Repository navigation
Open
Description
Follow-up to #45 (self-repair falsified: CE≈0 everywhere). The dip→recover of balef2026 Exp6 is passive redundancy, not active compensation. This issue is about the next question: what is the redundancy mechanism? Parked — no work planned yet.
Where we stand
- Ablating a layer dips a depth-
m+1tuned probe, but the answer is still decodable at the end → TE≈0. - Frozen-downstream DE shows no rewrite (CE = DE − TE ≈ 0; per-row unimodal at 0; CE~DE R²≤0.08; Mitra linear anchor agrees).
- So nobody compensates — the information was already duplicated / cheaply recomputable. Untested: how.
Why TFMs plausibly differ from LLMs: the input table stays in context → recompute is cheap → no need for backup-head-style active repair.
But "DE≈0" is not one thing
margin, |·| > 0.1:
| model | origin-redundant (DE≈0, TE≈0) | indirectly important (DE≈0, TE large) |
|---|---|---|
| TabFM | 78% | 10% |
| Mitra | 44% | 40% |
| TabICL | 41% | 40% |
| LimiX-2M | 16% | 53% |
- TabFM = cleanest subject (mostly pure redundancy).
- LimiX-2M = least redundant — early/mid layers do necessary feature work (dip does not recover).
- Any mechanism claim must be stated per model, not globally.
Hypotheses
- H1 — copies exist. The same information sits in the residual at many depths.
- test: linear probe at every depth / cross-layer representation similarity → near-ceiling decodability well before the load-bearing layers.
- H2 — recompute from the input table. After ablating
m, later layers re-derive rather than read another cached copy.- test: co-ablate layer
mand the attention path to the input rows → recover should vanish if H2, survive if H1.
- test: co-ablate layer
- H3 — k-layer joint ablation (cheapest, most decisive). One layer can't remove the information; how many can?
- genuine redundancy → TE has a threshold / phase transition in
k. - indirect importance → TE grows linearly from k=1.
- cost: reuses
inject_delta, just inject at k layers at once. No new method.
- genuine redundancy → TE has a threshold / phase transition in
- H4 — who writes the copies, attention vs MLP sublayers.
Suggested order
H3 (infra already exists) → H1 (probes already exist) → H2 → H4.
Links
- PR feat: path-patching DE/TE on the native head — DE–TE scatter (#44, Part 1) #45 — DE/TE path-patching, per-row, CE~DE law, results + method note
- Transfer LLM self-repair to TFMs — 3-part plan (path-patching DE → CE regression → LN artifact) #44 — original 3-part self-repair plan (now on its falsification branch)
- Path-patching direct effect: tell whether mid-late layers are redundant or self-repairing #37 — the earlier "redundant or self-repairing?" framing this answers
Metadata
Metadata
Assignees
Labels
No labels
Activity