The study evaluates synthetic pre‑pretraining (PPT) across a wide range of models (500 M–7 B parameters) and training budgets (up to 100 B tokens). Results show that PPT consistently improves downstream performance and token efficiency, saving at least 21 B tokens at the 3 B scale, but these gains do not appear to stem from a grammatical prior. Instead, PPT benefits arise from tasks that enhance long‑range retrieval, and the improvements remain robust across diverse data mixtures, diminishing only when web text is omitted.
By Atsuki Yamaguchi, Tatsuro Inaba, Joel Niklaus, Michal \v{S}tef\'anik, Aline Villavicencio, Nikolaos Aletras
The paper investigates how language models avoid overgeneralizations by distinguishing between two types of indirect negative evidence: preemption and entrenchment. Through controlled rearing experiments on models trained on child‑caregiver conversations, the authors find that models do not exhibit verb‑specific preemption but show weak abstract preemption. Analysis of training dynamics suggests that competing structures act as indirect positive evidence rather than negative in the verb‑specific condition.
By Yixuan Wang, Freda Shi, Kanishka Misra
arXiv:2606. 07479v1 Announce Type: cross Abstract: Turkish idiomatic light verb constructions (LVCs) are challenging for multiword expression processing because they often share the same surface form as fully literal verb-object combinations while functioning as a single, partially idiomatic predicate.
By Sercan Karaka\c{s}, Yusuf \c{S}im\c{s}ek
The paper introduces an experiment where a small masked language model (DeBERTa) is initialized with visual embeddings for tokens that correspond to image regions, following St. Augustine’s ostensive definition of word learning. The visual initialization leaves a measurable imprint that persists through training, yet it does not improve performance on most BabyLM benchmarks that test abstract grammatical knowledge. However, the seeded models show a consistent advantage in zero‑shot object‑property tasks and in a custom Visual‑Property Swap benchmark that probes color, material, size, and shape knowledge, with the advantage confined to the seeded words and transferable to newly seeded words.
By Lisa Bylinina
arXiv:2608.30315v1 Announce Type: new
Abstract: Token embeddings are the basic representational units that connect discrete tokens with continuous computation in language models. Although modern lang...
By Junjie Yao, Liangkai Hang, Zhi-Qin John Xu
arXiv:2607. 03478v1 Announce Type: new Abstract: Post-training of frontier language models is conducted on curated task suites, and inevitably leaves a distribution shift between training and deployment environments.
By Jou Barzdukas, Jack Peck, Julian Schulz, Paulius Rauba, Steven Basart, Lennie Wells
arXiv:2609.01151v1 Announce Type: new
Abstract: In the standard LM training pipeline, subword tokenisation is applied as a preprocessing step. Subword segmental language modelling is an alternative p...
By Francois Meyer
Neural Collapse predicts that balanced one-hot classification pushes model representations to be equally far from each other; a symmetric configuration that depends only on the output label and ignores any semantic similarity in the inputs. This creates a puzzle: next-token prediction language models are trained predominantly (as context length increases) with one-hot labels: the same context is very unlikely to appear twice in training with different labels.
arXiv:2606. 15521v1 Announce Type: cross Abstract: Tokenization introduces representational redundancy: under a fixed token vocabulary, every byte string admits many valid token encodings, or segmentations, that decode to the same surface string.
By Kanishk Jain, Matthew Day, Tankut Can
arXiv:2609.07876v1 Announce Type: cross
Abstract: Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linea...
By Arjun Patrawala, Jiahai Feng, Erik Jones, Jacob Steinhardt
arXiv:2606. 25987v1 Announce Type: cross Abstract: Large language models (LLMs) attain remarkable surface fluency on code, yet they neither formally guarantee the syntactic validity of their output nor leverage the hierarchical structure defining the target language.
By Alexandre Bouayad
The study examines how the geometric placement of synthetic data affects discourse‑pragmatic function classification. Using 410 annotated instances of the word "look" from the British National Corpus, synthetic examples were generated with Llama 3.1 and grouped by cosine distance from real data in RoBERTa space. Six training conditions were compared, showing that examples close to real data (NEAR) yield the largest macro‑F gain, while a distance‑balanced mix gives the highest accuracy, yet none improve AUC.
By Sara Sorahi, Kevin Tang, Reza Kazemian