Multimodal reasoning models can forget one face without forgetting the person: VKCR confines unlearning to the image channel
Synopsis
Multimodal large reasoning models can pull sensitive information back through cross-modal associations inside intermediate reasoning, so suppressing the final answer is not enough. VKCR reframes the forget target as one designated image-to-knowledge association: it localizes the edit in a compact association-conditioned low-rank parameter subspace, then counterfactually rebinds the affected reasoning branch, weakening target-bearing or unsupported continuations while favoring evidence-grounded non-sensitive alternatives. On CLEAR-R and MLLMU-R it reduces both image-conditioned target leakage and hallucinated reassociation while preserving non-sensitive visual reasoning and permitted text-only access to the same knowledge.
Interpretation
The forget object in multimodal large reasoning models shifts from knowledge content to the access association between one image and one piece of knowledge: the same knowledge should remain retrievable in permitted textual contexts, while the designated image should no longer serve as evidence for it. Earlier multimodal unlearning methods mainly targeted samples, concepts, or knowledge content without separating knowledge from the cross-modal association through which it is accessed; in reasoning models, target information may vanish from the final answer yet remain in intermediate reasoning, or be replaced by incorrect substitute knowledge. Under the 5% forgetting setting, VKCR reaches F-RSL 7.39 and F-RSH 18.36 on CLEAR-R, where R-MUSE gives 49.85 and 36.81; on MLLMU-R VKCR gives F-RSL 35.28 and F-RSH 21.25, against 96.15 and 24.68 for R-MUSE.
The edit lands in a compact parameter subspace responsive to the designated image condition, and then rewrites the affected branch according to the reasoning the model actually generates. VKCR carries the edit in a low-rank factorization while the original weights stay frozen; for homogeneous readouts it builds a fixed input-side basis from matched target-bearing and non-sensitive states, and for heterogeneous readouts it uses matched visual interventions to select editable modules. On CLEAR-R with Qwen2.5-VL-3B-Instruct it updates 81,920 parameters, 0.0022% of the model; on MLLMU-R with LLaVA-1.5-7B it updates 503,808 parameters, 0.0071%.
Suppressing the target is not enough; the model must also be stopped from swapping the forgotten object for an equally unsupported new identity or fact. Target suppression lowers support for the annotated target, while VKCR additionally suppresses the concrete relation value the model actually produced and, at the earliest reasoning prefix where target-bearing or unsupported information appears, lowers the average token log-likelihood of the invalid continuation, raises a valid alternative, and enforces a ranking margin. The ablation without reassociation control rises to F-RSH 59.00 against 21.25 for full VKCR; removing association-conditioned localization gives F-RSH 31.51 and R-RL 41.87, against R-RL 59.37 for the full method.
Perspective
The demonstrated scope is behavioral association unlearning in multimodal large reasoning models: on CLEAR-R and MLLMU-R, two reasoning-aware benchmarks extended from existing ones, with Qwen2.5-VL-3B-Instruct, LLaVA-1.5-7B-HF, and R1-Onevision-7B-RL as backbones, results are reported as means over three random-seed runs under the 5% forgetting setting, with additional 10%, 15%, and 1% ratios in the appendix.
The authors state that association unlearning here is behavioral rather than mechanistic and do not claim to identify or erase a unique internal neural pathway; they also acknowledge a moderate remaining drop in retain utility, attributed to parameter and representation coupling in current multimodal reasoning models, where the same layers and intermediate features jointly support multiple identities, relations, and reasoning trajectories.