ACL 2024 Papers

Neural Theory-of-Mind (N-ToM), machine’s ability to understand and keep track of the mental states of others, is pivotal in developing socially intelligent agents. However, prevalent N-ToM benchmarks have several shortcomings, including the presence of ambiguous and artificial narratives, absence of personality traits and preferences, a lack of questions addressing characters’ psychological mental states, and limited diversity in the questions posed. In response to these issues, we construct OpenToM, a new benchmark for assessing N-ToM with (1) longer and clearer narrative stories, (2) characters with explicit personality traits, (3) actions that are triggered by character intentions, and (4) questions designed to challenge LLMs’ capabilities of modeling characters’ mental states of both the physical and psychological world. Using OpenToM, we reveal that state-of-the-art LLMs thrive at modeling certain aspects of mental states in the physical world but fall short when tracking characters’ mental states in the psychological world.

The rapid development of Chinese large language models (LLMs) poses big challenges for efficient LLM evaluation. While current initiatives have introduced new benchmarks or evaluation platforms for assessing Chinese LLMs, many of these focus primarily on capabilities, usually overlooking potential alignment and safety issues. To address this gap, we introduce OpenEval, an evaluation testbed that benchmarks Chinese LLMs across capability, alignment and safety. For capability assessment, we include 12 benchmark datasets to evaluate Chinese LLMs from 4 sub-dimensions: NLP tasks, disciplinary knowledge, commonsense reasoning and mathematical reasoning. For alignment assessment, OpenEval contains 7 datasets that examines the bias, offensiveness and illegalness in the outputs yielded by Chinese LLMs. To evaluate safety, especially anticipated risks (e.g., power-seeking, self-awareness) of advanced LLMs, we include 6 datasets. In addition to these benchmarks, we have implemented a phased public evaluation and benchmark update strategy to ensure that OpenEval is in line with the development of Chinese LLMs or even able to provide cutting-edge benchmark datasets to guide the development of Chinese LLMs. In our first public evaluation, we have tested a range of Chinese LLMs, spanning from 7B to 72B parameters, including both open-source and proprietary models. Evaluation results indicate that while Chinese LLMs have shown impressive performance in certain tasks, more attention should be directed towards broader aspects such as commonsense reasoning, alignment, and safety.

The introduction of large language models has significantly advanced code generation. However, open-source models often lack the execution capabilities and iterative refinement of advanced systems like the GPT-4 Code Interpreter. To address this, we introduce OpenCodeInterpreter, a family of open-source code systems designed for generating, executing, and iteratively refining code. Supported by Code Feedback, a dataset featuring 68K multi-turn interactions, OpenCodeInterpreter integrates execution and human feedback for dynamic code refinement. Our comprehensive evaluation of OpenCodeInterpreter across key benchmarks such as HumanEval, MBPP, and their enhanced versions from EvalPlus reveals its exceptional performance. Notably, OpenCodeInterpreter-33B achieves an accuracy of 83.2 (76.4) on the average (and plus versions) of HumanEval and MBPP, closely rivaling GPT-4’s 84.2 (76.2) and further elevates to 91.6 (84.6) with synthesized human feedback from GPT-4. OpenCodeInterpreterbrings the gap between open-source code generation models and proprietary systems like GPT-4 Code Interpreter.

Open-Set Semi-Supervised Text Classification (OSTC) aims to train a classification model on a limited set of labeled texts, alongside plenty of unlabeled texts that include both in-distribution and out-of-distribution examples. In this paper, we revisit the main challenge in OSTC, i.e., outlier detection, from a measurement disagreement perspective and innovatively propose to improve OSTC performance by directly maximizing the measurement disagreements. Based on the properties of in-measurement and cross-measurements, we design an Adversarial Disagreement Maximization (ADM) model that synergeticly optimizes the measurement disagreements. In addition, we develop an abnormal example detection and measurement calibration approach to guarantee the effectiveness of ADM training. Experiment results and comprehensive analysis of three benchmarks demonstrate the effectiveness of our model.

This paper introduces the Open Ko-LLM Leaderboard and the Ko-H5 Benchmark as vital tools for evaluating Large Language Models (LLMs) in Korean. Incorporating private test sets while mirroring the English Open LLM Leaderboard, we establish a robust evaluation framework that has been well integrated in the Korean LLM community. We perform data leakage analysis that shows the benefit of private test sets along with a correlation study within the Ko-H5 benchmark and temporal analyses of the Ko-H5 score. Moreover, we present empirical support for the need to expand beyond set benchmarks. We hope the Open Ko-LLM Leaderboard sets precedent for expanding LLM evaluation to foster more linguistic diversity.

The emergence of large language models (LLMs) has increasingly drawn attention to the use of LLMs for human-like planning. Existing work on LLM-based planning either focuses on leveraging the inherent language generation capabilities of LLMs to produce free-style plans, or employs reinforcement learning approaches to learn decision-making for a limited set of actions within restricted environments. However, both approaches exhibit significant discrepancies from the open and executable requirements in real-world planning. In this paper, we propose a new planning task–open grounded planning. The primary objective of open grounded planning is to ask the model to generate an executable plan based on a variable action set, thereby ensuring the executability of the produced plan. To this end, we establishes a benchmark for open grounded planning spanning a wide range of domains. Then we test current state-of-the-art LLMs along with five planning approaches, revealing that existing LLMs and methods still struggle to address the challenges posed by grounded planning in open domains. The outcomes of this paper define and establish a foundational dataset for open grounded planning, and shed light on the potential challenges and future directions of LLM-based planning.

Contemporary practices in instruction tuning often hinge on enlarging data scaling without a clear strategy for ensuring data quality, inadvertently introducing noise that may compromise model performance. To address this challenge, we introduce Nuggets, a novel and efficient methodology that leverages one-shot learning to discern and select high-quality instruction data from extensive datasets. Nuggets assesses the potential of individual instruction examples to act as effective one-shot learning instances, thereby identifying those that can significantly improve performance across diverse tasks. Nuggets utilizes a scoring system based on the impact of candidate examples on the perplexity of a diverse anchor set, facilitating the selection of the most advantageous data for instruction tuning. Through rigorous evaluations on two benchmarks, namely MT-Bench and Alpaca-Eval, our study illustrates that instruction tuning with the top 1% of examples curated by Nuggets substantially outperforms conventional methods employing the entire dataset.

Evaluation of opinion summaries using conventional reference-based metrics often fails to provide a comprehensive assessment and exhibits limited correlation with human judgments. While Large Language Models (LLMs) have shown promise as reference-free metrics for NLG evaluation, their potential remains unexplored for opinion summary evaluation. Furthermore, the absence of sufficient opinion summary evaluation datasets hinders progress in this area. In response, we introduce the SUMMEVAL-OP dataset, encompassing 7 dimensions crucial to the evaluation of opinion summaries: fluency, coherence, relevance, faithfulness, aspect coverage, sentiment consistency, and specificity. We propose OP-I-PROMPT, a dimension-independent prompt, along with OP-PROMPTS, a dimension-dependent set of prompts for opinion summary evaluation. Our experiments demonstrate that OP-I-PROMPT emerges as a good alternative for evaluating opinion summaries, achieving an average Spearman correlation of 0.70 with human judgments, surpassing prior methodologies. Remarkably, we are the first to explore the efficacy of LLMs as evaluators, both on closed-source and open-source models, in the opinion summary evaluation domain.

Large language models (LLMs) possess immense capabilities but are susceptible to malicious exploitation. To mitigate the risk, safety alignment is employed to align LLMs with ethical standards. However, safety-aligned LLMs may remain vulnerable to carefully crafted jailbreak attacks, but these attacks often face high rejection rates and limited harmfulness. In this paper, we expose the vulnerabilities of safety alignment in open-access LLMs, which can significantly enhance the success rate and harmfulness of jailbreak attacks. Through reverse alignment, achieved by accessing model parameters, we show the feasibility of efficiently fine-tuning LLMs to undermine their inherent safeguards. We investigate two types of reverse alignment techniques: reverse supervised fine-tuning (RSFT) and reverse preference optimization (RPO). RSFT operates by supervising the fine-tuning of LLMs to reverse their inherent values. We also explore how to prepare data needed for RSFT. RPO optimizes LLMs to enhance their preference for harmful content, reversing the models’ safety alignment. Our extensive experiments reveal that open-access high-performance LLMs can be adeptly reverse-aligned to output harmful content, even in the absence of manually curated malicious datasets. Our research acts as a whistleblower for the community, emphasizing the need to pay more attention to safety of open-accessing LLMs. It also underscores the limitations of current safety alignment approaches and calls for research on robust safety alignment methods to counteract malicious fine-tuning attacks.

The incorporation of Denoising Diffusion Models (DDMs) in the Text-to-Speech (TTS) domain is rising, providing great value in synthesizing high quality speech. Although they exhibit impressive audio quality, the extent of their semantic capabilities is unknown, and controlling their synthesized speech’s vocal properties remains a challenge. Inspired by recent advances in image synthesis, we explore the latent space of frozen TTS models, which is composed of the latent bottleneck activations of the DDM’s denoiser. We identify that this space contains rich semantic information, and outline several novel methods for finding semantic directions within it, both supervised and unsupervised. We then demonstrate how these enable off-the-shelf audio editing, without any further training, architectural changes or data requirements. We present evidence of the semantic and acoustic qualities of the edited audio, and provide supplemental samples: https://latent-analysis-grad-tts.github.io/speech-samples/.

Retrieval augmented generation (RAG) exhibits outstanding performance in promoting the knowledge capabilities of large language models (LLMs) with retrieved documents related to user queries. However, RAG only focuses on improving the response quality of LLMs via enhancing queries indiscriminately with retrieved information, paying little attention to what type of knowledge LLMs really need to answer original queries more accurately. In this paper, we suggest that long-tail knowledge is crucial for RAG as LLMs have already remembered common world knowledge during large-scale pre-training. Based on our observation, we propose a simple but effective long-tail knowledge detection method for LLMs. Specifically, the novel Generative Expected Calibration Error (GECE) metric is derived to measure the “long-tailness” of knowledge based on both statistics and semantics. Hence, we retrieve relevant documents and infuse them into the model for patching knowledge loopholes only when the input query relates to long-tail knowledge. Experiments show that, compared to existing RAG pipelines, our method achieves over 4x speedup in average inference time and consistent performance improvement in downstream tasks.

Driven by the demand for cross-sentence and large-scale relation extraction, document-level relation extraction (DocRE) has attracted increasing research interest. Despite the continuous improvement in performance, we find that existing DocRE models which initially perform well may make more mistakes when merely changing the entity names in the document, hindering the generalization to novel entity names. To this end, we systematically investigate the robustness of DocRE models to entity name variations in this work. We first propose a principled pipeline to generate entity-renamed documents by replacing the original entity names with names from Wikidata. By applying the pipeline to DocRED and Re-DocRED datasets, we construct two novel benchmarks named Env-DocRED and Env-Re-DocRED for robustness evaluation. Experimental results show that both three representative DocRE models and two in-context learned large language models consistently lack sufficient robustness to entity name variations, particularly on cross-sentence relation instances and documents with more entities. Finally, we propose an entity variation robust training method which not only improves the robustness of DocRE models but also enhances their understanding and reasoning capabilities. We further verify that the basic idea of this method can be effectively transferred to in-context learning for DocRE as well.

The performance of modern language models (LMs) has been improved by chain-of-thought (CoT) reasoning, i.e., the process of generating intermediate results that guide the model towards a final answer. A possible explanation for this improvement is that CoT reasoning extends an LM’s computational power, as RNNs and transformers with additional scratch space are known to be Turing complete. Comparing LMs to Turing machines, however, introduces a category error—Turing machines decide language membership, whereas LMs define distributions over strings. To bridge this gap, we formalize CoT reasoning in a probabilistic setting. We present several results on the representational capacity of recurrent and transformer LMs with CoT reasoning, showing that they can represent the same family of distributions over strings as probabilistic Turing machines.

We examine the assumption that hidden-state vectors of recurrent neural networks (RNNs) tend to form clusters of semantically similar vectors, which we dub the clustering hypothesis. While this hypothesis has been assumed in RNN analyses in recent years, its validity has not been studied thoroughly on modern RNN architectures. We first consider RNNs that were trained to recognize regular languages. This enables us to draw on perfect ground-truth automata in our evaluation, against which we can compare the RNN’s accuracy and the distribution of the hidden-state vectors. Then, we consider context-free languages to examine if RNN states form clusters for more expressive languages.For our analysis, we fit (generalized) linear models to classify RNN states into automata states and we apply different unsupervised clustering techniques. With a new ambiguity score, derived from information entropy, we measure how well an abstraction function maps the hidden state vectors to abstract clusters. Our evaluation supports the validity of the clustering hypothesis for regular languages, especially if RNNs are well-trained, i.e., clustering techniques succeed in finding clusters of similar state vectors. However, the clustering accuracy decreases substantially for context-free languages. This suggests that clustering is not a reliable abstraction technique for RNNs used in tasks like natural language processing.

Web agents powered by Large Language Models (LLMs) have demonstrated remarkable abilities in planning and executing multi-step interactions within complex web-based environments, fulfilling a wide range of web navigation tasks. Despite these advancements, the potential for LLM-powered agents to effectively engage with sequential user instructions in real-world scenarios has not been fully explored. In this work, we introduce a new task of Conversational Web Navigation, which necessitates sophisticated interactions that span multiple turns with both the users and the environment, supported by a specially developed dataset named Multi-Turn Mind2Web (MT-Mind2Web). To tackle the limited context length of LLMs and the context-dependency issue of the conversational tasks, we further propose a novel framework, named self-reflective memory-augmented planning (Self-MAP), which employs memory utilization and self-reflection techniques. Extensive experiments are conducted to benchmark the MT-Mind2Web dataset, and validate the effectiveness of the proposed method.

Contrastive cross-modal models such as CLIP and CLAP aid various vision-language (VL) and audio-language (AL) tasks. However, there has been limited investigation of and improvement in their language encoder – the central component of encoding natural language descriptions of image/audio into vector representations. We extensively evaluate how unsupervised and supervised sentence embedding training affect language encoder quality and cross-modal task performance. In VL pretraining, we found that sentence embedding training enhances language encoder quality and aids in cross-modal tasks, improving contrastive VL models such as CyCLIP. Sentence embedding training benefits AL tasks when the amount of training data is large. We analyze the representation spaces to understand the strengths of sentence embedding training, and find that it improves text-space uniformity, at the cost of decreased cross-modal alignment.

Quantization and pruning form the foundation of compression for neural networks, enabling efficient inference for large language models (LLMs). Recently, various quantization and pruning techniques have demonstrated remarkable performance in a post-training setting. They rely upon calibration data, a small set of unlabeled examples that are used to generate layer activations. However, no prior work has systematically investigated how the calibration data impacts the effectiveness of model compression methods. In this paper, we present the first extensive empirical study on the effect of calibration data upon LLM performance. We trial a variety of quantization and pruning methods, datasets, tasks, and models. Surprisingly, we find substantial variations in downstream task performance, contrasting existing work that suggests a greater level of robustness to the calibration data. Finally, we make a series of recommendations for the effective use of calibration data in LLM quantization and pruning.

It is widely known that hallucination is a critical issue in Simultaneous Machine Translation (SiMT) due to the absence of source-side information. While many efforts have been made to enhance performance for SiMT, few of them attempt to understand and analyze hallucination in SiMT.Therefore, we conduct a comprehensive analysis of hallucination in SiMT from two perspectives: understanding the distribution of hallucination words and the target-side context usage of them.Intensive experiments demonstrate some valuable findings and particularly show that it is possible to alleviate hallucination by decreasing the over usage of target-side information for SiMT.

The Spoken Language Understanding Evaluation (SLUE) suite of benchmark tasks was recently introduced to address the need for openresources and benchmarking of complex spoken language understanding (SLU) tasks, including both classification and sequence generation tasks, on natural speech. The benchmark has demonstrated preliminary success in using pre-trained speech foundation models (SFM) for these SLU tasks. However, the community still lacks a fine-grained understanding of the comparative utility of different SFMs. Inspired by this, we ask: which SFMs offer the most benefits for these complex SLU tasks, and what is the most effective approach for incorporating these SFMs? To answer this, we perform an extensive evaluation of multiple supervised and self-supervised SFMs using several evaluation protocols: (i) frozen SFMs with a lightweight prediction head, (ii) frozen SFMs with a complex prediction head, and (iii) fine-tuned SFMs with a lightweight prediction head. Although the supervised SFMs are pre-trained on much more speech recognition data (with labels), they do not always outperform self-supervised SFMs; the latter tend to perform at least as well as, and sometimes better than, supervised SFMs, especially on the sequence generation tasks in SLUE. While there is no universally optimal way of incorporating SFMs, the complex prediction head gives the best performance for most tasks, although it increases the inference time. We also introduce an open-source toolkit and performance leaderboard, SLUE-PERB, for these tasks and modeling strategies.