Skip to main content

What a model believes, and how to tell when it changes.

Experiments, writing, and interactive pages I built to understand something. Most of what is here is provisional and says so, and the reading list marks which citations I have actually checked.

Projects

Notebook
edit-sliceMeasure whether weight-level knowledge edits (ROME/MEMIT) leave their own grounds intact and contradictory — the orphan category — and show that this backward failure is worse than the forward propagation these methods were tuned against.Phase 1 — Design
rome-neighborsWhen a fact is edited in a language model (e.g. "The Eiffel Tower is in Paris" → "The Eiffel Tower is in Rome"), the logically entailed neighbor facts should also update. They usually don't.Phase 2 — Implementation
claim-graphFeed real research papers in; get a queryable claim graph out — claims as nodes, typed edges (supports / contradicts / refines), every node carrying provenance back to its source sentence in a real paper.Phase 5 — LLM edge-typer

Watching a claim get smaller

I went to check a claim I liked against the benchmark most likely to have taken it. The claim survived, but smaller — and the smaller version is the one worth building on.

Papers
22
Read
14
To read
6
Citations unchecked
0
Bilinear representation mitigates reversal curse and enables consistent model editingThe structured-geometry predictor rome-neighbors adopts. Also bears on edit-slice: "reversal curse" is argument order under another name, so this is the closest published work to that half of the distinction. Read before E-009. next up
STEAM: A Semantic-Level Knowledge Editing Framework for Large Language ModelsLatent-space alignment for editing. Note the correction: this is an editing *method*, not an alignment-based *predictor* of propagation, so the "alignment arm" of the predictor comparison may not have a paper behind it yet. Check before building E-009 around it. next up
The Curious Case of Factual (Mis)Alignment between LLMs' Short- and Long-Form AnswersThe nearest existing method — mechanistic similarity predicting alignment at up to 78%. But the construct is short-form versus long-form answer agreement, not factual consistency in general, which makes it easier to differentiate from than the one-line note suggested. Still blocking, since it has to be cited either way. next up
MindThe Belief ProblemWhat it takes to give up a belief, and why removing one is harder than adding one. Walks expansion and contraction over the same belief set.
MindThe Researcher's StreamHow a line of inquiry actually moves — branching, parking, and coming back — rather than the tidy version that appears in a paper.
EconomicsOne CrateA supply chain told through its data problems. Four stops — measurement, Simpson's paradox, variance, the bullwhip effect — each reproducible from a seeded synthetic dataset.
AIClaim Knowledge Graph Explorer111 papers, 736 typed edges. Nodes are extracted claims rather than papers, so you can ask which claim supported or contradicted which.
MedicineApoB and the Cardiometabolic CascadeWhy the particle count matters more than the cholesterol number, traced through the cascade it sits in.
MedicineSpine Cracking — What Actually HappensThe mechanics behind the sound, and which of the common explanations survive contact with the evidence.