CoChat

Learning few-shot imitation as cultural transmission

Open access

The DOI matches its Crossref record; 83 of the 83 references it cites were found in a registry or an index; 21 of the 30 sources its claims cite were checked for support.

Abstract

Cultural transmission is the domain-general social skill that allows agents to acquire and use information from each other in real-time with high fidelity and recall. It can be thought of as the process that perpetuates fit variants in cultural evolution. In humans, cultural evolution has led to the accumulation and refinement of skills, tools and knowledge across generations. We provide a method for generating cultural transmission in artificially intelligent agents, in the form of few-shot imitation. Our agents succeed at real-time imitation of a human in novel contexts without using any pre-collected human data. We identify a surprisingly simple set of ingredients sufficient for generating cultural transmission and develop an evaluation methodology for rigorously assessing it. This paves the way for cultural evolution to play an algorithmic role in the development of artificial general intelligence.

Overview

CoChat Summary

TL;DR

The paper introduces GoalCycle3D and MEDAL-ADR, a deep RL recipe that lets agents learn few-shot imitation as cultural transmission from expert co-players in novel 3D tasks without pre-collected human data. Ablations show memory, expert presence, expert dropout, attention loss, and automatic domain randomisation are jointly necessary for robust cultural transmission, recall, and generalisation.

Key findings

  • MEDAL-ADR agents show cultural transmission: training CT rises to ~0.75 (pure following), approaches 1 (remembering and reproducing after the expert leaves), then falls toward 0 as the agent becomes independent and exceeds the expert’s score; a vanilla RL baseline (M) fails to score on the task.
  • Ablations show removing memory, expert demonstrations, attention loss, or expert dropout eliminates or degrades cultural transmission; ADR is crucial in complex worlds, with MEDAL-ADR achieving significantly higher scores and CT than MEDAL-DR and MEDAL.
  • Recall and generalisation: after expert dropout, MEDAL-ADR continues solving the task for the first trial while ME-AL cannot, and maintains performance for several trials despite only 1800-step training episodes; in a two-action task over 1000 trials, the agent’s preference matched the demonstrated option 100% of the time; it generalises across world and game spaces, including out-of-distribution tasks.
  • Interpretability: a “social neuron” encodes the presence or absence of the expert, and a “goal neuron” fires when the agent enters and remains within a goal sphere; the attention loss is at least partly responsible for socially-aware representations.

Methods

  • Deep RL with maximum a posteriori policy optimisation (MPO) in a distributed framework; the agent uses a single-layer LSTM of size 512 unrolled for 800 steps, LIDAR observations, and an auxiliary attention loss that predicts the egocentric relative position of other players via a privileged AVATAR sensor used only at training time.
  • GoalCycle3D is a Unity-based 3D physical task space with procedurally generated worlds and games; agents must visit goal spheres in a correct cyclic order, with +1 reward for correct goals, −1 for incorrect goals, and 0 for re-entering the last goal; evaluation uses held-out probe tasks, including a human proxy expert.
  • Expert dropout schemes include no dropout, full dropout, half dropout, and probabilistic dropout; ADR expands task parameter ranges when training CT exceeds 0.75 and contracts them below a lower threshold; the cultural transmission metric averages improvement with the expert present and after the expert departs, normalised by expert score; two-action task used 1000 trials, and linear probing identified the social neuron.

Limitations

  • Stated: cultural transmission was tested with only a single human player from the study team, so the authors cannot make statistically significant claims about robustness across a human population; task diversity is constrained by the navigation task; and the evaluation cannot distinguish whether agents memorise a geographical path or an abstraction over the correct sphere ordering.
  • Stated: the method assumes expert visibility, dropout, and procedural generation, with silent assumptions including easy initial reward, no fine-motor control, short episode timescale, no irreversibilities, visible goals, and a constant rewarding order within an episode; the agent also shows objective misgeneralisation by following an incorrect demonstration and reproducing it after the expert drops out.
  • Evident from the text: code was not released because it was developed in a proprietary context, and the human evaluation relies on a single human proxy rather than a range of human participants.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41467-023-42875-2

Cites

83 of 83 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 GoalCycle3D. A 3D physical simulated task space.Each task contains procedurally generated terrain, obstacles, and goal spheres, with parameters randomly sampled on task creation. Each agent is independently rewarded for visiting goals in a particular cyclic order, also randomly sampled on task creation. The correct order is not provided to the agent, so an agent must deduce the rewarding order either by experimentation or via cultural transmission from an expert. Our task space presents navigational challenges of open-ended complexity, parameterised by world size, obstacle density, terrain bumpiness and a number of goals. Our agent observes the world using LIDAR (see Supplementary Movie 30).
Fig. 1 GoalCycle3D. A 3D physical simulated task space.Each task contains procedurally generated terrain, obstacles, and goal spheres, with parameters randomly sampled on task creation. Each agent is independently rewarded for visiting goals in a particular cyclic order, also randomly sampled on task creation. The correct order is not provided to the agent, so an agent must deduce the rewarding order either by experimentation or via cultural transmission from an expert. Our task space presents navigational challenges of open-ended complexity, parameterised by world size, obstacle density, terrain bumpiness and a number of goals. Our agent observes the world using LIDAR (see Supplementary Movie 30).CC BY · PubMed Central
Fig. 2 Training without ADR. Training cultural transmission (left) and agent score (right) for training without ADR on 4-goal in a small empty world. Colours indicate four distinct phases of agent behaviour from left to right: (1) (red) startup and exploration, (2) (blue) learning to follow, (3) (yellow) learning to remember, (4) (purple) becoming independent from expert.
Fig. 2 Training without ADR. Training cultural transmission (left) and agent score (right) for training without ADR on 4-goal in a small empty world. Colours indicate four distinct phases of agent behaviour from left to right: (1) (red) startup and exploration, (2) (blue) learning to follow, (3) (yellow) learning to remember, (4) (purple) becoming independent from expert.CC BY · PubMed Central
Fig. 3 Ablations of MEDAL ingredients. Score (left), training cultural transmission (CT, centre), and evaluation CT on empty world 5-goal probe tasks (right) over the course of training. a Comparing MEDAL with three ablated agents, each trained without one crucial ingredient: without an expert (M―), memory (–EDAL), or attention loss (MED—). b Ablating the effect of expert dropout, comparing no dropout (ME—AL) with expert dropout (MEDAL). We report the mean performance for each across 10 initialisation seeds for agent parameters and task procedural generation. We also include the expert’s score and MEDAL’s best seed for scale and upper-bound comparisons. The shaded area on the graphs is one standard deviation.
Fig. 3 Ablations of MEDAL ingredients. Score (left), training cultural transmission (CT, centre), and evaluation CT on empty world 5-goal probe tasks (right) over the course of training. a Comparing MEDAL with three ablated agents, each trained without one crucial ingredient: without an expert (M―), memory (–EDAL), or attention loss (MED—). b Ablating the effect of expert dropout, comparing no dropout (ME—AL) with expert dropout (MEDAL). We report the mean performance for each across 10 initialisation seeds for agent parameters and task procedural generation. We also include the expert’s score and MEDAL’s best seed for scale and upper-bound comparisons. The shaded area on the graphs is one standard deviation.CC BY · PubMed Central
Fig. 4 Analysis of ADR parameter expansion and ablation of ADR ingredients. a The expansion of parameter ranges over training for one representative seed in MEDAL-ADR training. b Score (left), training Cultural Transmission (CT, centre), and evaluation CT on complex world probe tasks (right) over the course of training for the automatic (A) and domain randomisation (DR) ablations of MEDAL-ADR. We report the mean performance for each across 10 initialisation seeds for agent parameters and task procedural generation. We also include the expert’s score and the best MEDAL-ADR seed for scale and upper bound comparisons. The shaded area on the graphs is one standard deviation.
Fig. 4 Analysis of ADR parameter expansion and ablation of ADR ingredients. a The expansion of parameter ranges over training for one representative seed in MEDAL-ADR training. b Score (left), training Cultural Transmission (CT, centre), and evaluation CT on complex world probe tasks (right) over the course of training for the automatic (A) and domain randomisation (DR) ablations of MEDAL-ADR. We report the mean performance for each across 10 initialisation seeds for agent parameters and task procedural generation. We also include the expert’s score and the best MEDAL-ADR seed for scale and upper bound comparisons. The shaded area on the graphs is one standard deviation.CC BY · PubMed Central
Fig. 5 Agent recall. Score of MEDAL-ADR and ME-AL agents across trials since the expert dropped out. a Experts are scripted bots. b Experts are human trajectories. Supplementary Movie 10 shows MEDAL-ADR’s recall from a bot demonstration in a 3600-step (4 trial) episode. Supplementary Movie 31 shows MEDAL-ADR’s recall from a human demonstration in an 1800-step (2 trial) episode.
Fig. 5 Agent recall. Score of MEDAL-ADR and ME-AL agents across trials since the expert dropped out. a Experts are scripted bots. b Experts are human trajectories. Supplementary Movie 10 shows MEDAL-ADR’s recall from a bot demonstration in a 3600-step (4 trial) episode. Supplementary Movie 31 shows MEDAL-ADR’s recall from a human demonstration in an 1800-step (2 trial) episode.CC BY · PubMed Central
Fig. 6 Evidence of causality. Trajectory plots for MEDAL-ADR agent for a single episode. a The bot is absent for the whole episode. b The bot shows a correct trajectory in the first half of the episode and then drops out. c The bot shows an incorrect trajectory in the first half of the episode and then drops out. The coloured parts of the lines correspond to the colour of the goal sphere the agent and expert have entered and the × s correspond to when the agent entered the incorrect goal. Here, position refers to the agent’s position along the z-axis. Supplementary Movies 11–13 correspond to each plot respectively.
Fig. 6 Evidence of causality. Trajectory plots for MEDAL-ADR agent for a single episode. a The bot is absent for the whole episode. b The bot shows a correct trajectory in the first half of the episode and then drops out. c The bot shows an incorrect trajectory in the first half of the episode and then drops out. The coloured parts of the lines correspond to the colour of the goal sphere the agent and expert have entered and the × s correspond to when the agent entered the incorrect goal. Here, position refers to the agent’s position along the z-axis. Supplementary Movies 11–13 correspond to each plot respectively.CC BY · PubMed Central
Fig. 7 Task space generalisation. a A slice through the world space allows us to disentangle MEDAL-ADR’s generalisation capability across different world space parameters. b MEDAL-ADR generalises across the game space, demonstrating remembering capability both inside and outside the training distribution. We report the mean performance across 50 initialisation seeds for a and 20 initialisation seeds for b. The error bars on the graphs represent 95% confidence intervals. Supplementary Movies 14–20 demonstrate generalisation over the world space and game space.
Fig. 7 Task space generalisation. a A slice through the world space allows us to disentangle MEDAL-ADR’s generalisation capability across different world space parameters. b MEDAL-ADR generalises across the game space, demonstrating remembering capability both inside and outside the training distribution. We report the mean performance across 50 initialisation seeds for a and 20 initialisation seeds for b. The error bars on the graphs represent 95% confidence intervals. Supplementary Movies 14–20 demonstrate generalisation over the world space and game space.CC BY · PubMed Central
Fig. 8 Introspection of Agent’s Brain. a Activations for MEDAL-ADR’s social neuron. b We report the accuracy of three linear probing models trained to predict the expert’s presence based on the belief states of three agents (MED—, MEDAL, and MEDAL-ADR). We make two causal interventions (in green and purple) and a control check (in red) on the original test set (yellow). We report the mean performance across 10 different initialisation seeds. The small standard deviation error bars suggest a broad consensus across the 10 runs on which neurons encode social information. c Spikes in the goal neuron’s activations correlate with the time the agent remains inside a goal (illustrated by coloured shading). The goal neuron was identified using a variance analysis, rather than the linear probing method in b.
Fig. 8 Introspection of Agent’s Brain. a Activations for MEDAL-ADR’s social neuron. b We report the accuracy of three linear probing models trained to predict the expert’s presence based on the belief states of three agents (MED—, MEDAL, and MEDAL-ADR). We make two causal interventions (in green and purple) and a control check (in red) on the original test set (yellow). We report the mean performance across 10 different initialisation seeds. The small standard deviation error bars suggest a broad consensus across the 10 runs on which neurons encode social information. c Spikes in the goal neuron’s activations correlate with the time the agent remains inside a goal (illustrated by coloured shading). The goal neuron was identified using a variance analysis, rather than the linear probing method in b.CC BY · PubMed Central
Fig. 9 Ingredients of MEDAL-ADR. The minimal sufficient ingredients that comprise our methods, grouped by the timescale on which they operate.
Fig. 9 Ingredients of MEDAL-ADR. The minimal sufficient ingredients that comprise our methods, grouped by the timescale on which they operate.CC BY · PubMed Central
Fig. 10 Worlds and games used as probe tasks. a Empty world, 4-goal games. b Empty world, 5-goal games. c Complex world, 4/5-goal games. These cover a representative range of crossings and colour combinations. The empty world probe tasks have terrain of size 20 × 20 m2, while the complex world probe tasks have terrain of size 32 × 32 m2. The complex world probes require clear examples of jumping behaviours and navigation around vertical obstacles. The human movement pattern in all probes is always goal-directed and near-optimal, but clearly different from a scripted bot, taking some time to get situated in the first few seconds and not taking an identical path on repeated cycles, for instance. See Supplementary Movies 21–29.
Fig. 10 Worlds and games used as probe tasks. a Empty world, 4-goal games. b Empty world, 5-goal games. c Complex world, 4/5-goal games. These cover a representative range of crossings and colour combinations. The empty world probe tasks have terrain of size 20 × 20 m2, while the complex world probe tasks have terrain of size 32 × 32 m2. The complex world probes require clear examples of jumping behaviours and navigation around vertical obstacles. The human movement pattern in all probes is always goal-directed and near-optimal, but clearly different from a scripted bot, taking some time to get situated in the first few seconds and not taking an identical path on repeated cycles, for instance. See Supplementary Movies 21–29.CC BY · PubMed Central

Verification

Claims the paper makes

13 Valid · 8 Not valid · 5 Not checked · 98 Own results and interpretations

47 more claims · 1 with their cited sources checked

Create a free account to see all 124 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 44
  • It may state two separate claims as one. 36
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 83 references

Research works printed but not linked: 5

  1. Gabora, L. Meme and variations: a computer model of cultural evolution. In: Nadel, L. and Stein, D.(Eds.), 1993 Lectures in Complex Systems, Addison-Wesley. (1995).
  2. Sutton, R. S. & Barto, A. G. Reinforcement learning: An introduction (MIT press, 2018).
  3. Galef Jr., B. G. Imitation in animals: History, definition, and interpretation of data from the psychological laboratory. In Zentall, T. R. & Galef Jr., B. G. (eds.) In: Social learning: psychological and biological perspectives, 3–28 (Lawrence Erlbaum Associates, Inc., 1988).
  4. Bratman, M. E. Intention, plans, and practical reason. Mind 97, 632 (1988).
  5. Vezhnevets, A., Wu, Y., Eckstein, M., Leblond, R. & Leibo, J. Z. Options as responses: grounding behavioural hierarchies in multi-agent reinforcement learning. In: International Conference on Machine Learning, 9733–9742 (PMLR, 2020).

Cited by

See all 15

Metadata

Authors
Bhoopchand, A., Brownfield, B., Collister, A., Dal Lago, A., Edwards, A., Everett, R., Fréchette, A., Oliveira, Y. G., Hughes, E., Mathewson, K. W., Mendolicchio, P., Pawar, J., Pȋslar, M., Platonov, A., Senter, E., Singh, S., Zacherl, A., & Zhang, L. M.
Affiliations
Google DeepMind (United Kingdom) (GB)
Corresponding author
Hughes, Edward
Venue
Nature Communications
Published
2023-11-28
Volume
14
Issue
1
Pages
7536-7536
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4389080522
Cited by
14
References
53
Citations by year
2026: 4 · 2025: 4 · 2024: 5 · 2023: 1
FWCI
21.83
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
DeepMind
Topics
  • Social Sciences › Cultural Studies › Language and cultural evolution
  • Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
  • Physical Sciences › Computer Science › Artificial Intelligence › Evolutionary Algorithms and Applications
Keywords
cultural transmission, cultural evolution, artificial general intelligence, social learning, human-robot interaction, evaluation methodology

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Learning few-shot imitation as cultural transmission. https://app.cochat.ai/papers/10.1038/s41467-023-42875-2
The paper
Bhoopchand, A., Brownfield, B., Collister, A., Dal Lago, A., Edwards, A., Everett, R., Fréchette, A., Oliveira, Y. G., Hughes, E., Mathewson, K. W., Mendolicchio, P., Pawar, J., Pȋslar, M., Platonov, A., Senter, E., Singh, S., Zacherl, A., & Zhang, L. M. (2023). Learning few-shot imitation as cultural transmission. Nature Communications. https://doi.org/10.1038/s41467-023-42875-2

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.