time). To address this question, we analyze code generation LLMs of various sizes and make comparisons such as running a 70B model once vs. generating five outputs from a 13B model. We consider a standard unit-test setup, which can be used to select the correct output from the smaller model. Our findings reveal that the repeated use of smaller models can yield consistent improvements, with gains of up to 15% across five tasks. On the other hand, in scenarios where unit-tests are unavailable, a ranking-based selection of candidates from the smaller model falls short of the performance of a single output from larger ones. Our results highlight the potential of using smaller models instead of larger ones, and the importance of studying approaches for ranking LLM outputs.

Large language models (LLMs) acquire extensive knowledge during pre-training, known as their parametric knowledge. However, to remain up-to-date and align with human instructions, LLMs inevitably require external knowledge during interactions. This raises a crucial question: How will LLMs respond when external knowledge interferes with their parametric knowledge? To uncover the impacts systematically, we construct parametric knowledge graphs to reveal different LLM knowledge structures, and introduce external information through external knowledge distractors of varying degrees, methods, positions, and formats. Experiments on both closed and open-source models demonstrate that LLMs tend to believe in external knowledge sources, particularly when they direct conflict or make confounding changes within detailed contexts. We also discover while LLMs are sensitive to external knowledge veracity, they still get distracted by unrelated information. These findings highlight the mechanisms behind LLM's integration of external knowledge, even indirectly, during model-user interactions.

Language models have shown unprecedented capabilities, sparking debate over the source of their performance. Is it merely the outcome of learning syntactic patterns and surface level statistics, or do they extract semantics and a world model from the text? Prior work by Li et al. investigated this by training a GPT model on synthetic, randomly generated Othello games and found that the model learned an internal representation of the board state. We extend this work into the more complex domain of chess, training on real games and investigating our model's internal representations using linear probes and contrastive activations. The model is given no a priori knowledge of the game and is solely trained on next character prediction, yet we find evidence of internal representations of board state. We validate these internal representations by using them to make interventions on the model's activations and edit its internal board state. Unlike Li et al's prior synthetic dataset approach, our analysis finds that the model also learns to estimate latent variables like player skill to better predict the next character. We derive a player skill vector and add it to the model, improving the model's win rate by up to 2.6 times.