Pith. sign in

REVIEW 3 major objections 6 minor 300 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

Two informed neighbors flip a navigator from solo to social

2026-07-09 10:09 UTC pith:X6NL45UI

load-bearing objection The paper reports a genuinely new empirical finding — learned hybrid navigation strategies that combine individual landmark-based navigation with social following — but the central 'phase transition' claim lacks the statistical support needed to distinguish a real threshold from ES optimization noise. the 3 major comments →

arxiv 2607.07460 v1 pith:X6NL45UI submitted 2026-07-08 cs.NE

Social-spatial dependencies for learning visual navigation

Patrick Govoni , Pawel Romanczuk This is my paper
classification cs.NE
keywords social navigationcollective behaviorneural network agentsevolutionary strategiesphase transitionhybrid navigation strategyvisual raycastingspatial dependence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper trains individual neural-network agents to navigate toward a hidden target in a 2D arena where other agents—some walking directly to the target, some wandering randomly—are present. The central finding is that the number of informed (direct-walking) agents in the environment acts as a switch: below two informed agents, the trained agent learns to navigate on its own using wall landmarks; at two or more, it learns to depend on social information. The authors call this a phase transition in learned strategy. Beyond this threshold, a second phenomenon emerges: when many informed agents crowd the target, they physically block access, and the trained agent shifts to a hybrid strategy—navigating independently in the periphery and switching to social use only near the target, often adopting looping trajectories to avoid collisions. The paper identifies three strategy categories (individual, social following, hybrid) and shows their distribution is governed by the ratio of informed to uninformed agents and by spatial dynamics at the target. Three model extensions confirm that removing collisions restores the correlation between task and spatial social dependence, and that increasing the perceptibility of nearby agents boosts social following.

Core claim

The paper discovers that learned navigational strategy in a social context is not a smooth gradient but a step function: two direct-walking agents are the threshold above which a trained agent shifts from individual to social navigation, and this threshold holds regardless of how many random-walking agents are present. A second key finding is that high densities of informed agents cause a decorrelation between task performance (needing social information to finish) and spatial behavior (trajectories looking individual), because patch blocking by crowded informed agents forces hybrid strategies and collision-avoidance looping. The paper also shows that this decorrelation is an artifact of the

What carries the argument

A neural-network agent (CNN + perceptron + linear output) trained via evolutionary strategies, receiving 8-ray visual raycasting of walls and other agents, with a fitness function combining travel time and remaining distance to a hidden patch. Social dependence is measured by comparing performance in a non-social test environment versus a beacon-exploiter test environment, and spatial dependence is measured via Jensen-Shannon divergence on binned orientation distributions (directional divergence).

Load-bearing premise

The phase-transition and hybridization results depend on specific simulation choices: absorbing-boundary collisions that cause patch blocking, a cap of five untrained agents, 8-ray raycasting, and a particular fitness function. If collisions were modeled differently or the agent cap were higher, the observed strategy distributions—particularly the decorrelation between task and spatial dependence at high direct-agent densities—might not appear.

What would settle it

Train agents in an identical environment but with elastic (bouncing) rather than absorbing collision boundaries; if the step-function threshold at two direct agents disappears or the hybrid strategy category vanishes, the phase transition is an artifact of the collision model rather than a general property of social navigation learning.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the step-function threshold at two informed agents generalizes, it suggests a minimal quorum for social information use in navigation—a group needs at least two reliable signalers before an individual benefits from switching strategy.
  • The hybrid strategy (individual navigation far, social navigation near the target) implies that animals in predictable environments may segment their behavior spatially rather than committing to one strategy globally, which could be tested in foraging animals by observing strategy switches at distance thresholds from resources.
  • The finding that crowding at a resource drives collision-avoidance looping rather than direct following suggests that observed looping or circuitous paths in biological foragers may signal social crowding rather than poor navigation.
  • The decorrelation between task dependence and spatial dependence at high informed-agent densities means that studies measuring only one of these dimensions would miss hybrid strategies, supporting the paper's argument against individual-only analysis of social organisms.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The step-function threshold at two direct agents may reflect a reliability computation: a single informed agent is indistinguishable from noise, but two agents moving coherently toward the same location provide a signal that is statistically distinguishable from random motion. This connects to signal detection theory—two correlated signalers reduce false-positive risk below a threshold the agent c
  • If the hybrid strategy emerges because direct agents take time to reach the patch (creating a predictable temporal window), then the spatial interface between individual and social navigation should shift with the average transit time of informed agents—longer transit times should push the social-use boundary farther from the patch, and shorter times should compress it.
  • The collision-avoidance looping at high densities resembles bottleneck-queue dynamics in traffic flow and pedestrian evacuation; the paper's results suggest that embodied spatial constraints, not cognitive limitations, may drive much of the apparent strategy diversity in social navigation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This manuscript investigates how social context influences learned navigation strategies in visually-guided agents. The authors train neural network-controlled agents to navigate to a hidden patch in a bounded environment populated by untrained agents of varying skill (direct vs. random walkers). Using an evolutionary strategies (ES) optimization framework, they vary the ratio of direct to random agents and measure the trained agent's reliance on social information versus spatial landmarks. They introduce a two-metric approach: travel time difference (task dependence) and directional divergence (spatial dependence). The central findings are: (1) a step-function threshold where social dependence increases at two direct agents and plateaus thereafter, (2) a decorrelation between task and spatial dependence at high direct-agent densities due to patch blocking and collision-avoidance looping, and (3) the emergence of hybrid strategies combining individual and social navigation. Three model extensions (no collisions, collision sensing, close initialization) are tested to probe the robustness of the patch-blocking mechanism.

Significance. The paper addresses a relevant question at the intersection of collective behavior and computational neuroscience, namely how social information is integrated into individual sensorimotor control. The two-metric framework (task vs. spatial dependence) is a thoughtful methodological contribution, allowing the decoupling of navigational performance from trajectory-level social influence. The systematic exploration across 40 runs per condition with 5000 test initializations, and the inclusion of three model extensions (Fig. 4), demonstrate a serious experimental effort. The finding that high-quality social information can lead to behavioral hybridization rather than pure social following is a non-trivial result that challenges simplistic models of social learning. The work builds naturally on the authors' prior non-social navigation model [2] and visual raycasting approach [1].

major comments (3)
  1. §Task dependence, Fig. 2A: The central quantitative claim is a 'phase transition' or 'step function' at two direct agents. However, the data presented are per-condition medians over 40 runs with no error bars, confidence intervals, or statistical tests. The two testable components of this claim — (a) that the 1→2 direct-agent jump is statistically significant, and (b) that the 2→3, 3→4, 4→5 differences are not — are never formally tested. Given that ES is known to converge to different local optima across seeds, inter-run variance could be large relative to the median differences between adjacent conditions. Without variance reporting or pairwise comparisons, it is impossible to distinguish a genuine phase transition from median-smoothing over noisy optimization runs. This is load-bearing for the entire downstream narrative (behavioral categorization in Fig. 3, decorrelation claims, and,
  2. §Behavioral categorization, Fig. 3A: The classification into individual, follower, and hybrid strategies relies on two hard thresholds: travel time difference < 120 and remaining distance > 400. The text states these are 'drawn at histogram boundaries (Fig. S2),' but the sensitivity of the categorical distributions (Fig. 3B) to the exact threshold values is not assessed. If the step-function claim in Fig. 2A is noisy (see above), the categorical pie charts may not robustly reflect the stated trends (e.g., 'followers drop, absorbed by hybrids' from 3→5 direct agents). A sensitivity analysis on these thresholds, or a clustering-based classification, would substantially strengthen the claim that three distinct strategies are being learned.
  3. §Spatial dependence, Fig. 2B/C: The directional divergence metric uses Jensen-Shannon divergence on spatially binned distributions of agent orientations. The spatial bin size is a free parameter that affects the granularity and magnitude of the divergence values. The manuscript does not report the bin size or test sensitivity to it. If the bin size is too coarse, the decorrelation between task and spatial dependence at high direct-agent numbers could be an artifact of averaging over heterogeneous local behaviors. This is load-bearing for the hybridization claim, which depends on showing that agents navigate individually in the periphery but socially near the patch.
minor comments (6)
  1. §Model design: The fitness function combines travel time and remaining distance, but the relative weighting or trade-off between these terms is not specified. Clarifying this would help readers understand the optimization landscape.
  2. Fig. 2A: The y-axis label for the stair plot (bottom) is unclear. It appears to be travel time difference, but the units and scale should be explicitly stated.
  3. §Extending the Model, Fig. 4C: The initialize-close condition uses a different test perturbation (NS-OG rather than NS-BEt), making direct comparison with the other conditions difficult. The authors acknowledge this, but a brief discussion of how this affects interpretability would help.
  4. Fig. S1: The rationale for capping untrained agents at 5 is explicitly to keep travel times under 500 timesteps. This is a practical constraint, but it limits the generality of the density-dependent claims. A brief note in the main text acknowledging this ceiling effect would be appropriate.
  5. Typos: 'tranformations' (§Model design), 'bototm' (Fig. 3 caption), 'simluations' (§Extending the Model, Fig. 4C description), 'reguarly' (§Extending the Model, Fig. 4C).
  6. §Discussion: The claim that 'learning to follow other agents as a navigational strategy requires before movement' appears to be missing a word (perhaps 'requires perception before movement' or 'requires detecting before movement').

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive reading of our manuscript. The referee correctly identifies that our two-metric framework and the hybridization finding are central contributions, and we appreciate the recognition of the experimental effort. Below we address each major comment in turn. All three comments are well-taken and will be addressed in revision.

read point-by-point responses

  1. Referee: §Task dependence, Fig. 2A: No error bars, confidence intervals, or statistical tests for the 'phase transition' claim. The 1→2 jump significance and 2→3, 3→4, 4→5 non-significance are never formally tested. Inter-run variance from ES could be large relative to median differences.

    Authors: The referee is correct that we report only per-condition medians without variance estimates or formal hypothesis tests. This is an omission we will fix. In the revised manuscript, we will report interquartile ranges or bootstrap confidence intervals on the medians shown in Fig. 2A, and we will perform pairwise comparisons between adjacent conditions (1→2, 2→3, 3→4, 4→5 direct agents) using a non-parametric test appropriate for the 40 independent ES runs per condition (e.g., Mann-Whitney U with multiple-comparison correction). We will explicitly test both components of the step-function claim: (a) that the 1→2 jump is significant and (b) that subsequent differences are not. We agree that without these tests, the 'phase transition' language is not yet justified by the data as presented. If the tests reveal that some of the 2→3, 3→4, or 4→5 differences are in fact significant, we will revise the 'plateau' claim accordingly and soften the 'step function' characterization to match what the statistics support. We will retain the term only if the data warrant it. revision: yes

  2. Referee: §Behavioral categorization, Fig. 3A: Classification relies on hard thresholds (travel time difference < 120, remaining distance > 400) drawn at histogram boundaries. Sensitivity of categorical distributions to threshold values is not assessed. If the step-function claim is noisy, the pie charts may not robustly reflect stated trends.

    Authors: This is a fair point. The thresholds were chosen at visually identifiable histogram boundaries (Fig. S2), but we did not assess how sensitive the categorical distributions in Fig. 3B are to the exact threshold values. In the revision, we will add a sensitivity analysis: we will vary both thresholds over a reasonable range around their current values and report how the category proportions shift across conditions. We will also consider a data-driven clustering approach (e.g., k-means or Gaussian mixture on the two-dimensional metric space) as a complementary classification method, and show that the resulting categories align qualitatively with the threshold-based ones. If the trends we describe — particularly the shift from followers to hybrids at 3→5 direct agents — are not robust to threshold perturbation, we will revise the claim accordingly. We note that the referee's concern is partially contingent on Comment 1: if the step-function claim does not survive statistical testing, the downstream categorical claims would need to be tempered as well. We will ensure consistency between the two analyses. revision: yes

  3. Referee: §Spatial dependence, Fig. 2B/C: Directional divergence uses Jensen-Shannon divergence on spatially binned orientation distributions. Bin size is a free parameter that is not reported, and sensitivity to it is not tested. If too coarse, the decorrelation at high direct-agent numbers could be an artifact of averaging over heterogeneous local behaviors.

    Authors: The referee is correct that the spatial bin size is not reported in the manuscript and that this parameter could affect the magnitude and interpretation of the directional divergence metric. We will address this in two ways. First, we will explicitly state the bin size used in the current analysis in the revised text. Second, we will perform a sensitivity analysis by recomputing the directional divergence metric at two additional bin sizes (one finer, one coarser) and show the resulting Fig. 2B/C equivalents. This will allow the reader to assess whether the decorrelation between task and spatial dependence at high direct-agent numbers is robust to bin choice or whether it is an artifact of spatial averaging. We expect the qualitative pattern to hold — the decorrelation is driven by the spatial heterogeneity of social influence (individual navigation in the periphery, social dependence near the patch), which should persist across reasonable bin sizes — but we agree this must be demonstrated rather than assumed. If the decorrelation is not robust to bin size, we will revise the hybridization claim to reflect the appropriate level of confidence. revision: yes

Circularity Check

0 steps flagged

No significant circularity: the paper's central findings emerge from simulation experiments, not from definitions or self-citation chains.

full rationale

The paper trains neural network agents via evolutionary strategies in a simulation environment and reports empirical observations about learned navigation strategies. The central claims — phase transitions in social dependence, behavioral hybridization, decorrelation between task and spatial dependence — are outputs of simulation experiments (40 runs per condition, 5000 test initializations), not consequences of definitions or self-citations. The two self-citations ([1] Bastien & Romanczuk 2020 for visual raycasting inspiration; [2] Govoni & Romanczuk 2024 for the non-social navigation model and the directional divergence metric) provide methodological building blocks but do not constitute circular reasoning. The directional divergence metric (Jensen-Shannon divergence on spatially binned orientation distributions) is adapted from [2] but applied to a novel social comparison (NS vs. BEt test environments); it is not defined in terms of the results it claims to measure. The behavioral categorization thresholds (Fig. 3A, Fig. S2) are drawn at histogram boundaries of the empirical data, which is a classification choice rather than a circular definition — the categories (individual, follower, hybrid) are not defined to guarantee the paper's conclusions. The model extensions (Fig. 4) provide independent perturbation tests of the patch-blocking mechanism. The reader's concern about missing error bars and statistical tests is a correctness risk (underpowered claims), not circularity: the step-function claim could be noise, but it is not tautological. No step in the derivation chain reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities, particles, forces, or dimensions. The free parameters are primarily architectural and design choices for the simulation rather than fitted constants. The classification thresholds are post-hoc but are used for categorization rather than as fitted model parameters. The axioms are domain assumptions about the simulation design that collectively determine the observed results.

free parameters (7)
  • Number of rays (8) = 8
    Fixed architectural choice for the raycast visual system; not fitted to data but chosen by design.
  • Field of view limits (theta) = not specified
    Mentioned as between -theta and theta but exact value not stated in the main text.
  • Maximum number of untrained agents (5) = 5
    Explicitly chosen to keep travel times under 500 timesteps (Fig. S1); constrains the parameter space and affects patch-blocking dynamics.
  • Classification thresholds (travel time diff < 120, remaining distance > 400) = 120, 400
    Drawn at histogram boundaries (Fig. S2) to categorize agents as navigators, followers, or hybrids; post-hoc thresholds.
  • Spatial bin size for directional divergence = not specified
    Used for Jensen-Shannon divergence calculation but bin size not stated in main text.
  • Near-patch distance threshold (50 spatial units) = 50
    Used in trajectory integration (Fig. 3C) for near-initialization and near-patch categories.
  • Initialize-close radius (100 spatial units) = 100
    Used in the third model extension (Fig. 4C) for concentrating agent initialization.
axioms (5)
  • domain assumption Evolutionary strategies optimization is appropriate for training navigation policies in this setting
    §1 Model design: ES chosen over RL for its 'implicitly explorative population-based approach'; this choice affects the learned strategy distribution.
  • domain assumption The fitness function (travel time + remaining distance) adequately captures navigation ability
    §1 Model design: the second term (remaining distance) 'guides initial learning behavior'; the specific weighting affects which strategies are learned.
  • domain assumption Absorbing boundary conditions for agent-agent and agent-wall collisions are a reasonable model of physical embodiment
    §1 Model design: collisions are simulated with absorbing boundary conditions where only non-colliding directions are permitted; this directly causes the patch-blocking phenomenon central to the decorrelation finding.
  • domain assumption One-hot encoding of raycast collisions is sufficient visual input for learning the described strategies
    §1 Model design: visual input is one-hot encoded object identities at egocentric angles; the richness of this encoding constrains what the network can learn.
  • domain assumption The minimal square environment with four distinguishable walls is a valid testbed for social navigation dynamics
    §1 Model design: the environment is a minimal square arena; generalization to more complex environments is not tested.

pith-pipeline@v1.1.0-glm · 10012 in / 3454 out tokens · 568709 ms · 2026-07-09T10:09:00.599236+00:00 · methodology

0 comments
read the original abstract

Navigation for social organisms rarely is a fully independent activity. Group structure and dynamics, as well as embodied interactions, critically influence useful behavior. Individual neural network controlled agents are trained to navigate in different social contexts, where social dependence and behavioral strategy learned is determined by relative task performance and spatial effect. Increasing high quality social information drives phase transitions from individual to following navigational strategy, and to collision avoidance in response to a crowded foraging patch. Predictable, nonstationary environmental dynamics drive behavioral hybridization between individual and social navigation, far and near the patch. Our findings challenge the approach of only inspecting individual behavior for social organisms and highlight the importance of taking a bottom-up approach in understanding how organisms behave.

Figures

Figures reproduced from arXiv: 2607.07460 by Patrick Govoni, Pawel Romanczuk.

Figure 1
Figure 1. Figure 1: Agent flow & train-test methodology. Left: (clockwise from bottom left) visual encoding, information processing, action conversion, environment update. Visual encoding identifies walls and other agents corresponding to retinal angles of a raycast (between -θ & θ field of view limits). Visual information for the trained agent (blue) passes through convolutional neural network, perceptron, linear output laye… view at source ↗
Figure 2
Figure 2. Figure 2: Learned task & spatial dependencies, movement & neural activation maps. A: relative task performance between non-social (NS) & beacon-exploiter (BEt) test environments, as scatter plot for N=5 (top) & stair plot for N<=5 (bottom), where N is total number of untrained agents, varying the ratio between direct/random untrained agents per condition. All data (40 runs per condition, 5000 initializations average… view at source ↗
Figure 3
Figure 3. Figure 3: Categorizing learned behaviors. A: 2D plots replicating Fig. 2C, with the top separating agents into 3 categories: with navigators defined as less than 120 travel time difference (red dashed line), followers as above 400 remaining distance (shown below), and hybrids as those not fulfilling either criteria. B: pie charts organized into stair plots separated by training environment (top: categories, bototm: … view at source ↗
Figure 4
Figure 4. Figure 4: Model Extensions: Embodiment/Perceptibility. Results rerun with 3 changes to simulation parameters. Plots follow Fig. 2C/3B. A: no collisions between agents. B: focal agent receives extra binary input indicating current collision status to the linear output layer. C: untrained agents initialize within 100 spatial units of focal agent, where the test perturbations is NS-OG rather than NS-BEt (Fig. 1right-bo… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 300 canonical work pages · 24 internal anchors

  1. [1]

    Abels, Axel and Roijers, Diederik and Lenaerts, Tom and Now. Dynamic. Proceedings of the 36th

    work page

  2. [2]

    , year = 1980, publisher =

    Able, Kenneth P. , year = 1980, publisher =. Animal

    work page 1980

  3. [3]

    Tailoring

    Aceituno, Pau Vilimelis and Yan, Gang and Liu, Yang-Yu , year = 2020, month = sep, journal =. Tailoring. doi:10.1016/j.isci.2020.101440 , urldate =

    work page doi:10.1016/j.isci.2020.101440 2020

  4. [4]

    and Vuong, Cliff and Moore, Jason J

    Acharya, Lavanya and Aghajan, Zahra M. and Vuong, Cliff and Moore, Jason J. and Mehta, Mayank R. , year = 2016, month = jan, journal =. Causal. doi:10.1016/j.cell.2015.12.015 , urldate =

    work page doi:10.1016/j.cell.2015.12.015 2016

  5. [5]

    Physical Review E , volume =

    Thermodynamics of Evolutionary Games , author =. Physical Review E , volume =. doi:10.1103/PhysRevE.97.062136 , urldate =

    work page doi:10.1103/physreve.97.062136

  6. [6]

    Adamkiewicz, Michal and Chen, Timothy and Caccavale, Adam and Gardner, Rachel and Culbertson, Preston and Bohg, Jeannette and Schwager, Mac , year = 2022, month = apr, journal =. Vision-. doi:10.1109/LRA.2022.3150497 , urldate =

    work page doi:10.1109/lra.2022.3150497 2022

  7. [7]

    and Alexandropoulos, Stamatios-Aggelos N

    Adam, Stavros P. and Alexandropoulos, Stamatios-Aggelos N. and Pardalos, Panos M. and Vrahatis, Michael N. , editor =. No. Approximation and. doi:10.1007/978-3-030-12767-1_5 , urldate =

    work page doi:10.1007/978-3-030-12767-1_5

  8. [8]

    and Wheatleya, S.P

    Adams, R.R. and Wheatleya, S.P. and Gouldsworthy, A.M. and. Current Biology , volume =. doi:10.1016/S0960-9822(00)00673-4 , urldate =

    work page doi:10.1016/s0960-9822(00)00673-4

  9. [9]

    Melting Pot 2.0

    Agapiou, John P. and Vezhnevets, Alexander Sasha and. Melting. arXiv , file =:2211.13746 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Deep Reinforcement Learning at the Edge of the Statistical Precipice

    Agarwal, Rishabh and Schwarzer, Max and Castro, Pablo Samuel and Courville, Aaron and Bellemare, Marc G. , year = 2022, month = jan, number =. Deep. arXiv , langid =:2108.13264 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  11. [11]

    Learning Transferable Cooperative Behavior in Multi-Agent Teams

    Agarwal, Akshat and Kumar, Sumit and Sycara, Katia , year = 2019, month = jun, number =. Learning. doi:10.48550/arXiv.1906.01202 , urldate =. arXiv , langid =:1906.01202 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1906.01202 2019

  12. [12]

    Nature Neuroscience , volume =

    Impaired Spatial Selectivity and Intact Phase Precession in Two-Dimensional Virtual Reality , author =. Nature Neuroscience , volume =. doi:10.1038/nn.3884 , urldate =

    work page doi:10.1038/nn.3884

  13. [13]

    Current Opinion in Neurobiology , volume =

    Neuronal Strategies for Meeting the Right Partner during Brain Wiring , author =. Current Opinion in Neurobiology , volume =. doi:10.1016/j.conb.2020.01.002 , urldate =

    work page doi:10.1016/j.conb.2020.01.002 2020

  14. [14]

    Fungal Biology Reviews , volume =

    Branching out:. Fungal Biology Reviews , volume =. doi:10.1016/j.fbr.2015.03.001 , urldate =

    work page doi:10.1016/j.fbr.2015.03.001 2015

  15. [15]

    ISME Communications , volume =

    Network Traits Predict Ecological Strategies in Fungi , author =. ISME Communications , volume =. doi:10.1038/s43705-021-00085-1 , urldate =

    work page doi:10.1038/s43705-021-00085-1

  16. [16]

    eLife , volume =

    Perception as a Closed-Loop Convergence Process , author =. eLife , volume =. doi:10.7554/eLife.12830 , urldate =

    work page doi:10.7554/elife.12830

  17. [17]

    Journal of Experimental Biology , volume =

    Visual Navigation in Desert Ants. Journal of Experimental Biology , volume =. doi:10.1242/jeb.205.14.1971 , urldate =

    work page doi:10.1242/jeb.205.14.1971 1971

  18. [18]

    npj Digital Medicine , volume =

    Integrating Machine Learning and Multiscale Modeling---Perspectives, Challenges, and Opportunities in the Biological, Biomedical, and Behavioral Sciences , author =. npj Digital Medicine , volume =. doi:10.1038/s41746-019-0193-y , urldate =

    work page doi:10.1038/s41746-019-0193-y

  19. [19]

    Philosophical Transactions of the Royal Society B: Biological Sciences , volume =

    Expanding Theory, Methodology and Empirical Systems at the Spatial--Social Interface , author =. Philosophical Transactions of the Royal Society B: Biological Sciences , volume =. doi:10.1098/rstb.2022.0534 , urldate =

    work page doi:10.1098/rstb.2022.0534 2022

  20. [20]

    and Carstensen, Lucas C

    Alexander, Andrew S. and Carstensen, Lucas C. and Hinman, James R. and Raudies, Florian and Chapman, G. William and Hasselmo, Michael E. , year = 2020, journal =. Egocentric Boundary Vector Tuning of the Retrosplenial Cortex. doi:10.1126/sciadv.aaz2322 , urldate =

    work page doi:10.1126/sciadv.aaz2322 2020

  21. [21]

    Neuroscience , volume =

    Reinforcement Learning at the Interface of Artificial Intelligence and Cognitive Science , author =. Neuroscience , volume =. doi:10.1016/j.neuroscience.2025.09.004 , urldate =

    work page doi:10.1016/j.neuroscience.2025.09.004 2025

  22. [22]

    PNAS Nexus , volume =

    Nonlinear Social Evolution and the Emergence of Collective Action , author =. PNAS Nexus , volume =. doi:10.1093/pnasnexus/pgae131 , urldate =

    work page doi:10.1093/pnasnexus/pgae131

  23. [23]

    MicrobiologyOpen , volume =

    A Protocol for Recombinant Protein Quantification by Densitometry , author =. MicrobiologyOpen , volume =. doi:10.1002/mbo3.1027 , urldate =

    work page doi:10.1002/mbo3.1027

  24. [24]

    Cooperation and

    Aloor, Jasmine Jerry and Nayak, Siddharth Nagar and Dolan, Sydney and Balakrishnan, Hamsa , year = 2024, month = dec, journal =. Cooperation and. doi:10.1145/3702012 , urldate =

    work page doi:10.1145/3702012 2024

  25. [25]

    Project Sid: Many-agent simulations toward AI civilization

    AL, Altera and Ahn, Andrew and Becker, Nic and Carroll, Stephanie and Christie, Nico and Cortes, Manuel and Demirci, Arda and Du, Melissa and Li, Frankie and Luo, Shuying and Wang, Peter Y. and Willows, Mathew and Yang, Feitong and Yang, Guangyu Robert , year = 2024, month = oct, number =. Project. doi:10.48550/arXiv.2411.00114 , urldate =. arXiv , keywor...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2411.00114 2024

  26. [26]

    Trends in Cell Biology , volume =

    The Computational Sperm Cell , author =. Trends in Cell Biology , volume =. doi:10.1016/j.tcb.2013.10.004 , urldate =

    work page doi:10.1016/j.tcb.2013.10.004 2013

  27. [27]

    Alver, Safa and Precup, Doina , year = 2024, month = jun, journal =. An

    work page 2024

  28. [28]

    Autoencoder-Augmented Neuroevolution for Visual Doom Playing , booktitle =

    Alvernaz, Samuel and Togelius, Julian , year = 2017, month = aug, pages =. Autoencoder-Augmented Neuroevolution for Visual Doom Playing , booktitle =. doi:10.1109/CIG.2017.8080408 , urldate =

    work page doi:10.1109/cig.2017.8080408 2017

  29. [29]

    Alver, Safa and. Partial. doi:10.48550/arXiv.2405.16899 , urldate =. arXiv , keywords =:2405.16899 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.16899

  30. [30]

    Amato, Christopher , year = 2024, publisher =. An. doi:10.48550/ARXIV.2409.03052 , urldate =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2409.03052 2024

  31. [31]

    and Zahorik, Pavel , year = 2014, journal =

    Anderson, Paul W. and Zahorik, Pavel , year = 2014, journal =. Auditory/Visual Distance Estimation: Accuracy and Variability , shorttitle =

    work page 2014

  32. [32]

    Nature Reviews Neuroscience , volume =

    Circuit Modules Linking Internal States and Social Behaviour in Flies and Mice , author =. Nature Reviews Neuroscience , volume =. doi:10.1038/nrn.2016.125 , urldate =

    work page doi:10.1038/nrn.2016.125 2016

  33. [33]

    , year = 2010, month = aug, journal =

    Anderson, Michael L. , year = 2010, month = aug, journal =. Neural Reuse:. doi:10.1017/S0140525X10000853 , urldate =

    work page doi:10.1017/s0140525x10000853 2010

  34. [34]

    Scientific Reports , volume =

    Neural Signatures of Reinforcement Learning Correlate with Strategy Adoption during Spatial Navigation , author =. Scientific Reports , volume =. doi:10.1038/s41598-018-28241-z , urldate =

    work page doi:10.1038/s41598-018-28241-z

  35. [35]

    , year = 2021, month = jan, journal =

    Anobile, Giovanni and Arrighi, Roberto and Castaldi, Elisa and Burr, David C. , year = 2021, month = jan, journal =. A. doi:10.1016/j.tics.2020.10.009 , urldate =

    work page doi:10.1016/j.tics.2020.10.009 2021

  36. [36]

    Nature , volume =

    A Universal Biomolecular Integral Feedback Controller for Robust Perfect Adaptation , author =. Nature , volume =. doi:10.1038/s41586-019-1321-1 , urldate =

    work page doi:10.1038/s41586-019-1321-1

  37. [37]

    , year = 2014, month = feb, journal =

    Arbilly, Michal and Laland, Kevin N. , year = 2014, month = feb, journal =. The Local Enhancement Conundrum:. doi:10.1016/j.tpb.2013.09.006 , urldate =

    work page doi:10.1016/j.tpb.2013.09.006 2014

  38. [38]

    Roy and Eisenbach, Michael and Friedrich, Benjamin M

    Armon, Leah and Caplan, S. Roy and Eisenbach, Michael and Friedrich, Benjamin M. , year = 2012, month = mar, journal =. Testing. doi:10.1371/journal.pone.0032909 , urldate =

    work page doi:10.1371/journal.pone.0032909 2012

  39. [39]

    Nature , volume =

    Mapping of a Non-Spatial Dimension by the Hippocampal--Entorhinal Circuit , author =. Nature , volume =. doi:10.1038/nature21692 , urldate =

    work page doi:10.1038/nature21692

  40. [40]

    , year = 2022, month = dec, journal =

    Ashwood, Zoe and Jha, Aditi and Pillow, Jonathan W. , year = 2022, month = dec, journal =. Dynamic

    work page 2022

  41. [41]

    , year = 2001, month = oct, journal =

    Attwell, David and Laughlin, Simon B. , year = 2001, month = oct, journal =. An. doi:10.1097/00004647-200110000-00001 , urldate =

    work page doi:10.1097/00004647-200110000-00001 2001

  42. [42]

    and Mullally, Sin

    Auger, Stephen D. and Mullally, Sin. Retrosplenial. PLOS ONE , volume =. doi:10.1371/journal.pone.0043620 , urldate =

    work page doi:10.1371/journal.pone.0043620

  43. [43]

    Authors, Anonymous , langid =. An

    work page

  44. [44]

    Animal Behaviour , volume =

    Local Enhancement or Stimulus Enhancement?. Animal Behaviour , volume =. doi:10.1016/j.anbehav.2014.09.020 , urldate =

    work page doi:10.1016/j.anbehav.2014.09.020 2014

  45. [45]

    PLOS Computational Biology , volume =

    A Spectrum of Routing Strategies for Brain Networks , author =. PLOS Computational Biology , volume =. doi:10.1371/journal.pcbi.1006833 , urldate =

    work page doi:10.1371/journal.pcbi.1006833

  46. [46]

    Scientific Reports , volume =

    A Hippocampo-Cerebellar Centred Network for the Learning and Execution of Sequence-Based Navigation , author =. Scientific Reports , volume =. doi:10.1038/s41598-017-18004-7 , urldate =

    work page doi:10.1038/s41598-017-18004-7

  47. [47]

    Holistic Visual Encoding of Ant-like Routes:

    Baddeley, Bart and Graham, Paul and Philippides, Andrew and Husbands, Phillip , year = 2011, journal =. Holistic Visual Encoding of Ant-like Routes:

    work page 2011

  48. [48]

    Baddeley, Bart and Graham, Paul and Husbands, Philip and Philippides, Andrew , year = 2012, month = jan, journal =. A. doi:10.1371/journal.pcbi.1002336 , urldate =

    work page doi:10.1371/journal.pcbi.1002336 2012

  49. [49]

    Evolving

    Bah. Evolving. Proceedings of the 2015. doi:10.1145/2739480.2754790 , urldate =

    work page doi:10.1145/2739480.2754790 2015

  50. [50]

    Nature Neuroscience , volume =

    Neural Circuits for Evidence Accumulation and Decision Making in Larval Zebrafish , author =. Nature Neuroscience , volume =. doi:10.1038/s41593-019-0534-9 , urldate =

    work page doi:10.1038/s41593-019-0534-9

  51. [51]

    Baisero, Andrea , year = 2025, month = jun, series =. Role of. Proceedings of the 24th. doi:10.5555/3709347.3744047 , urldate =

    work page doi:10.5555/3709347.3744047 2025

  52. [52]

    eLife , volume =

    Spatial Modulation of Individual Behaviors Enables an Ordered Structure of Diverse Phenotypes during Bacterial Group Migration , author =. eLife , volume =. doi:10.7554/eLife.67316 , urldate =

    work page doi:10.7554/elife.67316

  53. [53]

    Proceedings of the National Academy of Sciences , volume =

    Stewardship of Global Collective Behavior , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.2025764118 , urldate =

    work page doi:10.1073/pnas.2025764118

  54. [54]

    Emergent Tool Use From Multi-Agent Autocurricula

    Baker, Bowen and Kanitscheider, Ingmar and Markov, Todor and Wu, Yi and Powell, Glenn and McGrew, Bob and Mordatch, Igor , year = 2020, month = feb, number =. Emergent. doi:10.48550/arXiv.1909.07528 , urldate =. arXiv , keywords =:1909.07528 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1909.07528 2020

  55. [55]

    doi:10.1101/2021.06.18.448989 , urldate =

    The Functional Specialization of Visual Cortex Emerges from Training Parallel Pathways with Self-Supervised Predictive Learning , author =. doi:10.1101/2021.06.18.448989 , urldate =

    work page doi:10.1101/2021.06.18.448989 2021

  56. [56]

    Organization Science , urldate =

    Balancing. Organization Science , urldate =

    work page

  57. [57]

    Balmant, Wellington and. A. PLOS ONE , volume =. doi:10.1371/journal.pone.0120307 , urldate =

    work page doi:10.1371/journal.pone.0120307

  58. [58]

    , year = 2019, month = mar, journal =

    Baltieri, Manuel and Buckley, Christopher L. , year = 2019, month = mar, journal =. doi:10.3390/e21030257 , urldate =

    work page doi:10.3390/e21030257 2019

  59. [59]

    , editor =

    Baltieri, Manuel and Buckley, Christopher L. , editor =. A. From. doi:10.1007/978-3-319-97628-0_2 , abstract =

    work page doi:10.1007/978-3-319-97628-0_2

  60. [60]

    Nature , volume =

    Vector-Based Navigation Using Grid-like Representations in Artificial Agents , author =. Nature , volume =. doi:10.1038/s41586-018-0102-6 , urldate =

    work page doi:10.1038/s41586-018-0102-6

  61. [61]

    doi:10.1007/978-3-319-24877-6 , urldate =

    Banisch, Sven , year = 2016, series =. doi:10.1007/978-3-319-24877-6 , urldate =

    work page doi:10.1007/978-3-319-24877-6 2016

  62. [62]

    Combining

    Bansal, Somil and Tolani, Varun and Gupta, Saurabh and Malik, Jitendra and Tomlin, Claire , year = 2020, month = may, pages =. Combining. Proceedings of the

    work page 2020

  63. [63]

    Bansal, Trapit and Pachocki, Jakub and Sidor, Szymon and Sutskever, Ilya and Mordatch, Igor , year = 2018, abstract =

    work page 2018

  64. [64]

    Current Opinion in Neurobiology , volume =

    Recurrent Neural Networks as Versatile Tools of Neuroscience Research , author =. Current Opinion in Neurobiology , volume =. doi:10.1016/j.conb.2017.06.003 , urldate =

    work page doi:10.1016/j.conb.2017.06.003 2017

  65. [65]

    Baratti, Greta and Potrich, Davide and Sovrano, Valeria Anna , year = 2020, month = apr, journal =. The. doi:10.1089/zeb.2019.1845 , urldate =

    work page doi:10.1089/zeb.2019.1845 2020

  66. [66]

    Barber. The. PLOS ONE , volume =. doi:10.1371/journal.pone.0143611 , urldate =

    work page doi:10.1371/journal.pone.0143611

  67. [67]

    doi:10.1101/2022.07.21.500962 , urldate =

    Flexible Selection of Task-Relevant Features through Population Gating , author =. doi:10.1101/2022.07.21.500962 , urldate =

    work page doi:10.1101/2022.07.21.500962 2022

  68. [68]

    Proceedings of the National Academy of Sciences , volume =

    Collective Cooperative Intelligence , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.2319948121 , urldate =

    work page doi:10.1073/pnas.2319948121

  69. [69]

    and Yang, Dongren and Lein, Pamela J

    Barnhart, Christopher D. and Yang, Dongren and Lein, Pamela J. , year = 2015, month = apr, journal =. Using the. doi:10.1371/journal.pone.0124521 , urldate =

    work page doi:10.1371/journal.pone.0124521 2015

  70. [70]

    Barreto, Andre and Borsa, Diana and Hou, Shaobo and Comanici, Gheorghe and Ayg. The. Advances in

    work page

  71. [71]

    Proceedings of the National Academy of Sciences , volume =

    Inhibitory Engrams in Perception and Memory , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.1701812114 , urldate =

    work page doi:10.1073/pnas.1701812114

  72. [72]

    Barton, N. H. and Turelli, Michael , year = 2011, month = sep, journal =. Spatial. doi:10.1086/661246 , urldate =

    work page doi:10.1086/661246 2011

  73. [73]

    Nature Neuroscience , volume =

    Learning-Induced Autonomy of Sensorimotor Systems , author =. Nature Neuroscience , volume =. doi:10.1038/nn.3993 , urldate =

    work page doi:10.1038/nn.3993

  74. [74]

    Nature Reviews Neuroscience , volume =

    On the Nature and Use of Models in Network Neuroscience , author =. Nature Reviews Neuroscience , volume =. doi:10.1038/s41583-018-0038-8 , urldate =

    work page doi:10.1038/s41583-018-0038-8

  75. [75]

    Nature Neuroscience , volume =

    Network Neuroscience , author =. Nature Neuroscience , volume =. doi:10.1038/nn.4502 , urldate =

    work page doi:10.1038/nn.4502

  76. [76]

    Science Advances , volume =

    A Model of Collective Behavior Based Purely on Vision , author =. Science Advances , volume =. doi:10.1126/sciadv.aay0792 , urldate =

    work page doi:10.1126/sciadv.aay0792

  77. [77]

    Nature , volume =

    The Orbitofrontal Cortex Maps Future Navigational Goals , author =. Nature , volume =. doi:10.1038/s41586-021-04042-9 , urldate =

    work page doi:10.1038/s41586-021-04042-9

  78. [78]

    IEEE Transactions on Aerospace and Electronic Systems , volume =

    Navigation Systems Based on Multiple Bearing Measurements , author =. IEEE Transactions on Aerospace and Electronic Systems , volume =. doi:10.1109/TAES.2015.140515 , urldate =

    work page doi:10.1109/taes.2015.140515 2015

  79. [79]

    Optimization and variability can coexist

    Optimization and Variability Can Coexist , author =. doi:10.48550/ARXIV.2505.23398 , urldate =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.23398

  80. [80]

    arXiv.org , urldate =

    Optimization and Variability Can Coexist , author =. arXiv.org , urldate =

    work page

Showing first 80 references.