Skip to main content
Search

# Weight Editing

2 related articles

Subscribe to updates
  1. arXiv

    Scientific feedback can rewrite weight-editing plans: HypoLoop cuts Burgers relative error from 0.040194 to 0.016559

    Scientific prediction models keep exposing new errors as experiments proceed, yet weight editing has typically refit the same parameter subset with the same objective. HypoLoop lets experimental feedback revise both which weights change and how errors drive the correction: fixed, numerically adaptive or LLM-guided policies propose an editing plan, numerical solvers compute the weight patch, and internal validation selects the deployed predictor. On Burgers, reallocating a matched editable-coordinate budget alone lowered relative error from 0.040194 to 0.016559 under DeepSeek and from 0.030975 to 0.016195 under Claude.

  2. arXiv

    TensorMutation skips retraining each mutant: a graph neural network predicts weight changes of pre-training mutations

    Deep-learning mutation testing has traded cost against realism: pre-training mutants resemble real faults more closely, but each one requires retraining the model. TensorMutation represents a trained model as a neural graph and uses a mutation-conditioned graph neural network to predict parameter changes directly, generating executable mutants without retraining. On 12 real-fault subjects its median behavioral similarity exceeds that of the post-training tool DeepMutation++, a difference that stays significant after Holm correction; differences from the pre-training tool DeepCrime and from a norm-matched random control are not significant.

You’re all caught up