ACL 2024 论文列表

This paper explores the image-sharing capability of Large Language Models (LLMs), such as GPT-4 and LLaMA 2, in a zero-shot setting. To facilitate a comprehensive evaluation of LLMs, we introduce the photochatplus dataset, which includes enriched annotations (ie intent, triggering sentence, image description, and salient information). Furthermore, we present the gradient-free and extensible Decide, Describe, and Retrieve () framework. With extensive experiments, we unlock the image-sharing capability of equipped with LLMs in zero-shot prompting, with ChatGPT achieving the best performance.Our findings also reveal the emergent image-sharing ability in LLMs under zero-shot conditions, validating the effectiveness of . We use this framework to demonstrate its practicality and effectiveness in two real-world scenarios: (1) human-bot interaction and (2) dataset augmentation. To the best of our knowledge, this is the first study to assess the image-sharing ability of various LLMs in a zero-shot setting. We make our source code and dataset publicly available at https://github.com/passing2961/DribeR.

Large language models (LLMs) are increasingly prevalent in conversational systems due to their advanced understanding and generative capabilities in general contexts. However, their effectiveness in task-oriented dialogues (TOD), which requires not only response generation but also effective dialogue state tracking (DST) within specific tasks and domains, remains less satisfying. In this work, we propose a novel approach FnCTOD for solving DST with LLMs through function calling. This method improves zero-shot DST, allowing adaptation to diverse domains without extensive data collection or model tuning. Our experimental results demonstrate that our approach achieves exceptional performance with both modestly sized open-source and also proprietary LLMs: with in-context prompting it enables various 7B or 13B parameter models to surpass the previous state-of-the-art (SOTA) achieved by ChatGPT, and improves ChatGPT’s performance beating the SOTA by 5.6% average joint goal accuracy (JGA). Individual model results for GPT-3.5 and GPT-4 are boosted by 4.8% and 14%, respectively. We also show that by fine-tuning on a small collection of diverse task-oriented dialogues, we can equip modestly sized models, specifically a 13B parameter LLaMA2-Chat model, with function-calling capabilities and DST performance comparable to ChatGPT while maintaining their chat capabilities. We have made the code publicly available at https://github.com/facebookresearch/FnCTOD.

In this paper, we uncover a positional bias in the evaluation paradigm of adopting large language models (LLMs), e.g., GPT-4, as a referee to score and compare the quality of responses generated by candidate models. We find that the quality ranking of candidate responses can be easily hacked by simply altering their order of appearance in the context. This manipulation allows us to skew the evaluation result, making one model appear considerably superior to the other, e.g., Vicuna-13B could beat ChatGPT on 66 over 80 tested queries with ChatGPT as an evaluator. We propose a simple yet effective calibration framework to address our discovered positional bias.To evaluate the effectiveness of our framework, we manually annotate the “win/tie/lose” outcomes of responses from ChatGPT and Vicuna-13B in the Vicuna Benchmark’s question prompt. Extensive experiments demonstrate that our approach successfully alleviates evaluation bias, resulting in closer alignment with human judgments.

Considerable efforts have been invested in augmenting the role-playing proficiency of open-source large language models (LLMs) by emulating proprietary counterparts. Nevertheless, we posit that LLMs inherently harbor role-play capabilities, owing to the extensive knowledge of characters and potential dialogues ingrained in their vast training corpora. Thus, we introduce Ditto, the first self-alignment method for role-play, which encourages an instruction-following LLM to simulate role-play dialogues as a variant of reading comprehension, and creates a role-play training set comprising 4000 characters, surpassing the scale of currently available datasets by tenfold regarding the number of roles. Subsequently, we fine-tune the LLM using this self-generated dataset to augment its role-playing capabilities. Upon evaluating our meticulously constructed role-play benchmark and the roleplay subset of MT-Bench, Ditto, in various parameter scales, consistently maintains a consistent role identity and provides accurate role-specific knowledge in multi-turn role-play conversations, outperforming all open-source role-play baselines. Furthermore, we present the first cross-supervision role-play experiment, revealing that the role-play styles can be easily acquired, while the intrinsic capabilities of LLMs confine the knowledge within role-play.

Recognizing fallacies is crucial for ensuring the quality and validity of arguments across various domains. However, computational fallacy recognition faces challenges due to the diverse genres, domains, and types of fallacies found in datasets. This leads to a highly multi-class, and even multi-label, setup with substantial class imbalance. In this study, we aim to enhance existing models for fallacy recognition by incorporating additional context and by leveraging large language models to generate synthetic data, thus increasing the representation of the infrequent classes. We experiment with GPT3.5 to generate synthetic examples and we examine the impact of prompt settings for this. Moreover, we explore zero-shot and few-shot scenarios to evaluate the effectiveness of using the generated examples for training smaller models within a unified fallacy recognition framework. Furthermore, we analyze the overlap between the synthetic data and existing fallacy datasets. Finally, we investigate the usefulness of providing supplementary context for detecting fallacy types that need such context, e.g., diversion fallacies. Our evaluation results demonstrate consistent improvements across fallacy types, datasets, and generators. The code and the synthetic datasets are all publicly available.

Advances in model editing through neuron pruning hold promise for removing undesirable concepts from large language models. However, it remains unclear whether models have the capacity to reacquire pruned concepts after editing. To investigate this, we evaluate concept relearning in models by tracking concept saliency and similarity in pruned neurons during retraining for named entity recognition tasks. Our findings reveal that models can quickly regain performance post-pruning by relocating advanced concepts to earlier layers and reallocating pruned concepts to primed neurons with similar semantics. This suggests that models exhibit polysemantic capacities and can blend old and new concepts in individual neurons. While neuron pruning provides interpretability into model concepts, our results highlight the challenges of permanent concept removal for improved model *safety*. Monitoring concept reemergence and developing techniques to mitigate relearning of unsafe concepts will be important directions for more robust model editing. Overall, our work strongly demonstrates the resilience and fluidity of concept representations in LLMs post concept removal.

Existing datasets for narrative understanding often fail to represent the complexity and uncertainty of relationships in real-life social scenarios. To address this gap, we introduce a new benchmark, Conan, designed for extracting and analysing intricate character relation graphs from detective narratives. Specifically, we designed hierarchical relationship categories and manually extracted and annotated role-oriented relationships from the perspectives of various characters, incorporating both public relationships known to most characters and secret ones known to only a few. Our experiments with advanced Large Language Models (LLMs) like GPT-3.5, GPT-4, and Llama2 reveal their limitations in inferencing complex relationships and handling longer narratives. The combination of the Conan dataset and our pipeline strategy is geared towards understanding the ability of LLMs to comprehend nuanced relational dynamics in narrative contexts.

While large language models (LLMs) have demonstrated remarkable reasoning capabilities, they are not without their flaws and inaccuracies. Recent studies have introduced various methods to mitigate these limitations. Temporal reasoning (TR), in particular, presents a significant challenge for LLMs due to its reliance on diverse temporal concepts and intricate temporal logic. In this paper, we propose TG-LLM, a novel framework towards language-based TR. Instead of reasoning over the original context, we adopt a latent representation, temporal graph (TG) that enhances the learning of TR. A synthetic dataset (TGQA), which is fully controllable and requires minimal supervision, is constructed for fine-tuning LLMs on this text-to-TG translation task. We confirmed in experiments that the capability of TG translation learned on our dataset can be transferred to other TR tasks and benchmarks. On top of that, we teach LLM to perform deliberate reasoning over the TGs via Chain-of-Thought (CoT) bootstrapping and graph data augmentation. We observed that those strategies, which maintain a balance between usefulness and diversity, bring more reliable CoTs and final results than the vanilla CoT distillation.

One major challenge for Large Language Models (LLMs) is completing complex tasks involving multiple entities, such as tool APIs. To tackle this, one approach is to retrieve relevant entities to enhance LLMs in task completion. A crucial issue here is obtaining accurate natural language representations for each entity to aid in retriever precision. In this paper, we propose the Natural Language Representation Optimization Problem, which aims to refine entity descriptions for improved retrieval and LLM utilization. We introduce the Learning to Represent with Natural Language method, which utilizes LLMs to optimize entity representations consisting of text patterns based on environmental feedback. We iteratively prompt LLMs to enhance or adjust patterns based on entity samples and evaluate their effectiveness through environmental feedback. Our method successfully learns human-understandable representations for classification tasks (e.g., instructions and documents) and API call tasks (e.g., APIbench and Virtual Home), significantly improving GPT-4’s task performance.

The development of large language models (LLMs) brings significant changes to the field of natural language processing (NLP), enabling remarkable performance in various high-level tasks, such as machine translation, question-answering, dialogue generation, etc., under end-to-end settings without requiring much training data. Meanwhile, fundamental NLP tasks, particularly syntactic parsing, are also essential for language study as well as evaluating the capability of LLMs for instruction understanding and usage. In this paper, we focus on analyzing and improving the capability of current state-of-the-art LLMs on a classic fundamental task, namely constituency parsing, which is the representative syntactic task in both linguistics and natural language processing. We observe that these LLMs are effective in shallow parsing but struggle with creating correct full parse trees. To improve the performance of LLMs on deep syntactic parsing, we propose a three-step approach that firstly prompts LLMs for chunking, then filters out low-quality chunks, and finally adds the remaining chunks to prompts to instruct LLMs for parsing, with later enhancement by chain-of-thought prompting. Experimental results on English and Chinese benchmark datasets demonstrate the effectiveness of our approach on improving LLMs’ performance on constituency parsing.

Large language models (LLMs) demonstrate remarkable multilingual capabilities without being pre-trained on specially curated multilingual parallel corpora.It remains a challenging problem to explain the underlying mechanisms by which LLMs process multilingual texts.In this paper, we delve into the composition of Transformer architectures in LLMs to pinpoint language-specific regions.Specially, we propose a novel detection method, language activation probability entropy (LAPE), to identify language-specific neurons within LLMs.Based on LAPE, we conduct comprehensive experiments on several representative LLMs, such as LLaMA-2, BLOOM, and Mistral. Our findings indicate that LLMs’ proficiency in processing a particular language is predominantly due to a small subset of neurons, primarily situated in the models’ top and bottom layers.Furthermore, we showcase the feasibility to “steer” the output language of LLMs by selectively activating or deactivating language-specific neurons. Our research provides important evidence to the understanding and exploration of the multilingual capabilities of LLMs.

Beam search decoding is the de-facto method for decoding auto-regressive Neural Machine Translation (NMT) models, including multilingual NMT where the target language is specified as an input. However, decoding multilingual NMT models commonly produces off-target translations – yielding translation outputs not in the intended language.In this paper, we first conduct an error analysis of off-target translations for a strong multilingual NMT model and identify how these decodings are produced during beam search. We then propose Language-informed Beam Search (LiBS), a general decoding algorithm incorporating an off-the-shelf Language Identification (LiD) model into beam search decoding to reduce off-target translations. LiBS is an inference-time procedure that is NMT-model agnostic and does not require any additional parallel data. Results show that our proposed LiBS algorithm on average improves +1.1 BLEU and +0.9 BLEU on WMT and OPUS datasets, and reduces off-target rates from 22.9% to 7.7% and 65.8% to 25.3% respectively.

To date, most investigations on surprisal and entropy effects in reading have been conducted on the group level, disregarding individual differences. In this work, we revisit the predictive power (PP) of different LMs’ surprisal and entropy measures on data of human reading times as a measure of processing effort by incorporating information of language users’ cognitive capacities. To do so, we assess the PP of surprisal and entropy estimated from generative language models (LMs) on reading data obtained from individuals who also completed a wide range of psychometric tests.Specifically, we investigate if modulating surprisal and entropy relative to cognitive scores increases prediction accuracy of reading times, and we examine whether LMs exhibit systematic biases in the prediction of reading times for cognitively high- or low-performing groups, revealing what type of psycholinguistic subjects a given LM emulates.Our study finds that in most cases, incorporating cognitive capacities increases predictive power of surprisal and entropy on reading times, and that generally, high performance in the psychometric tests is associated with lower sensitivity to predictability effects. Finally, our results suggest that the analyzed LMs emulate readers with lower verbal intelligence, suggesting that for a given target group (i.e., individuals with high verbal intelligence), these LMs provide less accurate predictability effect estimates.

Large Language Models (LLMs) have transformed NLP with their remarkable In-context Learning (ICL) capabilities. Automated assistants based on LLMs are gaining popularity; however, adapting them to novel tasks is still challenging. While colossal models excel in zero-shot performance, their computational demands limit widespread use, and smaller language models struggle without context. This paper investigates whether LLMs can generalize from labeled examples of predefined tasks to novel tasks. Drawing inspiration from biological neurons and the mechanistic interpretation of the Transformer architecture, we explore the potential for information sharing across tasks. We design a cross-task prompting setup with three LLMs and show that LLMs achieve significant performance improvements despite no examples from the target task in the context. Cross-task prompting leads to a remarkable performance boost of 107% for LLaMA-2 7B, 18.6% for LLaMA-2 13B, and 3.2% for GPT 3.5 on average over zero-shot prompting, and performs comparable to standard in-context learning. The effectiveness of generating pseudo-labels for in-task examples is demonstrated, and our analyses reveal a strong correlation between the effect of cross-task examples and model activation similarities in source and target input tokens. This paper offers a first-of-its-kind exploration of LLMs’ ability to solve novel tasks based on contextual signals from different task examples.

In this paper, we begin by illustrating that, when presented with a query, Large Language Models (LLMs) capable of providing accurate responses tend to exhibit a more uniform probability distribution compared to their less proficient counterparts. Building upon this observation, we introduce a novel self-assessment criterion termed ProbDiff for evaluating the performance of diverse LLMs. This method eliminates the need for training an additional evaluation model or relying on external proprietary models such as GPT-4 as a judger. Instead, it solely relies on the LLMs under evaluation to compute the probability discrepancy between the original response generation and its revised versions. A higher discrepancy in two LLMs for the same query suggests a relatively weaker ability. We discover that ProbDiff yields comparable results to mainstream GPT-4-based evaluations on various scenarios including NLG tasks like translation and summarization, as well as LLM evaluation benchmarks such as AlignBench, MT-Bench, and AlpacaEval, across LLMs of different sizes.

We propose RESTA to perform LLM realignment towards safety, which gets compromised due to downstream task fine-tuning. RESTA stands for REstoring Safety through Task Arithmetic. At its core, it involves a simple arithmetic addition of a safety vector to the weights of the compromised model. We demonstrate the effectiveness of RESTA in both parameter-efficient and full fine-tuning, covering a wide range of downstream tasks, including instruction following in Chinese, English, and Hindi, as well as problem-solving capabilities in Code and Math. We also showcase the generalizability of RESTA on three existing safety evaluation benchmarks and a multilingual benchmark dataset proposed as a part of this work, consisting of 550 harmful questions covering 11 categories, each with 5 sub-categories of harm. Overall, RESTA decreases the harmfulness of the compromised model from 18.6% to 5.1% and from 9.2% to 1.5% in parameter-efficient and full fine-tuning, respectively, while maintaining most of the model’s performance on the task. We release the source codes at: https://github.com/declare-lab/resta.

We argue that language-only models don’t learn the physical manifestation of language. We present an empirical investigation of visual-auditory properties of language through a series of tasks, termed H-Test.These tasks highlight a fundamental gap between human linguistic understanding and the sensory-deprived linguistic understanding of LLMs. In support of our hypothesis, 1. deliberate reasoning (Chain-of-Thought), 2. few-shot examples, or 3. stronger LLM from the same model family (LLaMA 2 13B -> LLaMA 2 70B) has no significant effect on H-Test performance. We bring in the philosophical case of Mary, who learns about the world in a sensory-deprived environment as a useful conceptual framework to understand how language-only models learn about the world (Jackson, 1986). Our experiments show that some of the strongest proprietary LLMs stay near random chance baseline accuracy of 50%, highlighting the limitations of linguistic knowledge acquired in the absence of sensory experience. Our code and data are available at <github.com/brucewlee/h-test>.

The ability (and inability) of large language models (LLMs) to perform arithmetic tasks has been the subject of much theoretical and practical debate. We show that LLMs are frequently able to correctly and confidently predict the first digit of n-digit by m-digit multiplication tasks without using chain of thought reasoning, despite these tasks require compounding operations to solve. Simultaneously, LLMs in practice often fail to correctly or confidently predict the last digit of an n-digit by m-digit multiplication, a task equivalent to 1-digit by 1-digit multiplication which can be easily learned or memorized. We show that the latter task can be solved more robustly when the LLM is conditioned on all of the correct higher-order digits, which on average increases the confidence of the correct last digit on 5-digit by 5-digit multiplication tasks using Llama 2-13B by over 230% (0.13→0.43) and Mistral-7B by 150% (0.22→0.55).

We investigate ways of using monolingual data in both the source and target languages for improving low-resource machine translation. As a case study, we experiment with translation from Finnish to Northern Sámi.Our experiments show that while conventional backtranslation remains a strong contender, using synthetic target-side data when training backtranslation models can be helpful as well.We also show that monolingual data can be used to train a language model which can act as a regularizer without any augmentation of parallel data.