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: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...
arXiv:2510.22752v2 Announce Type: replace-cross Abstract: In-context learning is governed by both temporal and semantic relationships, shaping how Large Language Models (LLMs) retrieve contextual inf...
The paper demonstrates that Large Language Models, despite their non‑linear components, exhibit a fundamental linearity property: when inputs from two distinct text streams are linearly combined, the model outputs a superposition of the individual next‑token distributions. This "Superposition Linearity Hypothesis" appears to be an intrinsic feature of the Transformer architecture, tends to weaken during pretraining, but can be largely restored with lightweight fine‑tuning. The authors also present a guided decoding method that separates the superposed outputs, allowing two coherent continuations to be generated from a single forward pass.
MGAL is a new multilingual benchmark for evaluating long‑context large language models, built from United Nations reports in six official UN languages and covering 8K to 128K tokens. It tests four linguistic granularities—word, sentence, paragraph, and document—while also stratifying examples by their position within the document (begin, middle, end). Experiments show that models excel at word‑level tasks but struggle with coarser granularity, and that closed‑source models outperform others in lower‑resource languages, revealing challenges such as local semantic crowding and a fluency‑consistency gap.
arXiv:2602. 00462v4 Announce Type: replace-cross Abstract: Transforming a large language model (LLM) into a vision-language model (VLM) can be achieved by mapping the visual tokens from a vision encoder into the embedding space of an LLM.
arXiv:2606. 00909v1 Announce Type: cross Abstract: This work presents MLLM-Microscope, a novel system designed for analyzing the hidden representations within Multimodal Large Language Models (MLLMs).
The paper proposes that large language models (LLMs) encode high‑level concepts as linear directions within their activation space and that they can use subspaces and vector algebra to perform tasks. By analyzing functional modules and residual streams during in‑context learning (ICL), the authors find that LLMs can create evidence‑accumulating subspaces and solve ICL tasks through simple algebraic operations within those subspaces.
The article examines whether large language models (LLMs) can effectively use representations they learn in-context to perform downstream tasks. Experiments with open‑weight LLMs on next‑token prediction and a new adaptive world modeling task reveal that these models struggle to deploy novel semantics defined in-context, even when such semantics are encoded in their latent representations. The study also shows that state‑of‑the‑art closed‑source reasoning models similarly fail to reliably leverage in‑context patterns, highlighting a gap in flexible representation usage.
The paper investigates how semantic meaning is arranged locally within large language models. By using two controlled probes—one cross‑lingual and one refinement probe—on four decoder‑only LLMs, it finds that after removing dominant variance directions, sentence pairs rank by semantic relatedness, semantic factors cluster around an invariant core with structured variation, refinement tightens neighborhoods, and sentence‑level versus token‑local readouts reveal complementary geometric patterns. These results suggest that LLM semantics form local anisotropic neighborhoods that contract when meaning is refined.
The paper compares latent representations in Selective State Space Models (SSMs) like Mamba and Transformers such as Pythia using Sparse Autoencoders. Across a 10‑million token corpus, 99.98% of Mamba features align closely with Pythia’s, supporting the Universality Hypothesis that core semantic representations are similar across architectures. A tiny 0.02% of features diverge, with Mamba’s recurrent bottleneck causing it to compress syntactic anomalies into polysemantic neurons, whereas Pythia’s attention can isolate distinct formatting edge‑cases.
arXiv:2605.16893v2 Announce Type: replace Abstract: Recent studies have introduced memory-augmented architectures that decouple knowledge storage from neural computation, enabling more direct access...
The paper introduces DarkLens, a lightweight method for detecting faithfulness hallucinations in large language models (LLMs). It estimates context reliance by projecting hidden states onto the dark space—directions weakly aligned with the unembedding matrix—and uses the resulting Dark Matter Ratio (DMR) as a direct score of context utilization. Combined with a parametric knowledge score, DarkLens operates within a single forward pass and outperforms many baselines on QA and summarization benchmarks across several open‑weight LLMs.