The paper introduces Exemplar Partitioning (EP), an unsupervised technique that constructs interpretable feature dictionaries from large language model activations by clustering streamed activations into Voronoi regions defined by exemplars and their averages. EP allows comparison of dictionaries across layers, checkpoints, and architectures, and demonstrates utility in interpreting model behavior, tracking training dynamics, detecting hidden concepts, and enabling targeted interventions. Experiments on Gemma‑2‑2B and Llama‑3.1‑8B show EP can reveal how instruction tuning reorganizes harmful prompt activations, facilitate interventions that alter model responses, and achieve high concept‑detection performance while requiring far fewer construction tokens than comparable methods.
By Jessica Rumbelow
arXiv:2604.17614v2 Announce Type: replace-cross
Abstract: Skills are a natural unit for describing what a language model can do and how its behavior can be changed. However, existing characterization...
By Feiyang Kang, Mahavir Dabas, Myeongseob Ko, Ruoxi Jia
arXiv:2608. 10537v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) have helped uncover mechanistic explanations for LLM behaviours such as reasoning, jailbreaking etc.
By Chuqiao Lin, Shivaji Sondhi, Xiao-Liang Qi
arXiv:2512. 07355v2 Announce Type: replace Abstract: Two traditions of interpretability have evolved side by side but seldom spoken to each other: Concept Bottleneck Models (CBMs), which prescribe what a concept should be, and Sparse Autoencoders (SAEs), which discover what concepts emerge.
By Alexandre Rocchi, Thomas Fel, Gianni Franchi
arXiv:2509. 25045v3 Announce Type: replace-cross Abstract: Despite their capabilities, Large Language Models (LLMs) remain opaque with limited understanding of their internal representations.
By Marco Bronzini, Carlo Nicolini, Bruno Lepri, Jacopo Staiano, Andrea Passerini
arXiv:2506. 14003v5 Announce Type: replace Abstract: Machine unlearning (MU) for large language models (LLMs), commonly referred to as LLM unlearning, seeks to remove specific undesirable data or knowledge from a trained model, while maintaining its performance on standard tasks.
By Yiwei Chen, Soumyadeep Pal, Yimeng Zhang, Qing Qu, Sijia Liu
arXiv:2509. 14704v3 Announce Type: replace Abstract: Benchmark saturation and training-data contamination increasingly obscure whether reported gains in large language models (LLMs) reflect genuine advances in reasoning or familiarity with recurring patterns in benchmark problems.
By Masaharu Mizumoto, Dat Nguyen, Zhiheng Han, Xingfu Li, Yo Nakawake, Le Minh Nguyen
The paper introduces CoT-Interpretability Alignment (CIA), a metric that quantifies how well a large language model’s chain-of-thought (CoT) explanations match its internal reasoning processes. Evaluated on two-hop question answering, hint intervention, and integer multiplication across three LLMs, the study finds limited alignment (44.8–75.9%) and demonstrates that post‑training with a reward combining task accuracy and parametric faithfulness can substantially improve CoT faithfulness without sacrificing accuracy. The authors provide a framework for auditing CoT faithfulness and a pathway to making explicit reasoning more trustworthy, with code and data publicly available.
By Yihuai Hong, Shauli Ravfogel, Chen Zhao, Eunsol Choi
arXiv:2510. 19990v2 Announce Type: replace Abstract: The reasoning paradigm, where language models reason before answering, has enabled breakthroughs on tasks such as mathematical problem-solving.
By Zachary Horvitz, Raghav Singhal, Hao Zou, Carles Domingo-Enrich, Zhou Yu, Rajesh Ranganath, Kathleen McKeown
arXiv:2608. 05624v1 Announce Type: new Abstract: Sycophantic responses are becoming pervasive in large language models (LLMs), and prior work has pointed out that some of them could be harmful.
By Bohan Jiang, Dawei Li, Yasin Silva, Huan Liu
arXiv:2607. 16448v1 Announce Type: cross Abstract: Interpretability methods for neural network activations span a wide cost spectrum, from cheap, training-free techniques (such as linear probes, PCA, SVD) to more expensive training-based ones (such as SAEs and activation oracles).
By Sriram Balasubramanian, Soheil Feizi
arXiv:2608. 15445v1 Announce Type: new Abstract: When a reward is correct on every training example yet consistent with more than one goal, a model can acquire an unintended one, a failure known as goal misgeneralization.
By Suyash Maniyar, Armaan Sandhu, Abhishek Mishra