Australian Researchers Fix AI Amnesia With SineProject's One-Line Projector Trick
Researchers trace catastrophic over-refusal in multimodal unlearning to ill-conditioned projector Jacobians, and fix it with a sinusoidal reparameterization that costs under 1% overhead.
ProTakeaways
- SineProject fixes over-refusal in unlearned multimodal LLMs by stabilizing the vision-language projector's Jacobian.
- Root cause: projector condition numbers grow 3-4 orders of magnitude during gradient-based unlearning, wrecking cross-modal alignment.
- Method wraps trainable weight deltas in sine:
W + sin(ΔW), keeping perturbations bounded to [-1, 1]. - Cuts Safe Answer Refusal Rate by 15% on LLaVA-7B and 8% on 13B while keeping 100% refusal on harmful queries.
- Adds under 1% overhead to time, memory, and FLOPs; works on MLP and attention projectors alike.
- Full paper and proofs at arxiv.org/abs/2511.18444, code to be released on acceptance.
Getting a multimodal model to forget something specific, a face, a piece of private data, or an unsafe association, has a nasty side effect: the model often forgets how to talk about anything at all. Ask a scrubbed LLaVA who Harriet Tubman was and it apologizes. Ask what somebody is holding in a photo and it refuses. Researchers at the Australian Institute for Machine Learning have traced this failure to a specific place in the network, the tiny projector that bridges vision and language, and proposed a fix called SineProject that amounts to a one-line change to the projector's weight parameterization.
When forgetting turns into refusing everything
Multimodal LLMs like LLaVA glue a vision encoder to a language backbone through a small two-layer MLP projector. That projector is the only path visual tokens take into the language model, so its geometry defines the alignment between the two modalities. When gradient-based unlearning methods designed for text-only LLMs are applied to MLLMs, they often fail catastrophically. SafeEraser reports over 100% Safe Answer Refusal Rate for gradient-based methods on LLaVA-1.5-7B, meaning the unlearned model refuses even benign queries at essentially all times.
The authors call this alignment drift, and identify three symptoms that appear together during unlearning:
- Jacobian condition numbers of projection layers increase by 3 to 4 orders of magnitude, which makes optimization numerically unstable.