HyperRepair
A hypernetwork-based framework for inference-time model repair. It generates a repair-side network from an activation fingerprint in order to restore the general capabilities of a fine-tuned model. The framework requires neither retraining nor target-domain data, performing repair during inference by conditioning on a compact representation of the model's behavioral change relative to its base model.
Timeline (last 90 days)
The HyperRepair framework requires only inference-stage cost, with no retraining or additional target-domain data, to simultaneously preserve domain specialization and restore general capability.
HyperRepair learns a hypernetwork that uses activation fingerprints to summarize the model's changes relative to the pretrained base and generates a customized Repair Side Network (RSN).
HyperRepair's hypernetwork can transfer its repair capability to unseen models that share the same pretrained base, even if those models come from unseen adaptation domains.
The key difference between HyperRepair and per-model repair approaches is that repair is no longer a one-off manual process but is automatically generated by the hypernetwork according to model features.
The paper evaluates on two model families, 12 LoRA-adapted datasets, and 15 general benchmarks.
Under the default split, HyperRepair achieves 93.4% of the base model's average general score while retaining 99.5% of the fine-tuned model's average domain score.
HyperRepair shifts general capability restoration from retraining to inference-time correction.
The paper's evaluation focuses on LoRA adaptation and two model families, and the method's applicability to full-parameter fine-tuning or other adaptation forms remains to be verified.
All Facts (8)
The HyperRepair framework requires only inference-stage cost, with no retraining or additional target-domain data, to simultaneously preserve domain specialization and restore general capability.
50%UnverifiedHyperRepair learns a hypernetwork that uses activation fingerprints to summarize the model's changes relative to the pretrained base and generates a customized Repair Side Network (RSN).
50%UnverifiedHyperRepair's hypernetwork can transfer its repair capability to unseen models that share the same pretrained base, even if those models come from unseen adaptation domains.
50%UnverifiedThe key difference between HyperRepair and per-model repair approaches is that repair is no longer a one-off manual process but is automatically generated by the hypernetwork according to model features.
50%UnverifiedThe paper evaluates on two model families, 12 LoRA-adapted datasets, and 15 general benchmarks.
50%UnverifiedUnder the default split, HyperRepair achieves 93.4% of the base model's average general score while retaining 99.5% of the fine-tuned model's average domain score.
50%UnverifiedHyperRepair shifts general capability restoration from retraining to inference-time correction.
50%UnverifiedThe paper's evaluation focuses on LoRA adaptation and two model families, and the method's applicability to full-parameter fine-tuning or other adaptation forms remains to be verified.
50%