United States

We find arithmetic ability resides within a limited number of attention heads, with each head specializing in distinct operations. To delve into the reason, we introduce the Comparative Neuron Analysis (CNA) method, which identifies an internal logic chain consisting of four distinct stages from input to prediction: feature enhancing with shallow FFN neurons, feature transferring by shallow attention layers, feature predicting by arithmetic heads, and prediction enhancing among deep FFN neurons. Moreover, we identify the human-interpretable FFN neurons within both feature-enhancing and feature-predicting stages. These findings lead us to investigate the mechanism of LoRA, revealing that it enhances prediction probabilities by amplifying the coefficient scores of FFN neurons related to predictions. Finally, we apply our method in model pruning for arithmetic tasks and model editing for reducing gender bias. Code is on https://github.com/zepingyu0512/arithmetic-mechanism.

EMNLP 2024

Interpreting Arithmetic Mechanism in Large Language Models through Comparative Neuron Analysis

contextual integrity

privacy laws

large language models

## Welcome to EMNLP 2024! 
We are excited to welcome you to one of the most prominent conferences in the field of Natural Language Processing. This year, EMNLP 2024 is being held in a hybrid format,
offering both virtual and in-person participation in beautiful Miami. Due to a record-breaking number of submissions, we've expanded the total number of accepted papers to accommodate more cutting-edge research from around the globe.
### [Conference Handbook](https://drive.google.com/file/d/1WPROgxjLAC96AJL7Ugy0tEnYm7dkrbHt/view?usp=sharing)

You are required to register for this event. **Please register [here](https://2024.emnlp.org/registration/).** The EMNLP 2024 event page on Underline will be open to public one week prior to the event.

Please register!

EMNLP 2024 will take place in Miami, Florida from Nov 12th to Nov 16th, 2024, at the Hyatt Regency Miami Hote and on Underline for remote participants.

technical paper

Pixel-based language models have emerged as a compelling alternative to subword-based language modelling, particularly because they can represent virtually any script. PIXEL, a canonical example of such a model, is a vision transformer that has been pre-trained on rendered text. While PIXEL has shown promising cross-script transfer abilities and robustness to orthographic perturbations, it falls short of outperforming monolingual subword counterparts like BERT in most other contexts. This discrepancy raises questions about the amount of linguistic knowledge learnt by these models and whether their performance in language tasks stems more from their visual capabilities than their linguistic ones. To explore this, we probe PIXEL using a variety of linguistic and visual tasks to assess its position on the vision-to-language spectrum. Our findings reveal a substantial gap between the model’s visual and linguistic understanding. The lower layers of PIXEL predominantly capture superficial visual features, whereas the higher layers gradually learn more syntactic and semantic abstractions. Additionally, we examine variants of PIXEL trained with different text rendering strategies, discovering that introducing certain orthographic constraints at the input level can facilitate earlier learning of surface-level features. With this study, we hope to provide insights that aid the further development of pixel-based language models.

Pixology: Probing the Linguistic and Visual Capabilities of Pixel-based Language Models

Large Language Models (LLMs) are typically shipped with tokenizers that *deterministically* encode text into so-called *canonical* token sequences, to which the LLMs assign probability values.
One common assumption is that the probability of a piece of text is the probability of its canonical token sequence.
However, the tokenization of a string is not unique: e.g., the Llama2 tokenizer encodes `Tokens` as `[Tok,ens]`, but `[Tok,en,s]` also represents the same text.
In this paper, we study non-canonical tokenizations.
We prove that, given a string, it is computationally hard to find the most likely tokenization for an autoregressive LLM, as well as to compute the marginal probability over all possible tokenizations.
We then show how the marginal is, in most cases, indistinguishable from the canonical probability.
Surprisingly, we then empirically demonstrate the existence of a significant amount of signal hidden within tokenization space.
Notably, by simply aggregating the probabilities of non-canonical tokenizations, we achieve improvements across a range of LLM evaluation benchmarks for a variety of architectures, including transformers and state space models.

Where is the signal in tokenization space?

Autoformalization is the task of automatically translating mathematical content written in natural language to a formal language expression. The growing language interpretation capabilities of Large Language Models (LLMs), including in formal languages, are lowering the barriers for autoformalization. However, LLMs alone are not capable of consistently and reliably delivering autoformalization, in particular as the complexity and specialization of the target domain grows. As the field evolves into the direction of systematically applying autoformalization towards large mathematical libraries, the need to improve syntactic, terminological and semantic control increases. This paper proposes the coordinated use of three mechanisms, most-similar retrieval augmented generation (MS-RAG), denoising steps, and auto-correction with syntax error feedback (Auto-SEF) to improve autoformalization quality. The empirical analysis, across different models, demonstrates that these mechanisms can deliver autoformalizaton results which are syntactically, terminologically and semantically more consistent. These mechanisms can be applied across different LLMs and have shown to deliver improve results across different model types.

Consistent Autoformalization for Constructing Mathematical Libraries

Claim: This work is not advocating the use of LLMs for paper (meta-)reviewing. Instead, we
present a comparative analysis to identify and distinguish LLM activities from human activities. Two research goals: i) Enable better recognition of instances when someone implicitly uses LLMs for reviewing activities; ii) Increase community awareness that LLMs, and AI in general, are currently inadequate for performing tasks that require a high level of expertise and nuanced judgment.

This work is motivated by two key trends. On one hand, large language models (LLMs) have shown remarkable versatility in various generative tasks such as writing, drawing, and question answering, significantly reducing the time required for many routine tasks. On the other hand, researchers, whose work is not only time-consuming but also highly expertise-demanding, face increasing challenges as they have to spend more time reading, writing, and reviewing papers. This raises the question: how can LLMs potentially assist researchers in alleviating their heavy workload?

This study focuses on the topic of LLMs as NLP Researchers, particularly examining the effectiveness of LLMs in assisting paper (meta-)reviewing and its recognizability. To address this, we constructed the ReviewCritique dataset, which includes two types of information: (i) NLP papers (initial submissions rather than camera-ready) with both human-written and LLM-generated reviews, and (ii) each review comes with "deficiency" labels and corresponding explanations for individual segments, annotated by experts. Using ReviewCritique, this study explores two threads of research questions: (i) "LLMs as Reviewers", how do reviews generated by LLMs compare with those written by humans in terms of quality and distinguishability? (ii) "LLMs as Metareviewers", how effectively can LLMs identify potential issues, such as Deficient or unprofessional review segments, within individual paper reviews? To our knowledge, this is the first work to provide such a comprehensive analysis.

LLMs Assist NLP Researchers: Critique Paper (Meta-)Reviewing

Recent advances in machine learning have significantly impacted the field of information extraction, with Language Models (LMs) playing a pivotal role in extracting structured information from unstructured text. Prior works typically represent information extraction as triplet-centric and use classical metrics such as precision and recall for evaluation. We reformulate the task to be entity-centric, enabling the use of diverse metrics that can provide more insights from various perspectives. We contribute to the field by introducing Structured Entity Extraction and proposing the Approximate Entity Set OverlaP (AESOP) metric, designed to appropriately assess model performance. Later, we introduce a new Multistage Structured Entity Extraction (MuSEE) model that harnesses the power of LMs for enhanced effectiveness and efficiency by decomposing the extraction task into multiple stages. Quantitative and human side-by-side evaluations confirm that our model outperforms baselines, offering promising directions for future advancements in structured entity extraction. 
Our source code is available at https://github.com/microsoft/Structured-Entity-Extraction.

Learning to Extract Structured Entities Using Language Models

This paper investigates Who's Harry Potter (WHP), a pioneering yet insufficiently understood method for LLM unlearning. We explore it in two steps. First, we introduce a new task of LLM targeted unlearning, where given an unlearning target (e.g., a person) and some unlearning documents, we aim to unlearn only the information about the target, rather than everything in the unlearning documents. We further argue that a successful unlearning should satisfy criteria such as not outputting gibberish, not fabricating facts about the unlearning target, and not releasing factual information under jailbreak attacks. Second, we construct a causal intervention framework for targeted unlearning, where the knowledge of the unlearning target is modeled as a confounder between LLM input and output, and the unlearning process as a deconfounding process. This framework justifies and extends WHP, deriving a simple unlearning algorithm that includes WHP as a special case. Experiments on existing and new datasets show that our approach, without explicitly optimizing for the aforementioned criteria, achieves competitive performance in all of them.

Revisiting Who’s Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective

The task of condensing large chunks of textual information into concise and structured tables has gained attention recently due to the emergence of Large Language Models (LLMs) and their potential benefit for downstream tasks, such as text summarization and text mining. Previous approaches often generate tables that directly replicate information from the text, limiting their applicability in broader contexts, as text-to-table generation in real-life scenarios necessitates information extraction, reasoning, and integration. However, there is a lack of both datasets and methodologies towards this task. In this paper, we introduce LiveSum, a new benchmark dataset created for generating summary tables of competitions based on real-time commentary texts. We evaluate the performances of state-of-the-art LLMs on this task in both fine-tuning and zero-shot settings, and additionally propose a novel pipeline called $T^3$
(Text-Tuple-Table) to improve their performances. Extensive experimental results demonstrate that LLMs still struggle with this task even after fine-tuning, while our approach can offer substantial performance gains without explicit training. Further analyses demonstrate that our method exhibits strong generalization abilities, surpassing previous approaches on several other text-to-table datasets. Our code
and data can be found at https://github.com/HKUST-KnowComp/LiveSum.

Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction

Entity Linking (EL) is the process of associating ambiguous textual mentions to specific entities in a knowledge base.
Traditional EL methods heavily rely on large datasets to enhance their performance, a dependency that becomes problematic in the context of few-shot entity linking, where only a limited number of examples are available for training. 
To address this challenge, we present OneNet, an innovative framework that utilizes the few-shot learning capabilities of Large Language Models (LLMs) without the need for fine-tuning. To the best of our knowledge, this marks a pioneering approach to applying LLMs to few-shot entity linking tasks. OneNet is structured around three key components prompted by LLMs: (1) an entity reduction processor that simplifies inputs by summarizing and filtering out irrelevant entities, (2) a dual-perspective entity linker that combines contextual cues and prior knowledge for precise entity linking, and (3) an entity consensus judger that employs a unique consistency algorithm to alleviate the hallucination in the entity linking reasoning.
Comprehensive evaluations across seven benchmark datasets reveal that OneNet outperforms current state-of-the-art entity linking methods.

OneNet: A Fine-Tuning Free Framework for Few-Shot Entity Linking via Large Language Model Prompting

We study LLMs' ability to extrapolate the behavior of various dynamical systems, including stochastic, chaotic, continuous, and discrete systems, whose evolution is governed by principles of physical interest. Our results show that LLaMA-2, a language model trained on text, achieves accurate predictions of dynamical system time series without fine-tuning or prompt engineering. Moreover, the accuracy of the learned physical rules increases with the length of the input context window, revealing an in-context version of a neural scaling law. Along the way, we present a flexible and efficient algorithm for extracting probability density functions of multi-digit numbers directly from LLMs.

LLMs learn governing principles of dynamical systems, revealing an in-context neural scaling law

Large Language Models (LLMs) are increasingly relied upon to evaluate text outputs of other LLMs, thereby influencing leaderboards and development decisions. However, concerns persist over the accuracy of these assessments and the potential for misleading conclusions. In this work, we investigate the effectiveness of LLMs as evaluators for text generation tasks. We propose FBI, a novel framework designed to examine the proficiency of Evaluator LLMs in assessing four critical abilities in other LLMs: factual accuracy, instruction following, coherence in long-form writing, and reasoning proficiency. By introducing targeted perturbations in answers generated by LLMs, that clearly impact one of these key capabilities, we test whether an Evaluator LLM can detect these quality drops. By creating a total of 2400 perturbed answers covering 22 perturbation categories, we conduct a comprehensive study using different evaluation strategies on five prominent LLMs commonly used as evaluators in the literature. Our findings reveal significant shortcomings in current Evaluator LLMs, which failed to identify quality drops in over 50% of cases on average. Single-answer and pairwise evaluations demonstrated notable limitations, whereas reference-based evaluations showed comparatively better performance. \textit{These results underscore the unreliable nature of current Evaluator LLMs and advocate for cautious implementation in practical applications.}

Downloads

Next from EMNLP 2024

Pixology: Probing the Linguistic and Visual Capabilities of Pixel-based Language Models

Similar lecture

Neuron-Level Knowledge Attribution in Large Language Models