In order to perform multimodal fusion of heterogeneous signals, we need to understand their interactions: how each modality individually provides information useful for a task and how this information changes in the presence of other modalities. In this paper, we perform a comparative study of how humans annotate two categorizations of multimodal interactions: (1) partial labels, where different annotators annotate the label given the first, second, and both modalities, and (2) counterfactual labels, where the same annotator annotates the label given the first modality before asking them to explicitly reason about how their answer changes when given the second. We further propose an alternative taxonomy based on (3) information decomposition, where annotators annotate the degrees of redundancy: the extent to which modalities individually and together give the same predictions, uniqueness: the extent to which one modality enables a prediction that the other does not, and synergy: the extent to which both modalities enable one to make a prediction that one would not otherwise make using individual modalities. Through experiments and annotations, we highlight several opportunities and limitations of each approach and propose a method to automatically convert annotations of partial and counterfactual labels to information decomposition, yielding an accurate and efficient method for quantifying multimodal interactions.

Language models (LMs) often exhibit undesirable text generation behaviors, including generating false, toxic, or irrelevant outputs. Reinforcement learning from human feedback (RLHF) - where human preference judgments on LM outputs are transformed into a learning signal - has recently shown promise in addressing these issues. However, such holistic feedback conveys limited information on long text outputs; it does not indicate which aspects of the outputs influenced user preference; e.g., which parts contain what type(s) of errors. In this paper, we use fine-grained human feedback (e.g., which sentence is false, which sub-sentence is irrelevant) as an explicit training signal. We introduce Fine-Grained RLHF, a framework that enables training and learning from reward functions that are fine-grained in two respects: (1) density, providing a reward after every segment (e.g., a sentence) is generated; and (2) incorporating multiple reward models associated with different feedback types (e.g., factual incorrectness, irrelevance, and information incompleteness). We conduct experiments on detoxification and long-form question answering to illustrate how learning with such reward functions leads to improved performance, supported by both automatic and human evaluation. Additionally, we show that LM behaviors can be customized using different combinations of fine-grained reward models. We release all data, collected human feedback, and codes at https://FineGrainedRLHF.github.io.

Recent advancements in scale have yielded large language models (LLMs) with extraordinary proficiency in nuanced reasoning with factual knowledge. Despite these achievements, LLMs are known to produce incorrect outputs, often referred to colloquially as "hallucinations" or "distractions" (Ji et al., 2023). Generally, hallucinations refer to the phenomenon that a model's outputs are syntactically and grammatically accurate but factually incorrect. There are various types of hallucinations, and the focus of this work is the "closeddomain" variety (Saparov and He, 2022; OpenAI, 2023), where the model predictions contain factually incorrect or made-up information according to a given context, regardless of their correctness in the real world. Perhaps surprisingly, such hallucinations can be observed even on simple algorithmic reasoning tasks. As a warmup, consider the queries shown in Figure 1 (and Appendix B.1), where we prompt LLMs to solve addition problems of various lengths. The responses simultaneously illustrate the following: 1. Nontrivial algorithmic generalization: In cases where the models succeed, it is unlikely that these exact numerical sequences appeared in the training data. To correctly output the first digit of the answer, the LLM must resolve a long dependency chain which generally depends on every digit in the input. Somewhere within these networks' internal representations, implementations of addition algorithms have emerged.

Large Language Models (LLMs) have shown promising performance in knowledge-intensive reasoning tasks that require a compound understanding of knowledge. However, deployment of the LLMs in real-world applications can be challenging due to their high computational requirements and concerns on data privacy. Previous studies have focused on building task-specific small Language Models (LMs) by fine-tuning them with labeled data or distilling LLMs. However, these approaches are ill-suited for knowledge-intensive reasoning tasks due to the limited capacity of small LMs in memorizing the knowledge required. Motivated by our theoretical analysis on memorization, we propose Knowledge-Augmented Reasoning Distillation (KARD), a novel method that fine-tunes small LMs to generate rationales obtained from LLMs with augmented knowledge retrieved from an external knowledge base. Moreover, we further propose a neural reranker to obtain documents relevant to rationale generation. We empirically show that KARD significantly improves the performance of small T5 and GPT models on the challenging knowledge-intensive reasoning datasets, namely MedQA-USMLE, StrategyQA, and OpenbookQA. Notably, our method makes the 250M T5 models achieve superior performance against the fine-tuned 3B models, having 12 times larger parameters, on both MedQA-USMLE and StrategyQA benchmarks.

Tabular data is prevalent across various industries, necessitating significant time and effort for users to understand and manipulate for their information-seeking purposes. The advancements in large language models (LLMs) have shown enormous potential to improve user efficiency. However, the adoption of LLMs in real-world applications for table information seeking remains underexplored. In this paper, we investigate the table-to-text capabilities of different LLMs using four datasets within two real-world information seeking scenarios. These include the LogicNLG and our newly-constructed LoTNLG datasets for data insight generation, along with the FeTaQA and our newly-constructed F2WTQ datasets for query-based generation. We structure our investigation around three research questions, evaluating the performance of LLMs in table-to-text generation, automated evaluation, and feedback generation, respectively. Experimental results indicate that the current high-performing LLM, specifically GPT-4, can effectively serve as a table-to-text generator, evaluator, and feedback generator, facilitating users' information seeking purposes in real-world scenarios. However, a significant performance gap still exists between other open-sourced LLMs (e.g., Tulu and LLaMA-2) and GPT-4 models. Our data and code are publicly available at https://github.com/yale-nlp/LLM-T2T.

We propose a novel multimodal video benchmark - the Perception Test - to evaluate the perception and reasoning skills of pre-trained multimodal models (e.g. Flamingo, SeViLA, or GPT-4). Compared to existing benchmarks that focus on computational tasks (e.g. classification, detection or tracking), the Perception Test focuses on skills (Memory, Abstraction, Physics, Semantics) and types of reasoning (descriptive, explanatory, predictive, counterfactual) across video, audio, and text modalities, to provide a comprehensive and efficient evaluation tool. The benchmark probes pre-trained models for their transfer capabilities, in a zero-shot / few-shot or limited finetuning regime. For these purposes, the Perception Test introduces 11.6k real-world videos, 23s average length, designed to show perceptually interesting situations, filmed by around 100 participants worldwide. The videos are densely annotated with six types of labels (multiple-choice and grounded video question-answers, object and point tracks, temporal action and sound segments), enabling both language and non-language evaluations. The fine-tuning and validation splits of the benchmark are publicly available (CC-BY license), in addition to a challenge server with a held-out test split. Human baseline results compared to state-of-the-art video QA models show a substantial gap in performance (91.4% vs 46.2%), suggesting that there is significant room for improvement in multimodal video understanding. Dataset, baseline code, and challenge server are available at https://github.com/deepmind/perception_test

We present The Vault, a dataset of high-quality code-text pairs in multiple programming languages for training large language models to understand and generate code. We present methods for thoroughly extracting samples that use both rule-based and deep learning-based methods to ensure that they contain high-quality pairs of code and text, resulting in a dataset of 43 million high-quality code-text pairs. Our extensive evaluations on common coding tasks including code generation, code search and code summarization show that when fine-tuning Code Large Language Models on The Vault, such models outperform the same models trained on other datasets such as CodeSearchNet. We also provide detailed analyses of our datasets to assess the effects of various programming languages and docstrings on the performance of such models.

Since the proposal of transformers, these models have been limited to bounded input lengths, because of their need to attend to every token in the input. In this work, we propose Unlimiformer: a general approach that wraps any existing pretrained encoder-decoder transformer, and offloads the cross-attention computation to a single k-nearest-neighbor (kNN) index, while the returned kNN distances are the attention dot-product scores. This kNN index can be kept on either the GPU or CPU memory and queried in sub-linear time; this way, we can index practically unlimited input sequences, while every attention head in every decoder layer retrieves its top-k keys, instead of attending to every key. We evaluate Unlimiformer on several long-document and book-summarization benchmarks, showing that it can process even 500k token-long inputs from the BookSum dataset, without any input truncation at test time. We demonstrate that Unlimiformer improves pretrained models such as BART and Longformer by extending them to unlimited inputs without additional learned weights and without modifying their code. We make our code and models publicly available at https://github.com/abertsch72/unlimiformer .

Dual use, the intentional, harmful reuse of technology and scientific artefacts, is a problem yet to be well-defined within the context of Natural Language Processing (NLP). However, as NLP technologies continue to advance and become increasingly widespread in society, their inner workings have become increasingly opaque. Therefore, understanding dual use concerns and potential ways of limiting them is critical to minimising the potential harms of research and development. In this paper, we conduct a survey of NLP researchers and practitioners to understand the depth and their perspective of the problem as well as to assess existing available support. Based on the results of our survey, we offer a definition of dual use that is tailored to the needs of the NLP community. The survey revealed that a majority of researchers are concerned about the potential dual use of their research but only take limited action toward it. In light of the survey results, we discuss the current state and potential means for mitigating dual use in NLP and propose a checklist that can be integrated into existing conference ethics-frameworks, e.g., the ACL ethics checklist.

Multi-agent artificial intelligence research promises a path to develop intelligent technologies that are more human-like and more human-compatible than those produced by "solipsistic" approaches, which do not consider interactions between agents. Melting Pot is a research tool developed to facilitate work on multi-agent artificial intelligence, and provides an evaluation protocol that measures generalization to novel social partners in a set of canonical test scenarios. Each scenario pairs a physical environment (a "substrate") with a reference set of co-players (a "background population"), to create a social situation with substantial interdependence between the individuals involved. For instance, some scenarios were inspired by institutional-economics-based accounts of natural resource management and public-good-provision dilemmas. Others were inspired by considerations from evolutionary biology, game theory, and artificial life. Melting Pot aims to cover a maximally diverse set of interdependencies and incentives. It includes the commonly-studied extreme cases of perfectly-competitive (zero-sum) motivations and perfectly-cooperative (shared-reward) motivations, but does not stop with them. As in real-life, a clear majority of scenarios in Melting Pot have mixed incentives. They are neither purely competitive nor purely cooperative and thus demand successful agents be able to navigate the resulting ambiguity. Here we describe Melting Pot 2.0, which revises and expands on Melting Pot. We also introduce support for scenarios with asymmetric roles, and explain how to integrate them into the evaluation protocol. This report also contains: (1) details of all substrates and scenarios; (2) a complete description of all baseline algorithms and results. Our intention is for it to serve as a reference for researchers using Melting Pot 2.0.