1
A Shared Subcircuit Lets LLMs Count Down Across Tasks
大模型内部发现通用倒数机制，助你在多种任务中精准控制输出长度
arXiv:2607.12279v1 Announce Type: cross Abstract: Writing a sentence of exactly twelve words; ending a DNA sequence at the right codon; formatting an …
2
The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs
研究发现对齐训练并非消除文化知识，而是通过抑制机制在拒绝时刻压制模型内部logit分布。
arXiv:2510.13154v2 Announce Type: replace Abstract: What happens inside a language model when alignment training conflicts with a cultural value it en…
3
Demystifying Variance in Circuit Discovery of LLMs
揭秘LLM电路发现中的方差来源，为可解释性研究提供新视角
arXiv:2606.16920v1 Announce Type: cross Abstract: Circuit discovery is a key technique in mechanistic interpretability to pinpoint the model component…
4
From Senses to Decisions: The Information Flow of Auditory and Visual Perception in Multimodal LLMs
多模态大模型能听会看，但音频和视觉信号如何内部流动最终影响回答？这篇论文深入追踪其信息路径。
arXiv:2606.10147v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) can listen and see, but how do audio and visual signals actua…
5
Inside the LLM Word Factory
揭秘大模型内部单词生成的完整流水线，从原理到图表示，一篇学术深研论文。
arXiv:2606.08562v1 Announce Type: new Abstract: Transformer language models process input provided as subword fragments, but natural language semantic…
6
MechLens: Late Crystallization of Factual Knowledge Explains Intervention Effectiveness in Language Models
新研究发现LLM事实知识在最后几层突然“结晶”，而非逐层涌现，解释了为何干预后期层能有效减少幻觉。
arXiv:2606.07978v1 Announce Type: new Abstract: Understanding where LLMs store factual knowledge is critical for hallucination mitigation. We systemat…
7
Emergent Causal-Geometric Dynamics Across Depth in Large Language Models
最新研究揭示大语言模型在不同深度层的因果几何动力学规律，为理解模型内部机制提供新视角。
arXiv:2602.04931v2 Announce Type: replace-cross Abstract: Geometric analyses of large language model (LLM) representations reveal structured variation…
8
How do LLMs Compute Verbal Confidence
揭秘LLMs如何内部计算口头自信度：是即时响应还是自动生成？这项研究打开黑箱，深入模型推理过程。
arXiv:2603.17839v3 Announce Type: replace-cross Abstract: Verbal confidence -- prompting LLMs to state their confidence as a number or category -- is …