arXiv:2610.29843v1 Announce Type: new Abstract: Large language models are typically trained with next-token prediction (NTP), which rewards predicting specific tokens rather than explicitly accountin...
Activation verbalizers translate the internal representations of large language models into human-readable text. The recently proposed natural language autoencoders (NLAs) learn this by reconstructing...
The paper investigates natural language autoencoders (NLAs) as tools for verbalizing activation differences in large language models. It examines whether NLAs capture compositional structure, generalize to fine‑tuned models, recover fine‑tuning objectives, and can be specialized for model diffing. Experiments across four models show that NLAs largely succeed in these tasks, though they do not yet match state‑of‑the‑art diffing methods on emergent misalignment detection.
The paper investigates how adjectival modifiers affect the semantic plausibility of events, using the Adept benchmark of 16,000 English sentence pairs that differ by a single adjective. Experiments show that sentence transformers, despite being conceptually suited to the task, underperform compared to models like RoBERTa. The authors provide an error analysis and discuss the implications of their findings for future work on balancing training and test data.
arXiv:2606. 00230v1 Announce Type: new Abstract: Grokking, the phenomenon in which neural networks generalize long after fitting their training data, has been studied in supervised settings on many epochs.
The paper introduces TokenAdapt, a model‑agnostic tokenizer transplantation method that uses a hybrid heuristic to initialize new token embeddings, and a novel pre‑tokenization learning approach for multi‑word Supertokens to improve compression. TokenAdapt combines local subword decomposition and global semantic similarity to preserve semantics while reducing retraining needs. Empirical results show that TokenAdapt outperforms existing baselines such as Transtokenizer and ReTok, achieving lower perplexity ratios and significant compression gains.
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.
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.
arXiv:2608. 14681v1 Announce Type: cross Abstract: Words recur constantly in natural language use, yet it remains unclear whether language models reactivate prior representations or re-evaluate repeated words afresh, and whether post-training changes this default behavior.
The study examines how the geometric placement of synthetic data affects discourse-pragmatic function classification. Using 410 annotated instances of the word "look" and synthetic examples generated by Llama 3.1, the authors partitioned the synthetic data by cosine distance in RoBERTa space and compared six training conditions. While all augmentation conditions improved macro‑F and accuracy over a real‑only baseline, the nearest synthetic examples (NEAR) yielded the largest macro‑F gain (0.113) and a distance‑balanced mix achieved the highest accuracy (0.748), though none improved AUC.
arXiv:2606.22473v2 Announce Type: replace-cross Abstract: Speech language models (SLMs) increasingly combine speech and text, often by interleaving their tokens within a single sequence. Yet how thes...
The study investigates whether pretrained transformer models encode functional words—such as pronouns and adverbs—in a way that mirrors human usage. By comparing embeddings of nouns with those of their functional counterparts in both isolated and parallel sentences, the authors find that functional words occupy a central yet distinct position in embedding space and that parallel lexicalized and functional sentences reside in different subspaces. Experiments show that only a mixed training set of functional and lexicalized sentences reveals shared syntactic and semantic structure, whereas training on either type alone fails to capture this parallelism.