PhylaFlow: Dual Adaptation Methods

• PhylaFlow is a framework that uses local adaptation and flow-based transport to meet global objectives in both hydraulic networks and phylogenetic tree spaces.
• In the hydraulic model, local chemical signals adjust conductance along graph edges to approximate gradient descent on a pressure-matching error with a clear satisfiability phase transition.
• In the phylogenetic model, a hybrid approach couples continuous BHV-geodesic velocity estimation with discrete topology transitions to guide tree proposals toward high-likelihood regions.

Searching arXiv for PhylaFlow and related papers to ground the article in current literature. PhylaFlow is a name used in two distinct arXiv contexts. In one usage, it denotes a decentralized, slime-mold–inspired algorithm for tuning a hydraulic network so that selected output-node pressures match prescribed target values through purely local, chemically mediated conductance adaptation (Anisetti et al., 2023). In a second, later usage, it denotes a hybrid flow-matching model in Billera–Holmes–Vogtmann (BHV) tree space for phylogenetic inference, coupling continuous branch-length transport with learned discrete topology transitions to move random start trees toward posterior-supported regions (Ektefaie et al., 21 May 2026). The shared name reflects a common emphasis on transport and adaptation, but the two constructions address different mathematical objects, objective functions, and application domains.

1. Dual usage and conceptual scope

A common source of ambiguity is that “PhylaFlow” does not identify a single method family in the literature. The 2023 work concerns biological flow networks and local adaptation in hydraulic graphs, with emphasis on conductance dynamics, Kirchhoff constraints, and a SAT–UNSAT phase transition in network tunability (Anisetti et al., 2023). The 2026 work concerns Bayesian phylogenetic inference in BHV tree space, with emphasis on geodesic supervision, flow matching, and geometry-aware proposal mechanisms for posterior exploration (Ektefaie et al., 21 May 2026).

The two usages nevertheless share a structural theme. In both cases, a local mechanism is trained or designed so that distributed updates realize a global objective. In the hydraulic setting, locally advected chemical signals implement an approximate descent step on a pressure-matching cost. In the phylogenetic setting, a learned vector field together with boundary-event and topology-transition modules approximates transport toward posterior-relevant regions. This suggests that the name has come to denote transport-based adaptation under hybrid constraints, although the underlying state spaces—weighted graphs versus BHV orthant complexes—are unrelated.

2. Hydraulic-network PhylaFlow: physical model and local adaptation

In the hydraulic formulation, the network topology is an undirected graph $G=(V,E)$ of $N$ nodes and $M$ edges (“tubes”), with time-dependent conductance $C_{ij}(t)>0$ on each edge $\{i,j\}\in E$ (Anisetti et al., 2023). The edge flow obeys

$Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$

and interior nodes satisfy Kirchhoff’s law,

$\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$

Boundary nodes carry prescribed inflow/outflow or fixed pressures $u(q_1),\ldots,u(q_j)$. In matrix form, the weighted graph Laplacian is defined by

$$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$

and the node-pressure vector $P$ is obtained from

$N$0

Adaptation is driven by output-node pressure errors. For each output node $N$1,

$N$2

If $N$3, the node releases a positive signal $N$4; if $N$5, it releases a negative signal $N$6. At release times $N$7, separated by $N$8 time units,

$N$9

These chemicals are advected through the network and split among downstream neighbors in proportion to the edge currents. Edge conductances are then updated locally according to

$M$0

or, in continuous time,

$M$1

The central claim of the construction is that each tube uses only locally available information while the collective dynamics minimize a global objective. In this model, the local transport of chemical error signals through the current flow field furnishes the coordination mechanism that substitutes for centralized optimization.

3. Global objective, approximate gradient descent, and phase behavior

The hydraulic PhylaFlow formulation defines a mean-squared-error objective over output pressures,

$M$2

and shows, via implicit differentiation of the Kirchhoff system, that the gradient with respect to an edge conductance takes the form

$M$3

The pressure sensitivity is written as

$M$4

and the paper states that the advected chemical from node $M$5 carries precisely the weight $M$6 along each path toward edge $M$7, so that the local rule approximates a descent step on $M$8 (Anisetti et al., 2023). In operational terms, each edge thickens or thins itself in proportion to the sum of downstream errors carried by the flow.

The same work characterizes a SAT–UNSAT phase transition for successful network tuning. It defines the “clause density” analogue

$M$9

and measures

$C_{ij}(t)>0$0

Empirically,

$C_{ij}(t)>0$1

Below $C_{ij}(t)>0$2, the network is almost always tunable; above $C_{ij}(t)>0$3, it almost never converges. For $C_{ij}(t)>0$4, the extracted scaling laws are

$C_{ij}(t)>0$5

Because $C_{ij}(t)>0$6 as $C_{ij}(t)>0$7, the transition sharpens in the thermodynamic limit, and plotting $C_{ij}(t)>0$8 against $C_{ij}(t)>0$9 yields a master-curve collapse. The practical guidance reported for parameter selection is to keep $\{i,j\}\in E$0 below the critical value, choose a moderate conductance-response rate $\{i,j\}\in E$1, use chemical-release gain $\{i,j\}\in E$2, set the release delay $\{i,j\}\in E$3 above the network’s advective propagation time—$\{i,j\}\in E$4–$\{i,j\}\in E$5 time units was sufficient for $\{i,j\}\in E$6–$\{i,j\}\in E$7—and stop updating an output once $\{i,j\}\in E$8 drops below a small tolerance such as $\{i,j\}\in E$9.

4. BHV-space PhylaFlow: geometric setting and model structure

The 2026 PhylaFlow is defined on Billera–Holmes–Vogtmann tree space, $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$0, where each fully resolved unrooted topology $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$1 on a fixed leaf set of size $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$2 corresponds to a Euclidean orthant

$Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$3

with coordinates given by the nonnegative lengths $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$4 of the $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$5 internal edges (Ektefaie et al., 21 May 2026). Orthants are glued along boundary faces where one or more $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$6 vanish, representing unresolved trees shared by adjacent topologies. The space is piecewise Euclidean and CAT(0), so any two trees are joined by a unique geodesic.

For trees $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$7 and $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$8, with split sets $Q_{ij}(t)=C_{ij}(t)\,\bigl[P_i(t)-P_j(t)\bigr],$9 and $\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$0, the BHV distance is

$\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$1

Geodesics are computed using the $\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$2 Owen–Provan algorithm, which contracts incompatible edges to zero at lower-dimensional faces and then expands the new splits. PhylaFlow uses these geodesics for supervision.

The model couples two update modes. First, within each orthant, a continuous vector field $\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$3 is trained to match BHV-geodesic velocity. A tree state $\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$4 is represented as graph- and token-embeddings, processed by a graph-transformer trunk, and converted to per-edge velocity predictions

$\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$5

The corresponding flow-matching loss is

$\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$6

Second, the model predicts discrete boundary events. A lightweight first-hit head assigns each internal edge a logit $\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$7; after a continuous update, edges with positive logits form the predicted collapse set $\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$8. Supervision uses the geodesically determined next-collapse set $\sum_{j:(k,j)\in E}Q_{kj}(t)=0.$9 through

$u(q_1),\ldots,u(q_j)$0

When selected collapses create a polytomy, an autoregressive topology head scores starter pairs, predicts target merge-set size, and scores component membership to decode the next split, inserted with a small “birth” length such as $u(q_1),\ldots,u(q_j)$1. The total objective combines $u(q_1),\ldots,u(q_j)$2, $u(q_1),\ldots,u(q_j)$3, $u(q_1),\ldots,u(q_j)$4, and additional terms with tuned weights $u(q_1),\ldots,u(q_j)$5.

5. Training protocol, inference dynamics, and split-guided refinement

Training begins by running a short MrBayes sketch to obtain posterior topologies and their frequencies for each dataset (Ektefaie et al., 21 May 2026). Random start trees $u(q_1),\ldots,u(q_j)$6 are then generated by inserting taxa one-by-one onto random edges. For each $u(q_1),\ldots,u(q_j)$7 pair, the BHV geodesic $u(q_1),\ldots,u(q_j)$8 is computed and decomposed into interior-orthant states $u(q_1),\ldots,u(q_j)$9 with target velocity $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$0 and next-collapse set $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$1, together with boundary states carrying target autoregressive actions $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$2. Supervised optimization iterates over these states and minimizes the combined velocity, first-hit, and autoregressive losses.

Inference proceeds in phases. From an initial tree $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$3, the model tokenizes the current tree, time $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$4, phase index $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$5, and optionally sequence conditioning $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$6. It predicts $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$7 and $$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$8, then computes the exact-boundary step

$$L_{ii}=\sum_j C_{ij}, \qquad L_{ij}=-C_{ij}\;(i\neq j),$$9

so that at least one predicted-collapse edge reaches zero. Branch lengths are updated by $P$0, selected edges are clamped to zero, and remaining edges are floored to a small $P$1. If a polytomy forms, the autoregressive topology head is invoked once, or up to $P$2 times, to insert a new split. The procedure terminates early if no negative-velocity edge is selected or if no valid polytomy action exists. An optional branch-length relaxation can then be applied, trained on short fixed-topology MrBayes warmups.

The same framework supplies a proposal mechanism for Bayesian refinement. A guide-bank of $P$3 PhylaFlow-generated trees is used to estimate split frequencies $P$4. MrBayes then mixes standard nearest-neighbor interchange proposals with split-guided proposals according to

$P$5

followed by the usual Metropolis–Hastings correction. The intended effect is to bias local proposals toward splits supported by the learned BHV transport.

6. Empirical evaluation, limitations, and interpretation

The phylogenetic PhylaFlow paper evaluates the method on eight standard posterior benchmarks, DS1–DS8, spanning viral, bacterial, and eukaryotic clades, with independent long-run MrBayes reference posteriors of more than $P$6 M generations used to define TreeKL and SplitKL metrics (Ektefaie et al., 21 May 2026). The reported metrics are Initial TreeKL, Final TreeKL after a finite-budget $P$7 K generations of MrBayes refinement, SplitKL, and the Optimized-Likelihood Gap.

The quantitative ranges reported in the paper are summarized below.

Quantity	Methods	Reported range
Initial TreeKL	Random	15.4–17.6
Initial TreeKL	Maximum parsimony	similar to random
Initial TreeKL	IQ-TREE ML	11.8–17.6
Initial TreeKL	PhylaFlow	0.5–11.3
Final TreeKL after 100 K MrBayes	Random	0.26–5.11
Final TreeKL after 100 K MrBayes	Best classical	0.18–4.97
Final TreeKL after 100 K MrBayes	PhylaFlow direct	0.05–5.04
Final TreeKL after 100 K MrBayes	PhylaFlow-MCMC	0.12–4.48

The paper states that PhylaFlow yields an $P$8–$P$9 reduction in Initial TreeKL relative to classical initializers, that direct PhylaFlow is best on five of eight datasets after finite-budget refinement, and that PhylaFlow-MCMC is best on seven of eight, with the strongest hard-case results on DS3, DS5, DS6, and DS7. Against specific controls, the best PhylaFlow variant outperforms short-warmup on seven of eight datasets and PhyloGFN on five of eight under the same refinement budget. Likelihood gaps remain small or negative for PhylaFlow starts, indicating that chains begin in a high-likelihood basin.

The reported limitations are equally specific. Sequence conditioning with per-taxon Phyla sequence embeddings $N$00 achieves low SplitKL under native conditioning, $N$01–$N$02, but higher TreeKL than dataset-specific training, $N$03–$N$04 versus $N$05–$N$06. Off-diagonal conditioning swaps degrade SplitKL by about $N$07 on average, and exact posterior topology recovery remains preliminary. Additional limitations are that autoregressive topology resolution can be slow when many small polytomies arise, per-dataset training still outperforms joint conditioning, and both the branch-length relaxer and geodesic oracle impose substantial overhead. A plausible implication is that the present gains derive primarily from geometry-aware amortization within specific posterior families rather than from broad cross-dataset transfer.

Taken together, the two PhylaFlow lines of work illustrate two sharply different instantiations of local-to-global adaptation. In hydraulic networks, local chemical transport approximates gradient descent on a pressure-matching objective and exhibits a satisfiability phase transition. In phylogenetic inference, supervised transport along BHV geodesics yields an amortized initializer and a split-informed proposal mechanism for finite-budget posterior exploration. The shared terminology should therefore be read as homonymous rather than as denoting a unified formalism.