acceptodds
Under review as a conference paper at ICLR 2027

How Humans Learn Unfamiliar Worlds: A Think-Aloud Dataset on ARC-AGI-3

Abstract

Learning through exploration and using newly acquired knowledge to achieve goals are central capabilities of general intelligence. Understanding these capabilities requires tracing how learners build, use, and revise world models during interaction. We construct the first dataset that records how humans discover and use the mechanisms of novel environments, using ARC-AGI-3 as a testbed: 144 trajectories from 64 participants across 25 public games, each pairing time-synchronized actions with think-aloud audio, plus 250 trajectories from large language model (LLM) agents with and without harnesses on the same games. We annotate mechanism triggers and recognition, action purposes, and errors and recovery, linking changes in understanding to the observations and actions that support them. We analyze characteristics of human cognition and compare them with those of LLM agents. Both groups anticipate mechanisms and use familiar analogies. Qualitative representations are more common in successful human trajectories than in successful LLM trajectories (50.0% versus 14.3%). Humans more often explore unfamiliar situations without a concrete hypothesis, while hypothesis testing is common in both groups. We also trace how learners recover through belief revision, replanning, and execution correction. As LLMs with well-designed harnesses reach saturated performance on public ARC-AGI-3 games, our cognitive analyses motivate next-generation benchmarks that challenge familiar priors, require accumulated evidence and abstraction, and test planning through subgoals, experimentation, and recovery.

Primary area:
datasets and benchmarks
Keywords:
World Model Induction, Human–AI Comparison, Agent Harnesses

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_uvb4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a dataset of human think-aloud reports aligned with gameplay trajectories on ARC-AGI-3, alongside LLM agent trajectories and mechanism-level annotations. It analyzes differences in prior knowledge, world-model representations, exploration, and error recovery between humans and LLM agents.

Critical strengths
  • The dataset provides useful human think-aloud records aligned with actions and environmental feedback, enabling more fine-grained analysis of interactive learning.
Critical weaknesses
  1. The analysis is largely descriptive, and many findings are relatively unsurprising, such as humans exhibiting more hypothesis-free exploration and LLM agents requiring fewer actions after recognizing relevant mechanisms. The paper does not establish whether these behavioral differences meaningfully affect learning efficiency, generalization, or task success. Without connecting the observations to measurable outcomes ( further agent improvement, designs, with results), the resulting insights remain limited.
  2. The main behavioral comparisons rely heavily on successful trajectories, introducing potential selection bias. Failed trajectories could reveal which exploration strategies or recovery behaviors distinguish successful learning from unsuccessful attempts. Moreover, although humans and LLMs are evaluated on the same 25 games, the paper does not provide sufficiently detailed task-matched comparisons to rule out differences in the composition of successful trajectories.
  3. The paper collects trajectories from multiple LLM agent configurations but does not sufficiently investigate how specific agent designs influence the observed behaviors. VISTA, Schema, and OPINE-World incorporate different mechanisms for memory, world modeling, and hypothesis testing, yet the analysis does not clearly isolate their contributions. More systematic agent-level comparisons or ablation studies would make the findings more actionable.
Questions for authors

The experiments are restricted to 25 public ARC-AGI-3 games, whose mechanisms may be relatively familiar to pretrained LLMs. How familiar are the models with these games? How do you adequately disentangle prior knowledge from knowledge acquired through interaction.

1 comment

We thank the reviewer for the constructive feedback on the dataset’s contribution and the behavioral comparisons. We have revised the manuscript accordingly, clarifying the dataset’s research aims and adding validation and robustness analyses.

W1. Connecting behavioral differences to learning outcomes and agent improvement

The findings in this paper are initial analyses illustrating part of the dataset’s research value. Other researchers can use the rich records to ask different questions and obtain further valuable findings. The dataset has three research uses.

For cognitive science, researchers studying human learning and reasoning are a primary audience for this dataset. The aligned reports, actions, and environmental feedback let researchers examine how people form explanations, choose informative actions, and revise beliefs in unfamiliar worlds.

For LLM-agent research, the strongest agents already solve the ARC-AGI-3 benchmark tasks studied here and use fewer actions on average than humans. Given this relative saturation, the dataset was not designed to study how to build a higher-scoring agent for these tasks. Human–agent differences can nevertheless suggest useful research questions. For example, WALL-E 2.0 learns explicit symbolic rules, while human reports frequently contain qualitative representations and exploration before a concrete hypothesis is formed. These contrasts motivate questions about the precision of world models and when agents formulate or test hypotheses.

For agent benchmarks, the findings suggest challenges for a harder evaluation: adapting beyond familiar priors, accumulating evidence, constructing abstractions, and learning from failed plans.

Appendices E–G add validation and robustness checks for the reported findings.

W2. Successful-trajectory selection and task-matched comparisons

Appendix F adds game-level comparisons and includes available unsuccessful agent trajectories. The main directions persist in most games and after their inclusion. Successful-human selection remains a limitation.

W3. Effects of individual agent designs and component ablations

We have added results for individual agent configurations and their action-purpose distributions (Appendix F). The main conclusions remain consistent across harness designs. The purpose of this dataset is to study human and LLM learning and action patterns and inform research on agent benchmarks, rather than compare agent designs or prescribe how to build an agent.

Q1. Distinguishing prior knowledge from learning through interaction

We agree, and our findings are consistent with your suggestion. Agents sometimes describe a mechanism correctly before it is first triggered, and use analogies to familiar games or objects (Section 4.1). Anticipation and analogy provide observable evidence of prior knowledge in their reasoning. The aligned records also show how these expectations are checked and revised through interaction.

Reviewer_Gz8uOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper releases 144 successful human ARC-AGI-3 trajectories with synchronized think-aloud reports, alongside 250 public agent trajectories. It annotates mechanism recognition, action purposes, and recovery to compare how humans and agents explore.

Critical strengths

The aligned speech and action records should be useful beyond success-rate comparisons. The examples make the annotations understandable. I appreciate the care taken not to equate missing verbal evidence with missing knowledge. Mechanism triggers provide a useful point of alignment between what a participant says and what happens in the game. This could support narrower analyses of exploration and recovery even if the broader human and agent comparison remains descriptive.

Critical weaknesses

The human and agent records are collected differently. Humans receive repeated prompts to verbalize, while agent traces contain notes and code shaped by their harnesses. The retained human trajectories are also all successful. Appendix D acknowledges these differences, but they still limit what the headline comparison can tell us. The 50.0% versus 14.3% qualitative-description gap could partly reflect reporting rather than reasoning.

The annotation checks are thin for this comparison. Action-purpose review covers 30 segments, and there is no independent agreement estimate for qualitative descriptions or error and recovery labels. The paper is explicit about these limitations, so I do not read the result as a causal claim. Still, the interpretation should stay close to what is observable in the transcripts. The existing recognition checks do not establish reliability for these other annotation categories.

Questions for authors

Does the qualitative-description gap hold within games and within each agent harness? Could a second annotator independently label a sample of those descriptions and error or recovery events? I am asking for sensitivity checks on the collected records, rather than another round of participant recruitment.

1 comment

We thank the reviewer for the helpful comments on comparability and annotation reliability. We have revised the manuscript in response, adding independent annotation validation and analyses across games, agent configurations, and reporting-verbosity groups.

W1. Different reporting conditions and successful-trajectory selection

We acknowledge that human and agent records were collected under different prompting conditions, which can affect the comparison. We have added analyses across agent configurations (Appendix F) and reporting-verbosity groups (Appendix G). The qualitative-description difference remains across the evaluated configurations and in both verbosity groups.

We agree that the conclusions may be affected by how fully participants verbalize their thinking. We have revised the opening of Section 4 to qualify the claims accordingly.

W2. Independent validation beyond mechanism recognition

We have extended validation beyond recognition timing. Five additional PhD students familiarized themselves with the games, annotated independently, and then discussed their annotations to establish reference labels. We checked the LLM annotations against these references for qualitative descriptions, action purposes, and errors and recovery (Appendix E). The results support the reliability of these categories on both human and agent records.

Q1. Within-game and within-harness comparisons and independent annotation

We have added within-game and configuration-specific comparisons in Appendix F.2. The qualitative-description difference holds in most games and for each evaluated configuration. We have also added independent human annotation of qualitative descriptions and errors and recovery (Appendices E.1 and E.3); the validation results support the reliability of these labels.

Reviewer_H7xzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents a dataset of humans, LLM agents, and LLM agents with harnesses playing games from the ARC-AGI-3 dataset while "thinking aloud". The ARC-AGI-3 tasks require the player to learn rules, controls, and goals by interacting with the environment. Thus, the agent's thoughts (for humans, a time-aligned voice recording; for LLMs, the reasoning records) catalog the process of exploring the game, forming and revising beliefs, responding to changes, etc. The authors catalog all the game mechanisms in the underlying testbed and insert detectors for these mechanisms in the game code. They then use an LLM to annotate the agent transcripts for human and LLM trajectories. Specifically, the annotator catalogs when a mechanism is correctly identified, the purposes of agents' actions, errors, and recovery following errors. From this dataset, the authors comment on some preliminary findings, comparing human and LLM behavior. For instance, they observe that humans more often rely on qualitative representations and are more likely to explore without an explicit hypothesis. Finally, the authors discuss implications for future work in designing benchmarks for interactive world model learning.

Critical strengths
  1. The dataset seems valuable and novel. Pairing the agents thought process during exploration and learning could provide pathways for many questions about how beliefs, goals, and plans are formed and revised. These are central and pressing questions.
  2. The unpacking of action purposes (pursuing a goal, collecting info, both) and further dissection of information-collecting actions (exploration, hypothesis testing, investigating unexpected outcomes) is particularly clean and useful. These categories are of natural interest and useful for later work that might utilize this dataset.
Critical weaknesses

Main points

  1. The data depends on agents immediately and thoroughly verbalizing their thoughts. This is understandable and hard to avoid, but I think the authors should be careful about over-claiming here. For instance, there may be a delay between an agent understanding a mechanism and them voicing that understanding. Likewise, the authors assert that "humans more often explore without a prior hypothesis", but it seems likely that human participants may have hypotheses or beliefs that are not vocalized. The authors even point this out in Section 4.4 ("Human concurrent verbal reports may not express every change in understanding"). Accordingly, they should be more measured about the claims made from the data.
  2. The analyses in section 4 (in particular those of 4.1 and 4.2, but also the others) are a bit shallow. They identify interesting phenomena but stop short of studying the impact of those phenomena on agent outcomes, learning, etc. The ability to do the latter seems to be one of the main strengths of a dataset like the one presented. For instance, do qualitative representations speed up learning, improve outcomes, decrease exploration without hypotheses, etc.? How do the different kinds of exploring affect how fast the agent learns and how they perform? For these questions, I also think the failed games (which are excluded from the data) could actually be especially useful and interesting.

Minor points

  1. The harnesses (and possibly the models themselves) have been designed for ARC-AGI-3. This might qualify certain human-LLM comparisons.
  2. Section 5 (Implications for Benchmark Design) seems a bit disconnected from the paper's main results and findings. I would suggest either abridging the section (which could free up room for the points above) or more clearly delineating how the three outlined directions follow from this work.
  3. Table 1 extends into the margin.
Questions for authors
  1. See questions in weakness 2
  2. Is it possible to infer recognition of a mechanism from agent behavior rather than from the transcript alone? If this is feasible (perhaps even only in specific instances), this could be a nice check on the reliance on verbalization pointed out in weakness 1.
  3. How much variance was there in the verbosity of different agents (humans and LLMs)? How do differences in baseline verbosity affect your different measures (e.g. recognition times, exploring without hypotheses, etc.)?
  4. In Section 4.3 you give the breakdown of human action purposes, but the three categories do not sum to 100%. What happened in the remaining ~15% of human actions? Was the purpose not identifiable? Were these execution errors?
2 comments

We thank the reviewer for the thoughtful comments on the interpretation and scope of our findings. We have revised the manuscript accordingly, adding supporting analyses, qualifying the conclusions, and clarifying how the benchmark design suggestions follow from the findings.

Ethics: AI use statement

We have expanded the AI Use Statement to include AI-based transcription, annotation and detector implementation, analysis code, and language polishing and translation. The statement also points to the human validation procedures.

W1. Qualifying claims based on verbalized thoughts

We agree that the conclusions may be affected by how fully participants verbalize their thinking. We have revised the opening of Section 4 to qualify the claims accordingly. Our annotations combine verbal reports with actions and observed outcomes, as suggested in your Question 2, and record recognition and action purposes only when the available evidence is sufficient (Appendices B.3–B.4).

W2. Depth of analysis and links to learning and task outcomes

The findings in this paper are initial analyses illustrating part of the dataset’s research value. Other researchers can use the rich records to ask different questions and obtain further valuable findings. The dataset has three research uses.

For cognitive science, researchers studying human learning and reasoning are a primary audience for this dataset. The aligned reports, actions, and environmental feedback let researchers examine how people form explanations, choose informative actions, and revise beliefs in unfamiliar worlds.

For LLM-agent research, the strongest agents already solve the ARC-AGI-3 benchmark tasks studied here and use fewer actions on average than humans. Given this relative saturation, the dataset was not designed to study how to build a higher-scoring agent for these tasks. Human–agent differences can nevertheless suggest useful research questions. For example, WALL-E 2.0 learns explicit symbolic rules, while human reports frequently contain qualitative representations and exploration before a concrete hypothesis is formed. These contrasts motivate questions about the precision of world models and when agents formulate or test hypotheses.

For agent benchmarks, the observations motivate more challenging tasks that require unfamiliar learning, accumulated evidence, abstraction, and recovery through repeated attempts. Our response to Weakness 4 explains how these design directions follow from the findings.

We have also included unsuccessful agent records in the added robustness analyses (Appendix F).

W3. Effects of ARC-specific models and harnesses on the comparison

We agree that models may benefit from familiarity with ARC-like games or task-specific development. Agents sometimes describe a rule correctly before it first takes effect and use analogies to familiar games; both observations are consistent with strong prior knowledge. Harness design also influences behavior and logs. Our study concerns ARC-AGI-3, and the added configuration-level analyses (Appendix F) show that the main conclusions remain consistent across the evaluated harnesses.

W4. Connecting benchmark design suggestions to the findings

Beyond its value for cognitive science, the dataset can inform a next generation of agent benchmarks that are more challenging than ARC-AGI-3. The connection to our findings is as follows.

We find that agents often recognize mechanisms before their first trigger. A benchmark could therefore introduce counterfactual rules and unfamiliar mechanisms whose effects must be discovered through testing and observation, creating a challenge for learning when existing expectations are unreliable.

We observe rapid mechanism inference by agents and more frequent qualitative representations in human reports. A benchmark could introduce stochastic effects, hidden states, and delayed consequences that require accumulated evidence. Environments too complex for exhaustive precise modeling could also test the construction of useful abstractions.

We observe short post-recognition action distances for agents and recovery from failed plans in human records. Tasks with deeper subgoal dependencies and uncertain combinations of local mechanisms could therefore require repeated attempts, learning from failed attempts, and revision of longer plans.

W5. Table 1 exceeds the text width

We have adjusted Table 1 to fit within the text width and checked the compiled PDF.

Q1. Questions about learning and task outcomes (see W2)

Please see our response to W2.

Q2. Inferring mechanism recognition from behavior

Yes. Our current annotations combine the aligned reports, action sequences, and observed outcomes (Appendix B.3). For example, in BP35 a participant encounters a blocked horizontal route, removes the obstructing green support, and then moves through the opened cell. The contrast between the blocked and successful attempts supports recognition of the obstruction rule, even without a complete verbal statement of that rule.

Q3. Effects of verbosity on the reported measures

We have added a verbosity analysis using the same successful trajectories as the main comparisons (Appendix G). Human records are grouped by spoken words per minute, and agent records by public reasoning or report words per action, separately within each harness configuration. The main differences in recognition timing, qualitative descriptions, and exploration without a concrete hypothesis remain in both verbosity groups, although some absolute measurements change.

Q4. Unclassified human action purposes

We have revised Section 4.3 to explicitly explain that these are mainly actions whose purpose could not be identified from the available evidence.

Reviewer_zdujOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how humans learn the rules and goals of unfamiliar games by recording their actions and spoken thoughts. It introduces a dataset of these recordings and compares them with LLM agents, focusing on exploration, rule discovery, and error correction. The authors check samples of the annotations manually and report differences in how humans and LLMs approach the games.

Critical strengths
  • The dataset pairs human actions with spoken thoughts, making it easier to examine why players explore, revise a rule, or change a plan. The aligned records and LLM trajectories provide useful material for studying learning processes beyond task success alone.
Critical weaknesses
  • Insufficient validation of cognitive annotations. Key labels, such as when a learner understands a mechanism and why an action is taken, require subjective interpretation. Although annotation procedures and manual spot checks are provided, the paper does not report independent inter-annotator agreement or establish that these labels are comparable across human spoken reports and agent logs.

  • Limited evidence for the robustness of the main findings. The comparisons rely mainly on aggregate percentages and selected examples. The paper does not show whether the findings hold across different games and agents. It also does not test how selecting successful trajectories or missing recognition evidence affects the results.

Questions for authors

None

1 comment

We thank the reviewer for the careful reading and constructive feedback. We have revised the manuscript in response to your comments, adding further annotation validation and robustness analyses.

W1. Reliability of cognitive annotations for humans and agents

We have conducted further validation with five additional PhD students (Appendix E). After familiarizing themselves with the games, they first annotated the records independently, then discussed their annotations to reach a consensus reference. We used these references to evaluate the LLM annotations across the main annotation categories, separately for human and agent trajectories. The results support the reliability of these labels on both sources.

W2. Robustness across games, agents, trajectory selection, and missing recognition evidence

We have added robustness checks in Appendix F for three findings: agents more often recognize mechanisms before their first trigger; human reports more often contain qualitative descriptions; and humans more often explore without a concrete hypothesis.

  1. Across games: we recomputed the comparisons within each game. These three findings hold in most games, with individual exceptions reported.

  2. Across agents: we report each configuration separately. All three findings hold across the ten configurations, including direct models, VISTA, Schema, OPINE-World, and Astra.

  3. Success selection: we added the available unsuccessful agent trajectories and compared the results with those using successful trajectories alone. All three findings remain.

We also acknowledge that recognition evidence is incomplete. Because we collect the data through think-aloud, participants do not always verbalize their thinking, and recognition time cannot be identified for every mechanism instance. This is an unavoidable shortcoming of this form of data collection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.