Skip to content

Latest commit

 

History

History
93 lines (71 loc) · 6.09 KB

File metadata and controls

93 lines (71 loc) · 6.09 KB

How Yaz compares to other knowledge-editing methods

Read this honestly. Yaz is a ≈807K-param research prototype on 50 synthetic facts. It is not state-of-the-art, not larger-scale, and not a unique capability. This page exists so you can place it accurately against the literature — including where it loses. Every Yaz number here is reproducible from this repo; every cross-method property is qualitative and from the public papers (cited where we're confident, marked otherwise).

The three families of knowledge editing

Family Idea Examples
Edit the base weights Change the model's own parameters so it "knows" the new fact ROME, MEMIT, MEND, KnowledgeNeurons
Add a side memory / adapter Leave the base frozen; route to an external store of edits SERAC, GRACE, WISE, MELO, PENME, RECIPE
Edit in context Put the correction in the prompt IKE and in-context variants

Yaz is in the side-memory family, with two unusual properties: its edit is a structural column swap (locality holds by construction, not just empirically), and it abstains on low routing confidence instead of answering with the base model.

Capability matrix (qualitative)

Method Edit lives in Retrain-free edit Locality Sequential editing Abstains when unsure? Model-agnostic
ROME base FFN weights (rank-1, closed-form) yes empirical; known to degrade under many sequential edits limited before quality loss no architecture-specific (GPT-style)
MEMIT base FFN weights (many layers) yes empirical scales to many edits, but selectivity erodes no architecture-specific
MEND base weights via a trained hypernetwork needs a trained hypernetwork empirical limited no per-model training
SERAC external memory + classifier + counterfactual model yes scoped by a learned classifier many no (routes to base) wrapper
GRACE added codebook of activations yes strong, but keyed on activations (paraphrase-sensitive) thousands (paper) no (routes to base) layer-specific
WISE added side memory yes strong designed for long edit streams no (routes to base) wrapper
PENME embedding-keyed adapter memory yes scoped by embedding many no wrapper
Yaz its own decoder columns (atoms) yes structural (disjoint columns) flat to 40 edits (tiny scale) yes (routing margin) no (intrinsic to Yaz)

The one column where Yaz stands alone is abstention: published editors treat low confidence as "route to the base model and answer"; Yaz declines ("I'm not sure which fact you mean"). This is a real, under-occupied feature — but it is a step, not a moat: selective prediction + a margin threshold is textbook, and any of these editors could add it.

Empirical head-to-head — selectivity vs ROME/MEMIT (our controlled study)

These numbers are from our own controlled experiments on the same tiny synthetic task (country→capital, a ≈256K-param Yaz, ROME/MEMIT applied to the model's unembedding). They are not a benchmark of ROME/MEMIT on production LLMs — they show how the mechanisms behave on an identical small task. The qualitative finding (weight-editing loses selectivity as edits accumulate) is independently well-established in the literature.

At 5 facts — equivalent: Yaz and ROME both ~0/4 side-effects. Weight editing is fine when edits are few.

At 50 facts (5 UPDATE edits, side-effects measured against the other 49 facts):

ROME MEMIT (joint) MEMIT-proper (cov) Yaz
Edits hitting target rank-1 4/5 4/5 4/5 4/5 (one PARTIAL)
Non-edit side-effect rate 22–67% 62% 67% 0%
Aggregate verdict DEAD 0/5 DEAD 0/5 DEAD 0/5 ALIVE 4/5

ROME/MEMIT land the new answer but corrupt a large fraction of the other facts. Yaz's structural locality keeps side-effects at 0/49.

At 200 facts (selectivity holds, but Yaz's own reliability degrades):

Yaz (256K, 200 facts)
Max side-effect rate per UPDATE 0–0.5% (≤1/199)
UPDATE reliability drops to 1/5 ALIVE, 5/5 PARTIAL — the tiny model can only memorize 101/200 facts

Honest read: Yaz's advantage is selectivity/locality (it edits one fact without touching others), which weight-editing loses at scale. Yaz's weakness is reliability and capacity — at 200 facts the 807K/256K model can't hold all the facts, so UPDATE success falls. ROME/MEMIT scale to far more real-world knowledge than Yaz ever has.

Yaz's published-model numbers (reproducible here)

The shipped model is the 50-fact semantic_v2 routing/abstention model (python scripts/scaling/s3_route_abstain.py):

  • Abstention risk-coverage AURC 0.004 (oracle 0.003; random 0.087); coverage @ ≤5% risk 96.6%.
  • Paraphrase routing reach 0.696 held-out (vs 0.216 surface routing).
  • Per-edit locality 0/10 collateral, bpc +0.000% across 40 sequential edits; retention flat 1.000.
  • CREATE passes a 4/4 battery (monosemantic / local / readable / deletable).

Caveats on this comparison

  • Provenance: the ROME/MEMIT selectivity numbers come from an earlier Yaz CRUD study (d_model=64, ROME-on-unembedding), a different experimental line from the shipped routing model. They're real and in this project's results, but they are a controlled-task illustration, not a production benchmark.
  • First-byte only: Yaz edits set the answer's first byte; full-word transfer ≈0.05. The selectivity comparison is on first-byte targets.
  • Tiny, synthetic: 50–200 country→capital facts. None of this is validated on open-vocabulary knowledge.
  • Not a moat: every mechanism above is published and copyable. Yaz is a clean recombination whose only genuine differentiators are sub-1M-param/CPU scale and abstention-as-refusal.

Cross-method specifics (exact max-edit counts, tool support in EasyEdit, etc.) are intentionally left qualitative here rather than risk citing an unverified number. See each method's paper for precise figures.