How LLMs Learn: Tracing Internal Representations with Sparse Autoencoders

Large Language Models (LLMs) demonstrate remarkable multilingual capabilities and broad knowledge. However, the internal mechanisms underlying the development of these capabilities remain poorly understood. To investigate this, we analyze how the information encoded in LLMs' internal representations evolves during the training process. Specifically, we train sparse autoencoders at multiple checkpoints of the model and systematically compare the interpretative results across these stages. Our findings suggest that LLMs initially acquire language-specific knowledge independently, followed by cross-linguistic correspondences. Moreover, we observe that after mastering token-level knowledge, the model transitions to learning higher-level, abstract concepts, indicating the development of more conceptual understanding.

Can Language Models Learn Typologically Implausible Languages?

Grammatical features across human languages show intriguing correlations often attributed to learning biases in humans. However, empirical evidence has been limited to experiments with highly simplified artificial languages, and whether these correlations arise from domain-general or language-specific biases remains a matter of debate. Language models (LMs) provide an opportunity to study artificial language learning at a large scale and with a high degree of naturalism. In this paper, we begin with an in-depth discussion of how LMs allow us to better determine the role of domain-general learning biases in language universals. We then assess learnability differences for LMs resulting from typologically plausible and implausible languages closely following the word-order universals identified by linguistic typologists. We conduct a symmetrical cross-lingual study training and testing LMs on an array of highly naturalistic but counterfactual versions of the English (head-initial) and Japanese (head-final) languages. Compared to similar work, our datasets are more naturalistic and fall closer to the boundary of plausibility. Our experiments show that these LMs are often slower to learn these subtly implausible languages, while ultimately achieving similar performance on some metrics regardless of typological plausibility. These findings lend credence to the conclusion that LMs do show some typologically-aligned learning preferences, and that the typological patterns may result from, at least to some degree, domain-general learning biases.

If Attention Serves as a Cognitive Model of Human Memory Retrieval, What is the Plausible Memory Representation?

Recent work in computational psycholinguistics has revealed intriguing parallels between attention mechanisms and human memory retrieval, focusing primarily on Transformer architectures that operate on token-level representations. However, computational psycholinguistic research has also established that syntactic structures provide compelling explanations for human sentence processing that word-level factors alone cannot fully account for. In this study, 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 model behavior and human processing difficulty. 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 retrieval algorithm, while highlighting the importance of incorporating syntactic structures as representational units.

Developmentally-plausible Working Memory Shapes a Critical Period for Language Acquisition

Large language models exhibit general linguistic abilities but significantly differ from humans in their efficiency of language acquisition. 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 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 models 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 underlying mechanisms of the critical period hypothesis in human language acquisition.

Large Language Models Are Human-Like Internally

Recent cognitive modeling studies have reported that larger language models (LMs) exhibit a poorer fit to human reading behavior, 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.

BabyLM Challenge: Exploring the Effect of Variation Sets on Language Model Training Efficiency

While current large language models have achieved a remarkable success, their data efficiency remains a challenge to overcome. Recently it has been suggested that child-directed speech (CDS) can improve training data efficiency of modern language models based on Transformer neural networks. However, it is not yet understood which specific properties of CDS are effective for training these models. In the context of the BabyLM Challenge, we focus on Variation Sets (VSs), sets of consecutive utterances expressing a similar intent with slightly different words and structures, which are ubiquitous in CDS. To assess the impact of VSs on training data efficiency, we augment CDS data with different proportions of artificial VSs and use these datasets to train an auto-regressive model, GPT-2. We find that the best proportion of VSs depends on the evaluation benchmark: BLiMP and GLUE scores benefit from the presence of VSs, but EWOK scores do not. Additionally, the results vary depending on multiple factors such as the number of epochs and the order of utterance presentation. Taken together, these findings suggest that VSs can have a beneficial influence on language models, while leaving room for further investigation.

Is Structure Dependence Shaped for Efficient Communication?: A Case Study on Coordination

Natural language exhibits various universal properties. But why do these universals exist? One explanation is that they arise from functional pressures to achieve efficient communication, a view which attributes cross-linguistic properties to domain-general cognitive abilities. This hypothesis has successfully addressed some syntactic universal properties such as compositionality and Greenbergian word order universals. However, more abstract syntactic universals have not been explored from the perspective of efficient communication. Among such universals, the most notable one is structure dependence, that is, the existence of grammar-internal operations that crucially depend on hierarchical representations. This property has traditionally been taken to be central to natural language and to involve domain-specific knowledge irreducible to communicative efficiency. In this paper, we challenge the conventional view by investigating whether structure dependence realizes efficient communication, focusing on coordinate structures. We design three types of artificial languages: (i) one with a structure-dependent reduction operation, which is similar to natural language, (ii) one without any reduction operations, and (iii) one with a linear (rather than structure-dependent) reduction operation. We quantify the communicative efficiency of these languages. The results demonstrate that the language with the structure-dependent reduction operation is significantly more communicatively efficient than the counterfactual languages. This suggests that the existence of structure-dependent properties can be explained from the perspective of efficient communication.

Can Language Models Induce Grammatical Knowledge from Indirect Evidence?

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 InDirect 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.

LLM-jp: A Cross-organizational Project for the Research and Development of Fully Open Japanese LLMs

This paper introduces LLM-jp, a cross-organizational project for the research and development of Japanese large language models (LLMs). LLM-jp aims to develop open-source and strong Japanese LLMs, and as of this writing, more than 1,500 participants from academia and industry are working together for this purpose. This paper presents the background of the establishment of LLM-jp, summaries of its activities, and technical reports on the LLMs developed by LLM-jp. For the latest activities, visit https://llm-jp.nii.ac.jp/en/.

Tree-Planted Transformers: Large Language Models with Implicit Syntactic Supervision

Large Language Models (LLMs) have achieved remarkable success thanks to scalability on large text corpora, but have some drawback in training efficiency. In contrast, Syntactic Language Models (SLMs) can be trained efficiently to reach relatively high performance thanks to syntactic supervision, but have trouble with scalability. Thus, given these complementary advantages of LLMs and SLMs, it is necessary to develop an architecture that integrates the scalability of LLMs with the training efficiency of SLMs, namely Syntactic Large Language Models (SLLM). In this paper, we propose a novel method dubbed tree-planting: implicitly "plant" trees into attention weights of Transformer LMs to reflect syntactic structures of natural language. Specifically, Transformer LMs trained with tree-planting will be called Tree-Planted Transformers (TPT), which learn syntax on small treebanks via tree-planting and then scale on large text corpora via continual learning with syntactic scaffolding. Targeted syntactic evaluations on the SyntaxGym benchmark demonstrated that TPTs, despite the lack of explicit syntactic supervision, significantly outperformed various SLMs with explicit syntactic supervision that generate hundreds of syntactic structures in parallel, suggesting that tree-planting and TPTs are the promising foundation for SLLMs.