Weight-Space Analysis

Follow this topic Ask about this topic

Weight-space analysis treats neural network parameters—not just their predictions—as objects whose geometry can reveal how models learn, differ, and store capabilities. Researchers study symmetries, spectra, and low-rank subspaces; compare checkpoint deltas and task vectors for merging or provenance; and train symmetry-aware encoders to retrieve or transform weights across architectures. These methods are tested on model zoos, language-model checkpoints and adapters, vision models, and implicit neural representations, for tasks such as model similarity, task composition, data inference, and compression. A central challenge is that equivalent functions can have very different parameters, while geometric similarity or a successful merge may depend on architecture, prompt, or training setup and need not provide a unique mechanistic explanation.

Latest papers

41 papers3 papers in the last four weeks, down 25% on the four weeks before. 0.0% of all new papers.

On the Sparsity of Direct Preference Optimization: Weight Disentanglement in the NTK Regime
Oct 27, 2026Kensuke Sasaki, Issei SatoDirect Preference Optimization LLM Alignment
Department of Computer Science, The University of Tokyo
Bifurcations from Weight Space: JEPAWG on Parametric Dynamical Systems
Oct 22, 2026Tobias Göbel, Miranda C. N. ChengWeight-Space Analysis Hypernetworks
Institute for Mathematics, Academia Sinica, Taipei, Taiwan · Department of Physics, National Taiwan University, Taipei, Taiwan · Institute of Physics, University of Amsterdam, Amsterdam, Netherlands
Interfaces and Modules: A Scalable Representation for Weight-Space Learning across Architectures
Oct 15, 2026Michał Kozyra, Michael BronsteinWeight-Space Learning Weight-Space Analysis
University of Oxford · AITHYRA
CrossGMN: Graph Metanetworks for Cross-Architecture Weight-Space Transformations
Oct 1, 2026Adir Dayan, Yam Eitan, Haggai MaronModel Compression Hypernetworks
Technion – Israel Institute of Technology · NVIDIA Research
A Mesoscopic View of Transformer Weights Through Row and Column Scale Fields
Sep 25, 2026Tiexin DingTransformer Weight-Space Analysis
Witness Overlap: Directional Provenance Inside Open-Weight Model Families
Sep 24, 2026Siyuan Li, Haoxuan Zeng, Xin Luo +6LLM Auditing Open-Weight Language Models
ModelOS · University of Michigan · Carnegie Mellon University
The Ups and Downs of Backprop Weights
Sep 18, 2026Giuseppe Chindemi, Benjamin F. GreweModular Neural Networks Dynamic Neural Networks
Institute of Neuroinformatics, UZH / ETH Zurich ETH AI Center Zurich, Switzerland
Spectral characteristics of autoencoder parameters as a vector representation of data
Sep 3, 2026Maria Nikitina, Anton Bishuk, Oleg BakhteevRepresentation Learning Autoencoders
GRRR: The Geometry of Reshaping, Rotation, and Routing in Decoder LLM post-training
Aug 26, 2026Jianing Qi, Hao Tang, Zhigang ZhuLLM Post-Training Weight-Space Analysis
CUNY Graduate Center · The City College of New York
Block-Level Weight-Space Structure Persists Under Post-Training: An Empirical Study Across LLM Families
Aug 23, 2026Zhaohui WangLanguage Model Post-Training Weight-Space Analysis
Viterbi School of Engineering, University of Southern California, Los Angeles, USA.
Beyond Endpoint Gains: A Weight-Delta Audit of Medical Specialization
Aug 21, 2026Praphul Singh, Shanu Kumar, Akshat AgarwalLLM Auditing LLM Interpretability
Oracle Health AI · IIT Kanpur · MBZUAI
When Do Task Vectors Interfere? Mapping the Validity Boundaries of Weight-Space Composition
Aug 10, 2026Chencheng Zhu, Xiaoyang Li, Taotao CaiPrompt Sensitivity Weight-Space Analysis
UNSW Sydney · University of Southern Queensland
Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space
Aug 7, 2026Yiwei Chen, Bingqi Shang, Sijia LiuLanguage Model Fingerprinting Open-Weight Language Models
Michigan State University
Finding Usable Weight Mechanisms with Tiled SVD
Aug 7, 2026Ash Manvi, Samreena TajreenLLM Interpretability Mechanistic Interpretability
Aquin Labs
Cross-Layer Interaction under Weight-Space Ablation: A Closed-Form Attention Jacobian Bound and a Test on a Real Pretrained Model
Aug 4, 2026Abdallah KhemaisTransformer Interpretability Weight-Space Analysis
ISITCOM, University of Sousse
A Theory of Conditional Collapse under Low-Rank Weight-Space Ablations: I. The Single-Block Theory and Synthetic Validation
Aug 4, 2026Abdallah KhemaisTransformer Interpretability Weight-Space Analysis
ISITCOM, University of Sousse
Are the High-weight Neurons the Important Ones in Image Classification Neural Networks?
Jul 28, 2026Qitao Chen, Dongfu Yin, F. Richard YuNeural Network Interpretability Weight-Space Analysis
WeBank Institute of FinTech, ShenZhen University, YueHai, ShenZhen, 518060, GuangDong, China. · Artificial Intelligence and Digital Economy Guangdong Provincial Laboratory, YuTang, ShenZhen, 518132, GuangDong, China.
Neural spectroscopy of AlphaFold2 reveals encoded protein conformational landscapes
Jul 17, 2026Kaustav MehtaProtein Structure Prediction Neural Network Interpretability
Independent researcher, India
Weight-Space Physics: Interpretable Hypernetworks for Lattice Quantum Field Theories
Jul 8, 2026Tobias Göbel, Julian R. Ebelt, Zier Mensch +2Phase Transitions Neural Network Interpretability
Institute of Physics, University of Amsterdam, Amsterdam, Netherlands · Department of Physics, National Taiwan University, Taipei, Taiwan · Department of Physics, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA +2
Observable- and Positional-Encoding-Dependent Symmetry Readout from Neural Network Weights
Jul 3, 2026Naoya Chiba, Satoshi Sugiyama, Yuki UranishiNeural Representation Geometry Implicit Neural Representations
D3 Center, The Unversity of Osaka
WARP: Weight-Space Analysis for Recovering Training Data Portfolios
Jul 2, 2026Tzu-Heng Huang, Aditya Goyal, John Cooper +1Weight-Space Analysis Training Data Attribution
University of Wisconsin-Madison
Weight-Space Geometry of Offline Reasoning Training
Jun 21, 2026Aleksandr Nikolich, Igor Kiselev, Vladimir Platonov +1RL for Language Models Direct Preference Optimization
1Keenable AI · 2Accenture · 3Yandex
Different Layers, Different Manifolds: Module-Wise Weight-Space Geometry in Transformer Optimization
Jun 11, 2026Kirato YoshiharaStiefel Manifold Optimization Neural Network Optimization
School of Engineering Science, The University of Osaka, Osaka, Japan.
Weibull Weight-Scale Parameter Evolution under AdamW Training Dynamics
Jun 11, 2026Tiexin DingWeight-Space Analysis Adaptive Gradient Methods
Independent Researcher.
Exploiting weight-space symmetries for approximating curvature
May 30, 2026Artem Artemev, Rui Xia, Benjamin M. Boyd +4Neural Network Optimization Second-Order Optimization
MediaTek, Cambridge, UK · University of Cambridge, Cambridge, UK · Vector Institute, Toronto, Canada
Access Sets Matter: Budgeting Expert Reads for Scalable Weight-Space Model Merging
May 28, 2026Yuanyi Wang, Yanggan Gu, Su Lu +5Weight-Space Analysis Language Model Merging
The Hong Kong Polytechnic University, PolyU · Hong Kong Polytechnic University Daya Bay Technology and Innovation Research Institute.
Extra-Merge: Tracing the Rank-1 Subspace of Model Merging in Language Model Pre-Training
May 26, 2026Wenjie Zhou, Bohan Wang, Hongtao Zhang +3Language Model Pretraining Neural Network Optimization
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, China · University of Chinese Academy of Sciences, China · Alibaba Group, China +2
What Linear Probes Miss: Multi-View Probing for Weight-Space Learning
May 22, 2026Eunwoo Heo, Kyeongkook Seo, Jaejun YooMulti-View Learning Weight-Space Analysis
Graduate School of Artificial Intelligence, Ulsan National Institute of Science and Technology, Ulsan, Republic of Korea.
A Two-Parameter Weibull Framework for Diagnosing Transformer Weight Distributions
May 17, 2026Tiexin DingTransformer Weight-Space Analysis
Independent Researcher
The Geometry of Learning: Read-Write Asymmetry Across Training
May 15, 2026Valeria Ruscio, Keiran ThompsonLanguage Model Pretraining Transformer Interpretability
Intuition Machines

Oct 27, 2026J/K move · Enter open · S save

On the Sparsity of Direct Preference Optimization: Weight Disentanglement in the NTK Regime

Kensuke Sasaki, Issei Sato

Department of Computer Science, The University of Tokyo

Direct Preference Optimization LLM Alignment Structured Sparsity Weight-Space Analysis

Open paper AskPDF arXiv

Direct Preference Optimization (DPO) has been widely adopted for aligning language models with human preferences. Recent empirical work has revealed that DPO induces remarkably sparse parameter updates, yet no theoretical explanation for this phenomenon has been offered. We analyze DPO's optimization dynamics under a neural tangent kernel linearization and show that the logistic structure of the DPO loss gives rise to an implicit maximum-margin bias. Building on this observation, we introduce the concept of dual sparsity: the DPO update direction is a sparse linear combination (data-level sparsity, driven by support vectors) of sparse gradient feature vectors (parameter-level sparsity, arising from gradient cancellation between preferred and dispreferred responses). We formalize this structure in our main theorem, bound the number of support pairs by a margin-dependent spectral quantity, and present a framework relating dual sparsity to weight disentanglement---a condition favorable for model merging. Experiments on GPT-2 Small and Medium and on 7--9B instruction-tuned models (Llama-3-8B, Mistral-7B, Gemma-2-9B) with the Anthropic HH-RLHF and UltraFeedback datasets confirm both sources of sparsity, and experiments on GPT-2 show that DPO reduces cross-task interference by up to

35{\times}

compared to supervised fine-tuning.