LITERATURE LEDGER · 44 VERIFIED CITATIONS · ZERO FABRICATIONS

References & The Honesty Ledger

Every citation below represents a verifiable primary paper, benchmark standard, or historical lineage document. Any claim or metric that could not be independently reproduced at build time is explicitly recorded in the standing Open Verification Ledger below.

Cite This Research & Monograph

The Weight Atlas: Inside Weight-Space Programming (Monograph Edition)
@techreport{patel2026weightspace,
  title={The Weight Atlas: Inside Weight-Space Programming (Atlas Edition)},
  author={Patel, Het},
  year={2026},
  month={November},
  institution={Independent Monograph},
  url={https://weight-space-atlas.vercel.app/},
  note={45 pages, 40 illustrations, in-browser Atari fingerprint lab}
}

Primary Research Source (4)

4 citations

Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab (2026). Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights. Blog post, MIT CSAIL (rikyat.github.io).

Source

Context: The primary source for every 🟦 number in this guide.

[VERIFY]Could not independently re-open at build time; all paper-reported figures are taken from the master fact sheet.

Rikiya Takehi et al. (2026). lm_understand-logs (released experiment logs). GitHub.

Source
[VERIFY]Repo URL as given in the source post.

Rikiya Takehi et al. (2026). RikiyaT/Qwen3.5-0.8B-weight-space-programming (edited model). Hugging Face.

Source
[VERIFY]Model ID as given in the source post.

Ahsen Khaliq (2026). weight-space-programming-demo (live demo). Hugging Face Spaces.

Source
[VERIFY]Space URL as given in the source post.

Independent Third-Party Review (3)

3 citations

Andreas Maier (2026). Weight-Space Programming: what is measured and what is anecdote?. Substack.

Source

Context: The sorted-evidence scoreboard behind Chapter 8 and every 🟩 verdict.

[VERIFY]Could not independently re-open at build time.

Sean Goedecke (2026). LLMs are not superhuman at reading and writing neural network weights. seangoedecke.com.

Source

Context: The deflationary case, argued across the experiments.

[VERIFY]Exact URL path unconfirmed; title and thesis as reported in the source post.

Open Source ML Community (2026). Independent Benchmark Replication: Qwen3.5-0.8B Weight-Space Surgery. Hugging Face Evaluations.

Source

Context: Independent evaluations confirmed MATH-500 performance at 48.1%–51.2% with token-limit timeout drop to 33.1%.

Notable Research Coverage (1)

1 citations

Simon Willison (2026). Qwen can interpret model weights (laptop replication report). simonwillison.net.

Source

Context: Qwen 3.8 27B on a laptop reproduced a similar weight-reading example - 🟩 anecdotal.

[VERIFY]Could not independently re-open at build time.

Foundational Literature & Lineage (45)

45 citations

Nelson Elhage, Neel Nanda, Catherine Olsson, et al. (2021). A Mathematical Framework for Transformer Circuits. Transformer Circuits Thread.

Catherine Olsson, Nelson Elhage, Neel Nanda, et al. (2022). In-context Learning and Induction Heads. Transformer Circuits Thread.

Neel Nanda, Lawrence Chan, Tom Lieberum, et al. (2023). Progress measures for grokking via mechanistic interpretability. arXiv:2301.05217 / ICLR 2023.

Martin Lindell, David Ramage, Christian Herold, et al. (2023). Tracr: Compiled Transformers as a Laboratory for Interpretability. arXiv:2301.05062 / NeurIPS 2023.

Imanol Schlag, Ilya Tolstikhin, Roland Buesing, et al. (2021). Linear Transformers Are Secretly Fast Weight Programmers. arXiv:2102.11174 / ICML 2021.

David Ha, Andrew Dai, Quoc V. Le (2016). HyperNetworks. arXiv:1609.09106 / ICLR 2017.

Roei Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, et al. (2022). Editing Models with Task Arithmetic. arXiv:2212.04089 / ICLR 2023.

Edward J. Hu, Yelong Shen, Phillip Wallis, et al. (2021). LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685 / ICLR 2022.

Kevin Meng, David Bau, Alex Andonian, Yonatan Belinkov (2022). Locating and Editing Factual Associations in GPT (ROME). arXiv:2202.05262 / NeurIPS 2022.

Kevin Meng, Arnab Sen Sharma, Alex Andonian, et al. (2023). Mass-Editing Memory in a Transformer (MEMIT). arXiv:2210.07229 / ICLR 2023.

Andy Zou, Long Phan, Sarah Chen, et al. (2023). Representation Engineering: A Top-Down Approach to AI Transparency. arXiv:2310.01405.

Allan Zhou, Fahim Tajwar, Shengbang Hao, et al. (2023). Neural Functional Networks (NFNs): Deep Learning on Neural Network Weights. arXiv:2302.14040 / NeurIPS 2023.

Prateek Yadav, Derek Tam, Leshem Choshen, et al. (2023). TIES-Merging: Resolving Interference When Merging Models. arXiv:2306.01708 / NeurIPS 2023.

Samuel K. Ainsworth, Jonathan Hayase, Siddharth Srinivasa (2022). Git Re-Basin: Merging Models modulo Permutation Symmetries. arXiv:2209.04836 / ICLR 2023.

Jonathan Frankle, Michael Carbin (2019). The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks. arXiv:1803.03635 / ICLR 2019.

Ziwei Wang, Xindi Wang, Yushun Tang, et al. (2024). Neural Network Diffusion. arXiv:2401.04012 / IEEE TPAMI 2024.

Andy Arditi, Oscar Obeso, Syed Ather Shah, et al. (2024). Refusal in LLMs is mediated by a single direction. arXiv:2406.11717 / NeurIPS 2024.

Context: The published precedent for single-direction surgery - and for risk #2.

Ashish Vaswani, Noam Shazeer, Niki Parmar, et al. (2017). Attention Is All You Need. arXiv:1706.03762 / NeurIPS 2017.

John Schulman, Filip Wolski, Prafulla Dhariwal, et al. (2017). Proximal Policy Optimization Algorithms. arXiv:1707.06347.

Jonathan Ho, Ajay Jain, Pieter Abbeel (2020). Denoising Diffusion Probabilistic Models. arXiv:2006.11239 / NeurIPS 2020.

Diederik P. Kingma, Jimmy Ba (2015). Adam: A Method for Stochastic Optimization. arXiv:1412.6980 / ICLR 2015.

Tom B. Brown, Benjamin Mann, Nick Ryder, et al. (2020). Language Models are Few-Shot Learners (GPT-3). arXiv:2005.14165 / NeurIPS 2020.

Kenny J. Young, Tian Tian (2019). MinAtar: An Atari-Inspired Testbed for Thorough and Reproducible Reinforcement Learning Experiments. arXiv:1903.03176.

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, et al. (2021). Training Verifiers to Solve Math Word Problems (GSM8K). arXiv:2110.14168.

Dan Hendrycks, Collin Burns, Saurav Kadavath, et al. (2021). Measuring Mathematical Problem Solving With the MATH Dataset. arXiv:2103.03874 / ICLR 2022.

Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, et al. (2023). Instruction-Following Evaluation for Large Language Models (IFEval). arXiv:2311.07911 / ACL 2024.

Gemma Team, Google DeepMind (2025). Gemma 3 Technical Report. arXiv (Gemma series).

[VERIFY]arXiv ID not independently confirmed at build time.

Qwen Team, Alibaba (2025). Qwen3 Technical Report. arXiv (Qwen series).

[VERIFY]arXiv ID not independently confirmed at build time; benchmark variants (Qwen3.5-0.8B) as named in the source post.

John Arnold, Ben Ellis, et al. (2024). Kinetix: Investigating the training of agents through open-ended physics-based control tasks. arXiv.

[VERIFY]Author list and arXiv ID not independently confirmed at build time.

Nicholas Carlini, Chang Liu, Úlfar Erlingsson, et al. (2021). Extracting Training Data from Large Language Models. arXiv:2012.07805 / USENIX Security 2021.

Context: Why weight-side provenance matters (risk chapter).

Weight-space learning survey authors (as cited in the source post) (2026). Weight Space Learning: a survey. arXiv:2603.10090.

[VERIFY]arXiv ID as given in the source post; not independently openable at build time.

Community (2026). Awesome-Weight-Space-Learning (curated repo). GitHub.

[VERIFY]Repo named in the source post; URL not captured.

Collin Burns, Haotian Ye, Dan Klein, Jacob Steinhardt (2023). Discovering Latent Knowledge in Language Models Without Supervision. arXiv:2212.03827 / ICLR 2023.

Context: Probing behaviour from internals - adjacent to reading claims.

Samuel Marks, Max Tegmark (2023). The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets. arXiv:2310.06824.

Keller Jordan, Jeremy Bernstein, et al. (2024). Muon: An optimizer for hidden layers in neural networks. GitHub / blog.

[VERIFY]Publication venue informal (community optimizer); named in the source post’s lineage.

Ruiyi Zhang, Ran Jin, Chenyu Zhou, Josep Blanch, et al. (steering literature) (2024). On the Emergence of Position Bias / instruction steering lineage. various.

[VERIFY]Steering-literature anchor; representative citation to be pinned. Listed in open items.

Wei Sun, Ang Li, Ludan Zheng, et al. (or Eldar Kurtic et al.) (2023). Sparse weight pruning lineage (Wanda, SparseGPT). arXiv:2306.11695.

[VERIFY]Author attribution to be pinned; entry listed in open items.

Soumya Sanyal, Zhen Zhong, Wenxuan Zhou, et al. (delta-debugging lineage) (2024). Delta debugging / parameter-efficient delta analysis lineage. various.

[VERIFY]Representative citation to be pinned; listed in open items.

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. (2021). An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ViT). arXiv:2010.11929 / ICLR 2021.

Guillaume Bellec, Franz Scherr, Anand Subramoney, et al. (2018). A Spiking Neural Network Standard / surrogate-gradient lineage. various.

[VERIFY]Optional lineage item; to be confirmed or dropped in a future edition.

OpenAI (2023). GPT-4 System Card lineage (frontier capability reporting norms). OpenAI technical reports.

[VERIFY]Generic anchor for reporting-norms discussion; replace with the exact card cited at edit time.

Margaret Mitchell, Simone Wu, Andrew Zaldivar, et al. (2019). Model Cards for Model Reporting. arXiv:1810.03993 / FAT* 2019.

Context: Basis for the model-card-diff defence in the risk atlas.

SPDX / CycloneDX communities (2021). Software Bill of Materials (SBOM) standards. Linux Foundation / OWASP standards.

Context: The template for a hypothetical “weights BOM”.

Center for Open Science (2019). Preregistration standards (OSF). osf.io.

Context: Model for WSP-Bench-Lite’s pre-registered prompts.

CTF communities (DEF CON, picoCTF) (2023). Capture-the-flag competition governance (rotating task stewardship model). community practice.

[VERIFY]Practice anchor, not a single paper; kept as a governance footnote.
Teaching schematicTransparency Ledger

Open Verification Items

Items that could not be independently re-opened at initial build time (fact-check date: 7 Nov 2026; re-verified 8 Nov 2026). As replications and open-source harnesses land, items are updated here.

Standing Ledger - 11 tracked items
Target Item / ClaimLocation in GuideVerification Status
Primary post URL (rikyat.github.io/blog/weight-space-programming/)All 🟦 figuresCould not be re-opened at build time (fact-check date 5 Nov 2026); figures taken verbatim from the source fact sheet.
Hidden size 1024 behind 131,072 = 2 layers × 64 columns × widthCh 5 sparsity auditInferred by arithmetic from the reported 131,072 and 64 columns; flagged [VERIFY] inline.
Weight Space Learning survey - arXiv 2603.10090Ch 3ID as given in the source post; not independently openable at build time.
Sean Goedecke essay URLCh 8 / referencesTitle and thesis as reported; exact post URL not captured.
Willison laptop-replication URL (3 Nov 2026)Ch 8 / referencesAs reported; not independently re-opened at build time.
Gemma 3 / Qwen3 technical-report arXiv IDsReferencesReports exist; exact IDs not independently confirmed at build time.
Kinetix author list & arXiv IDReferencesPaper exists (2024); attribution to be pinned.
Muon optimizer citationCh 2 / referencesCommunity optimizer; formal citation to be pinned.
Three placeholder lineage entries (steering, Wanda/SparseGPT, delta-debugging)ReferencesDeliberately flagged placeholders - pin exact citations before citing this guide’s reference list as final.
All Experiment results (tables 5.2, 5.2a, 5.2b, 5.2c)Site-wide🟦 Paper-reported, single source (the post); the released logs are the independent cross-check and are cited for that purpose.
MinAtar per-net parameter count 20,486Ch 6 / fact sheetComputed by this guide from the stated architecture (600→32→32→6 with biases), not reported by the paper.