A recent study (Kuribayashi et al., 2025) has shown that human sentence processing behavior, typically measured on syntactically unchallenging constructions, can be effectively modeled using surprisal from early layers of large language models (LLMs).This raises the question of whether such advantages of internal layers extend to more syntactically challenging constructions, where surprisal has been reported to underestimate human cognitive effort.In this paper, we begin by exploring internal layers that better estimate human cognitive effort observed in syntactic ambiguity processing in English.Our experiments show that, in contrast to naturalistic reading, later layers better estimate such a cognitive effort, but still underestimate the human data.This dual alignment sheds light on different modes of sentence processing in humans and LMs: naturalistic reading employs a somewhat weak prediction akin to earlier layers of LMs, while syntactically challenging processing requires more fully-contextualized representations, better modeled by later layers of LMs.Motivated by these findings, we also explore several probability-update measures using shallow and deep layers of LMs, showing a complementary advantage to single-layer's surprisal in reading time modeling. https://github.com/kuribayashi4/ internal_surprisal_targeted_assessmentPhenomena Example MVRR D + : The girl fed the lamb remained relatively calm before the sunset in silence.D -: The girl who was fed the lamb remained relatively calm before the sunset in silence.NPS D + : The girl found the lamb remained relatively calm near the wooden fence.D -: The girl found that the lamb remained relatively calm near the wooden fence.NPZ D + : When the girl attacked the lamb remained relatively calm despite the sudden noise.D -: When the girl attacked, the lamb remained relatively calm despite the sudden noise.RC D + : The bus driver that the kids followed waited patiently at dawn.D -: The bus driver that followed the kids waited patiently at dawn.Attachment D + : Janet charmed the executive of the assistants who decides almost everything during long weekly meetings.

Large Language Models (LLMs) remain substantially less data-efficient than humans.Prepretraining (PPT) on synthetic languages has been proposed to close this gap, with prior work emphasizing highly expressive formal languages such as k-Shuffle Dyck.Inspired by the Language Acquisition Device (LAD) hypothesis, which posits that innate constraints preemptively restrict the learner's hypothesis space to natural-language-like structure, we propose LAD-inspired PPT: pre-pretraining on MP-STRUCT, a formal language whose strings encode hierarchical composition, feature-based dependencies, and long-distance displacement via MERGE, AGREE, and MOVE.A brief 500step PPT with MP-STRUCT matches strong formal-language baselines in token efficiency while additionally imparting a human-like resistance to structurally implausible languages.Analyzing simplified variants, we find that MP-STRUCT CORE outperforms k-Shuffle Dyck despite not being definable in C-RASP (a formal bound on transformer expressivity), challenging the prior hypothesis that effective PPT languages must be both hierarchically expressive and circuit-theoretically learnable.We show that functional landmarks, which reduce dependency resolution ambiguity, are a key driver, suggesting that effective PPT design depends not only on expressivity but also on the accessibility of dependency resolution.

Surprisal theory hypothesizes that the difficulty of human sentence processing increases linearly with surprisal, the negative logprobability of a word given its context.Computational psycholinguistics has tested this hypothesis using language models (LMs) as proxies for human prediction.While surprisal derived from recent neural LMs generally captures human processing difficulty on naturalistic corpora that predominantly consist of simple sentences, it severely underestimates processing difficulty on sentences that require syntactic disambiguation (garden-path effects).This leads to the claim that the processing difficulty of such sentences cannot be reduced to surprisal, although it remains possible that neural LMs simply differ from humans in nextword prediction.In this paper, we investigate whether it is truly impossible to construct a neural LM that can explain garden-path effects via surprisal.Specifically, instead of evaluating offthe-shelf neural LMs, we fine-tune these LMs on garden-path sentences so as to better align surprisal-based reading-time estimates with actual human reading times.Our results show that fine-tuned LMs do not overfit and successfully capture human reading slowdowns on held-out garden-path items; they even improve predictive power for human reading times on naturalistic corpora and preserve their general LM capabilities.These results provide an existence proof for a neural LM that can explain both garden-path effects and naturalistic reading times via surprisal, but also raise a theoretical question: what kind of evidence can truly falsify surprisal theory?

Abstract Grammatical features across human languages exhibit intriguing correlations, often attributed to learning biases in humans. Language models (LMs) provide a scalable and naturalistic framework for studying artificial language learning—one not available in human research. We investigate how learnability varies across typologically plausible and implausible languages that closely follow the word order universals identified by linguistic typologists. Our study trains LMs on highly naturalistic counterfactual versions of English (head-initial) and Japanese (head-final). Compared to prior work, our datasets more precisely target the boundary between typological plausibility and implausibility. Our experiments show that LMs learn subtly implausible languages more slowly, though they eventually reach similar performance on some metrics regardless of typological plausibility. These findings suggest that LMs exhibit typologically aligned learning preferences and that certain typological patterns may emerge from general learning biases. https://github.com/sally-xu-42/Typological_Universals.

Large language models possess general linguistic abilities but acquire language less efficiently than humans. This study proposes a method for integrating the developmental characteristics of working memory during the critical period, a stage when human language acquisition is particularly efficient, into the training process of language models. The proposed method introduces a mechanism that initially constrains working memory during the early stages of training and gradually relaxes this constraint in an exponential manner as learning progresses. Targeted syntactic evaluation shows that the proposed method outperforms conventional methods without memory constraints or with static memory constraints. These findings not only provide new directions for designing data-efficient language models but also offer indirect evidence supporting the role of the developmental characteristics of working memory as the underlying mechanism of the critical period in language acquisition.

Recent work in computational psycholinguistics has revealed intriguing parallels between attention mechanisms and human memory retrieval, focusing primarily on vanilla Transformers that operate on token-level representations. However, computational psycholinguistic research has also established that syntactic structures provide compelling explanations for human sentence processing that token-level factors cannot fully account for. In this paper, we investigate whether the attention mechanism of Transformer Grammar (TG), which uniquely operates on syntactic structures as representational units, can serve as a cognitive model of human memory retrieval, using Normalized Attention Entropy (NAE) as a linking hypothesis between models and humans. Our experiments demonstrate that TG’s attention achieves superior predictive power for self-paced reading times compared to vanilla Transformer’s, with further analyses revealing independent contributions from both models. These findings suggest that human sentence processing involves dual memory representations—one based on syntactic structures and another on token sequences—with attention serving as the general memory retrieval algorithm, while highlighting the importance of incorporating syntactic structures as representational units.

Supervised fine-tuning (SFT) is a critical step in aligning large language models (LLMs) with human instructions and values, yet many aspects of SFT remain poorly understood.We trained a wide range of base models on a variety of datasets including code generation, mathematical reasoning, and general-domain tasks, resulting in 1,000+ SFT models under controlled conditions.We then identified the dataset properties that matter most and examined the layer-wise modifications introduced by SFT.Our findings reveal that some training-task synergies persist across all models while others vary substantially, emphasizing the importance of model-specific strategies.Moreover, we demonstrate that perplexity consistently predicts SFT effectiveness, often surpassing superficial similarity between the training data and the benchmark, and that mid-layer weight changes correlate most strongly with performance gains.We release these 1,000+ SFT models and benchmark results to accelerate further research.All resources are available at https://github.

Abstract Recent cognitive modeling studies have reported that larger language models (LMs) exhibit a poorer fit to human reading behavior (Oh and Schuler, 2023b; Shain et al., 2024; Kuribayashi et al., 2024), leading to claims of their cognitive implausibility. In this paper, we revisit this argument through the lens of mechanistic interpretability and argue that prior conclusions were skewed by an exclusive focus on the final layers of LMs. Our analysis reveals that next-word probabilities derived from internal layers of larger LMs align with human sentence processing data as well as, or better than, those from smaller LMs. This alignment holds consistently across behavioral (self-paced reading times, gaze durations, MAZE task processing times) and neurophysiological (N400 brain potentials) measures, challenging earlier mixed results and suggesting that the cognitive plausibility of larger LMs has been underestimated. Furthermore, we first identify an intriguing relationship between LM layers and human measures: Earlier layers correspond more closely with fast gaze durations, while later layers better align with relatively slower signals such as N400 potentials and MAZE processing times. Our work opens new avenues for interdisciplinary research at the intersection of mechanistic interpretability and cognitive modeling.1

Tatsuki Kuribayashi, Ryo Ueda, Ryo Yoshida, Yohei Oseki, Ted Briscoe, Timothy Baldwin. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

For nearly the past forty years, there has been discussion regarding whether symbolic representations are involved in morphological inflection, a debate commonly known as the Past Tense Debate. The previous literature has extensively explored whether neural models, which do not use symbolic representations can process morphological inflection like humans. However, current research interest has shifted towards whether neural models can acquire morphological inflection like humans. In this paper, we trained neural models, the recurrent neural network (RNN) with attention and the transformer, and a symbolic model, the Minimal Generalization Learner (MGL), under a human-like learning environment. Evaluating the models from the perspective of language acquisition, we found that while the transformer and the MGL exhibited some human-like characteristics, the RNN with attention did not demonstrate human-like behavior across all the evaluation metrics considered in this study. Furthermore, none of the models accurately inflected verbs in the same manner as humans in terms of morphological inflection direction. These results suggest that these models fall short as cognitive models of morphological inflection.

In Reinforcement Learning from Human Feedback (RLHF), explicit human feedback, such as rankings, is employed to align Natural Language Processing (NLP) models with human preferences. In contrast, the potential of implicit human feedback, encompassing cognitive processing signals like eye-tracking and brain activity, remains underexplored. These signals capture unconscious human responses but are often marred by noise and redundancy, complicating their application to specific tasks. To address this issue, we introduce the Cognitive Information Bottleneck (CIB), a method that extracts only the task-relevant information from cognitive processing signals. Grounded in the principles of the information bottleneck, CIB aims to learn representations that maximize the mutual information between the representations and targets while minimizing the mutual information between inputs and representations. By employing CIB to filter out redundant information from cognitive processing signals, our goal is to provide representations that are both minimal and sufficient. This approach enables more efficient fitting of models to inputs. Our results show that the proposed method outperforms existing methods in efficiently compressing various cognitive processing signals and significantly enhances performance on downstream tasks. Evaluated on public datasets, our model surpasses contemporary state-of-the-art models. Furthermore, by analyzing these compressed representations, we offer insights into how cognitive processing signals can be leveraged to improve performance.

Neural language models have exhibited outstanding performance in a range of downstream tasks. However, there is limited understanding regarding the extent to which these models internalize syntactic knowledge, so that various datasets have recently been constructed to facilitate syntactic evaluation of language models across languages. In this paper, we introduce JCoLA (Japanese Corpus of Linguistic Acceptability), which consists of 10,020 sentences annotated with binary acceptability judgments. Specifically, those sentences are manually extracted from linguistics textbooks, handbooks and journal articles, and split into in-domain data (86 %; relatively simple acceptability judgments extracted from textbooks and handbooks) and out-of-domain data (14 %; theoretically significant acceptability judgments extracted from journal articles), the latter of which is categorized by 12 linguistic phenomena. We then evaluate the syntactic knowledge of 9 different types of Japanese and multilingual language models on JCoLA. The results demonstrated that several models could surpass human performance for the in-domain data, while no models were able to exceed human performance for the out-of-domain data. Error analyses by linguistic phenomena further revealed that although neural language models are adept at handling local syntactic dependencies like argument structure, their performance wanes when confronted with long-distance syntactic dependencies like verbal agreement and NPI licensing.

In this paper, we propose a novel evaluation paradigm for Targeted Syntactic Evaluations, where we assess how well language models can recognize linguistic phenomena situated at different levels of the Chomsky hierarchy. Specifically, we create formal languages that abstract four syntactic phenomena in natural languages, each identified at a different level of the Chomsky hierarchy, and use these to evaluate the capabilities of language models: (1) (Adj)ˆn NP type, (2) NPˆn VPˆn type, (3) Nested Dependency type, and (4) Cross Serial Dependency type. We first train three different language models (LSTM, Transformer LM, and Stack-RNN) on language modeling tasks and then evaluate them using pairs of a positive and a negative sentence by investigating whether they can assign a higher probability to the positive sentence than the negative one. Our result demonstrated that all language models have the ability to capture the structural patterns of the (Adj)ˆn NP type formal language. However, LSTM and Transformer LM failed to capture NPˆn VPˆn type language and no architectures can recognize nested dependency and Cross Serial dependency correctly. Neural language models, especially Transformer LMs, have exhibited high performance across a multitude of downstream tasks, leading to the perception that they possess an understanding of natural languages. However, our findings suggest that these models may not necessarily comprehend the syntactic structures that underlie natural language phenomena such as dependency. Rather, it appears that they may extend grammatical rules equivalent to regular grammars to approximate the rules governing dependencies.

What kinds of and how much data is necessary for language models to induce grammatical knowledge to judge sentence acceptability?Recent language models still have much room for improvement in their data efficiency compared to humans.This paper investigates whether language models efficiently use indirect data (indirect evidence), from which they infer sentence acceptability.In contrast, humans use indirect evidence efficiently, which is considered one of the inductive biases contributing to efficient language acquisition.To explore this question, we introduce the Wug In-Direct Evidence Test (WIDET), a dataset consisting of training instances inserted into the pre-training data and evaluation instances.We inject synthetic instances with newly coined wug words into pretraining data and explore the model's behavior on evaluation data that assesses grammatical acceptability regarding those words.We prepare the injected instances by varying their levels of indirectness and quantity.Our experiments surprisingly show that language models do not induce grammatical knowledge even after repeated exposure to instances with the same structure but differing only in lexical items from evaluation instances in certain language phenomena.Our findings suggest a potential direction for future research: developing models that use latent indirect evidence to induce grammatical knowledge.

Language models (LMs) have been used in cognitive modeling as well as engineering studiesthey compute information-theoretic complexity metrics that simulate humans' cognitive load during reading.This study highlights a limitation of modern neural LMs as the model of choice for this purpose: there is a discrepancy between their context access capacities and that of humans.Our results showed that constraining the LMs' context access improved their simulation of human reading behavior.We also showed that LM-human gaps in context access were associated with specific syntactic constructions; incorporating syntactic biases into LMs' context access might enhance their cognitive plausibility.1

Tatsuki Kuribayashi, Yohei Oseki, Takumi Ito, Ryo Yoshida, Masayuki Asahara, Kentaro Inui. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

In computational linguistics, it has been shown that hierarchical structures make language models (LMs) more human-like.However, the previous literature has been agnostic about a parsing strategy of the hierarchical models.In this paper, we investigated whether hierarchical structures make LMs more human-like, and if so, which parsing strategy is most cognitively plausible.In order to address this question, we evaluated three LMs against human reading times in Japanese with head-final leftbranching structures: Long Short-Term Memory (LSTM) as a sequential model and Recurrent Neural Network Grammars (RNNGs) with top-down and left-corner parsing strategies as hierarchical models.Our computational modeling demonstrated that left-corner RNNGs outperformed top-down RNNGs and LSTM, suggesting that hierarchical and leftcorner architectures are more cognitively plausible than top-down or sequential architectures.In addition, the relationships between the cognitive plausibility and (i) perplexity, (ii) parsing, and (iii) beam size will also be discussed.1

The past decade has witnessed the happy marriage between natural language processing (NLP) and the cognitive science of language. Moreover, given the historical relationship between biological and artificial neural networks, the advent of deep learning has re-sparked strong interests in the fusion of NLP and the neuroscience of language. Importantly, this inter-fertilization between NLP, on one hand, and the cognitive (neuro)science of language, on the other, has been driven by the language resources annotated with human language processing data. However, there remain several limitations with those language resources on annotations, genres, languages, etc. In this paper, we describe the design of a novel language resource called BCCWJ-EEG, the Balanced Corpus of Contemporary Written Japanese (BCCWJ) experimentally annotated with human electroencephalography (EEG). Specifically, after extensively reviewing the language resources currently available in the literature with special focus on eye-tracking and EEG, we summarize the details concerning (i) participants, (ii) stimuli, (iii) procedure, (iv) data preprocessing, (v) corpus evaluation, (vi) resource release, and (vii) compilation schedule. In addition, potential applications of BCCWJ-EEG to neuroscience and NLP will also be discussed.