John Hewitt

Nelson F. Liu, Kevin Lin, John Hewitt et al. · stanford

While recent language models have the ability to take long contexts as input, relatively little is known about how well they use longer context. We analyze the performance of language models on two tasks that require identifying relevant information in their input contexts: multi-document question answering and key-value retrieval. We find that performance can degrade significantly when changing the position of relevant information, indicating that current language models do not robustly make use of information in long input contexts. In particular, we observe that performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models. Our analysis provides a better understanding of how language models use their input context and provides new evaluation protocols for future long-context language models.

John Hewitt, Christopher D. Manning, Percy Liang · stanford

Long samples of text from neural language models can be of poor quality. Truncation sampling algorithms--like top-$p$ or top-$k$ -- address this by setting some words' probabilities to zero at each step. This work provides framing for the aim of truncation, and an improved algorithm for that aim. We propose thinking of a neural language model as a mixture of a true distribution and a smoothing distribution that avoids infinite perplexity. In this light, truncation algorithms aim to perform desmoothing, estimating a subset of the support of the true distribution. Finding a good subset is crucial: we show that top-$p$ unnecessarily truncates high-probability words, for example causing it to truncate all words but Trump for a document that starts with Donald. We introduce $η$-sampling, which truncates words below an entropy-dependent probability threshold. Compared to previous algorithms, $η$-sampling generates more plausible long English documents according to humans, is better at breaking out of repetition, and behaves more reasonably on a battery of test distributions.

Ruth-Ann Armstrong, John Hewitt, Christopher Manning · stanford

JamPatoisNLI provides the first dataset for natural language inference in a creole language, Jamaican Patois. Many of the most-spoken low-resource languages are creoles. These languages commonly have a lexicon derived from a major world language and a distinctive grammar reflecting the languages of the original speakers and the process of language birth by creolization. This gives them a distinctive place in exploring the effectiveness of transfer from large monolingual or multilingual pretrained models. While our work, along with previous work, shows that transfer from these models to low-resource languages that are unrelated to languages in their training set is not very effective, we would expect stronger results from transfer to creoles. Indeed, our experiments show considerably better results from few-shot learning of JamPatoisNLI than for such unrelated languages, and help us begin to understand how the unique relationship between creoles and their high-resource base languages affect cross-lingual transfer. JamPatoisNLI, which consists of naturally-occurring premises and expert-written hypotheses, is a step towards steering research into a traditionally underserved language and a useful benchmark for understanding cross-lingual NLP.

John Hewitt, Nelson F. Liu, Percy Liang et al. · stanford

Instruction tuning commonly means finetuning a language model on instruction-response pairs. We discover two forms of adaptation (tuning) that are deficient compared to instruction tuning, yet still yield instruction following; we call this implicit instruction tuning. We first find that instruction-response pairs are not necessary: training solely on responses, without any corresponding instructions, yields instruction following. This suggests pretrained models have an instruction-response mapping which is revealed by teaching the model the desired distribution of responses. However, we then find it's not necessary to teach the desired distribution of responses: instruction-response training on narrow-domain data like poetry still leads to broad instruction-following behavior like recipe generation. In particular, when instructions are very different from those in the narrow finetuning domain, models' responses do not adhere to the style of the finetuning domain. To begin to explain implicit instruction tuning, we hypothesize that very simple changes to a language model's distribution yield instruction following. We support this by hand-writing a rule-based language model which yields instruction following in a product-of-experts with a pretrained model. The rules are to slowly increase the probability of ending the sequence, penalize repetition, and uniformly change 15 words' probabilities. In summary, adaptations made without being designed to yield instruction following can do so implicitly.

Matthew Finlayson, John Hewitt, Alexander Koller et al.

Despite their ubiquity in language generation, it remains unknown why truncation sampling heuristics like nucleus sampling are so effective. We provide a theoretical explanation for the effectiveness of the truncation sampling by proving that truncation methods that discard tokens below some probability threshold (the most common type of truncation) can guarantee that all sampled tokens have nonzero true probability. However, thresholds are a coarse heuristic, and necessarily discard some tokens with nonzero true probability as well. In pursuit of a more precise sampling strategy, we show that we can leverage a known source of model errors, the softmax bottleneck, to prove that certain tokens have nonzero true probability, without relying on a threshold. Based on our findings, we develop an experimental truncation strategy and the present pilot studies demonstrating the promise of this type of algorithm. Our evaluations show that our method outperforms its threshold-based counterparts under automatic and human evaluation metrics for low-entropy (i.e., close to greedy) open-ended text generation. Our theoretical findings and pilot experiments provide both insight into why truncation sampling works, and make progress toward more expressive sampling algorithms that better surface the generative capabilities of large language models.

Melody Ma, John Hewitt

We study reasoning for accessing world knowledge stored in a language model's parameters. For example, recalling that Canberra is Australia's capital may benefit from thinking through major cities and the concept of purpose-built capitals. While reasoning language models are trained via reinforcement learning to produce reasoning traces on tasks such as mathematics, they may not reason well for accessing their own world knowledge. We first find that models do not generate their best world knowledge reasoning by default: adding a simple "think step-by-step" cue demonstrates statistically significant improvement in knowledge recall but not math. Motivated by this, we propose training models to reason over their parametric knowledge using world-knowledge question answering as a verifiable reward. After reinforcement learning on TriviaQA (+9.9%), performance also improves on Natural Questions, HotpotQA, SimpleQA, and StrategyQA by 4.2%, 2.1%, 0.6%, and 3.0%, respectively. Reasoning models are under-optimized for parametric knowledge access, but can be easily trained to reason better.

Graham Neubig, Matthias Sperber, Xinyi Wang et al.

This paper describes XNMT, the eXtensible Neural Machine Translation toolkit. XNMT distin- guishes itself from other open-source NMT toolkits by its focus on modular code design, with the purpose of enabling fast iteration in research and replicable, reliable results. In this paper we describe the design of XNMT and its experiment configuration system, and demonstrate its utility on the tasks of machine translation, speech recognition, and multi-tasked machine translation/parsing. XNMT is available open-source at https://github.com/neulab/xnmt

Hao Sun, John Hewitt

The Backpack is a Transformer alternative shown to improve interpretability in English language modeling by decomposing predictions into a weighted sum of token sense components. However, Backpacks' reliance on token-defined meaning raises questions as to their potential for languages other than English, a language for which subword tokenization provides a reasonable approximation for lexical items. In this work, we train, evaluate, interpret, and control Backpack language models in character-tokenized Chinese, in which words are often composed of many characters. We find that our (134M parameter) Chinese Backpack language model performs comparably to a (104M parameter) Transformer, and learns rich character-level meanings that log-additively compose to form word meanings. In SimLex-style lexical semantic evaluations, simple averages of Backpack character senses outperform input embeddings from a Transformer. We find that complex multi-character meanings are often formed by using the same per-character sense weights consistently across context. Exploring interpretability-through control, we show that we can localize a source of gender bias in our Backpacks to specific character senses and intervene to reduce the bias.

George Morgulis, John Hewitt

Subliminal learning describes a student language model inheriting a behavioral bias by fine-tuning on seemingly innocuous data generated by a biased teacher model. Prior work has begun to characterize this phenomenon but leaves open questions about the scope of signals it can transfer, the mechanisms that explain it, and the precision with which a bias can be encoded by seemingly unrelated data. We tackle all three problems by introducing subliminal steering, a variant of subliminal learning in which the teacher's bias is implemented not via a system prompt, as in prior work, but through a steering vector trained to maximize the likelihood of a set of target samples. First, we show that subliminal steering transfers complex multi-word biases, whereas prior work focused on single-word preferences, demonstrating a large scope of subliminally transferrable signals. Second, we provide mechanistic evidence that subliminal learning transfers not only the target behavioral bias, but also the steering vector itself, localized to the layers at which the teacher was steered. Finally, we show that the bias is encoded with surprising precision. We train a new steering vector directly on the subliminally-laden dataset and find that it attains high cosine similarity with the original vector.

John Hewitt, Robert Geirhos, Been Kim

This position paper argues that, in order to understand AI, we cannot rely on our existing vocabulary of human words. Instead, we should strive to develop neologisms: new words that represent precise human concepts that we want to teach machines, or machine concepts that we need to learn. We start from the premise that humans and machines have differing concepts. This means interpretability can be framed as a communication problem: humans must be able to reference and control machine concepts, and communicate human concepts to machines. Creating a shared human-machine language through developing neologisms, we believe, could solve this communication problem. Successful neologisms achieve a useful amount of abstraction: not too detailed, so they're reusable in many contexts, and not too high-level, so they convey precise information. As a proof of concept, we demonstrate how a "length neologism" enables controlling LLM response length, while a "diversity neologism" allows sampling more variable responses. Taken together, we argue that we cannot understand AI using our existing vocabulary, and expanding it through neologisms creates opportunities for both controlling and understanding machines better.

John Hewitt, Sarah Chen, Lanruo Lora Xie et al. · stanford

We introduce model editing with canonical examples, a setting in which (1) a single learning example is provided per desired behavior, (2) evaluation is performed exclusively out-of-distribution, and (3) deviation from an initial model is strictly limited. A canonical example is a simple instance of good behavior, e.g., The capital of Mauritius is Port Louis) or bad behavior, e.g., An aspect of researchers is coldhearted). The evaluation set contains more complex examples of each behavior (like a paragraph in which the capital of Mauritius is called for.) We create three datasets and modify three more for model editing with canonical examples, covering knowledge-intensive improvements, social bias mitigation, and syntactic edge cases. In our experiments on Pythia language models, we find that LoRA outperforms full finetuning and MEMIT. We then turn to the Backpack language model architecture because it is intended to enable targeted improvement. The Backpack defines a large bank of sense vectors--a decomposition of the different uses of each word--which are weighted and summed to form the output logits of the model. We propose sense finetuning, which selects and finetunes a few ($\approx$ 10) sense vectors for each canonical example, and find that it outperforms other finetuning methods, e.g., 4.8% improvement vs 0.3%. Finally, we improve GPT-J-6B by an inference-time ensemble with just the changes from sense finetuning of a 35x smaller Backpack, in one setting outperforming editing GPT-J itself (4.1% vs 1.0%).

Been Kim, John Hewitt, Neel Nanda et al.

The era of Large Language Models (LLMs) presents a new opportunity for interpretability--agentic interpretability: a multi-turn conversation with an LLM wherein the LLM proactively assists human understanding by developing and leveraging a mental model of the user, which in turn enables humans to develop better mental models of the LLM. Such conversation is a new capability that traditional `inspective' interpretability methods (opening the black-box) do not use. Having a language model that aims to teach and explain--beyond just knowing how to talk--is similar to a teacher whose goal is to teach well, understanding that their success will be measured by the student's comprehension. While agentic interpretability may trade off completeness for interactivity, making it less suitable for high-stakes safety situations with potentially deceptive models, it leverages a cooperative model to discover potentially superhuman concepts that can improve humans' mental model of machines. Agentic interpretability introduces challenges, particularly in evaluation, due to what we call `human-entangled-in-the-loop' nature (humans responses are integral part of the algorithm), making the design and evaluation difficult. We discuss possible solutions and proxy goals. As LLMs approach human parity in many tasks, agentic interpretability's promise is to help humans learn the potentially superhuman concepts of the LLMs, rather than see us fall increasingly far from understanding them.

John Hewitt, Oyvind Tafjord, Robert Geirhos et al.

Humans invent new words when there is a rising demand for a new useful concept (e.g., doomscrolling). We explore and validate a similar idea in our communication with LLMs: introducing new words to better understand and control the models, expanding on the recently introduced neologism learning. This method introduces a new word by adding a new word embedding and training with examples that exhibit the concept with no other changes in model parameters. We show that adding a new word allows for control of concepts such as flattery, incorrect answers, text length, as well as more complex concepts in AxBench. We discover that neologisms can also further our understanding of the model via self-verbalization: models can describe what each new word means to them in natural language, like explaining that a word that represents a concept of incorrect answers means ``a lack of complete, coherent, or meaningful answers...'' To validate self-verbalizations, we introduce plug-in evaluation: we insert the verbalization into the context of a model and measure whether it controls the target concept. In some self-verbalizations, we find machine-only synonyms: words that seem unrelated to humans but cause similar behavior in machines. Finally, we show how neologism learning can jointly learn multiple concepts in multiple words.

Derry Wijaya, Brendan Callahan, John Hewitt et al.

Bilingual Lexicon Induction is the task of learning word translations without bilingual parallel corpora. We model this task as a matrix completion problem, and present an effective and extendable framework for completing the matrix. This method harnesses diverse bilingual and monolingual signals, each of which may be incomplete or noisy. Our model achieves state-of-the-art performance for both high and low resource languages.

John Hewitt, John Thickstun, Christopher D. Manning et al.

We present Backpacks: a new neural architecture that marries strong modeling performance with an interface for interpretability and control. Backpacks learn multiple non-contextual sense vectors for each word in a vocabulary, and represent a word in a sequence as a context-dependent, non-negative linear combination of sense vectors in this sequence. We find that, after training, sense vectors specialize, each encoding a different aspect of a word. We can interpret a sense vector by inspecting its (non-contextual, linear) projection onto the output space, and intervene on these interpretable hooks to change the model's behavior in predictable ways. We train a 170M-parameter Backpack language model on OpenWebText, matching the loss of a GPT-2 small (124Mparameter) Transformer. On lexical similarity evaluations, we find that Backpack sense vectors outperform even a 6B-parameter Transformer LM's word embeddings. Finally, we present simple algorithms that intervene on sense vectors to perform controllable text generation and debiasing. For example, we can edit the sense vocabulary to tend more towards a topic, or localize a source of gender bias to a sense vector and globally suppress that sense.

John Hewitt, Kawin Ethayarajh, Percy Liang et al.

Probing experiments investigate the extent to which neural representations make properties -- like part-of-speech -- predictable. One suggests that a representation encodes a property if probing that representation produces higher accuracy than probing a baseline representation like non-contextual word embeddings. Instead of using baselines as a point of comparison, we're interested in measuring information that is contained in the representation but not in the baseline. For example, current methods can detect when a representation is more useful than the word identity (a baseline) for predicting part-of-speech; however, they cannot detect when the representation is predictive of just the aspects of part-of-speech not explainable by the word identity. In this work, we extend a theory of usable information called $\mathcal{V}$-information and propose conditional probing, which explicitly conditions on the information in the baseline. In a case study, we find that after conditioning on non-contextual word embeddings, properties like part-of-speech are accessible at deeper layers of a network than previously thought.

Rishi Bommasani, Drew A. Hudson, Ehsan Adeli et al.

AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.

Benjamin Newman, Kai-Siang Ang, Julia Gong et al.

Targeted syntactic evaluation of subject-verb number agreement in English (TSE) evaluates language models' syntactic knowledge using hand-crafted minimal pairs of sentences that differ only in the main verb's conjugation. The method evaluates whether language models rate each grammatical sentence as more likely than its ungrammatical counterpart. We identify two distinct goals for TSE. First, evaluating the systematicity of a language model's syntactic knowledge: given a sentence, can it conjugate arbitrary verbs correctly? Second, evaluating a model's likely behavior: given a sentence, does the model concentrate its probability mass on correctly conjugated verbs, even if only on a subset of the possible verbs? We argue that current implementations of TSE do not directly capture either of these goals, and propose new metrics to capture each goal separately. Under our metrics, we find that TSE overestimates systematicity of language models, but that models score up to 40% better on verbs that they predict are likely in context.

Anna A. Ivanova, John Hewitt, Noga Zaslavsky

A major challenge in both neuroscience and machine learning is the development of useful tools for understanding complex information processing systems. One such tool is probes, i.e., supervised models that relate features of interest to activation patterns arising in biological or artificial neural networks. Neuroscience has paved the way in using such models through numerous studies conducted in recent decades. In this work, we draw insights from neuroscience to help guide probing research in machine learning. We highlight two important design choices for probes $-$ direction and expressivity $-$ and relate these choices to research goals. We argue that specific research goals play a paramount role when designing a probe and encourage future probing studies to be explicit in stating these goals.

John Hewitt, Michael Hahn, Surya Ganguli et al.

Recurrent neural networks empirically generate natural language with high syntactic fidelity. However, their success is not well-understood theoretically. We provide theoretical insight into this success, proving in a finite-precision setting that RNNs can efficiently generate bounded hierarchical languages that reflect the scaffolding of natural language syntax. We introduce Dyck-($k$,$m$), the language of well-nested brackets (of $k$ types) and $m$-bounded nesting depth, reflecting the bounded memory needs and long-distance dependencies of natural language syntax. The best known results use $O(k^{\frac{m}{2}})$ memory (hidden units) to generate these languages. We prove that an RNN with $O(m \log k)$ hidden units suffices, an exponential reduction in memory, by an explicit construction. Finally, we show that no algorithm, even with unbounded computation, can suffice with $o(m \log k)$ hidden units.

Benjamin Newman, John Hewitt, Percy Liang et al.

Extrapolation to unseen sequence lengths is a challenge for neural generative models of language. In this work, we characterize the effect on length extrapolation of a modeling decision often overlooked: predicting the end of the generative process through the use of a special end-of-sequence (EOS) vocabulary item. We study an oracle setting - forcing models to generate to the correct sequence length at test time - to compare the length-extrapolative behavior of networks trained to predict EOS (+EOS) with networks not trained to (-EOS). We find that -EOS substantially outperforms +EOS, for example extrapolating well to lengths 10 times longer than those seen at training time in a bracket closing task, as well as achieving a 40% improvement over +EOS in the difficult SCAN dataset length generalization task. By comparing the hidden states and dynamics of -EOS and +EOS models, we observe that +EOS models fail to generalize because they (1) unnecessarily stratify their hidden states by their linear position is a sequence (structures we call length manifolds) or (2) get stuck in clusters (which we refer to as length attractors) once the EOS token is the highest-probability prediction.

Ethan A. Chi, John Hewitt, Christopher D. Manning

Recent work has found evidence that Multilingual BERT (mBERT), a transformer-based multilingual masked language model, is capable of zero-shot cross-lingual transfer, suggesting that some aspects of its representations are shared cross-lingually. To better understand this overlap, we extend recent work on finding syntactic trees in neural networks' internal representations to the multilingual setting. We show that subspaces of mBERT representations recover syntactic tree distances in languages other than English, and that these subspaces are approximately shared across languages. Motivated by these results, we present an unsupervised analysis method that provides evidence mBERT learns representations of syntactic dependency labels, in the form of clusters which largely agree with the Universal Dependencies taxonomy. This evidence suggests that even without explicit supervision, multilingual masked language models learn certain linguistic universals.

John Hewitt, Percy Liang

Probes, supervised models trained to predict properties (like parts-of-speech) from representations (like ELMo), have achieved high accuracy on a range of linguistic tasks. But does this mean that the representations encode linguistic structure or just that the probe has learned the linguistic task? In this paper, we propose control tasks, which associate word types with random outputs, to complement linguistic tasks. By construction, these tasks can only be learned by the probe itself. So a good probe, (one that reflects the representation), should be selective, achieving high linguistic task accuracy and low control task accuracy. The selectivity of a probe puts linguistic task accuracy in context with the probe's capacity to memorize from word types. We construct control tasks for English part-of-speech tagging and dependency edge prediction, and show that popular probes on ELMo representations are not selective. We also find that dropout, commonly used to control probe complexity, is ineffective for improving selectivity of MLPs, but that other forms of regularization are effective. Finally, we find that while probes on the first layer of ELMo yield slightly better part-of-speech tagging accuracy than the second, probes on the second layer are substantially more selective, which raises the question of which layer better represents parts-of-speech.

Arshit Gupta, John Hewitt, Katrin Kirchhoff

With the advent of conversational assistants, like Amazon Alexa, Google Now, etc., dialogue systems are gaining a lot of traction, especially in industrial setting. These systems typically consist of Spoken Language understanding component which, in turn, consists of two tasks - Intent Classification (IC) and Slot Labeling (SL). Generally, these two tasks are modeled together jointly to achieve best performance. However, this joint modeling adds to model obfuscation. In this work, we first design framework for a modularization of joint IC-SL task to enhance architecture transparency. Then, we explore a number of self-attention, convolutional, and recurrent models, contributing a large-scale analysis of modeling paradigms for IC+SL across two datasets. Finally, using this framework, we propose a class of 'label-recurrent' models that otherwise non-recurrent, with a 10-dimensional representation of the label history, and show that our proposed systems are easy to interpret, highly accurate (achieving over 30% error reduction in SL over the state-of-the-art on the Snips dataset), as well as fast, at 2x the inference and 2/3 to 1/2 the training time of comparable recurrent models, thus giving an edge in critical real-world systems.