The information stored in large language models (LLMs) falls out of date quickly, and retraining from scratch is often not an option. This has recently given rise to a range of techniques for injecting new facts through updating model weights. Current evaluation paradigms are extremely limited, mainly validating the recall of edited facts, but changing one fact should cause rippling changes to the model's related beliefs. If we edit the UK Prime Minister to now be Rishi Sunak, then we should get a different answer to Who is married to the British Prime Minister? In this work, we present a benchmark, MQuAKE (Multi-hop Question Answering for Knowledge Editing), comprising multi-hop questions that assess whether edited models correctly answer questions where the answer should change as an entailed consequence of edited facts. While we find that current knowledge-editing approaches can recall edited facts accurately, they fail catastrophically on the constructed multi-hop questions. We thus propose a simple memory-based approach, MeLLo, which stores all edited facts externally while prompting the language model iteratively to generate answers that are consistent with the edited facts. While MQuAKE remains challenging, we show that MeLLo scales well with LLMs (e.g., OpenAI GPT-3.5-turbo) and outperforms previous model editors by a large margin.
Transformer-based language models (LMs) are powerful and widely-applicable tools, but their usefulness is constrained by a finite context window and the expensive computational cost of processing long text documents. We propose to adapt pre-trained LMs into AutoCompressors. These language models are capable of compressing long contexts into compact summary vectors, which are then accessible to the model as soft prompts. Summary vectors are trained with an unsupervised objective, whereby long documents are processed in segments, and summary vectors from all previous segments are used in language modeling. We fine-tune OPT and Llama-2 models on sequences of up to 30,720 tokens and show that AutoCompressors can utilize long contexts to improve perplexity. We evaluate AutoCompressors on in-context learning by compressing task demonstrations and find that summary vectors are good substitutes for plain-text demonstrations, increasing accuracy while reducing inference costs. Finally, we explore the benefits of pre-computing summary vectors for large corpora by applying summary vectors to retrievalaugmented language modeling and a passage re-ranking task. Overall, AutoCompressors emerge as a simple and inexpensive solution to extend the context window of LMs while speeding up inference over long contexts.
With the advance of large language models (LLMs), the research field of LLM applications becomes more and more popular and the idea of constructing pipelines to accomplish complex tasks by stacking LLM API calls come true. However, this kind of methods face two limitations: narrow information coverage and low fault tolerance. In this work, we propose a novel method called ALLIES. Given an input query, ALLIES leverages LLMs to iteratively generate new queries related to the original query, enabling an iterative reasoning process. By iteratively refining and expanding the scope of the original query, ALLIES captures and utilizes hidden knowledge that may not be directly obtainable through retrieval. We take zero-shot open-domain question answering (ODQA) as an application scene and evaluate ALLIES on the widely-used benchmarks, such as NQ, WebQ and TriviaQA. The experimental results demonstrate that ALLIES significantly outperforms other zero-shot baselines, indicating its effectiveness in tackling those challenges. Our code is available in https://github.com/microsoft/SimXNS/tree/main/ALLIES.
Large Language Models (LLMs) have seen an impressive wave of advances recently, with models now excelling in a variety of tasks, such as mathematical reasoning and program synthesis. However, their potential to effectively use tools via API calls remains unfulfilled. This is a challenging task even for today's state-of-the-art LLMs such as GPT-4, largely due to their inability to generate accurate input arguments and their tendency to hallucinate the wrong usage of an API call. We release Gorilla, a finetuned LLaMA-based model that surpasses the performance of GPT-4 on writing API calls. When combined with a document retriever, Gorilla demonstrates a strong capability to adapt to test-time document changes, enabling flexible user updates or version changes. It also substantially mitigates the issue of hallucination, commonly encountered when prompting LLMs directly. To evaluate the model's ability, we introduce APIBench, a comprehensive dataset consisting of HuggingFace, TorchHub, and TensorHub APIs. The successful integration of the retrieval system with Gorilla demonstrates the potential for LLMs to use tools more accurately, keep up with frequently updated documentation, and consequently increase the reliability and applicability of their outputs. Gorilla's code, model, data, and demo are available at https://gorilla.cs.berkeley.edu
In this paper, we present HuatuoGPT, a large language model (LLM) for medical consultation. The core recipe of HuatuoGPT is to leverage both \textit{distilled data from ChatGPT} and \textit{real-world data from doctors} in the supervised fine-tuned stage. The responses of ChatGPT are usually detailed, well-presented and informative while it cannot perform like a doctor in many aspects, e.g. for integrative diagnosis. We argue that real-world data from doctors would be complementary to distilled data in the sense the former could tame a distilled language model to perform like doctors. To better leverage the strengths of both data, we train a reward model to align the language model with the merits that both data bring, following an RLAIF (reinforced learning from AI feedback) fashion. To evaluate and benchmark the models, we propose a comprehensive evaluation scheme (including automatic and manual metrics). Experimental results demonstrate that HuatuoGPT achieves state-of-the-art results in performing medical consultation among open-source LLMs in GPT-4 evaluation, human evaluation, and medical benchmark datasets. It is worth noting that by using additional real-world data and RLAIF, the distilled language model (i.e., HuatuoGPT) outperforms its teacher model ChatGPT in most cases. Our code, data, and models are publicly available at \url{https://github.com/FreedomIntelligence/HuatuoGPT}. The online demo is available at \url{https://www.HuatuoGPT.cn/}.
Reinforcement Learning with Human Feedback (RLHF) is the most prominent method for Language Model (LM) alignment. However, RLHF is an unstable and data-hungry process that continually requires new high-quality LM-generated data for finetuning. We introduce Advantage-Leftover Lunch RL (A-LoL), a new class of offline policy gradient algorithms that enable RL training on any pre-existing data. By assuming the entire LM output sequence as a single action, A-LoL allows incorporating sequence-level classifiers or human-designed scoring functions as rewards. Subsequently, by using LM's value estimate, A-LoL only trains on positive advantage (leftover) data points, making it resilient to noise. Overall, A-LoL is an easy-to-implement, sample-efficient, and stable LM training recipe. We demonstrate the effectiveness of A-LoL and its variants with a set of four different language generation tasks. We compare against both online RL (PPO) and recent preference-based (DPO, PRO) and reward-based (GOLD) offline RL baselines. On the commonly-used RLHF benchmark, Helpful and Harmless Assistant (HHA), LMs trained with A-LoL methods achieve the highest diversity while also being rated more safe and helpful than the baselines according to humans. Additionally, in the remaining three tasks, A-LoL could optimize multiple distinct reward functions even when using noisy or suboptimal training data. We also release our experimental code. https://github.com/abaheti95/LoL-RL
Do the Spratly Islands belong to China, the Philippines, or Vietnam? A pretrained large language model (LLM) may answer differently if asked in the languages of each claimant country: Chinese, Tagalog, or Vietnamese. This contrasts with a multilingual human, who would likely answer consistently. In this paper, we show that LLMs recall certain geographical knowledge inconsistently when queried in different languages -- a phenomenon we term geopolitical bias. As a targeted case study, we consider territorial disputes, an inherently controversial and multilingual task. We introduce BorderLines, a dataset of territorial disputes which covers 251 territories, each associated with a set of multiple-choice questions in the languages of each claimant country (49 languages in total). We also propose a suite of evaluation metrics to precisely quantify bias and consistency in responses across different languages. We then evaluate various multilingual LLMs on our dataset and metrics to probe their internal knowledge and use the proposed metrics to discover numerous inconsistencies in how these models respond in different languages. Finally, we explore several prompt modification strategies, aiming to either amplify or mitigate geopolitical bias, which highlights how brittle LLMs are and how they tailor their responses depending on cues from the interaction context. Our code and data are available at https://github.com/manestay/borderlines
When people interpret text, they rely on inferences that go beyond the observed language itself. Inspired by this observation, we introduce a method for the analysis of text that takes implicitly communicated content explicitly into account. We use a large language model to produce sets of propositions that are inferentially related to the text that has been observed, then validate the plausibility of the generated content via human judgments. Incorporating these explicit representations of implicit content proves useful in multiple problem settings that involve the human interpretation of utterances: assessing the similarity of arguments, making sense of a body of opinion data, and modeling legislative behavior. Our results suggest that modeling the meanings behind observed language, rather than the literal text alone, is a valuable direction for NLP and particularly its applications to social science.
Pre-trained language models (PLMs) have ignited a surge in demand for effective fine-tuning techniques, particularly in low-resource domains and languages. Active learning (AL), a set of algorithms designed to decrease labeling costs by minimizing label complexity, has shown promise in confronting the labeling bottleneck. In parallel, adapter modules designed for parameter-efficient fine-tuning (PEFT) have demonstrated notable potential in low-resource settings. However, the interplay between AL and adapter-based PEFT remains unexplored. We present an empirical study of PEFT behavior with AL in low-resource settings for text classification tasks. Our findings affirm the superiority of PEFT over full-fine tuning (FFT) in low-resource settings and demonstrate that this advantage persists in AL setups. We further examine the properties of PEFT and FFT through the lens of forgetting dynamics and instance-level representations, where we find that PEFT yields more stable representations of early and middle layers compared to FFT. Our research underscores the synergistic potential of AL and PEFT in low-resource settings, paving the way for advancements in efficient and effective fine-tuning.
While automatic dialogue tutors hold great potential in making education personalized and more accessible, research on such systems has been hampered by a lack of sufficiently large and high-quality datasets. Collecting such datasets remains challenging, as recording tutoring sessions raises privacy concerns and crowdsourcing leads to insufficient data quality. To address this, we propose a framework to generate such dialogues by pairing human teachers with a Large Language Model (LLM) prompted to represent common student errors. We describe how we use this framework to collect MathDial, a dataset of 3k one-to-one teacher-student tutoring dialogues grounded in multi-step math reasoning problems. While models like GPT-3 are good problem solvers, they fail at tutoring because they generate factually incorrect feedback or are prone to revealing solutions to students too early. To overcome this, we let teachers provide learning opportunities to students by guiding them using various scaffolding questions according to a taxonomy of teacher moves. We demonstrate MathDial and its extensive annotations can be used to finetune models to be more effective tutors (and not just solvers). We confirm this by automatic and human evaluation, notably in an interactive setting that measures the trade-off between student solving success and telling solutions. The dataset is released publicly.
Using model weights pretrained on a high-resource language as a warm start can reduce the need for data and compute to obtain high-quality language models for other, especially low-resource, languages. However, if we want to use a new tokenizer specialized for the target language, we cannot transfer the source model's embedding matrix. In this paper, we propose FOCUS - Fast Overlapping Token Combinations Using Sparsemax, a novel embedding initialization method that initializes the embedding matrix effectively for a new tokenizer based on information in the source model's embedding matrix. FOCUS represents newly added tokens as combinations of tokens in the overlap of the source and target vocabularies. The overlapping tokens are selected based on semantic similarity in an auxiliary static token embedding space. We focus our study on using the multilingual XLM-R as a source model and empirically show that FOCUS outperforms random initialization and previous work in language modeling and on a range of downstream tasks (NLI, QA, and NER).
Transformer-based pretrained models like BERT, GPT-2 and T5 have been finetuned for a large number of natural language processing (NLP) tasks, and have been shown to be very effective. However, while finetuning, what changes across layers in these models with respect to pretrained checkpoints is under-studied. Further, how robust are these models to perturbations in input text? Does the robustness vary depending on the NLP task for which the models have been finetuned? While there exists some work on studying the robustness of BERT finetuned for a few NLP tasks, there is no rigorous study that compares this robustness across encoder only, decoder only and encoder-decoder models. In this paper, we characterize changes between pretrained and finetuned language model representations across layers using two metrics: CKA and STIR. Further, we study the robustness of three language models (BERT, GPT-2 and T5) with eight different text perturbations on classification tasks from the General Language Understanding Evaluation (GLUE) benchmark, and generation tasks like summarization, free-form generation and question generation. GPT-2 representations are more robust than BERT and T5 across multiple types of input perturbation. Although models exhibit good robustness broadly, dropping nouns, verbs or changing characters are the most impactful. Overall, this study provides valuable insights into perturbation-specific weaknesses of popular Transformer-based models, which should be kept in mind when passing inputs. We make the code and models publicly available [https://github.com/PavanNeerudu/Robustness-of-Transformers-models].
Contrast consistency, the ability of a model to make consistently correct predictions in the presence of perturbations, is an essential aspect in NLP. While studied in tasks such as sentiment analysis and reading comprehension, it remains unexplored in open-domain question answering (OpenQA) due to the difficulty of collecting perturbed questions that satisfy factuality requirements. In this work, we collect minimally edited questions as challenging contrast sets to evaluate OpenQA models. Our collection approach combines both human annotation and large language model generation. We find that the widely used dense passage retriever (DPR) performs poorly on our contrast sets, despite fitting the training set well and performing competitively on standard test sets. To address this issue, we introduce a simple and effective query-side contrastive loss with the aid of data augmentation to improve DPR training. Our experiments on the contrast sets demonstrate that DPR's contrast consistency is improved without sacrificing its accuracy on the standard test sets.
Semantic textual similarity (STS), a cornerstone task in NLP, measures the degree of similarity between a pair of sentences, and has broad application in fields such as information retrieval and natural language understanding. However, sentence similarity can be inherently ambiguous, depending on the specific aspect of interest. We resolve this ambiguity by proposing a novel task called Conditional STS (C-STS) which measures sentences' similarity conditioned on an feature described in natural language (hereon, condition). As an example, the similarity between the sentences "The NBA player shoots a three-pointer." and "A man throws a tennis ball into the air to serve." is higher for the condition "The motion of the ball" (both upward) and lower for "The size of the ball" (one large and one small). C-STS's advantages are two-fold: (1) it reduces the subjectivity and ambiguity of STS and (2) enables fine-grained language model evaluation through diverse natural language conditions. We put several state-of-the-art models to the test, and even those performing well on STS (e.g. SimCSE, Flan-T5, and GPT-4) find C-STS challenging; all with Spearman correlation scores below 50. To encourage a more comprehensive evaluation of semantic similarity and natural language understanding, we make nearly 19K C-STS examples and code available for others to train and test their models.
We investigate the predictability of large language model (LLM) capabilities: given records of past experiments using different model families, numbers of parameters, tasks, and numbers of in-context examples, can we accurately predict LLM performance on new experiment configurations? Answering this question has practical implications for LLM users (e.g., deciding which models to try), developers (e.g., prioritizing evaluation on representative tasks), and the research community (e.g., identifying hard-to-predict capabilities that warrant further investigation). We study the performance prediction problem on experiment records from BIG-bench. On a random train-test split, an MLP-based predictor achieves an $R^2$ score greater than 95%, indicating the presence of learnable patterns within the experiment records. We then formulate the problem of searching for "small-bench," an informative subset of BIG-bench tasks from which the performance on the full set can be maximally recovered. We find a subset as informative as BIG-bench Hard for evaluating new model families, while being $3\times$ smaller. Additionally, we find competitive subsets by clustering task representations learned by our MLP-based predictor and selecting tasks close to cluster centroids, highlighting the importance of task diversity in constructing "small-bench."
Large language models (LLMs) can use in-context demonstrations to improve performance on zero-shot tasks. However, selecting the best in-context examples is challenging because model performance can vary widely depending on the selected examples. We present a cross-entropy difference (CED) method for selecting in-context demonstrations. Our method is based on the observation that the effectiveness of in-context demonstrations negatively correlates with the perplexity of the test example by a language model that was finetuned on that demonstration. We utilize parameter efficient finetuning to train small models on training data that are used for computing the cross-entropy difference between a test example and every candidate in-context demonstration. This metric is used to rank and select in-context demonstrations independently for each test input. We evaluate our method on a mix-domain dataset that combines 8 benchmarks, representing 4 text generation tasks, showing that CED for in-context demonstration selection can improve performance for a variety of LLMs.
The answering quality of an aligned large language model (LLM) can be drastically improved if treated with proper crafting of prompts. In this paper, we propose ExpertPrompting to elicit the potential of LLMs to answer as distinguished experts. We first utilize In-Context Learning to automatically synthesize detailed and customized descriptions of the expert identity for each specific instruction, and then ask LLMs to provide answer conditioned on such agent background. Based on this augmented prompting strategy, we produce a new set of instruction-following data using GPT-3.5, and train a competitive open-source chat assistant called ExpertLLaMA. We employ GPT4-based evaluation to show that 1) the expert data is of significantly higher quality than vanilla answers, and 2) ExpertLLaMA outperforms existing open-source opponents and achieves 96\% of the original ChatGPT's capability. All data and the ExpertLLaMA model will be made publicly available at \url{https://github.com/OFA-Sys/ExpertLLaMA}.
With the rapid growth in model size, fine-tuning the large pre-trained language model has become increasingly difficult due to its extensive memory usage. Previous works usually focus on reducing the number of trainable parameters in the network. While the model parameters do contribute to memory usage, the primary memory bottleneck during training arises from storing feature maps, also known as activations, as they are crucial for gradient calculation. Notably, neural networks are usually trained using stochastic gradient descent. We argue that in stochastic optimization, models can handle noisy gradients as long as the gradient estimator is unbiased with reasonable variance. Following this motivation, we propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance, which only requires storing the sub-sampled activations for calculating the gradient. Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones. By replacing the linear operation with our approximated one in transformers, we can achieve up to 2.7$\times$ peak memory reduction with almost no accuracy drop and enables up to $6.4\times$ larger batch size. Under the same hardware, WTA-CRS enables better down-streaming task performance by applying larger models and/or faster training speed with larger batch sizes.
Despite outstanding performance in many tasks, language models are notoriously inclined to make factual errors in tasks requiring arithmetic computation. We address this deficiency by creating Calc-X, a collection of datasets that demonstrates the appropriate use of a calculator in reasoning chains. Calc-X is suitable for teaching language models to offload computations to a symbolic system. We survey and unify several existing chain-of-thought datasets into a proposed format, resulting in a standard collection of over 300,000 samples requiring arithmetic reasoning. Finally, we use the new Calc-X collection to train open-source calculator-using models we call Calcformers and show that these models approximately double the accuracy of generating correct results compared to vanilla language model baselines. We make all Calc-X datasets, source code and Calcformers models publicly available.
Recent developments in large pre-trained language models have enabled unprecedented performance on a variety of downstream tasks. Achieving best performance with these models often leverages in-context learning, where a model performs a (possibly new) task given one or more examples. However, recent work has shown that the choice of examples can have a large impact on task performance and that finding an optimal set of examples is non-trivial. While there are many existing methods for selecting in-context examples, they generally score examples independently, ignoring the dependency between them and the order in which they are provided to the model. In this work, we propose Retrieval for In-Context Learning (RetICL), a learnable method for modeling and optimally selecting examples sequentially for in-context learning. We frame the problem of sequential example selection as a Markov decision process and train an example retriever using reinforcement learning. We evaluate RetICL on math word problem solving and scientific question answering tasks and show that it consistently outperforms or matches heuristic and learnable baselines. We also use case studies to show that RetICL implicitly learns representations of problem solving strategies.