Myco-Opt uses the structural properties of fungal networks โ evolved solutions to resource-flow optimization under uncertainty โ as biological priors that guide transformer hyperparameter search, reducing the evaluation budget.
Training a large LLM costs millions. Finding the right hyperparameters (depth, experts, learning rate, routing) is still largely trial-and-error โ and every evaluation is extremely expensive.
A single frontier-model run can cost millions of dollars. Hyperparameter search multiplies that cost.
Our hypothesis: biological priors reach โฅ Bayesian Optimization quality with โค50% of the evaluation budget.
GlobalFungi (Nature Sci Data 2020) + MycoDB: open fungal interaction data to derive the priors.
Mycorrhizal networks connect plants through soil and transport resources. They are networks optimized by millions of years of evolution under uncertainty โ the same problem we face when designing AI architectures.
We do not claim "fungal intelligence". We use verifiable quantitative topology: modularity, degree distribution, small-world connectivity, diameter and clustering โ measured on real open data.
The mappings constrain the search space toward biologically plausible regions. The evolutionary loop with an LLM optimizer (OPRO-style) corrects the rest.
Each structural property of the network maps naturally onto the transformer design space.
| Network metric ๐ | What it captures in the forest | LLM hyperparameter ๐ง |
|---|---|---|
| Modularity Q | Fungi form specialized guilds | num_experts (MoE) |
| Degree distribution ฮณ (scale-free) | Few hubs concentrate flow | top_k routing |
| Small-world coefficient ฯ | Local signal + global shortcuts | num_heads / attention |
| Diameter / mean path L | How far carbon travels | num_layers |
| Clustering C | Local redundancy | dropout |
| Resilience to removal | The network survives failures | regularization / sparsity |
| Carbon-flow dynamics | Prioritizes by need | curriculum / LR schedule |
We validated the empirical foundation with dashAI (U. de Chile / CENIA): trained models on mycorrhizal network data to identify critical nodes (hubs), on both clean and noisy data.
The finding: under noisy data and few evaluations, agnostic optimization does not beat the baseline โ exactly the regime where a biological prior matters (hypothesis H3 of the paper).
Myco-Opt: Mycorrhizal-Network-Inspired Priors for LLM-Driven Hyperparameter Optimization
ยท arXiv / NeurIPS (draft v0.2) ยท Open code and data
"Hyperparameter optimization (HPO) and neural architecture search (NAS) remain sample-inefficient despite advances in Bayesian optimization (BO) and LLM-based optimizers such as OPRO. We introduce Myco-Opt, a framework that uses graph-theoretic properties of mycorrhizal networks as biological priors to constrain the search space of LLM-driven optimizers... We hypothesize that biological priors reduce the evaluation budget required to reach a target loss by up to 50%, while yielding structurally distinct, more interpretable configurations."
We are looking for collaborators to scale the benchmark: integrate real GlobalFungi data, build the OPRO loop with fungal priors, and compare against baselines. Everything is open source.