We investigate feature universality in Gemma-2 language models (Gemma-2-2B and Gemma-2-9B), asking whether models with a four-fold difference in scale still converge on comparable internal concepts. Using the Sparse Autoencoder (SAE) dictionary-learning pipeline, we utilize SAEs on each model's residual-stream activations, align the resulting monosemantic features via activation correlation, and compare the matched feature spaces with SVCCA and RSA. Middle layers yield the strongest overlap, while early and late layers show far less similarity. Preliminary experiments extend the analysis from single tokens to multi-token subspaces, showing that semantically similar subspaces interact similarly with language models. These results strengthen the case that large language models carve the world into broadly similar, interpretable features despite size differences, reinforcing universality as a foundation for cross-model interpretability.

Recent works have shown that transformers can solve contextual reasoning tasks by internally executing computational graphs called circuits. Circuits often use attention to logically match information from subspaces of the representation, e.g. using position-in-sequence to identify the previous token. In this work, we consider a semantic subspace to be any independent subspace of the latent representation that can fully determine an attention distribution. We show that Pre-Norm, the placement of normalisation layer used by state-of-the-art transformers, violates this ability unless the model learns a strict representation structure of orthogonal spheres. This is because it causes linear subspaces to interfere through their common normalisation factor. Theoretically, we analyse circuit stability by modelling this interference as random noise on the $L_2$-norms of the query/key/value vectors, predicting a phenomenon of circuit collapse when sparse-attention shifts to a different token. Empirically, we investigate the sensitivity of real-world models trained for mathematical addition, observing a 1% rate of circuit collapse when the norms are artificially perturbed by $\lesssim$10%. We contrast Pre-Norm with QKV-Norm, which places normalisation after the attention head's linear operators. Theoretically this relaxes the representational constraints. Empirically we observe comparable in-distribution but worse out-of-distribution performance.

Deep neural networks have exhibited substantial success in the field of Natural Language Processing and ensuring their safety and reliability is crucial: there are safety critical contexts where such models must be robust to variability or attack, and give guarantees over their output. Unlike Computer Vision, NLP lacks a unified verification methodology and, despite recent advancements in literature, they are often light on the pragmatical issues of NLP verification. In this paper, we attempt to distil and evaluate general components of an NLP verification pipeline, that emerges from the progress in the field to date. Our contributions are two-fold. Firstly, we give a general (i.e. algorithm-independent) characterisation of verifiable subspaces that result from embedding sentences into continuous spaces. We identify, and give an effective method to deal with, the technical challenge of semantic generalisability of verified subspaces; and propose it as a standard metric in the NLP verification pipelines (alongside with the standard metrics of model accuracy and model verifiability). Secondly, we propose a general methodology to analyse the effect of the embedding gap -- a problem that refers to the discrepancy between verification of geometric subspaces, and the semantic meaning of sentences which the geometric subspaces are supposed to represent. In extreme cases, poor choices in embedding of sentences may invalidate verification results. We propose a number of practical NLP methods that can help to quantify the effects of the embedding gap; and in particular we propose the metric of falsifiability of semantic subspaces as another fundamental metric to be reported as part of the NLP verification pipeline. We believe that together these general principles pave the way towards a more consolidated and effective development of this new domain.

The question of what kinds of linguistic information are encoded in different layers of Transformer-based language models is of considerable interest for the NLP community. Existing work, however, has overwhelmingly focused on word-level representations and encoder-only language models with the masked-token training objective. In this paper, we present experiments with semantic structural probing, a method for studying sentence-level representations via finding a subspace of the embedding space that provides suitable task-specific pairwise distances between data-points. We apply our method to language models from different families (encoder-only, decoder-only, encoder-decoder) and of different sizes in the context of two tasks, semantic textual similarity and natural-language inference. We find that model families differ substantially in their performance and layer dynamics, but that the results are largely model-size invariant.

If scientific discovery is one of the main driving forces of human progress, insight is the fuel for the engine, which has long attracted behavior-level research to understand and model its underlying cognitive process. However, current tasks that abstract scientific discovery mostly focus on the emergence of insight, ignoring the special role played by domain knowledge. In this concept paper, we view scientific discovery as an interplay between $thinking \ out \ of \ the \ box$ that actively seeks insightful solutions and $thinking \ inside \ the \ box$ that generalizes on conceptual domain knowledge to keep correct. Accordingly, we propose Mindle, a semantic searching game that triggers scientific-discovery-like thinking spontaneously, as infrastructure for exploring scientific discovery on a large scale. On this basis, the meta-strategies for insights and the usage of concepts can be investigated reciprocally. In the pilot studies, several interesting observations inspire elaborated hypotheses on meta-strategies, context, and individual diversity for further investigations.

Despite the great success of word embedding, sentence embedding remains a not-well-solved problem. In this paper, we present a supervised learning framework to exploit sentence embedding for the medical question answering task. The learning framework consists of two main parts: 1) a sentence embedding producing module, and 2) a scoring module. The former is developed with contextual self-attention and multi-scale techniques to encode a sentence into an embedding tensor. This module is shortly called Contextual self-Attention Multi-scale Sentence Embedding (CAMSE). The latter employs two scoring strategies: Semantic Matching Scoring (SMS) and Semantic Association Scoring (SAS). SMS measures similarity while SAS captures association between sentence pairs: a medical question concatenated with a candidate choice, and a piece of corresponding supportive evidence. The proposed framework is examined by two Medical Question Answering(MedicalQA) datasets which are collected from real-world applications: medical exam and clinical diagnosis based on electronic medical records (EMR). The comparison results show that our proposed framework achieved significant improvements compared to competitive baseline approaches. Additionally, a series of controlled experiments are also conducted to illustrate that the multi-scale strategy and the contextual self-attention layer play important roles for producing effective sentence embedding, and the two kinds of scoring strategies are highly complementary to each other for question answering problems.