Verified, then logged.
Sun 22 Nov 2026 · 00:33 KST / 15:33 UTC한국어 | ENSearch /

01 / AI · Claim vs. Check · Analysis

Study tests language model weight repair without retraining

A fixed projection matched live selection closely on Pythia, while both procedures used the same family of operations in a separate test with a compound fault.

A research monitor shows a damaged language-model weight matrix beside two repaired versions, with nearly equal sentence-completion accuracy bars below the repairs. AI-generated image

Researchers said they analysed 20 language model repair sessions to compare ways of editing model weights without further training.

The study examined what researchers gained by choosing a repair in response to the target model’s behaviour. On Pythia, a fixed projection and live selection produced similar sentence completion accuracy.

The researchers explained that an external language model chose edits based on predictions and weight statistics. Each projection recipe specifies which mathematical operations to apply to the weights.

Fixed projection on Pythia

In the researchers’ published results for Pythia, the fixed projection and live selection achieved similar sentence completion accuracy. A search that used outcome scores to choose edits scored lower than both.

In the same published results, live selection and the score-based search performed equally well on SmolLM2 and Qwen2.5.

A shared family of operations for the compound fault

In a separate test with a compound fault, the researchers compared a scripted procedure with one in which a language model selected the edits. Both used the same family of mathematical operations.

In the researchers’ published results, average accuracy rose from 24.89% to 71.25% with each procedure.

The researchers also tested scripted projections on models without injected faults. Their published results showed small gains or losses. The abstract did not disclose scores for individual models in that test.