Fly Language Model (FLM) Wires the Full Fruit Fly Connectome Into a Frozen 1.2B LLM, and Its Own Controls Show the Wiring Does Not Help 果蝇语言模型（FLM）将完整果蝇连接组接入冻结的1.2B大语言模型，但其自身控制实验显示这种接入并无帮助

TL;DR

• FLM couples the complete MaleCNS v1.0 fruit fly connectome (166,700 nodes, 25.6M edges) to a frozen LiquidAI LFM2.5-1.2B-Instruct backbone via a minimal 278,528-parameter readout (~0.024% of backbone parameters)
• The fly-reservoir readout reduces NLL by 0.0222 nats/token vs. the frozen backbone alone, but a direct-input control without the graph outperforms it in all tested seeds
• The connectome's state contracts by at most 0.6 per token, proving it adds no long-range memory; context remains entirely backbone-derived
• Structural controls confirm the graph participates: zeroing edge weights exactly reproduces backbone performance, while relabeling nodes without retraining breaks the readout
• The project is MIT-licensed and locally deployable, but study artifacts are private, limiting independent reproducibility

Why It Matters

This work represents a novel attempt to integrate biological neural architecture into language model inference, offering a concrete benchmark for how much anatomical connectome structure can influence synthetic language processing. For researchers exploring neuromorphic computing or biologically-inspired AI, FLM provides an open, deployable reference system that rigorously quantifies both the contributions and limitations of connectome-based reservoirs. The finding that a connectome does not outperform a simple direct-input control has direct implications for the broader field of brain-computer interface design and hybrid bio-digital systems.

Technical Details

• Architecture (GPF — Generative Pre-trained Fly): A frozen LFM2.5-1.2B-Instruct backbone is coupled to the full MaleCNS v1.0 connectome via a reservoir computer interface. Token embeddings (2,048-dim) are compressed through a fixed Gaussian projection to 128 channels, each injected into one of 166,700 reservoir nodes with a random sign.
• Reservoir dynamics: Each node updates via x = tanh(W(0.6x + 0.4Bc)), where W encodes incoming-normalized anatomical contact counts from the connectome. States are pooled into 128 bins, passed through two trained bias-free matrices (U: 128×128, V: 2,048×128), and projected as a bounded residual (RMS capped at 0.25) added to the backbone logits.
• Training: Only the 278,528 readout parameters are trained; the graph, backbone, and all projection matrices are frozen. Evaluation used 32 SmolTalk everyday-conversation dialogues (1,236 target tokens).
• Benchmark results (NLL in nats/token): Frozen backbone: 1.381995 | Fly readout: 1.359816 ± 0.000110 | Direct-input control: 1.359328 ± 0.000108 | Relabeled no-refit: 1.381265 ± 0.000802 | No edges: 1.381995. Paired bootstrap interval (+0.000005 to +0.00104) confirms no statistically significant fly-specific gain.
• Memory analysis: The recurrence contraction bound of 0.6 per token decays to ~0.0000366 after 20 tokens, mathematically proving the connectome cannot sustain long-range dependencies beyond what the backbone already provides.

Industry Insight

• The study sets a methodological precedent for rigorously evaluating biological-reservoir hybrids by including strict controls (no edges, relabeled nodes, direct-input comparator), a standard other bio-AI proposals should adopt before claiming novel gains from anatomical structure.
• The negligible performance contribution of 166,700 biological neurons over a 1.2B-parameter frozen LM suggests that near-term investments in connectome-scale reservoir augmentation should be tempered by expectations; simpler projection strategies may capture most of the benefit at lower complexity.
• The gap between public code and private study artifacts highlights a reproducibility risk for emerging hybrid AI systems—full transparency of training data and fitted weights will be essential before connectome-language model claims can be independently validated or built upon.

TL;DR

• 完整果蝇连接组（166,700节点）接入冻结语言模型骨干（LFM2.5-1.2B-Instruct），仅训练278,528参数读出层（占骨干0.0238%）
• 果蝇读出层将困惑度降低0.0222 nats/token，但无图对照实验在所有种子中均表现更优，表明连接组未提供实质性提升
• 果蝇图状态遗忘率高达0.6/token，无法承载长期记忆，上下文仍完全依赖骨干网络
• 代码以MIT许可证开源可本地运行，但研究数据未公开，结果暂不可独立复现
• 作者明确否认是首个连接组语言模型，承认此前小规模原型工作（ngxson/fly-hf）

为什么值得看

该研究探索了将生物神经连接组嵌入大语言模型的可行性，结果显示连接组本身未带来显著性能提升，但对生物启发AI设计提供了重要警示；同时，其开源代码与未公开数据的对比凸显了AI研究中可复现性的关键价值。

技术解析

架构采用"水库计算"机制，将果蝇连接组作为固定动态系统（25,582,938条有向边）接入预训练语言模型骨干，通过高斯投影将2048维token嵌入压缩至128通道，随机符号分配到各节点，图状态更新公式为x = tanh(W(0.6x + 0.4Bc))，其中W为入边归一化解剖接触计数。参数效率极高：仅训练278,528参数的读出层（两个无偏置矩阵U[128×128]和V[2048×128]），作为有界残差（RMS≤0.25）添加到骨干logits，其余部分全部冻结。实验验证通过控制组确认机制：设置W=0使残差归零复现骨干损失，节点重标记后NLL回归基线，证明读出发音依赖学习对齐而非拓扑优势；无图直接输入控制在三组种子中均略胜一筹。理论分析证明图状态以0.6比率每token收缩遗忘，10 token后差异≤0.006，20 token后≤0.0000366，表明连接组无法提供长期上下文。

行业启示

生物神经结构的直接移植未必提升AI性能，设计应聚焦于功能等价而非解剖复制，避免陷入"结构决定论"陷阱；开源代码与私有数据的分离可能阻碍科学验证，研究团队需建立数据共享规范以平衡透明度与知识产权保护；参数效率优化策略在生物混合架构中效果有限，需开发更有效的微调机制以释放异构系统潜力。