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:2605.01609v2 Announce Type: replace-cross Abstract: We find that transformer concept representations systematically anti-concentrate in the spectral tail of the unembedding covariance, encoding...
The paper examines how Large Language Models’ internal representations become sparser as task difficulty rises, finding that harder inputs lead to more concentrated activation patterns in the last hidden state. This sparsity–difficulty relationship holds across various models and domains, indicating that models learn dense representations for familiar patterns but default to sparse activations for unfamiliar ones. Leveraging this insight, the authors propose Sparsity‑Guided Curriculum In‑Context Learning (SG‑ICL), which selects few‑shot demonstrations based on sparsity and outperforms standard CoT and Auto‑CoT baselines on the MATH‑500 benchmark.
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.
arXiv:2603. 06592v2 Announce Type: replace-cross Abstract: Contemporary studies in mechanistic interpretability have uncovered many puzzling phenomena in the neural information processing of Transformer-based language models, such as induction heads, function vectors, and the Hydra effect.
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:2607. 23054v1 Announce Type: cross Abstract: Multi-head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value pairs through a shared low-rank bottleneck (cKV), achieving 81% KV-cache reduction during inference.
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:2606. 01060v1 Announce Type: cross Abstract: Preference alignment has substantially improved the observable behavior of large language models, yet it remains unclear what alignment changes internally.
arXiv:2608. 11197v1 Announce Type: new Abstract: Shani et al.
arXiv:2607. 08605v1 Announce Type: cross Abstract: Sparse autoencoders (SAEs) have emerged as a promising technique for mechanistic interpretability by learning a set of sparse latent features in large models, each of which encodes a distinct concept.
The paper extends mechanistic interpretability of large language models by modeling concepts as low‑dimensional non‑linear manifolds rather than linear subspaces. It introduces a concept‑based alignment (CBA) score to compare these manifolds across layers and models, revealing block structures in intermediate layers, a shift from syntax‑dominated to mixed syntactic‑semantic concepts, and training‑dependent multilingual sharing. The study also shows that alignment patterns differ across model families and training stages, with adjacent stages aligning more closely than distant ones.
arXiv:2601.02978v3 Announce Type: replace-cross Abstract: Recent work in Mechanistic Interpretability (MI) has enabled the identification and intervention of internal features in Large Language Model...