A rapidly growing number of applications rely on a small set of closed-source language models (LMs). This dependency might introduce novel security risks if LMs develop self-recognition capabilities. Inspired by human identity verification methods, we propose a novel approach for assessing self-recognition in LMs using model-generated "security questions". Our test can be externally administered to keep track of frontier models as it does not require access to internal model parameters or output probabilities. We use our test to examine self-recognition in ten of the most capable open- and closed-source LMs currently publicly available. Our extensive experiments found no empirical evidence of general or consistent self-recognition in any examined LM. Instead, our results suggest that given a set of alternatives, LMs seek to pick the "best" answer, regardless of its origin. Moreover, we find indications that preferences about which models produce the best answers are consistent across LMs. We additionally uncover novel insights on position bias considerations for LMs in multiple-choice settings.

Humor is a fundamental facet of human cognition and interaction. Yet, despite recent advances in natural language processing, humor detection remains a challenging task that is complicated by the scarcity of datasets that pair humorous texts with similar non-humorous counterparts. In our work, we investigate whether large language models (LLMs), can generate synthetic data for humor detection via editing texts. We benchmark LLMs on an existing human dataset and show that current LLMs display an impressive ability to 'unfun' jokes, as judged by humans and as measured on the downstream task of humor detection. We extend our approach to a code-mixed English-Hindi humor dataset, where we find that GPT-4's synthetic data is highly rated by bilingual annotators and provides challenging adversarial examples for humor classifiers.

We ask whether multilingual language models trained on unbalanced, English-dominated corpora use English as an internal pivot language -- a question of key importance for understanding how language models function and the origins of linguistic bias. Focusing on the Llama-2 family of transformer models, our study uses carefully constructed non-English prompts with a unique correct single-token continuation. From layer to layer, transformers gradually map an input embedding of the final prompt token to an output embedding from which next-token probabilities are computed. Tracking intermediate embeddings through their high-dimensional space reveals three distinct phases, whereby intermediate embeddings (1) start far away from output token embeddings; (2) already allow for decoding a semantically correct next token in the middle layers, but give higher probability to its version in English than in the input language; (3) finally move into an input-language-specific region of the embedding space. We cast these results into a conceptual model where the three phases operate in "input space", "concept space", and "output space", respectively. Crucially, our evidence suggests that the abstract "concept space" lies closer to English than to other languages, which may have important consequences regarding the biases held by multilingual language models.

The field of AI alignment aims to steer AI systems toward human goals, preferences, and ethical principles. Its contributions have been instrumental for improving the output quality, safety, and trustworthiness of today's AI models. This perspective article draws attention to a fundamental challenge inherent in all AI alignment endeavors, which we term the "AI alignment paradox": The better we align AI models with our values, the easier we make it for adversaries to misalign the models. We illustrate the paradox by sketching three concrete example incarnations for the case of language models, each corresponding to a distinct way in which adversaries can exploit the paradox. With AI's increasing real-world impact, it is imperative that a broad community of researchers be aware of the AI alignment paradox and work to find ways to break out of it, in order to ensure the beneficial use of AI for the good of humanity.

Large language models (LLMs) have significantly evolved, moving from simple output generation to complex reasoning and from stand-alone usage to being embedded into broader frameworks. In this paper, we introduce \emph{Fleet of Agents (FoA)}, a novel framework utilizing LLMs as agents to navigate through dynamic tree searches, employing a genetic-type particle filtering approach. FoA spawns a multitude of agents, each exploring autonomously, followed by a selection phase where resampling based on a heuristic value function optimizes the balance between exploration and exploitation. This mechanism enables dynamic branching, adapting the exploration strategy based on discovered solutions. We experimentally validate FoA using two benchmark tasks, "Game of 24" and "Mini-Crosswords". FoA outperforms the previously proposed Tree-of-Thoughts method in terms of efficacy and efficiency: it significantly decreases computational costs (by calling the value function less frequently) while preserving comparable or even superior accuracy.

An edit summary is a succinct comment written by a Wikipedia editor explaining the nature of, and reasons for, an edit to a Wikipedia page. Edit summaries are crucial for maintaining the encyclopedia: they are the first thing seen by content moderators and help them decide whether to accept or reject an edit. Additionally, edit summaries constitute a valuable data source for researchers. Unfortunately, as we show, for many edits, summaries are either missing or incomplete. To overcome this problem and help editors write useful edit summaries, we propose a model for recommending edit summaries generated by a language model trained to produce good edit summaries given the representation of an edit diff. This is a challenging task for multiple reasons, including mixed-quality training data, the need to understand not only what was changed in the article but also why it was changed, and efficiency requirements imposed by the scale of Wikipedia. We address these challenges by curating a mix of human and synthetically generated training data and fine-tuning a generative language model sufficiently small to be used on Wikipedia at scale. Our model performs on par with human editors. Commercial large language models are able to solve this task better than human editors, but would be too expensive to run on Wikipedia at scale. More broadly, this paper showcases how language modeling technology can be used to support humans in maintaining one of the largest and most visible projects on the Web.

The remarkable and ever-increasing capabilities of Large Language Models (LLMs) have raised concerns about their potential misuse for creating personalized, convincing misinformation and propaganda. To gain insights into LLMs' persuasive capabilities without directly engaging in experimentation with humans, we propose studying their performance on the related task of detecting convincing arguments. We extend a dataset by Durmus & Cardie (2018) with debates, votes, and user traits and propose tasks measuring LLMs' ability to (1) distinguish between strong and weak arguments, (2) predict stances based on beliefs and demographic characteristics, and (3) determine the appeal of an argument to an individual based on their traits. We show that LLMs perform on par with humans in these tasks and that combining predictions from different LLMs yields significant performance gains, even surpassing human performance. The data and code released with this paper contribute to the crucial ongoing effort of continuously evaluating and monitoring the rapidly evolving capabilities and potential impact of LLMs.

Emoji have become ubiquitous in written communication, on the Web and beyond. They can emphasize or clarify emotions, add details to conversations, or simply serve decorative purposes. This casual use, however, barely scratches the surface of the expressive power of emoji. To further unleash this power, we present Emojinize, a method for translating arbitrary text phrases into sequences of one or more emoji without requiring human input. By leveraging the power of large language models, Emojinize can choose appropriate emoji by disambiguating based on context (eg, cricket-bat vs bat) and can express complex concepts compositionally by combining multiple emoji (eq, "Emojinize" is translated to input-latin-letters right-arrow grinning-face). In a cloze test--based user study, we show that Emojinize's emoji translations increase the human guessability of masked words by 55%, whereas human-picked emoji translations do so by only 29%. These results suggest that emoji provide a sufficiently rich vocabulary to accurately translate a wide variety of words. Moreover, annotating words and phrases with Emojinize's emoji translations opens the door to numerous downstream applications, including children learning how to read, adults learning foreign languages, and text understanding for people with learning disabilities.

Large language models (LLMs) have been shown to perform better when asked to reason step-by-step before answering a question. However, it is unclear to what degree the model's final answer is faithful to the stated reasoning steps. In this paper, we perform a causal mediation analysis on twelve LLMs to examine how intermediate reasoning steps generated by the LLM influence the final outcome and find that LLMs do not reliably use their intermediate reasoning steps when generating an answer. To address this issue, we introduce FRODO, a framework to tailor small-sized LMs to generate correct reasoning steps and robustly reason over these steps. FRODO consists of an inference module that learns to generate correct reasoning steps using an implicit causal reward function and a reasoning module that learns to faithfully reason over these intermediate inferences using a counterfactual and causal preference objective. Our experiments show that FRODO significantly outperforms four competitive baselines. Furthermore, FRODO improves the robustness and generalization ability of the reasoning LM, yielding higher performance on out-of-distribution test sets. Finally, we find that FRODO's rationales are more faithful to its final answer predictions than standard supervised fine-tuning.

Traditional language models, adept at next-token prediction in text sequences, often struggle with transduction tasks between distinct symbolic systems, particularly when parallel data is scarce. Addressing this issue, we introduce \textit{symbolic autoencoding} ($\Sigma$AE), a self-supervised framework that harnesses the power of abundant unparallel data alongside limited parallel data. $\Sigma$AE connects two generative models via a discrete bottleneck layer and is optimized end-to-end by minimizing reconstruction loss (simultaneously with supervised loss for the parallel data), such that the sequence generated by the discrete bottleneck can be read out as the transduced input sequence. We also develop gradient-based methods allowing for efficient self-supervised sequence learning despite the discreteness of the bottleneck. Our results demonstrate that $\Sigma$AE significantly enhances performance on transduction tasks, even with minimal parallel data, offering a promising solution for weakly supervised learning scenarios.