Sebastian Schuster

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Aaron Mueller, Robert Frank, Tal Linzen et al.

Relations between words are governed by hierarchical structure rather than linear ordering. Sequence-to-sequence (seq2seq) models, despite their success in downstream NLP applications, often fail to generalize in a hierarchy-sensitive manner when performing syntactic transformations - for example, transforming declarative sentences into questions. However, syntactic evaluations of seq2seq models have only observed models that were not pre-trained on natural language data before being trained to perform syntactic transformations, in spite of the fact that pre-training has been found to induce hierarchical linguistic generalizations in language models; in other words, the syntactic capabilities of seq2seq models may have been greatly understated. We address this gap using the pre-trained seq2seq models T5 and BART, as well as their multilingual variants mT5 and mBART. We evaluate whether they generalize hierarchically on two transformations in two languages: question formation and passivization in English and German. We find that pre-trained seq2seq models generalize hierarchically when performing syntactic transformations, whereas models trained from scratch on syntactic transformations do not. This result presents evidence for the learnability of hierarchical syntactic information from non-annotated natural language text while also demonstrating that seq2seq models are capable of syntactic generalization, though only after exposure to much more language data than human learners receive.

Sebastian Schuster, Tal Linzen

Understanding longer narratives or participating in conversations requires tracking of discourse entities that have been mentioned. Indefinite noun phrases (NPs), such as 'a dog', frequently introduce discourse entities but this behavior is modulated by sentential operators such as negation. For example, 'a dog' in 'Arthur doesn't own a dog' does not introduce a discourse entity due to the presence of negation. In this work, we adapt the psycholinguistic assessment of language models paradigm to higher-level linguistic phenomena and introduce an English evaluation suite that targets the knowledge of the interactions between sentential operators and indefinite NPs. We use this evaluation suite for a fine-grained investigation of the entity tracking abilities of the Transformer-based models GPT-2 and GPT-3. We find that while the models are to a certain extent sensitive to the interactions we investigate, they are all challenged by the presence of multiple NPs and their behavior is not systematic, which suggests that even models at the scale of GPT-3 do not fully acquire basic entity tracking abilities.

Jennifer Hu, Roger Levy, Judith Degen et al.

Scalar inferences (SI) are a signature example of how humans interpret language based on unspoken alternatives. While empirical studies have demonstrated that human SI rates are highly variable -- both within instances of a single scale, and across different scales -- there have been few proposals that quantitatively explain both cross- and within-scale variation. Furthermore, while it is generally assumed that SIs arise through reasoning about unspoken alternatives, it remains debated whether humans reason about alternatives as linguistic forms, or at the level of concepts. Here, we test a shared mechanism explaining SI rates within and across scales: context-driven expectations about the unspoken alternatives. Using neural language models to approximate human predictive distributions, we find that SI rates are captured by the expectedness of the strong scalemate as an alternative. Crucially, however, expectedness robustly predicts cross-scale variation only under a meaning-based view of alternatives. Our results suggest that pragmatic inferences arise from context-driven expectations over alternatives, and these expectations operate at the level of concepts.

Zilu Tang, Qiao Zhao, Gabriel Franco et al.

Entity tracking (ET), the ability to keep track of states, is a fundamental skill that underlies complex reasoning. An increasing amount of work investigates how transformer language models (LMs) solve entity binding $\textit{without}$ state changes. However, there is limited understanding of how non-toy LMs address ET problems of realistic difficulties expressed in natural language. To this end, we investigate the mechanisms underlying ET in more complex scenarios featuring multiple state-changing operations. We find that LMs do not incrementally track world states across tokens or query-relevant states across layers, but simply aggregate relevant information in parallel at the last token when the query becomes evident. We further investigate mechanisms of individual operations ($\texttt{PUT}$, $\texttt{REMOVE}$, $\texttt{MOVE}$) to characterize this non-incremental ET mechanism. Surprisingly, LMs implement the $\texttt{REMOVE}$ operation with a fragile global suppression tag; this global removal mechanism predicts various failure modes that we confirm behaviorally. We provide a mechanistic solution of nullifying this tag to partially address this issue. Overall, our findings reveal that LMs solve a fundamentally sequential task using a non-sequential strategy. More broadly, our work illustrates how behavioral and mechanistic analyses can fruitfully interact. Behavioral results inform mechanistic hypotheses, and insights from mechanistic analyses help build stronger behavioral evaluations by predicting failure modes missing from existing evaluations.

Gaurav Kamath, Sreenath Madathil, Sebastian Schuster et al.

Human reasoning often involves working over limited information to arrive at probabilistic conclusions. In its simplest form, this involves making an inference that is not strictly entailed by a premise, but rather only likely given the premise. While reasoning LLMs have demonstrated strong performance on logical and mathematical tasks, their behavior on such open-ended, non-deterministic inferences remains largely unexplored. We introduce ProbCOPA, a dataset of 210 handcrafted probabilistic inferences in English, each annotated for inference likelihood by 25--30 human participants. We find that human responses are graded and varied, revealing probabilistic judgments of the inferences in our dataset. Comparing these judgments with responses from eight state-of-the-art reasoning LLMs, we show that models consistently fail to produce human-like distributions. Finally, analyzing LLM reasoning chains, we find evidence of a common reasoning pattern used to evaluate such inferences. Our findings reveal persistent differences between humans and LLMs, and underscore the need to evaluate reasoning beyond deterministic settings.

Nicholas Edwards, Sebastian Schuster

As Large Language Model (LLM) agents are increasingly deployed in open-ended domains like software engineering, they frequently encounter underspecified instructions that lack crucial context. While human developers naturally resolve underspecification by asking clarifying questions, current agents are largely optimized for autonomous execution. In this work, we systematically evaluate the clarification-seeking abilities of LLM agents on an underspecified variant of SWE-bench Verified. We propose an uncertainty-aware multi-agent scaffold that explicitly decouples underspecification detection from code execution. Our results demonstrate that this multi-agent system using OpenHands + Claude Sonnet 4.5 achieves a 69.40% task resolve rate, significantly outperforming a standard single-agent setup (61.20%) and closing the performance gap with agents operating on fully specified instructions. Furthermore, we find that the multi-agent system exhibits well-calibrated uncertainty, conserving queries on simple tasks while proactively seeking information on more complex issues. These findings indicate that current models can be turned into proactive collaborators, where agents independently recognize when to ask questions to elicit missing information in real-world, underspecified tasks.

Gaurav Kamath, Sebastian Schuster, Sowmya Vajjala et al.

Sentences containing multiple semantic operators with overlapping scope often create ambiguities in interpretation, known as scope ambiguities. These ambiguities offer rich insights into the interaction between semantic structure and world knowledge in language processing. Despite this, there has been little research into how modern large language models treat them. In this paper, we investigate how different versions of certain autoregressive language models -- GPT-2, GPT-3/3.5, Llama 2 and GPT-4 -- treat scope ambiguous sentences, and compare this with human judgments. We introduce novel datasets that contain a joint total of almost 1,000 unique scope-ambiguous sentences, containing interactions between a range of semantic operators, and annotated for human judgments. Using these datasets, we find evidence that several models (i) are sensitive to the meaning ambiguity in these sentences, in a way that patterns well with human judgments, and (ii) can successfully identify human-preferred readings at a high level of accuracy (over 90% in some cases).

Nicholas Edwards, Yukyung Lee, Yujun Audrey Mao et al.

Agents based on Large Language Models (LLMs) have shown promise for performing sophisticated software engineering tasks autonomously. In addition, there has been progress towards developing agents that can perform parts of the research pipeline in machine learning and the natural sciences. We argue that research extension and its implementation is a critical capability for such systems, and introduce RExBench to support the evaluation of this capability. RExBench is a benchmark consisting of 12 realistic research experiment implementation tasks that aim to investigate research hypotheses that have not previously been implemented. Each task is set up as an extension to an existing research paper and codebase, accompanied by domain expert-written instructions. RExBench is robust to data contamination, and supports an automatic evaluation infrastructure that executes agent outputs to determine whether the success criteria are met. We use this benchmark to evaluate nine LLM agents implemented using three different frameworks: aider, Claude Code, and OpenHands. We find that all agents evaluated fail to autonomously implement the majority of the extensions. Although the success rate improves with additional human-written hints, the best performance under this setting remains below 40%. This indicates that current agents are still short of being able to handle realistic research extension tasks without substantial human guidance.

Sarah Breckner, Sebastian Schuster

Diffusion LLMs have been proposed as an alternative to autoregressive LLMs, excelling especially at complex reasoning tasks with interdependent sub-goals. Curiously, this is particularly true if the generation length, i.e., the number of tokens the model has to output, is set to a much higher value than is required for providing the correct answer to the task, and the model pads its answer with end-of-sequence (EoS) tokens. We hypothesize that diffusion models think EoS-by-EoS, that is, they use the representations of EoS tokens as a hidden scratchpad, which allows them to solve harder reasoning problems. We experiment with the diffusion models LLaDA1.5, LLaDA2.0-mini, and Dream-v0 on the tasks Addition, Entity Tracking, and Sudoku. In a controlled prompting experiment, we confirm that adding EoS tokens improves the LLMs' reasoning capabilities. To further verify whether they serve as space for hidden computations, we patch the hidden states of the EoS tokens with those of a counterfactual generation, which frequently changes the generated output to the counterfactual. The success of the causal intervention underscores that the EoS tokens, which one may expect to be devoid of meaning, carry information on the problem to solve. The behavioral experiments and the causal interventions indicate that diffusion LLMs can indeed think EoS-by-EoS.

Najoung Kim, Sebastian Schuster

Keeping track of how states of entities change as a text or dialog unfolds is a key prerequisite to discourse understanding. Yet, there have been few systematic investigations into the ability of large language models (LLMs) to track discourse entities. In this work, we present a task probing to what extent a language model can infer the final state of an entity given an English description of the initial state and a series of state-changing operations. We use this task to first investigate whether Flan-T5, GPT-3 and GPT-3.5 can track the state of entities, and find that only GPT-3.5 models, which have been pretrained on large amounts of code, exhibit this ability. We then investigate whether smaller models pretrained primarily on text can learn to track entities, through finetuning T5 on several training/evaluation splits. While performance degrades for more complex splits, we find that even when evaluated on a different set of entities from training or longer operation sequences, a finetuned model can perform non-trivial entity tracking. Taken together, these results suggest that language models can learn to track entities but pretraining on text corpora alone does not make this capacity surface.

Alicia Parrish, Sebastian Schuster, Alex Warstadt et al.

Understanding language requires grasping not only the overtly stated content, but also making inferences about things that were left unsaid. These inferences include presuppositions, a phenomenon by which a listener learns about new information through reasoning about what a speaker takes as given. Presuppositions require complex understanding of the lexical and syntactic properties that trigger them as well as the broader conversational context. In this work, we introduce the Naturally-Occurring Presuppositions in English (NOPE) Corpus to investigate the context-sensitivity of 10 different types of presupposition triggers and to evaluate machine learning models' ability to predict human inferences. We find that most of the triggers we investigate exhibit moderate variability. We further find that transformer-based models draw correct inferences in simple cases involving presuppositions, but they fail to capture the minority of exceptional cases in which human judgments reveal complex interactions between context and triggers.

Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter et al.

Universal Dependencies is an open community effort to create cross-linguistically consistent treebank annotation for many languages within a dependency-based lexicalist framework. The annotation consists in a linguistically motivated word segmentation; a morphological layer comprising lemmas, universal part-of-speech tags, and standardized morphological features; and a syntactic layer focusing on syntactic relations between predicates, arguments and modifiers. In this paper, we describe version 2 of the guidelines (UD v2), discuss the major changes from UD v1 to UD v2, and give an overview of the currently available treebanks for 90 languages.

Sebastian Schuster, Yuxing Chen, Judith Degen

Pragmatic inferences often subtly depend on the presence or absence of linguistic features. For example, the presence of a partitive construction (of the) increases the strength of a so-called scalar inference: listeners perceive the inference that Chris did not eat all of the cookies to be stronger after hearing "Chris ate some of the cookies" than after hearing the same utterance without a partitive, "Chris ate some cookies." In this work, we explore to what extent neural network sentence encoders can learn to predict the strength of scalar inferences. We first show that an LSTM-based sentence encoder trained on an English dataset of human inference strength ratings is able to predict ratings with high accuracy (r=0.78). We then probe the model's behavior using manually constructed minimal sentence pairs and corpus data. We find that the model inferred previously established associations between linguistic features and inference strength, suggesting that the model learns to use linguistic features to predict pragmatic inferences.

Sebastian Schuster, Sonal Gupta, Rushin Shah et al.

One of the first steps in the utterance interpretation pipeline of many task-oriented conversational AI systems is to identify user intents and the corresponding slots. Since data collection for machine learning models for this task is time-consuming, it is desirable to make use of existing data in a high-resource language to train models in low-resource languages. However, development of such models has largely been hindered by the lack of multilingual training data. In this paper, we present a new data set of 57k annotated utterances in English (43k), Spanish (8.6k) and Thai (5k) across the domains weather, alarm, and reminder. We use this data set to evaluate three different cross-lingual transfer methods: (1) translating the training data, (2) using cross-lingual pre-trained embeddings, and (3) a novel method of using a multilingual machine translation encoder as contextual word representations. We find that given several hundred training examples in the the target language, the latter two methods outperform translating the training data. Further, in very low-resource settings, multilingual contextual word representations give better results than using cross-lingual static embeddings. We also compare the cross-lingual methods to using monolingual resources in the form of contextual ELMo representations and find that given just small amounts of target language data, this method outperforms all cross-lingual methods, which highlights the need for more sophisticated cross-lingual methods.

Sebastian Schuster, Joakim Nivre, Christopher D. Manning

Sentences with gapping, such as Paul likes coffee and Mary tea, lack an overt predicate to indicate the relation between two or more arguments. Surface syntax representations of such sentences are often produced poorly by parsers, and even if correct, not well suited to downstream natural language understanding tasks such as relation extraction that are typically designed to extract information from sentences with canonical clause structure. In this paper, we present two methods for parsing to a Universal Dependencies graph representation that explicitly encodes the elided material with additional nodes and edges. We find that both methods can reconstruct elided material from dependency trees with high accuracy when the parser correctly predicts the existence of a gap. We further demonstrate that one of our methods can be applied to other languages based on a case study on Swedish.