EMNLP 2024 Papers

Evaluating Large Language Models (LLMs) on reasoning benchmarks demonstrates their ability to solve compositional questions. However, little is known of whether these models engage in genuine logical reasoning or simply rely on implicit cues to generate answers. In this paper, we investigate the transitive reasoning capabilities of two distinct LLM architectures, LLaMA 2 and Flan-T5, by manipulating facts within two compositional datasets: QASC and Bamboogle. We controlled for potential cues that might influence the models’ performance, including (a) word/phrase overlaps across sections of test input; (b) models’ inherent knowledge during pre-training or fine-tuning; and (c) Named Entities. Our findings reveal that while both models leverage (a), Flan-T5 shows more resilience to experiments (b and c), having less variance than LLaMA 2. This suggests that models may develop an understanding of transitivity through fine-tuning on knowingly relevant datasets, a hypothesis we leave to future work.

A diverse array of reasoning strategies has been proposed to elicit the capabilities of large language models. However, in this paper, we point out that traditional evaluations which focus solely on performance metrics miss a key factor: the increased effectiveness due to additional compute. By overlooking this aspect, a skewed view of strategy efficiency is often presented. This paper introduces a framework that incorporates the compute budget into the evaluation, providing a more informative comparison that takes into account both performance metrics and computational cost. In this budget-aware perspective, we find that complex reasoning strategies often don’t surpass simpler baselines purely due to algorithmic ingenuity, but rather due to the larger computational resources allocated. When we provide a simple baseline like chain-of-thought self-consistency with comparable compute resources, it frequently outperforms reasoning strategies proposed in the literature. In this scale-aware perspective, we find that unlike self-consistency, certain strategies such as multi-agent debate or Reflexion can become worse if more compute budget is utilized.

Despite recent advances demonstrating vision- language models’ (VLMs) abilities to describe complex relationships among objects in images using natural language, their capability to quantitatively reason about object sizes and distances remains underexplored. In this work, we introduce a manually annotated benchmark of 241 questions across five categories specifically designed for quantitative spatial reasoning, and systematically investigate the performance of SoTA VLMs on this task. Our analysis reveals that questions involving reasoning about distances between objects are particularly challenging for SoTA VLMs; however, some VLMs perform significantly better at this task than others, with an almost 40 points gap between the two best performing models. We also make the surprising observation that the success rate of the top-performing VLM increases by 19 points when a reasoning path using a reference object emerges naturally in the response. Inspired by this observation, we develop a zero-shot prompting technique, SpatialPrompt, that encourages VLMs to answer quantitative spatial questions using references objects as visual cues. Specifically, we demonstrate that instruct- ing VLMs to use reference objects in their reasoning paths significantly improves their quantitative spatial reasoning performance, bypassing the need for external data, architectural modifications, or fine-tuning. Remarkably, by solely using SpatialPrompt, Gemini 1.5 Pro, GPT-4V, and GPT-4o improve by 56.2, 28.5, and 6.7 points on average in Q-Spatial Bench without the need for more data, model architectural modifications, or fine-tuning.

Advanced models such as OpenAI o1 exhibit impressive problem-solving capabilities through step-by-step reasoning. However, they may still falter on more complex problems, making errors that disrupt their reasoning paths. We attribute this to the expansive solution space, where each step has the risk of diverging into mistakes. To enhance language model reasoning, we introduce a specialized training framework called Reasoning Paths Optimization (RPO), which enables learning to reason and explore from diverse paths. Our approach encourages favorable branches at each reasoning step while penalizing unfavorable ones, enhancing the model’s overall problem-solving performance. Reasoning Paths Optimization does not rely on large-scale human-annotated rationales or outputs from closed-source models, making it scalable and data-efficient. We focus on multi-step reasoning tasks, such as math word problems and science-based exam questions. The experiments demonstrate that our framework significantly enhances the reasoning performance of large language models, with up to 3.1% and 4.3% improvement on GSM8K and MMLU (STEM) respectively. Our data and code can be found at https://reasoning-paths.github.io.

Media storms, dramatic outbursts of attention to a story, are central components of media dynamics and the attention landscape. Despite their importance, there has been little systematic and empirical research on this concept due to issues of measurement and operationalization. We introduce an iterative human-in-the-loop method to identify media storms in a large-scale corpus of news articles. The text is first transformed into signals of dispersion based on several textual characteristics. In each iteration, we apply unsupervised anomaly detection to these signals; each anomaly is then validated by an expert to confirm the presence of a storm, and those results are then used to tune the anomaly detection in the next iteration. We make available the resulting media storm dataset. Both the method and dataset provide a basis for comprehensive empirical study of media storms.

The latest advancements in large language models (LLMs) have sparked interest in their potential for software vulnerability detection. However, there is currently a lack of research specifically focused on vulnerabilities in the PHP language, and challenges in data sampling and processing persist, hindering the model’s ability to effectively capture the characteristics of specific vulnerabilities. In this paper, we present RealVul, the first LLM-based framework designed for PHP vulnerability detection, addressing these issues. By improving code sampling methods and employing normalization techniques, we can isolate potential vulnerability triggers while streamlining the code and eliminating unnecessary semantic information, enabling the model to better understand and learn from the generated vulnerability samples. We also address the issue of insufficient PHP vulnerability samples by improving data synthesis methods. To evaluate RealVul’s performance, we conduct an extensive analysis using five distinct code LLMs on vulnerability data from 180 PHP projects. The results demonstrate a significant improvement in both effectiveness and generalization compared to existing methods, effectively boosting the vulnerability detection capabilities of these models.

We present a comprehensive evaluation of large language models for multilingual readability assessment. Existing evaluation resources lack domain and language diversity, limiting the ability for cross-domain and cross-lingual analyses. This paper introduces ReadMe++, a multilingual multi-domain dataset with human annotations of 9757 sentences in Arabic, English, French, Hindi, and Russian, collected from 112 different data sources. This benchmark will encourage research on developing robust multilingual readability assessment methods. Using ReadMe++, we benchmark multilingual and monolingual language models in the supervised, unsupervised, and few-shot prompting settings. The domain and language diversity in ReadMe++ enable us to test more effective few-shot prompting, and identify shortcomings in state-of-the-art unsupervised methods. Our experiments also reveal exciting results of superior domain generalization and enhanced cross-lingual transfer capabilities by models trained on ReadMe++. We will make our data publicly available and release a python package tool for multilingual sentence readability prediction using our trained models at: https://github.com/tareknaous/readme

Graphical User Interfaces (GUIs) are central to our interaction with digital devices and growing efforts have been made to build models for various GUI understanding tasks. However, these efforts largely overlook an important GUI-referring task: screen reading based on user-indicated points, which we name the Screen Point-and-Read (ScreenPR) task. Currently, this task is predominantly handled by rigid accessible screen reading tools, in great need of new models driven by advancements in Multimodal Large Language Models (MLLMs). In this paper, we propose a Tree-of-Lens (ToL) agent, utilizing a novel ToL grounding mechanism, to address the ScreenPR task. Based on the input point coordinate and the corresponding GUI screenshot, our ToL agent constructs a Hierarchical Layout Tree. Based on the tree, our ToL agent not only comprehends the content of the indicated area but also articulates the layout and spatial relationships between elements. Such layout information is crucial for accurately interpreting information on the screen, distinguishing our ToL agent from other screen reading tools. We also thoroughly evaluate the ToL agent against other baselines on a newly proposed ScreenPR benchmark, which includes GUIs from mobile, web, and operating systems. Last but not least, we test the ToL agent on mobile GUI navigation tasks, demonstrating its utility in identifying incorrect actions along the path of agent execution trajectories. Code and data: https://screen-point-and-read.github.io.

Recently, there has been increasing interest in using Large Language Models (LLMs) to construct complex multi-agent systems to perform tasks such as compiling literature reviews, drafting consumer reports, and planning vacations. Many tools and libraries exist for helping create such systems, however none support *recursive* multi-agent systems—where the models themselves flexibly decide when to delegate tasks and how to organize their delegation structure. In this work, we introduce ReDel: a toolkit for recursive multi-agent systems that supports custom tool-use, delegation schemes, event-based logging, and interactive replay in an easy-to-use web interface. We show that, using ReDel, we are able to achieve significant performance gains on agentic benchmarks and easily identify potential areas of improvements through the visualization and debugging tools. Our code, documentation, and PyPI package are open-source at https://github.com/zhudotexe/redel, and free to use under the MIT license.

The rapid scaling of large language models (LLMs) has raised concerns about the transparency and fair use of the data used in their pretraining. Detecting such content is challenging due to the scale of the data and limited exposure of each instance during training. We propose ReCaLL (Relative Conditional Log-Likelihood), a novel membership inference attack (MIA) to detect LLMs’ pretraining data by leveraging their conditional language modeling capabilities. ReCaLL examines the relative change in conditional log-likelihoods when prefixing target data points with non-member context. Our empirical findings show that conditioning member data on non-member prefixes induces a larger decrease in log-likelihood compared to non-member data. We conduct comprehensive experiments and show that ReCaLL achieves state-of-the-art performance on the WikiMIA dataset, even with random and synthetic prefixes, and can be further improved using an ensemble approach. Moreover, we conduct an in-depth analysis of LLMs’ behavior with different membership contexts, providing insights into how LLMs leverage membership information for effective inference at both the sequence and token level.

Gender-Based Violence (GBV) is an increasing problem online, but existing datasets fail to capture the plurality of possible annotator perspectives or ensure the representation of affected groups. We revisit two important stages in the moderation pipeline for GBV: (1) manual data labelling; and (2) automated classification. For (1), we examine two datasets to investigate the relationship between annotator identities and attitudes and the responses they give to two GBV labelling tasks. To this end, we collect demographic and attitudinal information from crowd-sourced annotators using three validated surveys from Social Psychology. We find that higher Right Wing Authoritarianism scores are associated with a higher propensity to label text as sexist, while for Social Dominance Orientation and Neosexist Attitudes, higher scores are associated with a negative tendency to do so.For (2), we conduct classification experiments using Large Language Models and five prompting strategies, including infusing prompts with annotator information. We find: (i) annotator attitudes affect the ability of classifiers to predict their labels; (ii) including attitudinal information can boost performance when we use well-structured brief annotator descriptions; and (iii) models struggle to reflect the increased complexity and imbalanced classes of the new label sets.

To enhance the reasoning capabilities of off-the-shelf Large Language Models (LLMs), we introduce a simple, yet general and effective prompting method, RE2, i.e., Re-Reading the question as input. Unlike most thought-eliciting prompting methods, such as Chain-of-Thought (CoT), which aim to elicit the reasoning process in the output, RE2 shifts the focus to the input by processing questions twice, thereby enhancing the understanding process. Consequently, RE2 demonstrates strong generality and compatibility with most thought-eliciting prompting methods, including CoT. Crucially, RE2 facilitates a “bidirectional” encoding in unidirectional decoder-only LLMs because the first pass could provide global information for the second pass. We begin with a preliminary empirical study as the foundation of RE2, illustrating its potential to enable “bidirectional” attention mechanisms. We then evaluate RE2 on extensive reasoning benchmarks across 14 datasets, spanning 112 experiments, to validate its effectiveness and generality. Our findings indicate that, with the exception of a few scenarios on vanilla ChatGPT, RE2 consistently enhances the reasoning performance of LLMs through a simple re-reading strategy. Further analyses reveal RE2’s adaptability, showing how it can be effectively integrated with different LLMs, thought-eliciting prompting, and ensemble strategies.

Finetuning language agents with reasoning-action trajectories is effective, but obtaining these trajectories from human annotations or stronger models is costly and sometimes impractical. In this paper, we investigate the use of self-training in language agents, which can generate supervision from the agent itself, offering a promising alternative without relying on human or stronger model demonstrations. Self-training, however, requires high-quality model-generated samples, which are hard to obtain for challenging language agent tasks. To address this, we present Reflection-Reinforced Self-Training (Re-ReST), which uses a reflector to refine low-quality generated samples during self-training. The reflector takes the agent’s output and feedback from an external environment (e.g., unit test results in code generation) to produce improved samples. This technique enhances the quality of inferior samples and efficiently enriches the self-training dataset with higher-quality samples. We conduct extensive experiments on open-source language agents across tasks, including multi-hop question answering, sequential decision-making, code generation, visual question answering, and text-to-image generation. The results demonstrate the effectiveness of self-training and Re-ReST in language agent tasks, with self-training improving baselines by 7.6% on HotpotQA and 28.4% on AlfWorld, and Re-ReST further boosting performance by 2.0% and 14.1%, respectively. Our studies also confirm the efficiency of using a reflector to generate high-quality samples for self-training. Moreover, we demonstrate a method to employ reflection during inference without ground-truth feedback, addressing the limitation of previous reflection work.

Recent advances in large language models (LLMs) have enabled autonomous agents with complex reasoning and task-fulfillment capabilities using a wide range of tools. However, effectively identifying the most relevant tools for a given task becomes a key bottleneck as the toolset size grows, hindering reliable tool utilization. To address this, we introduce Re-Invoke, an unsupervised tool retrieval method designed to scale effectively to large toolsets without training. Specifically, we first generate a diverse set of synthetic queries that comprehensively cover different aspects of the query space associated with each tool document during the tool indexing phase. Second, we leverage LLM’s query understanding capabilities to extract key tool-related context and underlying intents from user queries during the inference phase. Finally, we employ a novel multi-view similarity ranking strategy based on intents to pinpoint the most relevant tools for each query. Our evaluation demonstrates that Re-Invoke significantly outperforms state-of-the-art alternatives in both single-tool and multi-tool scenarios, all within a fully unsupervised setting. Notably, on the ToolE datasets, we achieve a 20% relative improvement in nDCG@5 for single-tool retrieval and a 39% improvement for multi-tool retrieval.

Automatic evaluation approaches (ROUGE, BERTScore, LLM-based evaluators) have been widely used to evaluate summarization tasks. Despite the complexities of script differences and tokenization, these approaches have been indiscriminately applied to summarization across multiple languages. While previous works have argued that these approaches correlate strongly with human ratings in English, it remains unclear whether the conclusion holds for other languages. To answer this question, we construct a small-scale pilot dataset containing article-summary pairs and human ratings in English, Chinese and Indonesian. To measure the strength of summaries, our ratings are measured as head-to-head comparisons with resulting Elo scores across four dimensions. Our analysis reveals that standard metrics are unreliable measures of quality, and that these problems are exacerbated in Chinese and Indonesian. We advocate for more nuanced and careful considerations in designing a robust evaluation framework for multiple languages.

We propose an end-to-end differentiable training paradigm for stable training of a rationalized transformer classifier. Our approach results in a single model that simultaneously classifies a sample and scores input tokens based on their relevance to the classification. To this end, we build on the widely-used three-player-game for training rationalized models, which typically relies on training a rationale selector, a classifier and a complement classifier. We simplify this approach by making a single model fulfill all three roles, leading to a more efficient training paradigm that is not susceptible to the common training instabilities that plague existing approaches. Further, we extend this paradigm to produce class-wise rationales while incorporating recent advances in parameterizing and regularizing the resulting rationales, thus leading to substantially improved and state-of-the-art alignment with human annotations without any explicit supervision.

Answer verification identifies correct solutions among candidates generated by large language models (LLMs). Current approaches typically train verifier models by labeling solutions as correct or incorrect based solely on whether the final answer matches the gold answer. However, this approach neglects any flawed rationale in the solution yielding the correct answer, undermining the verifier’s ability to distinguish between sound and flawed rationales. We empirically show that in StrategyQA, only 19% of LLM-generated solutions with correct answers have valid rationales, thus leading to an unreliable verifier. Furthermore, we demonstrate that training a verifier on valid rationales significantly improves its ability to distinguish valid and flawed rationale. To make a better verifier without extra human supervision, we introduce REPS (Rationale Enhancement through Pairwise Selection), a method for selecting valid rationales from candidates by iteratively applying pairwise self-evaluation using the same LLM that generates the solutions. Verifiers trained on solutions selected by REPS outperform those trained using conventional training methods on three reasoning benchmarks (ARC-Challenge, DROP, and StrategyQA). Our results suggest that training reliable verifiers requires ensuring the validity of rationales in addition to the correctness of the final answers, which would be critical for models assisting humans in solving complex reasoning tasks.

Human feedback is essential for building human-centered AI systems across domains where disagreement is prevalent, such as AI safety, content moderation, or sentiment analysis. Many disagreements, particularly in politically charged settings, arise because raters have opposing values or beliefs. Vicarious annotation is a method for breaking down disagreement by asking raters how they think others would annotate the data. In this paper, we explore the use of vicarious annotation with analytical methods for moderating rater disagreement. We employ rater cohesion metrics to study the potential influence of political affiliations and demographic backgrounds on raters’ perceptions of offense. Additionally, we utilize CrowdTruth’s rater quality metrics, which consider the demographics of the raters, to score the raters and their annotations. We study how the rater quality metrics influence the in-group and cross-group rater cohesion across the personal and vicarious levels.

Major search engine providers are rapidly incorporating Large Language Model (LLM)-generated content in response to user queries. These *conversational search engines* operate by loading retrieved website text into the LLM context for summarization and interpretation. Recent research demonstrates that LLMs are highly vulnerable to jailbreaking and prompt injection attacks, which disrupt the safety and quality goals of LLMs using adversarial strings. This work investigates the impact of prompt injections on the ranking order of sources referenced by conversational search engines. To this end, we introduce a focused dataset of real-world consumer product websites and formalize conversational search ranking as an adversarial problem. Experimentally, we analyze conversational search rankings in the absence of adversarial injections and show that different LLMs vary significantly in prioritizing product name, document content, and context position. We then present a tree-of-attacks-based jailbreaking technique which reliably promotes low-ranked products. Importantly, these attacks transfer effectively to state-of-the-art conversational search engines such as *perplexity.ai*. Given the strong financial incentive for website owners to boost their search ranking, we argue that our problem formulation is of critical importance for future robustness work.