Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

Weights and activations to language

Diff Interpretation Tuning, LatentQA, Activation Oracles, the Jacobian lens and LoRA.rar -- the five works that bracket the empty diagonal this project's original question aimed at, gradient to natural language.

Why these five sit on one page

CONTEXT.md section 2 lays out a grid and points at the hole in it. "Gradient → natural language" is empty: "The gradient interpretability literature runs the other way (integrated gradients, saliency: gradients used to explain outputs). Weight-diff → language exists (Diff Interpretation Tuning, 2510.05092); activation → language exists (LatentQA 2412.08686, Activation Oracles 2512.15674, NLA); gradient → causal attribution exists (Jacobian lens, 2607.15495). The diagonal is empty." LoRA.rar (2412.05148) is named separately as the nearest precedent for a learned weight-space transform rather than a closed-form one.

That empty diagonal is the origin of this project. The trait zoo is what the question turned into once the gradient experiments said what they said -- see The drift experiment — weight-space control of a trait fails.

Diff Interpretation Tuning

Avichal Goel, Yoon Kim, Nir Shavit, Tony T. Wang. Learning to Interpret Weight Differences in Language Models. arXiv:2510.05092 [cs.LG, cs.AI, cs.CL], submitted 6 October 2025.

Weight diff in, natural-language description of what the finetuning changed out. The method, Diff Interpretation Tuning (DIT), trains a DIT-adapter on synthetic labelled weight diffs; applied to a compatible finetuned model, the adapter makes that model describe how it has changed. Two proof-of-concept settings: reporting hidden behaviours, and summarising finetuned knowledge. The motivation is that finetuning datasets are often unavailable or too large to inspect, so the weights have to speak for themselves.

This is the published work closest to reading a trait adapter. The reading list flags a specific confusion worth not repeating: DIT is not the LoRAcles paper that a previous session first mistook it for, and it is prior work rather than the target.

LatentQA

Alexander Pan, Lijie Chen, Jacob Steinhardt. LatentQA: Teaching LLMs to Decode Activations Into Natural Language. arXiv:2412.08686, submitted 11 December 2024; v2 23 March 2026. ICLR 2026.

Open-ended natural-language question answering about a model's activations, trained by a pipeline that generates the QA data automatically. The decoder beats probing baselines on supervised reading tasks (uncovering hidden prompts, knowledge extraction) and is "precise enough to steer the target model to exhibit behaviors unseen during training".

The reading-list entry marks one convention as unexamined: "The decoder LLM is always initialized as a copy of the target LLM", with no justification found and no ablation comparing a clone against an independent decoder in this paper or in Activation Oracles. If that ablation exists somewhere, it changes the design answer for any reader built on the zoo.

Activation Oracles

Adam Karvonen, James Chua, Clement Dumas, Kit Fraser-Taliente, Subhash Kantamneni, Julian Minder, Euan Ong, Arnab Sen Sharma, Daniel Wen, Owain Evans, Samuel Marks. Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers. arXiv:2512.15674 [cs.CL, cs.AI, cs.LG], submitted 17 December 2025; last revised 6 January 2026.

LatentQA taken generalist. Activation Oracles are LatentQA-trained models evaluated far out of distribution, with a study of how performance scales with training-data diversity. The headline that matters here: AOs "can recover information fine-tuned into a model (e.g., biographical knowledge or malign propensities) that does not appear in the input text, despite never being trained with activations from a fine-tuned model." Their best oracles match or exceed white-box baselines on all four downstream tasks.

Mechanism detail worth carrying, from the reading list: activations are added, not replaced, at placeholder token positions at layer 1 of the oracle, and norm-matched. Replacement blew the activation norms up. The oracle accepts vectors from several source depths but the sites are still chosen by hand.

Clement Dumas is a co-author here and on Persona Cartography (Baines et al., 2026).

The Jacobian lens

Wes Gurnee, Nicholas Sofroniew, Adam Pearce, Mateusz Piotrowski, Isaac Kauvar, Runjin Chen, Anna Soligo, Paul Bogdan, Euan Ong, Rowan Wang, Ben Thompson, David Abrahams, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson, Jack Lindsey. Verbalizable Representations Form a Global Workspace in Language Models. arXiv:2607.15495 [cs.CL, cs.AI, cs.LG], submitted 16 July 2026. Also published as a Transformer Circuits Thread article, 6 July 2026 -- the blog post precedes the arXiv posting by ten days.

The Jacobian lens computes, per layer, the average linearised effect of an activation on the model's likelihood of producing each vocabulary token; the resulting ranking over tokens reads as a description of the activation. The paper positions it against the older lenses: the logit lens applies the unembedding matrix directly to the intermediate residual stream, and the J-lens "can be understood as the principled correction". The representations it surfaces behave like a global workspace -- reportable, deliberately summonable, usable in reasoning steps, concentrated in intermediate layers, holding on the order of tens of concepts. Post-training shapes this workspace toward an assistant perspective, which is the connection to The Assistant Axis (Lu et al., 2026). The paper also introduces counterfactual reflection training.

This is CONTEXT.md's "gradient → causal attribution" corner: it takes derivatives and emits attribution, not language. The reading list points at Appendix A.9.2, an oracle lens that does emit free-form phrases, whose reconstructor is "held frozen during RL training" and which the authors suspect is "less likely to produce confabulations than NLAs". Code for the J-lens is open; the oracle lens was not released.

LoRA.rar

Donald Shenaj, Ondrej Bohdal, Mete Ozay, Pietro Zanuttigh, Umberto Michieli. LoRA.rar: Learning to Merge LoRAs via Hypernetworks for Subject-Style Conditioned Image Generation. arXiv:2412.05148, submitted 6 December 2024; v2 10 August 2025. ICCV 2025.

A hypernetwork pretrained on diverse content-style LoRA pairs that emits merge coefficients for a new pair, reported as over 4000x faster than optimisation-based merging while improving content and style fidelity. Vision personalisation, not language.

CONTEXT.md section 2 names it as the nearest precedent for the one thing that literature does not have: "a learned transform rather than closed-form projection". It is the contrast case for The gradient-projection lineage -- same job, learned rather than derived -- and the reason that gap was judged genuinely open rather than merely unsearched.

What this project takes, and where it departs

The zoo exists because reading turned out to work and control turned out not to. CONTEXT.md section 3.4 reports a fit-free probe -- no trained classifier -- in which a single gradient step per document carries the document's content: pooled effect +0.049 at z=25.1, and at the best layer +0.104, z=42.0, retrieval MRR 0.508, meaning the right fact out of 200 about half the time from one gradient step. Section 3.5 found traits more legible than planted facts: the register effect is +0.090 at z=232, roughly twice the planted-fact identity effect.

Section 3.6 is the constraint every reader in this family has to live with. "A gradient carries a trait while the trait is being ACQUIRED, and goes blind once the model has learned it." The register effect by training quartile runs +0.0150, +0.0061, +0.0040, +0.0004. The consequence stated there: "gradient-based monitoring is available; gradient-based forensics is not. You can watch persona drift happen; you cannot audit an already-drifted model from its gradients."

And one warning to carry into any variance-explained number reported by this family: CONTEXT.md section 4 records natural-language autoencoders reaching 0.6-0.8 variance explained while recovering none of 988 planted facts. Variance explained and content recovered came apart almost completely.

Verification note

Abstract only, all five. Each arXiv abstract page was fetched 2026-09-07 for title, authors, dates, subject class and venue. The reading list at ~/projects/agent-harness/memory/library/persona-cartography-reading-list.md marks LatentQA, Activation Oracles and the workspace paper [read] and Diff Interpretation Tuning [abstract only]; the mechanism details attributed to the reading list above are its readings, not re-verified here. LoRA.rar was not in the reading list and is abstract-only from both sources. None of the five appears in qwen35/paper_notes.md.