← Back to all news
arXiv AI By Jacob Solinsky (University of Minnesota), Serguei Pakhomov (University of Minnesota)

Encoding and Propagation of Grammatical Number in Large Language Models

Read the original on arXiv AI →

The study uses sparse autoencoders (SAEs) to probe how the Pythia-70M language model encodes grammatical number. By identifying and manipulating SAE‑derived features such as "plural noun" and "generate a plural verb," the authors show that removing these features forces the model to produce singular verbs, and vice versa. Similar asymmetries were observed in larger models (Gemma‑2‑2b and Llama‑3.1‑8b), indicating a privative representation of number in English.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

Related stories

arXiv Computation and Language
Sep 25

Scoring Both Directions: LLMs realize the MRS they cannot reliably parse

The paper evaluates two large language models, Claude Sonnet 4.5 and Claude Opus 5, on the bidirectional English Resource Grammar (ERG) tasks of generating English from Minimal Recursion Semantics (MRS) and parsing English into MRS. In generation, Opus achieves 76.3 BLEU—surpassing a 72k‑pair trained system and matching a million‑pair system—while Sonnet scores 65.7 BLEU, rising to 69.6 when selecting from ACE’s candidates. In parsing, both models lag behind ACE, attaining only 57.2 and 65.5 F₁ respectively, with exact‑match on about 1 % of sentences, highlighting that high generation scores do not guarantee accurate semantic parsing.

By Soham Dan
llmsbenchmarks
More like this →
arXiv Computation and Language
Oct 22

Why is AI-generated language so hard to understand?

The paper investigates why AI-generated text has become harder to understand, focusing on psycholinguistic features linked to human processing difficulty such as object relative clauses, long syntactic dependencies, and low-frequency n-grams. Using a new corpus of AI-generated text from diverse domains, the study finds that frontier and near-frontier language models employ these difficult features more frequently than paraphrases from round-trip translation and human writing, including more inanimate subjects with animate verbs, longer dependencies, stacked nouns, and reduced object relatives. The authors hypothesize that AI models prioritize information density for tasks like math, code, and reasoning, diverging from human communication goals, and demonstrate this by fine‑tuning a model with a length penalty, which reduces human‑like linguistic behavior.

By Thomas Hikaru Clark, Matias B. Martin, Kyle Mahowald
llmsfine-tuning
More like this →
arXiv Computation and Language
Sep 18

Generalization through Lexical Abstraction in Transformer Models: The Case of Functional Words

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.

By Giuseppe Samo, Vivi Nastase, Paola Merlo
llmsrag
More like this →
arXiv Computation and Language
Oct 27

Large Language Model probabilities cannot reliably distinguish between possible and impossible language

arXiv:2509.15114v2 Announce Type: replace Abstract: A controversial test for Large Language Models concerns the ability to discern possible from impossible language. While some evidence attests to th...

By Evelina Leivada, Raquel Montero, Paolo Morosi, Natalia Moskvina, Tamara Serrano, Nikoleta Pantelidou, Marcel Aguilar, Fritz Guenther
llmsbenchmarks
More like this →