acceptodds
Under review as a conference paper at ICLR 2027

Talking to a Connectome: Language-Model Agents Teach and Interrogate a Whole-Brain Fly Model

Abstract

Whole-brain connectome models make it possible to run behavioural experiments in silico. We let language-model agents experiment on such a model as an experimenter would: presenting odours, pairing them with sugar or shock, and measuring the fly’s approach valence. The model is a leaky integrate-and-fire network of the adult Drosophila brain on the FlyWire connectome (138,639 neurons) with dopamine-gated plasticity at Kenyon-cell output synapses. Manipulating the model shows that an odour recruits its own dopamine through Kenyon-cell-to-dopamineneuron synapses, so that presenting it alone changes its memory, that learning changes this dopamine, and that a memory cannot be undone to within tolerance. In open-loop teaching, one pairing per target meets 0.80 of targets (doing nothing meets 0.71), a zero-shot frontier model writes nearly the same protocols, and teachers optimised on a fast surrogate of the fly exploit the surrogate’s errors (0.99 predicted, 0.75 measured). In closed-loop diagnosis of a hidden training history from twelve measurements, the main difficulty is generalisation between odours that share Kenyon cells. Claude Sonnet 5 scores 0.832 against 0.793 for a thresholding rule because it respects the stated prior that at most two odours were trained; the rule with that prior scores 0.846. Given a one-page summary of the model’s noise and generalisation between odours, and told how answers are scored, Claude reaches 0.894, level with a Bayes player that knows the same (0.912), and gets all seven labels right in 73% of flies instead of 35%. Qwen3-32B cannot use the summary, although thirty steps of multi-turn GRPO raise it from 0.765 to 0.830, Claude’s zero-shot level. Language-model representations of odour names predict

Primary area:
applications to neuroscience & cognitive science
Keywords:
connectome, Drosophila, mushroom body, language-model agents, experimental design, GRPO, reward hacking, spiking neural networks

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_SbmMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper builds a closed-loop environment for language-model agents in a whole-brain LIF model of adult Drosophila (FlyWire connectome), with dopamine-gated plasticity at KC→MBON synapses. Three tasks—teach (open-loop), diagnose (closed-loop), and erase—reveal three substrate properties (odor self-recruits dopamine via KC→DAN synapses; learning alters that dopamine; memories cannot be truly erased), and are used to benchmark agents including threshold rules, an approximate Bayesian player, Claude Sonnet 5 (zero-shot), and Qwen3-32B (zero-shot and multi-round GRPO).

Critical strengths
  1. The novelty of the environment is solid: this is the first closed-loop environment built on a genuine whole-brain spiking connectome model (as opposed to text-based or probabilistic generative environments) in which LLM agents can manipulate the system through experiments, with outcomes exactly scored by the simulator.

  2. The experimental design is complete: it includes three controls—shuffled features, Morgan molecular fingerprints, and character n-grams. After the preregistered test failed (median rank of 45/111 for the key compound), the exploratory follow-up (letting the model name the compound first, then having the fly smell it; median rank of 1) cleanly explains the mechanism—the interface operates through the model's chemical knowledge rather than a representation shared with the fly. The logical loop is fully closed.

Critical weaknesses
  1. The evaluation scale is small, and several key comparisons fail to reach statistical significance. Only 60 test flies, 74 teaching configurations, and 7 odors are used; the authors themselves concede that "a difference of 0.04 is at the edge of the resolution of this sample size."

  2. The evidence for the scale and stability of the RPO training is insufficient: only 30 steps and a single training run; the brain-model reward training for the teach task was terminated early at steps 59/12/16 ("when we moved to the closed-loop tasks," Appendix C); and the first GRPO run for the diagnose task was discarded after 5 steps due to temperature and learning-rate issues (Appendix D). The claim that "30 steps of GRPO elevates Qwen3-32B to Claude's zero-shot level" is one of the highlights of the paper, yet it is supported by neither multi-seed training variance nor convergence analysis, leaving the robustness of the conclusion unclear.

Questions for authors

See weaknesses.

1 comment

Thank you for raising the question of scale. The seven-odour panel is small, but it lets us cover every single teaching target and every ordered pair with one odour to move up and one to move down. Those cases and 18 three-target cases make up the 74 profiles. The clearest teaching result is what happened to the protocols selected by search. The surrogate predicted a success rate of 0.988 for them. When we ran those same protocols on the spiking model, they scored 0.749.

In diagnosis, the 60 flies leave considerable uncertainty around Sonnet’s 0.039 lead over the threshold rule (paired 95% CI −0.002 to +0.079). We can say more about how they made their decisions. Sonnet always respected the instruction that at most two odours had been trained. Giving the rule the same cap reduced false alarms from 102 to 34 among the 111 untrained odour–fly pairs whose responses had shifted. On 60 fresh flies the cap again reduced false alarms, though it also missed more trained odours. The balanced score on those flies was 0.797 with the cap and 0.819 without it.

For GRPO, Appendix C reports the shorter teaching runs. The diagnosis run in Appendix D used 240 training flies. We stopped an initial five-step pilot on the basis of its training behaviour, before testing its trained policy. In the subsequent run, we evaluated each checkpoint on the same 60 held-out flies. At steps 0, 10, 20 and 30, Qwen named too many odours as trained on 42, 32, 20 and 14 flies, respectively. Its balanced score rose from 0.765 to 0.830 (paired gain 0.065, 95% CI 0.013 to 0.121). We will describe the result in the abstract as a gain from 0.765 to 0.830 in one 30-step training run. Sonnet’s zero-shot score of 0.832 gives readers a reference on those same flies.

Reviewer_r61YOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds an experimental environment and a whole-brain LIF model of the fly, adding dopamine-gated plasticity. Agents interact with the model by presenting odours, pairing them with sugar or shock, and measuring a valence readout. The authors defined three tasks: open-loop teaching, closed-loop diagnosis of a hidden training history, and erasure.

The authors first describe properties of the model that constrain these tasks: odours recruit their own dopamine, learning changes that dopamine, and memories can’t be reversed. They then compare heuristics, a surrogate-based optimiser, an approximate Bayes player, Claude Sonnet 5 and Qwen3 (zero-shot and with GRPO).

In diagnosis, Claude’s advantage over a threshold rule comes from respecting the stated prior. Given a summary of noise and generalisation, Claude approaches the Bayes player while Qwen does not.

Critical strengths

The environment is new and well thought out. Having exact ground truth on a realistic simulator is useful for evaluating agents and as a reinforcement learning reward source.

The surrogate overoptimisation result is clear, and the authors trace it to a specific mechanism (the dopamine feedback the surrogate ignores).

The statistics are careful (paired tests, bootstrap intervals), and the limitations section is honest, including which analyses were added later.

A novel use of AI agents. With some limitations addressed, this framework has a lot of potential for advancing scientific discoveries.

Critical weaknesses
  1. The sample seems small (60 flies, 74 profiles). Several key comparisons are borderline, and “level with” claims rest on overlapping intervals rather than equivalence tests.

  2. Several of the conditions that support the main conclusions (the ≤2 rule, the balanced Bayes decision, the brief, the scoring rule) were added after seeing the results.

  3. Only one frontier model is evaluated, with no control over sampling and apparently one run per fly. Qwen runs without thinking mode, which makes the comparison between the two models hard to interpret.

In diagnosis, the agent can only measure, and the gains come mostly from the decision rule. No agent learns the generalisation structure itself, so the framing of agents “interrogating” the brain feels stronger than what is shown.

The RL experiments are short (30 GRPO steps for diagnosis, and teaching runs with the real model as reward were stopped early), which limits what can be concluded from them.

Questions for authors
  1. What happens if Claude’s decision step is applied to random or threshold-rule measurement sequences?

  2. What is the run-to-run variance for Claude on the same flies? Was extended reasoning enabled?

  3. How does Qwen3-32B do with thinking enabled, and how would other frontier models perform?

  4. What does an episode cost in simulation time? Is it feasible to scale up the number of test flies?

1 comment

Thank you for suggesting that we separate choosing a measurement from deciding what it means. We ran the reverse of your proposed test. Using the saved episodes, we gave the threshold rule the measurements Sonnet had chosen. When the rule called at most two odours trained, it gave exactly the same seven answers as Sonnet on 48 of 60 flies. The rule scored 0.8325 and Sonnet scored 0.8322. We used the logged valences and counted unmeasured odours as untrained. On its own twelve measurements, the rule scored 0.793 without the cap and 0.846 with it. We have not run Sonnet on the rule’s or a random player’s transcripts. We will describe diagnosis more precisely. The agents chose which odours to probe, then used those measurements to infer the training history. In the brief condition, we supplied the pattern of generalisation between odours.

Each Sonnet condition used one episode per fly. The CLI logs do not tell us whether extended thinking was active, and we have no repeat estimate on the same fly. We do have the Qwen comparison you asked for. Turning thinking on changed the balanced score from 0.765 to 0.701, while the fraction of flies with all seven labels correct rose from 7% to 17%. On 120 of 544 turns Qwen reached the 2,048-token limit while it was still reasoning. We also tested Opus 5.5. It scored 0.809 zero-shot and 0.881 with the brief and scoring rule on the same 60 flies.

Scaling up the simulator is feasible. A trial took 0.87 seconds on one Apple M3 core, or roughly 20 seconds of simulator time for a full diagnosis episode. For this response, four local workers completed 60 new threshold-rule episodes in about ten minutes after startup. Including model calls, a zero-shot Sonnet episode took a median of 229 seconds.

We agree about the wording of the comparisons. Sonnet with the brief and scoring rule scored 0.894 and the Bayes player scored 0.912 (paired difference −0.018, 95% CI −0.074 to +0.037). We will give those scores in place of “level with” and mark the capped rule, the balanced Bayes decision and the prompt changes as post hoc in the abstract. The GRPO result is a gain of 0.065 on held-out flies after one 30-step diagnosis run (paired 95% CI 0.013 to 0.121). The shorter runs with rewards from the brain model were for teaching.

Reviewer_8cMuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds a simulation of fruit fly brain and adds a dopamine based learning rule, so the simulated fly can learn to like or avoid odours. Language model agents and baselines then experiment on it. They teach the fly target preferences, or work out a hidden training history from 12 measurements.

Contributions are:

  1. Odours trigger their own dopamine, and memories in the model cannot be erased.
  2. For teaching, one pairing per target works best. Claude finds this zero-shot, while optimising on a fast surrogate model fails.
  3. For diagnosis, the main difficulty is that training one odour also shifts similar odours. Claude beats a simple rule mainly by respecting the at most two odours prior.

All validation is in simulation, with no comparison to real fly data.

Critical strengths
  1. A brain simulation as a test environment with exact answers. The simulator knows the true training history, so every agent’s conclusion can be scored exactly. This makes it possible to compare AI agents fairly, and to use the simulator as a reward signal for reinforcement learning. This matters to my assessment because it is the paper’s main new contribution, and other researchers could build on it.
  2. The authors do not just report that Claude scores higher than a simple rule. They show where its advantage comes from, and that the simple rule with the same prior does just as well. The control experiments and the error breakdowns support the claims well. This matters to my assessment because it makes the conclusions more trustworthy.
Critical weaknesses
  1. All results come from a simulation that uses four hand-tuned weight changes, and the authors note that these are not a biological validation. Apart from matching Kenyon cell sparseness, the model’s learning behaviour is not tested against real flies. For example, the central effect, that training one odour also shifts similar odours, has been studied in behavioural experiments on real flies, but the paper does not compare its results with such data. Some findings are also true by construction like memories cannot be undone because the learning rule can only weaken synapses. This matters to my assessment because without a check against real fly data, it is unclear whether the findings describe fly brains or only this particular model.

  2. The statistical evidence is weak. With only 60 test flies, Claude’s advantage over the simple rule is not significant. Only two language models were tested, each run once. This matters to my assessment because the paper’s main claims about how agents compare rest on small differences that this setup cannot reliably support.

Questions for authors
  1. Can the model’s generalisation between similar odours be compared with behavioural data from real flies? Even one comparison would help show whether the findings describe fly brains or only this model.
  2. Could the post-hoc methods be tested on a new set of flies that was not used during development? Please also mark these results as post-hoc in the abstract.
  3. How much do the language model results vary between runs? Each model was run once per fly. Repeated runs, or results from other models, would show whether the differences between agents are reliable.
1 comment

Thank you for suggesting a comparison with real flies. Campbell et al. paired pentyl acetate with shock. Later, when the flies could choose between butyl acetate and ethyl lactate, they chose butyl acetate less often than controls did. The effect also held when the two acetates switched roles. Imaging showed that they activated many of the same Kenyon cells. In our model, a single shock pairing with ethyl acetate lowered valence by 0.35 for that odour and by 0.34 for untrained methyl acetate. Acetic acid shifted by only 0.05. We see a similar transfer in the model. The odour pairs and readouts differ, so we can compare the direction of the effect. Campbell et al., Fig. 4H

You are also right about the learning rule. A weakened synapse cannot regain its old weight under Eq. 1. Whether the fly’s measured response can return to baseline is a different question, and it came back within tolerance in 5 of 20 cases when we chose the best corrective protocol for each fly. We will use that result in §4.3 and explain in the main text how the fixed gains affect the number of responding Kenyon cells and allow us to measure both approach and avoidance.

We tested your suggestion on 60 new flies. We kept the threshold fixed at ±0.10 and compared the two rules on exactly the same measurements. Allowing at most two “trained” labels cut false alarms from 94 to 45 across 337 untrained odour–fly pairs. It also missed more trained odours. The balanced score was 0.819 without the cap and 0.797 with it (paired difference −0.022, 95% CI −0.067 to +0.020). The cap helps explain why Sonnet made fewer false calls on the original flies. Whether it improves the score depends on how many true calls it loses. The new test covers the capped rule. We will identify the post-hoc analyses in the abstract, including the brief and the Bayes decision rule.

We also have results for a second model on the original 60 flies. Opus 5.5 got all seven labels right on 35% of flies with the original prompt and 62% when given the circuit brief and scoring rule. Sonnet went from 35% to 73% under the same conditions. For Opus, the gain in balanced score was 0.073 (paired 95% CI 0.015 to 0.133). Each model ran once per fly per condition, so these intervals describe variation across flies. Repeated runs on the same fly would answer the sampling question.

Reviewer_ScmnOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents the results of an olfactory learning experiment where the experimenters are LLMs instead of humans. The fly model is also simulated, using a leaky integrate-and-fire network of the Drosophila brain with the FlyWire connectome. The main results of their experiments are that odors recruit their own dopamine neurons, learning changes those dopamine neurons, and memories cannot be undone.

Critical strengths

One possible strength of the work might be motivating more experiments to be conducted by LLMs to test hypotheses and make predictions that can be verified in the lab, assuming an LLM might generate better hypotheses and/or experimental designs to test them.

Critical weaknesses

The main objectives of the paper are not clearly stated: whether it aims to compare different LLMs to see which is a better experimenter or use LLMs to make new predictions. If it is the former, how the performance of LLMs is measured is not clear in the paper. If it is the latter, it is not clear what the main hypothesis is or which of the predictions made in these experiments are new.

Moreover, the paper is poorly written, and most of the sections are incomprehensible, partly due to very long, ambiguous sentences containing many numbers, even in the abstract.

Questions for authors

I recommend rejecting this paper in its current form; however, clarifying the answers to the following questions will improve a future submission.

  1. What is the motivation/justification behind the learning rule in equation 1, as that is critical to this work?

  2. It is stated that there are four corrections made to the connectome (Appendix A); what is the motivation for these corrections, and how much do those corrections affect the results? What would the results look like if these corrections were not made? The answer to this question might require new experiments.

  3. How much of the results presented here are driven by the learned synapses between KC and MBON alone, rather than the entire connectome; in other words, would the results be different if the rest of the connectome were randomly generated synapses, instead of the synapses based on FlyWire? The answer to this question might require new experiments, too. This clarification becomes even more important when one of the corrections made to the connectome is multiplying all KC-to-MBON synapses by 3, L502-503.

  4. What is the motivation behind the erase experiment (section 4.3)? It is also stated in the paper that because the learning rule only depressed synapses between KCs and MBONs, a memory can be masked by an opposing one but not undone; it is unclear what undone means here. If undone here means a valence close to zero, that seems possible judging from the definition of valence provided in L138; if the new training (mask) has an opposite but relatively close magnitude of approach/avoidance (A and B in the equation) compared to before.

  5. The paper mentions fitting a fast surrogate of the fly in some of the experiments; what is the fast surrogate?

  6. What is the main point of showing results on cutting KC-to-DAN connections in Figure 2a? The description of Figure 2a in the text (L166–176) does not seem to match the figure.

Minor comments (these are non-exhaustive, exemplary comments):

Some wording is confusing, especially in section titles, such as in L164: "ODOURS RECRUIT THEIR OWN DOPAMINE"; I believe this should refer to dopamine neurons rather than just dopamine. Additionally, in L207, "generalisation through shared Kenyon cells makes every extra trial a side effect" is unclear; what does "side effect" mean in this context? Ect.

Some abbreviations are never explained, like PN and ORN.

Some sections are very confusing and full of numbers, which could be easily seen in figures/tables (e.g., L200–210, L273–284).

1 comment

Thank you for asking what we want the agent to do. In one of our test flies, only two odours had been paired with sugar, but five odours changed enough to cross the detection threshold. Could an agent work out which two had been trained? We know the answer because the simulator records the training history. We score how well the agent identifies trained odours and their sign, giving equal weight to how well it recognises untrained odours. In teaching, the agent chooses how to train the fly. Its score is the fraction of odours that reach their targets, including those that should stay unchanged. We will use this example in the introduction. The agents carry out the tasks. We designed the circuit experiments in §4 to understand their results.

  1. Eq. 1 follows the finding of Hige et al. (2015) that pairing an odour with dopamine-neuron activation can weaken the response to that odour at Kenyon-cell outputs. In our model, a synapse weakens when the Kenyon cell (KC) responds and dopamine neurons in the same compartment are active. The learning rate and the minimum weight are choices we made for the model.

  2. Appendix A gives the reason for each change. With the original signs on antennal-lobe connections, stimulating one glomerulus activated all 330 uniglomerular projection neurons. We changed the local-neuron outputs to limit that spread. We also made dopamine act through the learning rule alone. Doubling the input from projection neurons to KCs brought the fraction of KCs responding to an odour from under 3% to about 7%. At the original KC output strength, only three mushroom-body output neurons (MBONs) fired. Tripling that strength let us read out both approach and avoidance. The original checks for feeding, grooming and escape still passed. These comparisons concern the circuit responses. We have not scored the agents with all four changes removed.

3 and 6. Fig. 2a tests what the wiring contributes when the learning rule and the threefold KC output gain stay fixed. The rightmost bars show what happened when we presented ethyl acetate and reinforcement in separate trials. With the KC inputs to dopamine neurons intact, valence fell from +0.19 to −0.04. When we cut those inputs, it stayed at +0.187. The fly still learned from shock paired with the odour. This singles out a pathway outside the synapses that store the memory. We have not tested how much of the effect would survive randomising the rest of the wiring. The panel shows valence, and each dot is one seed. The neuron counts mentioned in §4.1 come from Appendix A. We will make that distinction clear in the text.

  1. You are right that two opposing changes can cancel. In the erase task we ask whether an odour returns to its own value before training, which can be quite far from zero. The learning rule only lets weights decrease, but the measured response can recover. Even choosing the best of 13 corrective protocols separately for each fly, we recovered the trained odour’s response to within 0.15 of baseline in 5 of 20 cases. We will change the heading to “Counter-conditioning seldom restores the original valence.”

  2. The surrogate is a shortcut for scoring a teaching protocol. It stores the firing rates of KCs and dopamine neurons in an untrained fly, applies Eq. 1, and uses a fitted slope for each MBON to predict its firing rate. On 40 held-out random protocols, predicted and measured valence changes correlated at r = 0.88. The shortcut mattered when we searched for the best protocols. Those scored 0.988 on the surrogate and 0.749 on the spiking model, where training can change the dopamine response on the next trial.

We will simplify the abstract, define ORN (olfactory receptor neuron) and PN (projection neuron) on first use, and change the §4.1 heading to “Odours activate dopamine neurons.” By “side effect” at line 207 we mean that an extra pairing can shift the response to an odour the agent was asked to leave unchanged. We will say that directly and leave the detailed scores to the figures and tables.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.