Large language models (LLMs) have recently been used for sequential decision making in interactive environments. However, leveraging environment reward signals for continual LLM actor improvement is not straightforward. We propose Skill Set Optimization (SSO) for improving LLM actor performance through constructing and refining sets of transferable skills. SSO constructs skills by extracting common subtrajectories with high rewards and generating subgoals and instructions to represent each skill. These skills are provided to the LLM actor in-context to reinforce behaviors with high rewards. Then, SSO further refines the skill set by pruning skills that do not continue to result in high rewards. We evaluate our method in the classic videogame NetHack and the text environment ScienceWorld to demonstrate SSO's ability to optimize a set of skills and perform in-context policy improvement. SSO outperforms baselines by 40% in our custom NetHack task and outperforms the previous state-of-the-art in ScienceWorld by 35%.
Amidst growing concerns of large language models (LLMs) being misused for generating misinformation or completing homework assignments, watermarking has emerged as an effective solution for distinguishing human-written and LLM-generated text. A prominent watermarking strategy is to embed a signal into generated text by upsampling a (pseudorandomly-chosen) subset of tokens at every generation step. Although this signal is imperceptible to a human reader, it is detectable through statistical testing. However, implanting such signals alters the model's output distribution and can have unintended effects when watermarked LLMs are used for downstream applications. In this work, we evaluate the performance of watermarked LLMs on a diverse suite of tasks, including text classification, textual entailment, reasoning, question answering, translation, summarization, and language modeling. We find that watermarking has negligible impact on the performance of tasks posed as k-class classification problems in the average case. However, the accuracy can plummet to that of a random classifier for some scenarios (that occur with non-negligible probability). Tasks that are cast as multiple-choice questions and short-form generation are surprisingly unaffected by watermarking. For long-form generation tasks, including summarization and translation, we see a drop of 15-20% in the performance due to watermarking. Our findings highlight the trade-offs that users should be cognizant of when using watermarked models, and point to cases where future research could improve existing trade-offs.
The inevitable appearance of spurious correlations in training datasets hurts the generalization of NLP models on unseen data. Previous work has found that datasets with paired inputs are prone to correlations between a specific part of the input (e.g., the hypothesis in NLI) and the label; consequently, models trained only on those outperform chance. Are these correlations picked up by models trained on the full input data? To address this question, we propose a new evaluation method, Counterfactual Attentiveness Test (CAT). CAT uses counterfactuals by replacing part of the input with its counterpart from a different example (subject to some restrictions), expecting an attentive model to change its prediction. Using CAT, we systematically investigate established supervised and in-context learning models on ten datasets spanning four tasks: natural language inference, reading comprehension, paraphrase detection, and visual & language reasoning. CAT reveals that reliance on such correlations is mainly data-dependent. Surprisingly, we find that GPT3 becomes less attentive with an increased number of demonstrations, while its accuracy on the test data improves. Our results demonstrate that augmenting training or demonstration data with counterfactuals is effective in improving models' attentiveness. We show that models' attentiveness measured by CAT reveals different conclusions from solely measuring correlations in data.
Large text corpora are the backbone of language models. However, we have a limited understanding of the content of these corpora, including general statistics, quality, social factors, and inclusion of evaluation data (contamination). In this work, we propose What's In My Big Data? (WIMBD), a platform and a set of sixteen analyses that allow us to reveal and compare the contents of large text corpora. WIMBD builds on two basic capabilities -- count and search -- at scale, which allows us to analyze more than 35 terabytes on a standard compute node. We apply WIMBD to ten different corpora used to train popular language models, including C4, The Pile, and RedPajama. Our analysis uncovers several surprising and previously undocumented findings about these corpora, including the high prevalence of duplicate, synthetic, and low-quality content, personally identifiable information, toxic language, and benchmark contamination. For instance, we find that about 50% of the documents in RedPajama and LAION-2B-en are duplicates. In addition, several datasets used for benchmarking models trained on such corpora are contaminated with respect to important benchmarks, including the Winograd Schema Challenge and parts of GLUE and SuperGLUE. We open-source WIMBD's code and artifacts to provide a standard set of evaluations for new text-based corpora and to encourage more analyses and transparency around them: github.com/allenai/wimbd.
Large language models primarily rely on incontext learning to execute tasks. We introduce EchoPrompt, a simple yet effective approach to prompt the model to rephrase its queries before answering them. EchoPrompt is inspired by self-questioning, a cognitive strategy humans use to vocalize queries before providing answers, thereby reducing misconceptions. Experimental results demonstrate that EchoPrompt leads to substantial improvements in both zero-shot and few-shot in-context learning with standard and chain-of-thought prompting on four families of causal language models. These improvements are observed across various numerical reasoning (GSM8K, SVAMP, MultiArith, SingleOp), reading comprehension (DROP, SQuAD), and logical reasoning (Shuffled Objects, Date Understanding, Coin Flipping) tasks. On average, EchoPrompt improves the Zero-shot-CoT performance of code-davinci-002 by 5% in numerical tasks and 13% in reading comprehension tasks. We investigate the effectiveness of EchoPrompt through ablation studies, which reveal the significance of both original and rephrased queries for EchoPrompt's efficacy. Our empirical results show that EchoPrompt is an effective technique that can easily augment in-context learning for better performance.
Bias amplification is a phenomenon in which models increase imbalances present in the training data. In this paper, we study bias amplification in the text-to-image domain using Stable Diffusion by comparing gender ratios in training vs. generated images. We find that the model appears to amplify gender-occupation biases found in the training data (LAION). However, we discover that amplification can largely be attributed to discrepancies between training captions and model prompts. For example, an inherent difference is that captions from the training data often contain explicit gender information while the prompts we use do not, which leads to a distribution shift and consequently impacts bias measures. Once we account for various distributional differences between texts used for training and generation, we observe that amplification decreases considerably. Our findings illustrate the challenges of comparing biases in models and the data they are trained on, and highlight confounding factors that contribute to bias amplification.
Large language models (LLMs) are being applied as actors for sequential decision making tasks in domains such as robotics and games, utilizing their general world knowledge and planning abilities. However, previous work does little to explore what environment state information is provided to LLM actors via language. Exhaustively describing high-dimensional states can impair performance and raise inference costs for LLM actors. Previous LLM actors avoid the issue by relying on hand-engineered, task-specific protocols to determine which features to communicate about a state and which to leave out. In this work, we propose Brief Language INputs for DEcision-making Responses (BLINDER), a method for automatically selecting concise state descriptions by learning a value function for task-conditioned state descriptions. We evaluate BLINDER on the challenging video game NetHack and a robotic manipulation task. Our method improves task success rate, reduces input size and compute costs, and generalizes between LLM actors.
Although the prevention of AI vulnerabilities is critical to preserve the safety and privacy of users and businesses, educational tools for robust AI are still underdeveloped worldwide. We present the design, implementation, and assessment of Maestro. Maestro is an effective open-source game-based platform that contributes to the advancement of robust AI education. Maestro provides goal-based scenarios where college students are exposed to challenging life-inspired assignments in a competitive programming environment. We assessed Maestro's influence on students' engagement, motivation, and learning success in robust AI. This work also provides insights into the design features of online learning tools that promote active learning opportunities in the robust AI domain. We analyzed the reflection responses (measured with Likert scales) of 147 undergraduate students using Maestro in two quarterly college courses in AI. According to the results, students who felt the acquisition of new skills in robust AI tended to appreciate highly Maestro and scored highly on material consolidation, curiosity, and mastery in robust AI. Moreover, the leaderboard, our key gamification element in Maestro, has effectively contributed to students' engagement and learning. Results also indicate that Maestro can be effectively adapted to any course length and depth without losing its educational quality.
Content Warning: This paper contains examples of misgendering and erasure that could be offensive and potentially triggering. Gender bias in language technologies has been widely studied, but research has mostly been restricted to a binary paradigm of gender. It is essential also to consider non-binary gender identities, as excluding them can cause further harm to an already marginalized group. In this paper, we comprehensively evaluate popular language models for their ability to correctly use English gender-neutral pronouns (e.g., singular they, them) and neo-pronouns (e.g., ze, xe, thon) that are used by individuals whose gender identity is not represented by binary pronouns. We introduce MISGENDERED, a framework for evaluating large language models' ability to correctly use preferred pronouns, consisting of (i) instances declaring an individual's pronoun, followed by a sentence with a missing pronoun, and (ii) an experimental setup for evaluating masked and auto-regressive language models using a unified method. When prompted out-of-the-box, language models perform poorly at correctly predicting neo-pronouns (averaging 7.6% accuracy) and gender-neutral pronouns (averaging 31.0% accuracy). This inability to generalize results from a lack of representation of non-binary pronouns in training data and memorized associations. Few-shot adaptation with explicit examples in the prompt improves the performance but plateaus at only 45.4% for neo-pronouns. We release the full dataset, code, and demo at https://tamannahossainkay.github.io/misgendered/
The remarkable capabilities of large language models have been accompanied by a persistent drawback: the generation of false and unsubstantiated claims commonly known as "hallucinations". To combat this issue, recent research has introduced approaches that involve editing and attributing the outputs of language models, particularly through prompt-based editing. However, the inference cost and speed of using large language models for editing currently bottleneck prompt-based methods. These bottlenecks motivate the training of compact editors, which is challenging due to the scarcity of training data for this purpose. To overcome these challenges, we exploit the power of large language models to introduce corruptions (i.e., noise) into text and subsequently fine-tune compact editors to denoise the corruptions by incorporating relevant evidence. Our methodology is entirely unsupervised and provides us with faux hallucinations for training in any domain. Our Petite Unsupervised Research and Revision model, PURR, not only improves attribution over existing editing methods based on fine-tuning and prompting, but also achieves faster execution times by orders of magnitude.