Problem-Conditioned Interface Fusion: A Modular Language Model Framework for Multi-Objective Branch-Specific Post-Training

Multi-objective post-training may cause interference in shared parameters, yet assigning objectives to private branches does not ensure that their outputs can be combined directly. We propose problem-conditioned interface fusion: after joint pretraining and partitioned specialization, the backbone and experts are frozen, while a fusion layer learns candidate selection, low-rank input adaptation, branch-message reading, and bounded residual synthesis. The framework maintains a single main hidden-state stream and uses predictions of branch contributions and complementarity as auxiliary learning objectives. We derive local correction bounds, zero-message response, permutation invariance for the same candidate set, and pure-path preservation under explicit cache conditions. We also discuss cross-layer perturbation propagation under additional smoothness assumptions. A potential conflict between reasoning and literary expression serves as a motivating example, not an established universal causal relationship. No model training or capability validation has been conducted; the proposed benefits, learnability, and extension behavior remain hypotheses to be tested.Theoretical preprint. No model implementation, training, or empirical evaluation is reported. The record contains English and Chinese editions of the same work. ChatGPT assisted the writing and preparation; zhanghuatao is the sole author and an independent researcher.中文摘要:多目标后训练可能在共享参数上产生干扰,而将目标分配给私有分支又不能保证其输出适合直接组合。本文提出问题条件化接口融合框架:在共同预训练和分区专业训练之后,冻结主干与专家,由融合层学习候选选择、低秩输入适配、分支消息读取和有界残差合成。框架仅维护一条主隐藏状态流,并将分支贡献与互补预测作为辅助学习目标。本文给出局部修正界、零消息响应、同集合排列不变性和带缓存条件的纯路径保持,并讨论在额外光滑性假设下的跨层扰动传播。逻辑与文学表达的潜在冲突用作动机案例,不被视为已确立的普遍因果规律。本文没有模型训练或能力验证;所提出的收益、可学性与扩展效果均属于待检验假说。