Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking

When writing and talking, people sometimes pause to think. Although reasoning-focused works have often framed reasoning as a method of answering questions or completing agentic tasks, reasoning is implicit in almost all written text. For example, this applies to the steps not stated between the lines of a proof or to the theory of mind underlying a conversation. In the Self-Taught Reasoner (STaR, Zelikman et al. 2022), useful thinking is learned by inferring rationales from few-shot examples in question-answering and learning from those that lead to a correct answer. This is a highly constrained setting -- ideally, a language model could instead learn to infer unstated rationales in arbitrary text. We present Quiet-STaR, a generalization of STaR in which LMs learn to generate rationales at each token to explain future text, improving their predictions. We address key challenges, including 1) the computational cost of generating continuations, 2) the fact that the LM does not initially know how to generate or use internal thoughts, and 3) the need to predict beyond individual next tokens. To resolve these, we propose a tokenwise parallel sampling algorithm, using learnable tokens indicating a thought's start and end, and an extended teacher-forcing technique. Encouragingly, generated rationales disproportionately help model difficult-to-predict tokens and improve the LM's ability to directly answer difficult questions. In particular, after continued pretraining of an LM on a corpus of internet text with Quiet-STaR, we find zero-shot improvements on GSM8K (5.9%$\rightarrow$10.9%) and CommonsenseQA (36.3%$\rightarrow$47.2%) and observe a perplexity improvement of difficult tokens in natural text. Crucially, these improvements require no fine-tuning on these tasks. Quiet-STaR marks a step towards LMs that can learn to reason in a more general and scalable way.

Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models

We study how to apply large language models to write grounded and organized long-form articles from scratch, with comparable breadth and depth to Wikipedia pages. This underexplored problem poses new challenges at the pre-writing stage, including how to research the topic and prepare an outline prior to writing. We propose STORM, a writing system for the Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking. STORM models the pre-writing stage by (1) discovering diverse perspectives in researching the given topic, (2) simulating conversations where writers carrying different perspectives pose questions to a topic expert grounded on trusted Internet sources, (3) curating the collected information to create an outline. For evaluation, we curate FreshWiki, a dataset of recent high-quality Wikipedia articles, and formulate outline assessments to evaluate the pre-writing stage. We further gather feedback from experienced Wikipedia editors. Compared to articles generated by an outline-driven retrieval-augmented baseline, more of STORM's articles are deemed to be organized (by a 25% absolute increase) and broad in coverage (by 10%). The expert feedback also helps identify new challenges for generating grounded long articles, such as source bias transfer and over-association of unrelated facts.

Class Incremental Learning via Likelihood Ratio Based Task Prediction

Class incremental learning (CIL) is a challenging setting of continual learning, which learns a series of tasks sequentially. Each task consists of a set of unique classes. The key feature of CIL is that no task identifier (or task-id) is provided at test time for each test sample. Predicting the task-id for each test sample is a challenging problem. An emerging theoretically justified and effective approach is to train a task-specific model for each task in a shared network for all tasks based on a task-incremental learning (TIL) method to deal with forgetting. The model for each task in this approach is an out-of-distribution (OOD) detector rather than a conventional classifier. The OOD detector can perform both within-task (in-distribution (IND)) class prediction and OOD detection. The OOD detection capability is the key for task-id prediction during inference for each test sample. However, this paper argues that using a traditional OOD detector for task-id prediction is sub-optimal because additional information (e.g., the replay data and the learned tasks) available in CIL can be exploited to design a better and principled method for task-id prediction. We call the new method TPLR (Task-id Prediction based on Likelihood Ratio}). TPLR markedly outperforms strong CIL baselines.

Class-Incremental Learning based on Label Generation

Despite the great success of pre-trained language models, it is still a challenge to use these models for continual learning, especially for the class-incremental learning (CIL) setting due to catastrophic forgetting (CF). This paper reports our finding that if we formulate CIL as a continual label generation problem, CF is drastically reduced and the generalizable representations of pre-trained models can be better retained. We thus propose a new CIL method (VAG) that also leverages the sparsity of vocabulary to focus the generation and creates pseudo-replay samples by using label semantics. Experimental results show that VAG outperforms baselines by a large margin.

ACCENT: An Automatic Event Commonsense Evaluation Metric for Open-Domain Dialogue Systems

Commonsense reasoning is omnipresent in human communications and thus is an important feature for open-domain dialogue systems. However, evaluating commonsense in dialogue systems is still an open challenge. We take the first step by focusing on event commonsense that considers events and their relations, and is crucial in both dialogues and general commonsense reasoning. We propose ACCENT, an event commonsense evaluation metric empowered by commonsense knowledge bases (CSKBs). ACCENT first extracts event-relation tuples from a dialogue, and then evaluates the response by scoring the tuples in terms of their compatibility with the CSKB. To evaluate ACCENT, we construct the first public event commonsense evaluation dataset for open-domain dialogues. Our experiments show that ACCENT is an efficient metric for event commonsense evaluation, which achieves higher correlations with human judgments than existing baselines.

Continual Learning of Language Models

Language models (LMs) have been instrumental for the rapid advance of natural language processing. This paper studies continual learning of LMs, in particular, continual domain-adaptive pre-training (or continual DAP-training). Existing research has shown that further pre-training an LM using a domain corpus to adapt the LM to the domain can improve the end-task performance in the domain. This paper proposes a novel method to continually DAP-train an LM with a sequence of unlabeled domain corpora to adapt the LM to these domains to improve their end-task performances. The key novelty of our method is a soft-masking mechanism that directly controls the update to the LM. A novel proxy is also proposed to preserve the general knowledge in the original LM. Additionally, it contrasts the representations of the previously learned domain knowledge (including the general knowledge in the pre-trained LM) and the knowledge from the current full network to achieve knowledge integration. The method not only overcomes catastrophic forgetting, but also achieves knowledge transfer to improve end-task performances. Empirical evaluation demonstrates the effectiveness of the proposed method.

Adapting a Language Model While Preserving its General Knowledge

Domain-adaptive pre-training (or DA-training for short), also known as post-training, aims to train a pre-trained general-purpose language model (LM) using an unlabeled corpus of a particular domain to adapt the LM so that end-tasks in the domain can give improved performances. However, existing DA-training methods are in some sense blind as they do not explicitly identify what knowledge in the LM should be preserved and what should be changed by the domain corpus. This paper shows that the existing methods are suboptimal and proposes a novel method to perform a more informed adaptation of the knowledge in the LM by (1) soft-masking the attention heads based on their importance to best preserve the general knowledge in the LM and (2) contrasting the representations of the general and the full (both general and domain knowledge) to learn an integrated representation with both general and domain-specific knowledge. Experimental results will demonstrate the effectiveness of the proposed approach.

LUNA: Language Understanding with Number Augmentations on Transformers via Number Plugins and Pre-training

Transformers are widely used in NLP tasks. However, current approaches to leveraging transformers to understand language expose one weak spot: Number understanding. In some scenarios, numbers frequently occur, especially in semi-structured data like tables. But current approaches to rich-number tasks with transformer-based language models abandon or lose some of the numeracy information - e.g., breaking numbers into sub-word tokens - which leads to many number-related errors. In this paper, we propose the LUNA framework which improves the numerical reasoning and calculation capabilities of transformer-based language models. With the number plugin of NumTok and NumBed, LUNA represents each number as a whole to model input. With number pre-training, including regression loss and model distillation, LUNA bridges the gap between number and vocabulary embeddings. To the best of our knowledge, this is the first work that explicitly injects numeracy capability into language models using Number Plugins. Besides evaluating toy models on toy tasks, we evaluate LUNA on three large-scale transformer models (RoBERTa, BERT, TabBERT) over three different downstream tasks (TATQA, TabFact, CrediTrans), and observe the performances of language models are constantly improved by LUNA. The augmented models also improve the official baseline of TAT-QA (EM: 50.15 -> 59.58) and achieve SOTA performance on CrediTrans (F1 = 86.17).