Human-Timescale Adaptation in an Open-Ended Task Space

1. Introduction

• AdA adapts within minutes on challenging held-out sparse-reward 3D tasks, matches trained human players’ adaptation timescales, and improves with zero-shot first-person demonstrations.The agent uses hypothesis-driven exploration and refines its policy toward near-optimal behavior without prompts, fine-tuning, or offline datasets.Source: §1 · pp. 2–3

2.1. Open-ended task space: XLand 2.0

• XLand 2.0 creates a vast, diverse adaptation space by combining procedurally generated games, worlds, co-players, and production-rule dynamics with partially hidden information.Tasks require capabilities including experimentation, tool use, navigation, coordination, division of labour, and coping with irreversibility.Source: §2.1 · pp. 3–4, Figure 3 · p. 4

2.2. Meta-RL

• Meta-RL trains memory as an implicit online learning algorithm across repeated trials, enabling test-time policy adaptation without updating neural-network weights.Tasks are partially observable Markov decision processes, with memory retained across trial boundaries and reset only between episodes.Source: §2.2 · p. 5

2.3. Auto-curriculum learning

• Adaptive curricula select tasks near the agent’s capability frontier because uniform sampling would mostly produce tasks too difficult or too easy to support learning.The proposed no-op filtering and prioritized level replay extensions improve performance and sample efficiency while producing an emergent curriculum of increasing complexity.Source: §2.3 · pp. 5–6

2.4. RL agent

• Attention-based memory addresses fast adaptation in partially observable 3D tasks: both RNN-with-attention and Transformer-XL outperform vanilla RNNs, with Transformer-XL performing best.Transformer-XL supports longer, variable-length context windows and is the default memory architecture; subsampling further extends effective memory without added computational cost.Source: §2.4 · pp. 7–8

2.5. Distillation

• Distillation kickstarts AdA with a pretrained teacher during its first four billion training steps, amortizing early training costs and helping avoid harmful initial representations.The teacher has 23M Transformer parameters, whereas multi-agent AdA has 265M; AdA acts using its own policy while matching the teacher’s action distributions.Source: §2.5 · p. 8

3.1. AdA shows human-timescale adaptation

• AdA improves on over 80% of held-out single-agent tasks with additional trials and adapts on human timescales, with score slopes as steep as or steeper than humans.On 30 hand-authored tasks, both AdA and humans improve with more trials; some tasks remain selectively solvable by only one group.Source: §3.1 · p. 10, Figure 6 · p. 10
• AdA uses early trials for structured hypothesis-driven exploration, then refines successful strategies; with 8 trials, its performance approaches that of a fine-tuned agent.The pattern is reported consistently across held-out probe tasks, including the Wrong Pair Disappears task.Source: §3.1 · pp. 10–11, Figure 7 · p. 11
• In cooperative multi-agent tasks, adaptation emerges across almost 90% of held-out tasks, while self-play substantially outperforms random-action co-play through division of labour and physical coordination.Agents explore independently or together, then coordinate paths and roles after finding solutions, despite no explicit training incentive beyond cooperative reward.Source: §3.1 · pp. 11–12, Figure 8 · p. 12

3.2. Architecture influences performance

• Transformer-XL performs best among matched architectures, while adding multi-head attention to an RNN recovers most Transformer performance.The comparison includes Transformer-XL, GRU-based RNN, and RNN with Attention at matched total network size.Source: §3.2 · p. 13, Figure 10 · p. 13

3.3. Auto-curriculum learning improves performance

• No-op filtering and PLR strongly outperform uniform task sampling, with PLR especially improving longer-term adaptation; both curricula are also more sample-efficient.The curricula produce emergent increases in task complexity despite not explicitly optimizing the reported complexity metrics.Source: §3.3 · p. 13, Figure 10 · p. 13, Figure 11 · p. 14

3.4. Scaling the agent increases performance

• Larger Transformer models, longer memory, and larger task pools improve test performance, especially lower-percentile scores and adaptation at higher trial counts.Model-size relationships appear roughly linear on log-log axes; memory scaling benefits effective horizons beyond a single task’s maximum trial duration, while larger models benefit more from larger pools.Source: Figure 12 · p. 14, §3.4 · p. 15, Figure 13 · p. 15

3.6. Distillation improves performance and enables scaling agents

• Kickstarting a 265M-parameter model with 4 billion distilled frames reverses its disadvantage against a similarly trained 23M model, while same-size distillation also yields large gains.Distillation’s gains are largest in the first round and diminish thereafter; the authors attribute them partly to avoiding poor early representations.Source: §3.6 · pp. 17–18, Figure 15 · p. 17, Figure 16 · p. 17

3.7. Training on more trials with skip memory enables many-shot adaptation

• Increasing training trials improves later-trial adaptation and teaches better exploration and refinement, while scaling memory can help, but additional memory alone does not improve performance without training to use it.Source: §3.7 · p. 18

3.8. AdA can leverage prompting with first-person demonstrations

• Increasing training trials has the largest effect on 20th-percentile test scores and is required before memory sub-sampling improves adaptation over more than 30 minutes.Adding an RNN encoder to enable sub-sampling does not substantially improve performance by itself.Source: Figure 17 · p. 19
• Prompted AdA outperforms the unprompted baseline across all trial counts but scores below the teacher demonstration, despite no training on off-policy first-person demonstrations.This shows that its in-context learning generalizes to demonstration prompts, though it cannot exactly mimic the teacher.Source: §3.8 · p. 19, Figure 17 · p. 19
• Prompting succeeds on some hand-authored tasks, but human demonstrations cannot overcome inherent limitations of AdA’s training task distribution.Results include all single-agent hand-authored tasks and early human-demonstration experiments on a subset.Source: §3.8 · p. 19

Overall positioning

• The approach combines in-context meta-RL with automated curation of a vast task distribution, differing from OEL Team et al. (2021) and Akkaya et al. (2019) by targeting adaptation with large Transformers rather than zero-shot performance or task-specific sim-to-real control.This combination is framed as integrating two pillars of an AI generating algorithm: in-context learning and automated training-task curation.Source: §4 · pp. 19–20

Procedural environment generation

• Procedural content generation supplies diverse RL testbeds and open-ended worlds, but Minecraft lacks full simulation-stack control, limiting task-space smoothness relative to XLand.The cited literature spans game-generation research, generalization benchmarks, and 3D open-ended environments such as Avalon and Minecraft.Source: §4 · p. 20

Open-ended learning

• Prior open-ended learning and UED methods curate task distributions through adversaries, regret, or curricula, whereas AdA discovers unobserved dynamics at test time through hypothesis-driven exploration over multiple trials.The related methods commonly target a downstream task, singleton environment, or uniformly sampled tasks, while AdA focuses on general adaptation without test-task knowledge.Source: §4 · p. 20

Adaptation and Transformers

• The work builds on memory-based meta-RL, where persistent memory over rewards and terminations can implement learning algorithms, and uses Transformer-XL to scale online adaptation beyond prior small-scale successes.Related approaches include meta-gradients and latent-variable methods; prior Transformer-based meta-RL and Alchemy results were limited to simpler or smaller-scale settings, while this work reports online meta-RL scaling laws.Source: §4 · p. 21

• AdA rapidly adapts across a vast open-ended task space at a timescale similar to human players, using structured exploration, few interactions, and first-person prompting.Its performance scales with parameter count, context length, and the richness of the training task distribution.Source: §5 · p. 21
• The work shows that black-box meta-RL can scale when automatic curricula shape sufficiently informative data distributions for learning to learn.This challenges the view that black-box meta-RL is inherently difficult to scale in open-ended task spaces.Source: §5 · p. 21
• Attention-based architectures exploit the curriculum signal more effectively than purely recurrent networks, while distillation enables large-scale Transformer architectures.The results emphasize co-adapting the training distribution and agent architecture to facilitate rapid adaptation.Source: §5 · p. 22
• AdA provides a recipe for training a 500M parameter model with scalable adaptation across tool use, experimentation, division of labour, and navigation.The authors suggest that scaling trends could make such models foundations for few-shot adaptation and fine-tuning on real-world control problems.Source: §5 · p. 22

6.1. Core contributors

• Core contributors covered technical and research leadership, curriculum and agent research, infrastructure, evaluation, task authoring, scaling, analysis, visualization, human experiments, and paper writing.The listed roles span the paper’s central research, engineering, evaluation, and communication activities.Source: §6.1 · p. 22

6.2. Partial contributors

• Partial contributors supported project management, QA testing, evaluation, agent research and analysis, XLand development, task authoring, human experiments, and paper writing.These contributions are distributed across project coordination, experimentation, environment development, and research support.Source: §6.2 · pp. 22–23

6.3. Sponsors

• Sponsors provided strategic advice to the project.Satinder Baveja and Karl Tuyls are listed as sponsors with this role.Source: §6.3 · p. 23

A.1. XLand 2.0

• XLand 2.0 differs from the original XLand environment of OEL Team et al. (2021) through changes to its configuration space.The appendix introduces these changes as environment-level modifications.Source: §A.1 · p. 33
• XLand 2.0 adds touching(a,b), removes floor-based relations, supports permanently frozen objects as anchors, and restricts predicates to single relations or their negations.Touching uses Unity collision detection with a 1 millimeter distance threshold; frozen objects can appear in goals or production rules.Source: §A.1 · p. 33
• Although predicates are limited to single relations or negations, production rules still support sequential goals and multiple ways to reach a desired state.This preserves task expressivity through rule structure rather than conjunctions or disjunctions within predicates.Source: §A.1 · p. 33
• Appendix tables enumerate the shapes, colours, and predicates used for XLand objects, goals, and production rules.Tables A.1, A.2, and A.3 provide the corresponding vocabularies.Source: §A.1 · p. 33, Table A.1 · p. 33, Table A.2 · p. 33

A.2. Pre-sampling tasks for training

• The task generator samples from at least 10^40 distinct goal-and-production-rule tasks, then pre-samples a training subset for practicality.The estimate is a lower bound under restrictive bounds on shapes, colours, predicates, rules, objects, and hiding options.Source: §A.2 · p. 34, §B.1 · p. 34
• Single-player games chain one to four production rules toward a sampled goal, add up to 10 distractors, and include randomly hidden rule components.Dead-end rules spawn only distractors and can make games unsolvable, discouraging indiscriminate rule triggering.Source: §A.2 · p. 34
• Multi-player training games are fully cooperative two-player variants with potentially asymmetric rule information and an equilibrium-selection challenge.They are created by transforming a sampled single-player game, copying its goal, and independently masking player two’s rules.Source: §A.2 · p. 34

B.1. Test scores

• Agents are evaluated on 1,000 held-out tasks from the training distribution but with held-out world topologies and no duplicated goal-rule games.Rejection sampling prevents any test game from appearing in training.Source: §B.1 · p. 34
• Test scores divide each trial’s total reward by a task-specific normalizer estimated from fine-tuned AdA, producing an approximate [0, 1] scale that can exceed 1.Scores above 1 can result from evaluation noise or the evaluated agent outperforming the normalizer agent.Source: §B.1 · pp. 34–35
• Evaluation reports the final trial’s reward across k ∈ {1, 2, 3, 5, 8, 13}, testing whether additional experience improves adaptation, including held-out trial counts 8 and 13.Across games, a fixed percentile summarizes performance and emphasizes coverage of many tasks rather than outlier performance.Source: §B.1 · p. 35

B.2. Hand-authored probe tasks

• Because sampled test tasks may be difficult for humans to interpret, the study additionally examines 30 single-agent and 28 multi-agent hand-authored probe tasks.These probes target intuitive situations and qualitatively different behaviours that can be inspected directly; representative examples appear in Figures B.1–B.3.Source: §B.2 · p. 35

B.3. Adaptation metric

• The adaptation metric ranks agents by their few-shot performance across hand-authored probe tasks and trial counts.It uses normalized last-trial total rewards, aggregates task scores by the median, and compares agents using Pareto improvement rather than trading off zero-shot and few-shot performance.Source: §B.3 · p. 35
• The metric values both high-quality zero-shot generalization at k=1 and efficient improvement when k>1.An agent ranks above another only when its task-aggregated scores are a Pareto improvement across the evaluated trial counts.Source: §B.3 · p. 35
• The authors convert the Pareto criterion into a scalar using the Nash average method of Balduzzi et al. (2018).They form an agents-versus-k meta-game, compute its maximum-entropy Nash equilibrium, and use the Nash payoff to rank agents; non-dominating agents receive equal payoffs.Source: §B.3 · p. 35
• The adaptation metric selects hyperparameters for training the best-performing agent reported in Section 3.1.This connects the metric directly to model-selection decisions rather than using it only for post hoc evaluation.Source: §B.3 · p. 35

B.4. Human data collection

• Human benchmarking used 100 players completing 30 single-agent probe tasks after a graded 23-task training curriculum.Players experienced varied trial counts, with each player assigned a unique task ordering and similar tasks separated to reduce knowledge transfer; 3.4% of 3000 episodes were incomplete.Source: §B.4 · p. 37
• The human protocol sampled k∈{1, 2, 3, 5, 8} while preserving each player’s short-term memory across episode boundaries.Each participant attempted each probe task for a single k, producing an average of 19.3 samples per task-and-k pair and at least 17 for every pair.Source: §B.4 · p. 37

C. Agent Details

• The agent-details section specifies the agent architecture, observation space, and action space.The supplied passage introduces these technical specifications without providing their component values or design details.Source: C. Agent Details · p. 38

C.1. Agent Architecture

• The architecture encodes first-person RGB observations with a ResNet, goals with an 8-dimensional embedding, and production rules with a 256-dimensional embedding.RGB processing uses [16, 32, 32] convolutional channels, while goals pass through a 3-layer [8, 8, 8] MLP and production rules through [512, 256].Source: §C.1 · p. 38
• Encoded visual, goal, rule, scalar, and vector observations are concatenated and projected to the Transformer memory’s hidden dimension.Remaining inputs include previous reward, previous action, proprioception, and trial timing information.Source: §C.1 · p. 38
• Transformer-XL provides causally masked memory with an 80-step learner context and 300 cached activations, equivalent to 1800 effective timesteps.The actor context window is 1; the model also uses pre-attention layer normalization, gated feedforward layers, relative positional embeddings, and GELU activations.Source: §C.1 · p. 38
• The Transformer output initializes a Muesli LSTM whose policy and value heads produce six action-group outputs, binned value predictions, and multi-step reward predictions.The Muesli model uses width-1000 MLPs, a 601-unit binned value prediction, four additional LSTM steps, and a 601-unit binned reward prediction.Source: §C.1 · p. 38

C.2. Observations

• The agent receives a structured observation set summarised in Table C.1, including inherited legacy-reason fields and trial-duration information.For a task, all trials share the same duration, including the final and next trial durations.Source: Table C.1 · p. 39, §C.2 · pp. 39–40

D.1. Meta-RL

• AdA uses meta-RL episodes containing multiple task trials, resetting the environment at trial boundaries while retaining memory until the episode ends.A fixed nonzero discount factor applies throughout the episode, including across trial boundaries.Source: §D.1 · p. 40
• Training samples MDP-trial pairs from procedurally generated XLand using an automatic curriculum, with trial counts 1–6 supplied as conditioning inputs.After training, the agent adapts few-shot across held-out MDPs outside the training distribution and across episodes with more than six trials.Source: §D.1 · p. 40

Single-agent training

• The single-agent setup trained AdA from scratch for approximately 5 weeks on 64 Google TPUv3 devices, yet the model had not converged.Training comprised 1 week for the teacher and 4 weeks for AdA, illustrating the benefits of open-ended learning methods.Source: §D.2 · p. 40
• Single-agent training used an earlier XLand 2.0 version without frozen objects, which were also excluded from test and hand-authored probe task sets.Source: §D.2 · p. 40
• The single-agent experiments used a task pool generated as described in Section 2.1 and a distillation teacher summarized in Table D.1.Source: §D.2 · p. 40, Table D.1 · p. 40

Task pool and task construction

• Training uses a 50%-50% mixture of single-player and two-player tasks generated from a 25B-sized task pool, with some initially spawned objects permanently frozen.Frozen objects are introduced with 50% probability, and each spawned object type is frozen with 20% probability while preserving solvability.Source: §D.3 · pp. 40–41

Co-player training

• A fictitious-self-play pool samples co-player policies, beginning with random actions and adding learning-agent snapshots every 500M training frames.This progressively introduces more capable co-players during training.Source: §D.3 · p. 41

Curriculum and distillation teacher

• PLR auto-curriculum curates worlds, games, and co-players using the agent’s TD-error-based fitness, while Table D.2 summarizes the multi-agent distillation teacher.The table describes the distillation teacher used in the multi-agent experiments.Source: §D.3 · p. 41, Table D.2 · p. 41

D.4. Architecture experiments

• The experiments compare different memory architectures in Section 3.2 using the setup reported in Table D.3.Source: Table D.3 · p. 41, §D.4 · p. 41

D.5. Auto-curriculum learning

• No-op filtering admits tasks only when no-op and agent performance, divergence from no-op, and agent-score variance satisfy threshold criteria.Relative thresholds are used for duration-dependent quantities; admitted tasks run for all specified trials across 30 episodes, with experience sent to the learner.Source: §D.5 · p. 41
• PLR maintains a fixed-size archive that replays tasks according to fitness and staleness, adding proposals whose fitness exceeds the archive minimum.The adapted fitness uses normalized last-trial scores, and last-trial fitness performs better and is more sample-efficient than first- or average-trial fitness.Source: §D.5 · p. 42
• TD-error and action-model fitness outperform value-model fitness, while TD-error is selected for PLR because it has better asymptotic zero-shot and few-shot performance.Action-model fitness matches TD-error in few-shot performance but has weaker zero-shot performance.Source: §D.5 · p. 43, Figure D.1 · p. 43
• PLR and No-op filtering substantially improve training sample efficiency over uniform task sampling, with PLR better early in learner steps and later in FLOPs.No-op has a slight early FLOPs advantage, while PLR later outperforms it; No-op evaluates each task 10 times versus PLR’s single evaluation.Source: §D.5 · p. 44, Figure D.2 · p. 44

D.6. Distillation teacher for scaling experiments

• Scaling experiments distill every policy from an identical teacher snapshot, which kickstarts agents during their first 4B training frames.The teacher configuration is documented separately for comparability across scaling experiments.Source: §D.6 · p. 44, Table D.7 · p. 45

D.7. Scaling the network

• Network-size scaling fixes Transformer depth and varies only embedding size, preserving the effective Transformer-XL memory horizon while increasing parameters.The feed-forward size is fixed at 4d_model, and the effective horizon is M x L for M cached activations and L attention layers.Source: §D.7 · pp. 45–46, Table D.8 · p. 45, Table D.9 · p. 46

D.8. Scaling the memory length

• Memory-length scaling reports effective memory timesteps as cached activations multiplied by the six Transformer blocks.The experiments and their configurations are listed in Table D.10.Source: §D.8 · p. 46, Table D.10 · p. 46

D.9. Scaling the size of the task pool

• Task-pool scaling uses the experimental setup documented in Table D.11 to vary the training task-pool size.The setup corresponds to the task-pool scaling experiments in Section 3.5.Source: §D.9 · p. 46, Table D.11 · p. 46

D.10. Scaling the complexity of the task pool

• Task-complexity scaling compares agents with different distillation teachers, so distillation and task complexity cannot be disentangled, although results remain indicative of task-complexity importance.The compared teachers differ across multi-topology and single-room task distributions.Source: §D.10 · pp. 46–47, Table D.12 · p. 47

D.11. Distillation enables scaling agents

• The distillation experiments use the experimental setup reported in Table D.13.The supplied passage identifies this setup as supporting the distillation experiments in Section 3.6.Source: §D.11 · p. 47, Table D.13 · p. 47

D.12. Training on more trials with skip memory

• The skip-memory experiments train on more trials while varying training steps, memory architecture, and training trials relative to the reported setup.Table D.14 documents the experimental setup for these memory-scaling experiments.Source: §D.12 · p. 47, Table D.14 · p. 47

E.1. Multi-agent adaptation

• AdA’s last-trial reward increases with more trials across most of the multi-agent task distribution, demonstrating test-time adaptation in cooperative self-play.A representative probe task shows the same performance increase as the number of trials grows.Source: Figure E.1 · p. 48, Figure E.2 · p. 48, §E.1 · p. 48

E.2. Conditioning on number of shots doesn’t affect agents’ performance

• AdA’s first-trial scores are almost identical with one versus eight trials, indicating that trial-count conditioning does not alter behavior enough to affect performance.The expected lower first-trial score under an exploratory policy with eight total trials is not observed.Source: §E.2 · p. 49, Figure E.3 · p. 49
• The unchanged early behavior may be optimal for XLand 2.0 or a training limitation, because all episode rewards and discounting may discourage delayed exploration.A Bayes-optimal policy could sacrifice early stochastic reward to secure deterministic later reward, but this behavior was not directly incentivized.Source: §E.2 · p. 49

E.3. Scaling complexity of the task pool

• The task-complexity comparison preserves goals, production rules, and objects while replacing diverse world topologies with an empty room and adding 5,000 initialization points.The added points proxy for 4,000 worlds but are intentionally less diverse and complex than the main training pool.Source: §E.3 · p. 49
• With complex world-topology distributions, larger models continue improving through at least 75M parameters, whereas empty-room training begins degrading beyond 42M parameters.This indicates that low environment complexity can bottleneck model scaling even when the total number of tasks is increased.Source: §E.3 · p. 50, Figure E.4 · p. 50

E.4. Computational cost

• Under matched total FLOPs, performance declines beyond a model-size sweet spot around 57M Transformer parameters, while longer memory contexts remain beneficial.The FLOPs-matched comparisons account for learner and actor computation.Source: Figure E.5 · p. 51, Figure E.6 · p. 51, §E.4 · p. 52
• When only learner-step FLOPs are matched, performance still increases with model size and memory length, though less steeply than under sample-efficiency comparisons.The authors suggest actor-step FLOPs may be poorly optimized for single-step Transformer queries and may be affected by curriculum-evaluation differences.Source: §E.4 · p. 52, Figure E.7 · p. 52, Figure E.8 · p. 52

E.5. Repeated distillation

• Repeated distillation produces additional performance gains across three generations, but successive generations show diminishing returns.The third generation was distilled from the second after 25 billion frames, corresponding to 50 billion frames of total experience including the teacher.Source: §E.5 · p. 53, Figure E.9 · p. 54

F. Human-Timescale Adaptation

• The human-timescale adaptation section provides further details regarding the paper’s claim of human-timescale adaptation.Source: F. Human-Timescale Adaptation · p. 54

F.1. Probe tasks

• Probe tasks span complex single- and multi-agent worlds with many objects and fully hidden production rules, testing exploration, manipulation, navigation, memory, hierarchy, and coordination.Examples include tool use, chained crafting, distractor-heavy exploration, object permanence, forced multi-agent coordination, information asymmetry, and repeated machine-based production.Source: §F.1 · pp. 54–58

F.2. Comparing human and agent scores on every probe task

• Figure F.1 compares AdA with 19 human players, a zero-shot baseline, and a fine-tuned agent across 30 held-out single-agent probe tasks.The zero-shot baseline tests whether adaptation sacrifices zero-shot generalisation, while the fine-tuned agent indicates the maximum reward achievable in one trial.Source: §F.2 · p. 59, Figure F.1 · p. 60

F.3. Quantifying stochasticity

• 50 repetitions reduce the maximum standard deviation of the mean task score from 0.43 with one repetition to 0.06 with 50 repetitions.Averaging repeated trials addresses variance from random avatar spawn rotations and stochastic policy behavior.Source: §F.3 · pp. 59, 61
• Across five initialization seeds, the maximum standard deviation is 0.04 for the median and 0.02 for the 20th percentile, supporting single-seed ablations.These values are reported for the 76M-parameter AdA version after accounting for task variation.Source: §F.3 · p. 61

F.4. Prompting through first-person demonstrations

• Prompting helps short navigation and tasks with up to two production rules, but longer demonstrations and solutions with three or more rules leave performance near the unprompted baseline.Long demonstrations may exceed memory context, while multi-rule tasks cause AdA to remember only a subset and continue exploring.Source: §F.4 · pp. 61–62
• Across 30 hand-authored tasks, prompting improves AdA on some tasks, often producing perfect or near-perfect scores, and harms performance in only one case.The comparison covers prompted and unprompted agents on the full probe set.Source: §F.4 · p. 61, Figure F.4 · p. 62
• Prompting effectiveness depends on world configuration: displaced distractors and subtle spatial or tool-use requirements hinder recall, whereas favorable object placement can support optimal behavior.AdA succeeds on Irreversible production despite three rules but fails to immediately infer hiding objects in Antimatter.Source: §F.4 · p. 62
• Prompts can overcome training-induced biases, as shown by higher prompted performance on Object permanence: yellow cube, where unprompted AdA tends toward the wrong black cube.The yellow-cube task requires navigating right, whereas the otherwise identical black-cube task requires navigating left.Source: §F.4 · pp. 62–63, Figure F.6 · p. 63
• Human demonstrations show the same selective prompting pattern: AdA approaches optimal performance on some tasks but fails on others, and successful demonstrations do not unlock capabilities outside its training distribution.Human prompting succeeds on Navigation: hold up high and Object permanence: yellow cube, but fails on Navigation: find the cube with teaser, Spacer tool, and Transform to transport.Source: Figure F.5 · p. 63, §F.4 · pp. 63–64