Biomimetic perception systems integrate heterogeneous cues selectively rather than treating all available information as equally reliable. Inspired by this principle, this study proposes QGFace-LLaVA, a multimodal large language model (MLLM)-centered framework for robust face analysis under imperfect metadata. A pretrained MLLM serves as the shared prompt-conditioned visual–language reasoning backbone, while structured side information, including age, gender, confidence cues, and availability indicators, is regulated through task-aware quality estimation, reliability-guided metadata calibration, gated residual correction, and counterfactual metadata reliability regularization (CMRR). Experiments on FER2013, CelebA-40, and UTKFace cover facial expression recognition, facial attribute recognition, and age estimation under clean, noisy, missing, shuffled, naturally erroneous, and counterfactual metadata conditions. The results show that metadata utility depends jointly on task relevance, metadata reliability, and fusion strategy, and that clean-setting gains do not necessarily imply robustness. QGFace-LLaVA reduces harmful dependence on unreliable metadata, while CMRR provides additional stability under corruption and mismatch. Overall, the framework transfers biomimetic selective cue integration into MLLM-based face analysis by treating metadata as reliability-controlled auxiliary evidence rather than a uniformly beneficial input.

Biological vision integrates cues according to their reliability and context and uses coarse contextual information to form initial hypotheses before directing attention to local evidence. Inspired by these principles, we propose ADCCR (Adaptive Description Composition and Coarse-to-Fine Refinement), a language-guided framework for flexible human keypoint localization. A multi-granularity bank describes each keypoint by its name, anatomical definition, inter-joint relation, and visual appearance. A reliability-aware composer predicts sample-specific gates from the person image, target identity, and candidate cues, replacing predefined sampling probabilities. Counterfactual cue-consistency supervision suppresses inconsistent cues, while a sparse cue-budget regularizer reduces redundancy. The selected description guides a multimodal large language model to estimate a coarse coordinate from the whole-person image. A category-aware crop of the resized person image then focuses text-conditioned heatmap refinement on the candidate region. ADCCR thus provides a functional abstraction of reliability-aware cue integration and selective global-to-local inspection. Trained on COCO, ADCCR achieves 80.1 AP on the validation set, outperforming PoseLLM by 2.5 points. Results on MPII and Human-Art further demonstrate strong cross-dataset generalization.