Watching a claim get smaller

These are lab notes, not results. I'm starting a project called edit-slice and writing down what I learn while it is still provisional — including the parts that turn out to be wrong. Today's entry is about a claim of mine shrinking, which was the most useful thing that happened all day.

The setup

When you edit a fact into a language model's weights — ROME (Meng et al., 2022), MEMIT (Meng et al., 2023), and friends — you change one thing and disturb an unknown amount of other things. The standard way to evaluate this is to check the edit's consequences. Move the Eiffel Tower to Rome, then ask what country it's in, what language people speak nearby, what else is in Rome. Does the change ripple outward correctly?

The thing I got interested in runs the other way. The Eiffel Tower is in Paris partly because of a set of other facts — it was built for the 1889 Paris Exposition, it stands on the Champ de Mars. Those aren't consequences of its location. They're closer to the reasons the location was true in the first place.

Edit the tower into Rome and those facts are still sitting there, untouched, quietly implying the thing you just deleted. The model now holds a belief and a complete set of grounds for its negation. I've been calling that an orphan, and the claim I started with was roughly: everyone probes forward, nobody probes backward.

The suspicion

That claim was too comfortable, which is usually a bad sign.

The obvious place for it to break is the RippleEdits benchmark (Cohen et al., 2024). One of its evaluation criteria is Logical Generalization, and I knew it covered symmetric and inverse relations. Symmetric relations sound backward. If Logical Generalization is already a grounds probe, my claim isn't narrow — it's taken.

I could have gone on citing a summary. Instead I went to read the definitions.

Two things I found

First, a bookkeeping thing that isn't really about the argument but is worth saying out loud. I had a note in another repo describing this paper as EMNLP 2023. It's TACL 2024 (arXiv 2307.12976). I don't know how many times I'd read past that. Secondhand notes decay silently, and I'd been treating mine as though they couldn't.

Second, the actual definition. Logical Generalization says that relations satisfy logical constraints — Sibling is symmetric, so if (e, Sibling, o) holds then (o, Sibling, e) holds too — and checks whether those constraints survive the edit. Location gets checked for transitivity the same way.

So it does probe something that looks backward. Edit (Eiffel, located-in, Rome) and it will ask what Rome contains.

But that's the same triple with its arguments swapped. It's the edit restated, not the reasons for the edit. And the transitive case is sharper: to get from "Eiffel is in Rome" to "Eiffel is in Italy" you have to use the fact that Rome is in Italy — but the benchmark only ever checks the conclusion. Whether the premise survived is never asked.

That distinction turned out to be the whole thing, so I now write it down every time:

• Argument order — swap the arguments of the edited fact. Covered, by RippleEdits, since 2024.
• Justification order — the distinct facts whose truth was a premise for the edited fact. Not covered by anything I can find.

"The Eiffel Tower was built for the 1889 Paris Exposition" isn't the location fact inverted. It's a different fact about a different thing, and it's the one that goes stale in silence.

The claim, smaller

I can't say backward probing is unexplored. I can say:

Existing ripple evaluation probes the logical closure of the edited triple. Inverse and symmetric probes exist and are backward in argument order. Nothing probes backward in justification order.

That's a much less exciting sentence and a much more defensible one. It also cost me nothing except an afternoon, which is the point: I'd rather find the counterexample to my own claim than have a reviewer find it for me, in a room, in front of people. Same information, very different afternoon.

What I'd take from it

Three things, none of them profound, all of them things I apparently needed to relearn:

1. Read the primary source before designing around a summary. Both errors here — the wrong venue and the almost-scooped claim — came from trusting notes.
2. Go looking for the strongest counterexample to your own idea, early. If it holds, you've lost a week. If it doesn't, you've gained a defensible sentence.
3. A claim getting smaller is progress. It felt like a setback for about ten minutes and then it felt like the first solid ground the project has had.

Still entirely open: whether "the grounds of a fact" is even a well-posed notion in a system that stores everything in superposition and reasons probabilistically. I have a design and no data. More when there's a number.

edit-slice is a research instrument for measuring directional propagation of weight-level knowledge edits. It's early. These notes are written to be read over the shoulder, not cited.

References

1. Cohen, R., Biran, E., Yoran, O., Globerson, A., Geva, M (2024). Evaluating the Ripple Effects of Knowledge Editing in Language Models. TACL.
2. Meng, K., Bau, D., Andonian, A., Belinkov, Y (2022). Locating and Editing Factual Associations in GPT. NeurIPS.
3. Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., Bau, D (2023). Mass-Editing Memory in a Transformer. ICLR.