JNO Fixes How LLMs Handle Fact Updates Without Breaking Related Knowledge
A new framework called JNO jointly handles the two competing pressures of knowledge editing in LLMs, improving propagation and preservation metrics by at least 7%
- The problem: Single-fact edits in LLMs trigger ripple effects that either fail to propagate to related facts or corrupt unrelated ones.
- The insight: Propagation and preservation pressures are coupled, not independent; existing methods like ROME and MEMIT treat them separately and pay the price.
- The solution: JNO (Joint Neighborhood Optimization) jointly optimizes neighborhood targets under coupled constraints before any weights are modified.
- Key mechanism: A semantic pre-execution gate rejects high-risk edit plans before parameter execution, acting as a safety filter.
- Results: At least 7% improvement on both propagation and preservation metrics on the RippleEdits benchmark, stable across Qwen2.5-1.5B, GPT-J-6B, and LLaMA3-8B.
- Limitation: Evaluated on single-edit scenarios only; no code or weights released; sequential/batch editing behavior is untested.
When you update a fact inside a large language model, the change rarely stays contained. Tell the model that Kevin Durant now plays for the Houston Rockets instead of the Phoenix Suns, and you want it to automatically infer that his team's city, his teammates, and his league affiliation have all changed too. But you also need it to leave unrelated facts, like his height or his college career, completely untouched. This tension is the core challenge of knowledge editing, and it has been quietly resisting clean solutions for years.
A new paper, "Revisiting Ripple Effects in Knowledge Editing through Pressure-Aware Joint Neighborhood Optimization", argues that the field has been solving this problem the wrong way: by treating propagation and preservation as two separate concerns. The authors propose that these two forces are fundamentally coupled, and that ignoring their interaction is exactly why existing methods keep failing.
The Edit That Breaks Everything Nearby
LLMs accumulate substantial factual knowledge during pretraining, but this information often becomes outdated, compromising model reliability. Given the prohibitive cost of retraining, knowledge editing has emerged as a vital technique to update LLMs by injecting new information or modifying existing parameters. The appeal is obvious: surgical, cheap, no full retraining required.
The problem is what happens downstream from the edit. Injecting one fact, such as "Jack Depp is the son of Johnny Depp," introduces a ripple effect in the form of additional facts that the model needs to update, such as "Jack Depp is the sibling of Lily-Rose Depp." These downstream implications are called ripple effects, and they split into two categories:
- Desirable propagation: facts logically implied by the edit that should also update (e.g., team city, league membership)
- Unintended perturbation: unrelated facts that get corrupted as a side effect of the edit
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
