Leaving food and coming back: the whole circuit, with what is proven and what is not

A hungry fly touches a drop of sugar, walks away in the dark for a minute, and comes back to it. To do that it must know which way it is facing, how fast it is moving, add those up into a running estimate of where the sugar is, turn that estimate into a direction to walk, and steer its legs. The diagram is the wiring that does each of those jobs: the compass on the left, velocity and travel direction next, then the memory that holds the sugar's position, then the goal-to-steering comparison, then the descending neurons. The first, second, fifth and sixth steps have neurons with recordings and causal tests. The memory in the middle does not, and that is where the open questions and most of the findings sit.

Proven: causal + physiology, several labs Physiology recorded, causal role partial Structural only: this work, connectome Missing: behaviour demands it, no neuron Controversial: wiring contradicts the model Finding number (click); C = confirmation of published work
1 · Sense2 · Heading3 · Velocity → travel direction 4 · Integrate, store, choose a goal5 · Steer6 · Move Side branches Visual cues → ER ringsER4d, ER4m, ER2 … Wind → ER ringsER1, ER3w Internal state → ER5sleep drive Motor copy: PS196_bunnamed source Legs → ascending → FB3A12 % of PFNd input Sugar → search stateCorfas 2019 Self-deposited scentChen 2024, Titova 2023 EPG compass ringheading bump+ Δ7 inhibition PEN shiftersL / R rotation GLNOrotational velocity Cue-learning gatesEL octopamine: which synapsesExR2 dopamine: when read write ±1 learn 1 C 4 hΔBworld travel direction PFNd / PFNvvelocity × heading LNO2, SpsPvelocity scaling PFR_a / PFR_btravel + heading sum (Lyu 2022) C heading 20,165 FB4M / FB1H dopaminedriven by walking → hΔ stores hΔH/A/I/G weightshΔB → hΔ synapsessum the path? OA-VPM3reward octopamine→ hΔB, hΔJ, PFR hΔM (after FC2)178° inverterthe only one found Integrateno physiology;model works Store zerore-zero at foodno physiology Invertstores arrivesign-preserved FC2 goal (A / B / C subtypes)A: self-motion · B: MB relay, memoryC: arousal layers · FB5A inhibitorno direct hΔB input (<150 synapses) hΔK / PFGs loopworking memory FB6A / FB6Minhibitory gates FR1 self-loop→ MB dopamine via hΔC / hΔF 2 PFL3 / PFL2heading vs goal→ L / R error hΔMgoal + 180°← FC2 hΔA / hΔItravel-derived← hΔB, MB heading (via Δ7) goal 7,690 3,983 7,349 3 LAL interneuronsget most of PFL3 outputcontralateral, inhibitory Descending neuronsDNa02, DNa03, DNa15, DNg04PFL3 → DNa02 direct: 1.5 % Legs turnVNC not in graph LAL013 / DNa11 (Feng)PFL input: ~0 C Mushroom body (value)MBON21/09 → FB4R → hΔBMBON25/34 → FB4G/H → PFL3 FS outputsFB state → SMP, oviIN oviIN egg-laying gate↔ FC2, FS1A (~3,500) FB4R → hΔB feedback to goal C C

Stage by stage

1 · Sense

ProvenVisual and wind cues reach the compass through ring-neuron classes whose synapses onto EPG are learned (Kim 2019; Fisher 2019; Okubo 2020). Sugar contact switches the fly into local search through several sensory routes (Corfas 2019). Flies also mark rewarded spots with their own scent and use it (Chen 2024; Titova 2023). NewThe self-motion sources are named for the first time: PS196_b feeds the rotational velocity neuron GLNO and the dopamine/feedback cells; FB3A carries an ascending (leg-derived) signal into PFNd and supplies 12 % of its input. 4 MissingWhich signal tells the navigation system "food is here". Candidates only: the dopamine tangentials FB4M/FB4L (target hΔB, PFNd, PFR) and FB1H (targets the goal input hΔJ).

2 · Heading

ProvenEPG ring attractor, PEN shifters, Δ7 inhibition, GLNO velocity input, EL octopamine and ExR2 dopamine for cue learning (Seelig 2015; Kim 2017; Green 2017; Turner-Evans 2017/2020; Hulse 2023; Plitt 2025; Fisher 2022). The shifted PEN wiring is confirmed cell by cell in both connectomes. ControversialThe shifters receive three times more compass input at the position they write to than at the position they read from. Every published model omits this; in simulation it brakes rotation. Either the synapses are weak, axo-axonic, or cancelled by inhibition, and each answer is a different compass. 1 NewEL feedback is ring-class specific, so cue learning should depend on octopamine for some cue types and not others. ExR4/ExR6 provide a global glutamatergic feedback that could set bump amplitude. C

3 · Velocity → travel direction

ProvenPFNd/PFNv combine heading with body velocity; hΔB sums them into world-centred travel direction (Lu 2022; Lyu 2022). Airflow vectors in PFNa, including sign inversion by calcium spikes (Currier 2020; Ishida 2026). NewFB3A and PS196_b as inputs (stage 1). The mushroom-body output FB4R modulates hΔB directly (8 % of its input). C MissingNothing here is missing for the instantaneous computation; what is missing is what happens to it over time (stage 4).

4 · Integrate, store, choose a goal

ProvenBehaviourally only: flies integrate distance and re-zero at food (Kim 2017; Behbahani 2021; Titova 2023; D'Atri 2025), and the PFNd/PFNv/hΔB/PFR pathway is required. FC2 holds the goal and drives steering (Mussells Pires 2024). The hΔK/PFGs loop holds an odor working memory (Kathman 2026; Lanz 2025). MissingPhysiology for all three red boxes. No recording has shown a population integrating travel direction, no recording has shown the zero being written at food, and the inversion is anatomical inference. hΔB has no direct route to FC2 (fewer than 150 synapses), so the goal is built through hΔ inputs to FC2, none of which has been recorded during a return. NewA systematic search finds no population with recurrence strong enough to hold a vector in activity. Several hΔ types (hΔH, hΔA, hΔI, hΔG; hΔJ with cancellation) have the ingredients of a synaptic integrator: hΔB input, walking-driven dopamine (FB4M, FB1H), and the reward octopamine neuron OA-VPM3. But anchored to the published column conventions, every one of them delivers the stored displacement to FC2 or PFL3 with its sign preserved, i.e. as the outbound direction; only hΔM (FC2 → PFL3 at 178°) and the hΔK/PFGs loop invert. A closed-loop model returns only with that inversion. PFR sums travel direction and heading (Lyu 2022), not a memory. FR1 is a scalar accumulator feeding MB dopamine. Gorko & Kim 2026 independently found a stored FB vector that is not held in presynaptic calcium. 2 3 C

5 · Steer

ProvenPFL3 compares heading with the FC2 goal and outputs a lateralised error; PFL2 scales the response with error size (Mussells Pires 2024; Westeinde 2024). The ±67° PFL3 phase offset is visible in every cell's anatomy. NewPFL3 also receives the goal inverted by 180° through hΔM (about half the direct route) and a travel-direction-derived input through hΔA/hΔI that outweighs any single FC2 subtype. The published steering model has neither. 3 MissingWhat PFL3 does with three goal-like inputs at once, and how arrival (home vector near zero) stops the steering.

6 · Move

ProvenDNa02 activation turns the fly and receives PFL3 input (Rayshubskiy 2020; Westeinde 2024). A separate LAL steering hierarchy (LAL013, DNa03, DNa11) generates turns of all sizes (Feng 2024). ControversialThe direct PFL3 → DNa02 link is 1.5 % of DNa02's input. Most of the steering signal goes through contralateral, largely inhibitory LAL interneurons that fan out to five descending neurons, and Feng's hierarchy receives no PFL input at all. Hulse 2021 flagged the LAL interneurons; the steering papers did not follow it up. C MissingThe ventral nerve cord is not in this connectome, so the leg-level command is out of reach here.

Side branches

NewThree routes tie the mushroom body to the navigation system (value into hΔB, value into PFL3, travel direction back to MB dopamine). The goal population and its FS1A readouts send ~3,500 synapses to oviIN, the egg-laying gate, which projects back. C C MissingAny physiology on the FS output neurons, which broadcast fan-shaped-body state to the rest of the brain.

What the picture says

Every box on the left and right is green or blue: sensing, heading, velocity, steering and descending control each have named neurons with recordings and, mostly, causal tests. The middle of the flow, where the behaviour requires memory, has no physiology. The fly demonstrably integrates, stores, inverts and stops. The orange chain in the middle is the one wiring-consistent account found by searching every fan-shaped-body population: the memory is synaptic, at hΔB's synapses onto a family of hΔ types, written while walking under dopamine, erased at food by octopamine, delivered to the goal layer as the outbound direction, and inverted into a return direction by hΔM. The experiments that would promote it to blue are on finding 9 and finding 6. The two purple outlines mark where the wiring disagrees with the models of stages everyone considers solved.

Sources: findings; repository docs behavioral-task-map.md, uncharacterised-neurons.md, findings.md. Data: MaleCNS v1.0 (2026) and hemibrain v1.2 (2020). Colours reflect the state of evidence as read for this project, not a consensus.