We look at reasoning on GSM8k, a dataset of short texts presenting primary school, math problems. We find, with Mirzadeh et al. (2024), that current LLM progress on the data set may not be explained by better reasoning but by exposure to a broader pretraining data distribution. We then introduce a novel information source for helping models with less data or inferior training reason better: discourse structure. We show that discourse structure improves performance for models like Llama2 13b by up to 160%. Even for models that have most likely memorized the data set, adding discourse structural information to the model still improves predictions and dramatically improves large model performance on out of distribution examples.

We examine two in context learning (ICL) tasks with mathematical functions in several train and test settings for transformer models. Our study generalizes work on linear functions by showing that small transformers, even models with attention layers only, can approximate arbitrary polynomial functions and hence continuous functions under certain conditions. Our models also can approximate previously unseen classes of polynomial functions, as well as the zeros of complex functions. Our models perform far better on this task than LLMs like GPT4 and involve complex reasoning when provided with suitable training data and methods. Our models also have important limitations; they fail to generalize outside of training distributions and so don't learn class forms of functions. We explain why this is so.

In-context learning (ICL) has emerged as a powerful paradigm for easily adapting Large Language Models (LLMs) to various tasks. However, our understanding of how ICL works remains limited. We explore a simple model of ICL in a controlled setup with synthetic training data to investigate ICL of univariate linear functions. We experiment with a range of GPT-2-like transformer models trained from scratch. Our findings challenge the prevailing narrative that transformers adopt algorithmic approaches like linear regression to learn a linear function in-context. These models fail to generalize beyond their training distribution, highlighting fundamental limitations in their capacity to infer abstract task structures. Our experiments lead us to propose a mathematically precise hypothesis of what the model might be learning.

This paper explores the much discussed, possible explanatory link between attention weights (AW) in transformer models and predicted output. Contrary to intuition and early research on attention, more recent prior research has provided formal arguments and empirical evidence that AW are not explanatorily relevant. We show that the formal arguments are incorrect. We introduce and effectively compute efficient attention, which isolates the effective components of attention matrices in tasks and models in which AW play an explanatory role. We show that efficient attention has a causal role (provides minimally necessary and sufficient conditions) for predicting model output in NLP tasks requiring contextual information, and we show, contrary to [7], that efficient attention matrices are probability distributions and are effectively calculable. Thus, they should play an important part in the explanation of attention based model behavior. We offer empirical experiments in support of our method illustrating various properties of efficient attention with various metrics on four datasets.

In this paper, we study whether transformer-based language models can extract predicate argument structure from simple sentences. We firstly show that language models sometimes confuse which predicates apply to which objects. To mitigate this, we explore two tasks: question answering (Q/A), and first order logic (FOL) translation, and two regimes, prompting and finetuning. In FOL translation, we finetune several large language models on synthetic datasets designed to gauge their generalization abilities. For Q/A, we finetune encoder models like BERT and RoBERTa and use prompting for LLMs. The results show that FOL translation for LLMs is better suited to learn predicate argument structure.

When engaging in collaborative tasks, humans efficiently exploit the semantic structure of a conversation to optimize verbal and nonverbal interactions. But in recent "language to code" or "language to action" models, this information is lacking. We show how incorporating the prior discourse and nonlinguistic context of a conversation situated in a nonlinguistic environment can improve the "language to action" component of such interactions. We fine tune an LLM to predict actions based on prior context; our model, NeBuLa, doubles the net-action F1 score over the baseline on this task of Jayannavar et al.(2020). We also investigate our model's ability to construct shapes and understand location descriptions using a synthetic dataset.

This paper provides the first discourse parsing experiments with a large language model (LLM) finetuned on corpora annotated in the style of SDRT (Asher, 1993; Asher and Lascarides, 2003). The result is a discourse parser, LLaMIPa (LLaMA Incremental Parser), which is able to more fully exploit discourse context, leading to substantial performance gains over approaches that use encoder-only models to provide local, context-sensitive representations of discourse units. Furthermore, it is able to process discourse data incrementally, which is essential for the eventual use of discourse information in downstream tasks.

A central goal of eXplainable Artificial Intelligence (XAI) is to assign relative importance to the features of a Machine Learning (ML) model given some prediction. The importance of this task of explainability by feature attribution is illustrated by the ubiquitous recent use of tools such as SHAP and LIME. Unfortunately, the exact computation of feature attributions, using the game-theoretical foundation underlying SHAP and LIME, can yield manifestly unsatisfactory results, that tantamount to reporting misleading relative feature importance. Recent work targeted rigorous feature attribution, by studying axiomatic aggregations of features based on logic-based definitions of explanations by feature selection. This paper shows that there is an essential relationship between feature attribution and a priori voting power, and that those recently proposed axiomatic aggregations represent a few instantiations of the range of power indices studied in the past. Furthermore, it remains unclear how some of the most widely used power indices might be exploited as feature importance scores (FISs), i.e. the use of power indices in XAI, and which of these indices would be the best suited for the purposes of XAI by feature attribution, namely in terms of not producing results that could be deemed as unsatisfactory. This paper proposes novel desirable properties that FISs should exhibit. In addition, the paper also proposes novel FISs exhibiting the proposed properties. Finally, the paper conducts a rigorous analysis of the best-known power indices in terms of the proposed properties.

Previous studies have shown that it is possible to map brain activation data of subjects viewing images onto the feature representation space of not only vision models (modality-specific decoding) but also language models (cross-modal decoding). In this work, we introduce and use a new large-scale fMRI dataset (~8,500 trials per subject) of people watching both images and text descriptions of such images. This novel dataset enables the development of modality-agnostic decoders: a single decoder that can predict which stimulus a subject is seeing, irrespective of the modality (image or text) in which the stimulus is presented. We train and evaluate such decoders to map brain signals onto stimulus representations from a large range of publicly available vision, language and multimodal (vision+language) models. Our findings reveal that (1) modality-agnostic decoders perform as well as (and sometimes even better than) modality-specific decoders (2) modality-agnostic decoders mapping brain data onto representations from unimodal models perform as well as decoders relying on multimodal representations (3) while language and low-level visual (occipital) brain regions are best at decoding text and image stimuli, respectively, high-level visual (temporal) regions perform well on both stimulus types.

Despite great performance on many tasks, language models (LMs) still struggle with reasoning, sometimes providing responses that cannot possibly be true because they stem from logical incoherence. We call such responses \textit{strong hallucinations} and prove that they follow from an LM's computation of its internal representations for logical operators and outputs from those representations. Focusing on negation, we provide a novel solution in which negation is treated not as another element of a latent representation, but as \textit{an operation over an LM's latent representations that constrains how they may evolve}. We show that our approach improves model performance in cloze prompting and natural language inference tasks with negation without requiring training on sparse negative data.