Repair-side Network
A lightweight side network dynamically generated by a hypernetwork, equipped with a prompt-conditioned gating mechanism. It is used to selectively correct the behavior of a fine-tuned model at inference time. Because the network is generated rather than trained directly, corrections can be applied without modifying the original model parameters.
Timeline (last 90 days)
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).
The Repair Side Network (RSN) carries a prompt-conditioned gating mechanism that can selectively apply corrections.
The effectiveness of the repair side depends on whether the repair patterns the hypernetwork learns on the source model are general enough, and its transferability across base architectures also needs more evidence.
All Facts (3)
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).
50%UnverifiedThe Repair Side Network (RSN) carries a prompt-conditioned gating mechanism that can selectively apply corrections.
50%UnverifiedThe effectiveness of the repair side depends on whether the repair patterns the hypernetwork learns on the source model are general enough, and its transferability across base architectures also needs more evidence.
50%