'It feels like a gold rush," says Timnit Gebru. "In fact, it is a gold rush. And a lot of the people who are making money are not the people actually in the midst of it. But it's humans who decide whether all this should be done or not. We should remember that we have the agency to do that." Gebru is talking about her specialised field: artificial intelligence. On the day we speak via a video call, she is in Kigali, Rwanda, preparing to host a workshop and chair a panel at an international conference on AI. It will address the huge growth in AI's capabilities, as well as something that the frenzied conversation about AI misses out: the fact that many of its systems may well be built on a huge mess of biases, inequalities and imbalances of power. This gathering, the clunkily titled International Conference on Learning Representations, marks the first time people in the field have come together in an African country – which makes a powerful point about big tech's neglect of the global south. When Gebru talks about the way that AI "impacts people all over the world and they don't get to have a say on how they should shape it", the issue is thrown into even sharper relief by her backstory. In her teens, Gebru was a refugee from the war between Ethiopia, where she grew up, and Eritrea, where her parents were born. After a year in Ireland, she made it to the outskirts of Boston, Massachusetts, and from there to Stanford University in northern California, which opened the way to a career at the cutting edge of the computing industry: Apple, then Microsoft, followed by Google. But in late 2020, her work at Google came to a sudden end. As the co-leader of Google's small ethical AI team, Gebru was one of the authors of an academic paper that warned about the kind of AI that is increasingly built into our lives, taking internet searches and user recommendations to apparently new levels of sophistication and threatening to master such human talents as writing, composing music and analysing images. The clear danger, the paper said, is that such supposed "intelligence" is based on huge data sets that "overrepresent hegemonic viewpoints and encode biases potentially damaging to marginalised populations". Put more bluntly, AI threatens to deepen the dominance of a way of thinking that is white, male, comparatively affluent and focused on the US and Europe. In response, senior managers at Google demanded that Gebru either withdraw the paper, or take her name and those of her colleagues off it. This triggered a run of events that led to her departure. Google says she resigned; Gebru insists that she was fired. What all this told her, she says, is that big tech is consumed by a drive to develop AI and "you don't want someone like me who's going to get in your way.

Augmented Language Models (ALMs) blend the reasoning capabilities of Large Language Models (LLMs) with tools that allow for knowledge retrieval and action execution. Existing ALM systems trigger LLM thought processes while pulling observations from these tools in an interleaved fashion. Specifically, an LLM reasons to call an external tool, gets halted to fetch the tool's response, and then decides the next action based on all preceding response tokens. Such a paradigm, though straightforward and easy to implement, often leads to huge computation complexity from redundant prompts and repeated execution. This study addresses such challenges for the first time, proposing a modular paradigm ReWOO (Reasoning WithOut Observation) that detaches the reasoning process from external observations, thus significantly reducing token consumption. Comprehensive evaluations across six public NLP benchmarks and a curated dataset reveal consistent performance enhancements with our proposed methodology. Notably, ReWOO achieves 5x token efficiency and 4% accuracy improvement on HotpotQA, a multi-step reasoning benchmark. Furthermore, ReWOO demonstrates robustness under tool-failure scenarios. Beyond prompt efficiency, decoupling parametric modules from non-parametric tool calls enables instruction fine-tuning to offload LLMs into smaller language models, thus substantially reducing model parameters. Our illustrative work offloads reasoning ability from 175B GPT3.5 into 7B LLaMA, demonstrating the significant potential for truly efficient and scalable ALM systems.

Much attention and concern has been raised recently about bias and the use of machine learning algorithms in healthcare, especially as it relates to perpetuating racial discrimination and health disparities. Following an initial system dynamics workshop at the Data for Black Lives II conference hosted at MIT in January of 2019, a group of conference participants interested in building capabilities to use system dynamics to understand complex societal issues convened monthly to explore issues related to racial bias in AI and implications for health disparities through qualitative and simulation modeling. In this paper we present results and insights from the modeling process and highlight the importance of centering the discussion of data and healthcare on people and their experiences with healthcare and science, and recognizing the societal context where the algorithm is operating. Collective memory of community trauma, through deaths attributed to poor healthcare, and negative experiences with healthcare are endogenous drivers of seeking treatment and experiencing effective care, which impact the availability and quality of data for algorithms. These drivers have drastically disparate initial conditions for different racial groups and point to limited impact of focusing solely on improving diagnostic algorithms for achieving better health outcomes for some groups.

A prominent weakness of modern language models (LMs) is their tendency to generate factually incorrect text, which hinders their usability. A natural question is whether such factual errors can be detected automatically. Inspired by truth-seeking mechanisms in law, we propose a factuality evaluation framework for LMs that is based on cross-examination. Our key idea is that an incorrect claim is likely to result in inconsistency with other claims that the model generates. To discover such inconsistencies, we facilitate a multi-turn interaction between the LM that generated the claim and another LM (acting as an examiner) which introduces questions to discover inconsistencies. We empirically evaluate our method on factual claims made by multiple recent LMs on four benchmarks, finding that it outperforms existing methods and baselines, often by a large gap. Our results demonstrate the potential of using interacting LMs for capturing factual errors.

In this work, we conduct a detailed analysis on the performance of legal-oriented pre-trained language models (PLMs). We examine the interplay between their original objective, acquired knowledge, and legal language understanding capacities which we define as the upstream, probing, and downstream performance, respectively. We consider not only the models' size but also the pre-training corpora used as important dimensions in our study. To this end, we release a multinational English legal corpus (LeXFiles) and a legal knowledge probing benchmark (LegalLAMA) to facilitate training and detailed analysis of legal-oriented PLMs. We release two new legal PLMs trained on LeXFiles and evaluate them alongside others on LegalLAMA and LexGLUE. We find that probing performance strongly correlates with upstream performance in related legal topics. On the other hand, downstream performance is mainly driven by the model's size and prior legal knowledge which can be estimated by upstream and probing performance. Based on these findings, we can conclude that both dimensions are important for those seeking the development of domain-specific PLMs.

For instance, backgrounds in marketing, social media marketing, social work, education, public health, and journalism can contribute fresh perspectives and expertise. Second, diversity and inclusion should be covered in training and development programs via mentorships, job shadowing, simulation exercises, and contact with diverse end user panels. Third, partnerships with academic, civil society and public sector institutions should be established to contribute to holistic and pan-disciplinary reviews of AI systems, diversity and inclusion audits, and assessment of social impacts. Fourth, a workplace culture of belonging should be created and periodically assessed via both open and confidential feedback mechanisms which include diversity markers.

Multi-label text classification (MLC) is a challenging task in settings of large label sets, where label support follows a Zipfian distribution. In this paper, we address this problem through retrieval augmentation, aiming to improve the sample efficiency of classification models. Our approach closely follows the standard MLC architecture of a Transformer-based encoder paired with a set of classification heads. In our case, however, the input document representation is augmented through cross-attention to similar documents retrieved from the training set and represented in a task-specific manner. We evaluate this approach on four datasets from the legal and biomedical domains, all of which feature highly skewed label distributions. Our experiments show that retrieval augmentation substantially improves model performance on the long tail of infrequent labels especially so for lower-resource training scenarios and more challenging long-document data scenarios.

Neural language models (LMs) represent facts about the world described by text. Sometimes these facts derive from training data (in most LMs, a representation of the word "banana" encodes the fact that bananas are fruits). Sometimes facts derive from input text itself (a representation of the sentence "I poured out the bottle" encodes the fact that the bottle became empty). We describe REMEDI, a method for learning to map statements in natural language to fact encodings in an LM's internal representation system. REMEDI encodings can be used as knowledge editors: when added to LM hidden representations, they modify downstream generation to be consistent with new facts. REMEDI encodings may also be used as probes: when compared to LM representations, they reveal which properties LMs already attribute to mentioned entities, in some cases making it possible to predict when LMs will generate outputs that conflict with background knowledge or input text. REMEDI thus links work on probing, prompting, and LM editing, and offers steps toward general tools for fine-grained inspection and control of knowledge in LMs.

Academic and policy proposals on algorithmic accountability often seek to understand algorithmic systems in their socio-technical context, recognising that they are produced by 'many hands'. Increasingly, however, algorithmic systems are also produced, deployed, and used within a supply chain comprising multiple actors tied together by flows of data between them. In such cases, it is the working together of an algorithmic supply chain of different actors who contribute to the production, deployment, use, and functionality that drives systems and produces particular outcomes. We argue that algorithmic accountability discussions must consider supply chains and the difficult implications they raise for the governance and accountability of algorithmic systems. In doing so, we explore algorithmic supply chains, locating them in their broader technical and political economic context and identifying some key features that should be understood in future work on algorithmic governance and accountability (particularly regarding general purpose AI services). To highlight ways forward and areas warranting attention, we further discuss some implications raised by supply chains: challenges for allocating accountability stemming from distributed responsibility for systems between actors, limited visibility due to the accountability horizon, service models of use and liability, and cross-border supply chains and regulatory arbitrage

Although large language models have achieved impressive zero-shot ability, the huge model size generally incurs high cost. Recently, semi-parametric language models, which augment a smaller language model with an external retriever, have demonstrated promising language modeling capabilities. However, it remains unclear whether such semi-parametric language models can perform competitively well as their fully-parametric counterparts on zero-shot generalization to downstream tasks. In this work, we introduce $\text{Zemi}$, a zero-shot semi-parametric language model. To our best knowledge, this is the first semi-parametric language model that can demonstrate strong zero-shot performance on a wide range of held-out unseen tasks. We train $\text{Zemi}$ with a novel semi-parametric multitask prompted training paradigm, which shows significant improvement compared with the parametric multitask training as proposed by T0. Specifically, we augment the multitask training and zero-shot evaluation with retrieval from a large-scale task-agnostic unlabeled corpus. In order to incorporate multiple potentially noisy retrieved augmentations, we further propose a novel $\text{augmentation fusion}$ module leveraging perceiver resampler and gated cross-attention. Notably, our proposed $\text{Zemi}_\text{LARGE}$ outperforms T0-3B by 16% on all seven evaluation tasks while being 3.9x smaller in model size.