arXiv:2605.09129v2 Announce Type: replace Abstract: Circuit study aims to explain how language models (LMs) implement a task by localizing and interpreting a circuit: a computational subgraph respons...
arXiv:2608. 12334v1 Announce Type: cross Abstract: Despite the impressive multilingual capabilities of Large Language Models, the latent dynamics dictating language selection remain poorly understood.
arXiv:2607. 12279v1 Announce Type: cross Abstract: Writing a sentence of exactly twelve words; ending a DNA sequence at the right codon; formatting an ASCII table.
Writing a sentence of exactly twelve words; ending a DNA sequence at the right codon; formatting an ASCII table. These are all tasks that language models can do that requires tracking how many tokens remain before a target.
The paper introduces Declarative Attention (DA), a protocol that lets language models explicitly declare which parts of their context to focus on during generation. By partitioning decoding into full-context, region-specific, and recent-output-only modes, the inference engine can skip large portions of the KV cache, dramatically reducing attended tokens. Experiments on 15 long-context tasks with off-the-shelf models show significant savings (52.0% and 31.1% reductions) with only modest accuracy drops that diminish as model size increases.
The paper investigates whether linguistic and arithmetic tasks in large language models (LLMs) rely on shared neural circuits. Using mechanistic interpretability techniques on five LLMs, the authors find that circuit similarity aligns primarily with task domain rather than input/output vocabularies or syntactic structure. This suggests that LLMs may employ largely separate neural circuits for language and arithmetic, indicating a modular organization of abilities.
The study investigates how multilingual large language models (LLMs) handle negation across languages. Using causal interventions on synthetic and FLORES-200 derived datasets, the authors find that negation representations are causally controllable and largely shared across English, French, Vietnamese, and Chinese. They demonstrate that a single sparse autoencoder feature can control polarity across languages and that these representations generalize to additional languages.
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.
arXiv:2603.29396v2 Announce Type: replace Abstract: Standard evaluations of Large language models (LLMs) focus on task performance, offering limited insight into whether correct behavior reflects app...
The paper introduces a method to embed a controllable latent variable into natural-looking text by steering a teacher LLM along eight sparse autoencoder directions that follow a ring-shaped Markov chain. A small transformer trained on this data successfully tracks the Bayesian posterior of the planted variable and arranges the eight states in the same ring order, linking belief states to concept geometry. This demonstrates that LLMs can model latent variables and that concept geometry may arise from the statistical dynamics of these variables.
The paper introduces LLM-Microscope, a toolkit for measuring how Large Language Models encode contextual information at the token level. It shows that seemingly minor tokens—such as determiners, stopwords, and punctuation—carry surprisingly high contextual weight, and removing them degrades performance on benchmarks like MMLU and BABILong-4k. The study also finds a strong link between contextualization and linearity, indicating that the transformation between layers can be approximated by a single linear mapping when tokens are well contextualized.
arXiv:2509. 24808v2 Announce Type: replace Abstract: Explaining why a language model produces a particular output requires local, input-level explanations.