在工业级大语言模型（LLM）应用中，动态适配任务与保留既有能力的 "自进化" 需求日益迫切。真实场景中，不同领域语言模式差异显著，LLM 需在学习新场景 合规规则的同时，不丢失旧场景的判断能力。这正是大模型自进化核心诉求，即 "自主优化跨任务知识整合，适应动态环境而无需大量外部干预"。 为解决此问题，北邮百家 AI 团队与腾讯 AI Lab 团队提出参数高效的对抗性混合专家架构 MoE-CL，专门用于 LLM 的自进化持续指令微调。其核心设计在于 "解 耦 LoRA 专家" 与 "GAN 对抗降噪" 的结合：为每个任务配置专属 LoRA 专家以保留任务特定知识，避免参数更新相互干扰；同时设置共享 LoRA 专家，通过生成 对抗网络（GAN）中的任务感知鉴别器抑制无关噪声，确保跨任务知识高效且精准传递，最终实现 "知识保留" 与 "跨任务泛化" 的平衡，这也是 LLM 自进化的核 心逻辑。 从实验效果来看，MoE-CL 的自进化能力已在实际场景与基准测试中得到验证。在腾讯真实业务场景 A/B 测试中，它将人工介入成本降低 15.3%；在公开 MTL5 跨域基准与工业级 Tencent3 基准测试中，其平均准确率 ...

Core Insights - The article introduces FlowDrive, a novel end-to-end driving framework that integrates energy-based flow field representation, adaptive anchor trajectory optimization, and motion-decoupled trajectory generation to enhance safety and interpretability in autonomous driving [4][45]. Group 1: Introduction and Background - End-to-end autonomous driving has gained attention for its potential to simplify traditional modular pipelines and leverage large-scale data for joint learning of perception, prediction, and planning tasks [4]. - A mainstream research direction involves generating Bird's Eye View (BEV) representations from multi-view camera inputs, which provide structured spatial views beneficial for downstream planning tasks [4][6]. Group 2: FlowDrive Framework - FlowDrive introduces energy-based flow fields in the BEV space to explicitly model geometric constraints and rule-based semantics, enhancing the effectiveness of BEV representations [7][15]. - The framework includes a flow-aware anchor trajectory optimization module that aligns initial trajectories with safe and goal-oriented areas, improving spatial effectiveness and intention consistency [15][22]. - A task-decoupled diffusion planner separates high-level intention prediction from low-level trajectory denoising, allowing for targeted supervision and flow field conditional decoding [9][27]. Group 3: Experimental Results - Experiments on the NAVSIM v2 benchmark dataset demonstrate that FlowDrive achieves state-of-the-art performance, with an Extended Predictive Driver Model Score (EPDMS) of 86.3, surpassing previous benchmark methods [3][40]. - FlowDrive shows significant advantages in safety-related metrics such as Drivable Area Compliance (DAC) and Time to Collision (TTC), indicating superior adherence to driving constraints and hazard avoidance capabilities [40][41]. - The framework's performance is validated through ablation studies, showing that removing any core component leads to significant declines in overall performance [43][47]. Group 4: Technical Details - The flow field learning module encodes dense, physically interpretable spatial gradients to provide fine-grained guidance for trajectory planning [20][21]. - The perception module utilizes a Transformer-based architecture to effectively fuse multi-modal sensor inputs into a compact and semantically rich BEV representation [18][37]. - The training process involves a composite loss function that supervises trajectory planning, anchor trajectory optimization, flow field modeling, and auxiliary perception tasks [30][31][32][34].

Core Viewpoint - The article introduces the LoRI technology, which demonstrates that significantly reducing the trainable parameters of LoRA can still maintain strong model performance, achieving comparable or superior results to full fine-tuning and other methods while using only 5% of LoRA's parameters [1][9]. Summary by Sections LoRA and Its Limitations - LoRA is widely adopted for parameter-efficient fine-tuning (PEFT) but still incurs significant memory overhead, especially in large models [3][4]. - Recent research indicates substantial redundancy in incremental parameters, prompting the development of LoRI, which reduces the number of trainable parameters while preserving model knowledge [4]. LoRI Methodology - LoRI keeps the low-rank matrix A fixed as a random projection and uses a task-specific sparse mask to train matrix B, allowing for significant parameter reduction [4][13]. - Even with 90% sparsity in B, LoRI maintains good performance, indicating that the adaptation process does not require updating A [4][17]. Multi-Task Learning and Adapter Merging - Multi-task learning is essential for creating versatile models, but training on mixed datasets is costly. LoRI allows for the merging of existing models without retraining, effectively combining LoRA adapters for multi-task capabilities [7]. - Directly merging heterogeneous LoRA can lead to parameter interference, but LoRI mitigates this by mapping task-specific adapters to nearly orthogonal subspaces [7][20]. Continuous Learning and Safety - LoRI provides a lightweight continuous learning method that maintains safety while adapting to new tasks, addressing the challenge of catastrophic forgetting [8][22]. - The two-phase training process for safety adapters shows that LoRI-S outperforms other methods in retaining safety alignment, even under aggressive sparsity [22][23]. Performance Evaluation - Extensive experiments on various benchmarks show that LoRI achieves or exceeds the performance of full fine-tuning and other PEFT methods while using 95% fewer trainable parameters [9][19]. - In single-task performance, LoRI variants demonstrate competitive results across natural language understanding, mathematics, programming, and safety tasks [19][20]. Conclusion - Overall, LoRI presents an effective and lightweight approach to building safe adapters that support downstream task adaptation while maintaining alignment [23].