Bryan Sanchez

Bryan Sanchez

Alignment-tuned language models frequently suppress factual log-probabilities on politically sensitive topics despite retaining the knowledge in their hidden representations. We show that a 786K-parameter (approximately 0.02% of the base model) post-transformer adapter, trained on frozen hidden states, corrects this suppression on 31 ideology-discriminating facts across Qwen3-4B, 8B, and 14B. The adapter memorizes all 15 training facts and generalizes to 11--39% of 16 held-out facts across 5 random splits per scale, with zero knowledge regressions via anchored training. Both gated (SwiGLU) and ungated (linear bottleneck) adapters achieve comparable results; neither consistently outperforms the other (Fisher exact p > 0.09 at all scales). On instruct models, the adapter corrects log-probability rankings. When applied at all token positions during generation, the adapter produces incoherent output; however, when applied only at the current prediction position (last-position-only), the adapter produces coherent, less censored text. A logit-space adapter operating after token projection fails to produce coherent generation at any application mode, suggesting hidden-state intervention is the correct level for generation correction. A previously undocumented silent gradient bug in Apple MLX explains all null results in earlier iterations of this work: the standard pattern nn.value_and_grad(model, fn)(model.parameters()) returns zero gradients without error; the correct pattern nn.value_and_grad(model, fn)(model, data) resolves this. We provide a minimal reproduction and discuss implications for other adapter research using MLX.

Mingren Shen, Guanzhao Li, Dongxia Wu et al.

Electron microscopy is widely used to explore defects in crystal structures, but human detecting of defects is often time-consuming, error-prone, and unreliable, and is not scalable to large numbers of images or real-time analysis. In this work, we discuss the application of machine learning approaches to find the location and geometry of different defect clusters in irradiated steels. We show that a deep learning based Faster R-CNN analysis system has a performance comparable to human analysis with relatively small training data sets. This study proves the promising ability to apply deep learning to assist the development of automated microscopy data analysis even when multiple features are present and paves the way for fast, scalable, and reliable analysis systems for massive amounts of modern electron microscopy data.