Sparse domain-conditioned neuron activation in multimodal transformer networks: computational analysis and cross-backbone empirical validation

Multimodal transformer networks couple a vision encoder, a learnable projector, and a language decoder to reason jointly across text and image inputs. Dense activation of every feed-forward neuron at every layer makes per-inference computational cost increase with model size even when the task does not require the network's full representational capacity. We present NEUROBALANCE, a domain-conditioned sparse-activation framework combining per-head attention reweighting, sparse feed-forward neuron gating trained with the straight-through estimator, and Elastic-Weight-Consolidation-inspired retention regularisation. A single soft domain embedding computed from input tokens drives all three mechanisms at inference. We characterise the resulting savings theoretically through a closed-form floating-point-operations (FLOPs) analysis as a function of active-neuron density and empirically on LLaVA-NeXT-7B and InstructBLIP across PMC-VQA, PathVQA, and LingoQA. NEUROBALANCE reduces forward-pass FLOPs by 36.4% and 36.6%, A100 wall-clock latency by 32.4% and 36.1%, and peak memory by 18.1% and 25.8%, respectively. Benchmark accuracy is preserved, with mean gains of +0.6 percentage points on LLaVA-NeXT and +0.5 percentage points on InstructBLIP; all per-benchmark differences are statistically significant under a paired sign-flip permutation test after Holm--Bonferroni adjustment. Per-inference energy consumption falls by at least 32.4% under a conservative latency-proportional estimate and by 43.9% when measured power is used for both dense and sparse terms. The estimated saving is 58.5 kg CO2-equivalent per million inferences at the stated global-average grid intensity. Cross-backbone consistency indicates that the computational savings arise from architectural conditioning rather than a particular backbone.