🤖 AI

Humans' ALMANAC: A Human Collaboration Dataset of Action-Level Mental Model Annotations for Agent Collaboration

The paper introduces ALMANAC, a dataset of 2,987 action-level mental model annotations derived from human dyadic collaboration on the Map Task, designed to bridge the gap in training and evaluating LLM agents' ability to maintain aligned mental models and simulate human collaborative behaviors.

Original authors: Jiaju Chen, Yuxuan Lu, Jiayi Su, Chaoran Chen, Songlin Xiao, Zheng Zhang, Yun Wang, Yunyao Li, Jian Zhao, Tongshuang Wu, Toby Jia-Jun Li, Dakuo Wang, Bingsheng Yao

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiaju Chen, Yuxuan Lu, Jiayi Su, Chaoran Chen, Songlin Xiao, Zheng Zhang, Yun Wang, Yunyao Li, Jian Zhao, Tongshuang Wu, Toby Jia-Jun Li, Dakuo Wang, Bingsheng Yao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to draw a picture of a treasure map with a friend, but you are in different rooms. You can only talk to each other through a text chat. One person (the Guide) has the original map with the correct path drawn on it. The other person (the Follower) has a blank map with just the landmarks (like a mountain or a bridge) but no path.

To succeed, they have to work together: the Guide describes the path, and the Follower tries to draw it.

This is the core of a new research project called ALMANAC. Here is the simple breakdown of what the researchers did and why it matters, using everyday analogies.

1. The Problem: Robots are Good at "Doing," but Bad at "Thinking Together"

Currently, AI agents (like advanced chatbots) are great at following orders. If you say, "Book a flight," they do it. But real human collaboration isn't just about following orders; it's about mental alignment.

Think of a human team like a jazz band. They aren't just reading sheet music; they are constantly listening to each other, guessing what the other musician is about to play, and adjusting their own rhythm in real-time. They have a "shared mental model."

Self-reasoning: "Why am I playing this note?"
Partner intent: "What is my bandmate trying to do?"
Team goal: "Are we still trying to play a jazz tune, or did we drift into rock?"

The paper argues that current AI is like a robot that only knows how to play its own instrument perfectly but has no idea what the rest of the band is thinking. Existing datasets only record what people said and what they did, but they miss the why inside their heads.

2. The Solution: ALMANAC (The "Mind-Reading" Dataset)

The researchers built a new dataset called ALMANAC to fix this. They took the "Map Task" described above and added a special layer of observation.

How they collected the data:

The Game: 50 people played the Map Task in pairs (25 pairs total).
The "Checkpoints": Every time the pair reached 25%, 50%, and 75% of the task, the game paused briefly. The players had to quickly speak into a microphone and answer: "What do you think our goal is right now?" "What do you think your partner is trying to do?" and "Why did you just do that?"
The Replay: After the game, the players watched a replay of their own actions. For every single move they made (drawing a line, erasing a mistake, sending a message), they had to label their mental state again.

The Result:
They ended up with 2,987 specific actions, and for every single one, they have a record of:

The action taken (e.g., "Drew a line").
The Team Goal (e.g., "We are trying to connect the bridge to the mountain").
The Partner Intent (e.g., "I think my partner is confused about the direction").
The Self-Reasoning (e.g., "I drew this because I thought the mountain was to the left").
A Rationale (a free-form explanation of their thought process).

3. The Experiment: Can AI "Read Minds"?

The researchers tested six different AI models (including big ones like GPT-5.5 and Claude) to see if they could use this new dataset to act like a human collaborator. They gave the AI two jobs:

Predict the Next Move: "Based on what happened so far, what will the human do next?"
Predict the Mental State: "Based on what happened so far, what is the human thinking right now?"

The Findings:

The "What" is easy, the "Why" is hard: The AI models were pretty good at predicting the next action (e.g., "The human will probably draw a line").
The "Mind" is tricky: The models struggled to predict the internal thoughts (the mental models). They could guess the shared goal (e.g., "We are building a bridge") fairly well, but they were terrible at guessing the private, personal reasoning of the human (e.g., "I'm drawing this because I'm nervous").
The Role Matters:
- Guides (who give instructions) were harder for the AI to predict mentally because their thoughts involve complex spatial planning that isn't always spoken out loud.
- Followers (who draw) were easier to predict mentally because their thoughts are directly shaped by the Guide's explicit instructions.
Training Helps: When they "taught" (fine-tuned) a smaller AI model using this specific dataset, it got much better at simulating human behavior, almost catching up to the giant, expensive models.

4. The Bottom Line

The paper concludes that to make AI true collaborators, we can't just train them on how to finish tasks. We need to train them on how humans think while working together.

ALMANAC is the first dataset that pairs human actions with these "action-level mental models." It shows that while AI is getting better at simulating human behavior, it still struggles to understand the invisible, internal reasoning that makes human teamwork work. The dataset provides the "training wheels" needed to teach AI how to align its mental model with a human's, rather than just blindly following orders.

Technical Summary: ALMANAC

Problem Statement

Recent advances in Large Language Model (LLM) agents have endowed them with complex cognitive capabilities, such as multi-step reasoning and tool use, positioning them as potential human collaborators. However, effective collaboration requires more than task completion; it necessitates the continuous maintenance and alignment of mental models—specifically, self-reasoning about one's own actions, perceived partner intent, and understanding of shared team goals.

Current human-agent collaboration research and benchmarks (e.g., ToolBench, WebArena, MultiAgentBench) primarily evaluate agents on task-directed exchanges and independent task completion. They lack authentic human collaboration data that includes action-level mental model annotations. Existing human-human datasets capture observable interactions (dialogues, outcomes) but omit the critical cognitive content underlying collaborative behaviors. Consequently, the community lacks the data necessary to guide agents toward process-level collaborative competence and the ability to simulate human collaborative behaviors and infer underlying mental states.

Methodology

Dataset Construction: ALMANAC

To address this gap, the authors present ALMANAC (Action-Level Mental model ANnotations for Agent Collaboration), a dataset derived from the Map Task, a classic dyadic routing task from social science.

Task Design: Two participants (a Guide and a Follower) collaborate to reproduce a route on a map via text-based communication and workspace actions (drawing, erasing). The Guide sees the full route; the Follower sees only landmarks.
Experimental Conditions: The study manipulates canvas visibility as a between-subjects factor:
- $C_{not\_visible}$ : The Guide cannot see the Follower's real-time drawing.
- $C_{visible}$ : The Guide can view the Follower's live canvas.
Annotation Framework: Grounded in theories of teamwork processes, situation awareness, and common ground, the framework captures mental models at the action level through a two-step process:
1. In-Session Checkpoints: At 25%, 50%, and 75% of task progress, participants provide brief, real-time reports on team goals, partner intent, and self-reasoning via voice recording.
2. Post-Session Retrospective Labeling: Immediately after the task, participants review their action trajectory and screenshots. For every action, they provide:
  - Free-form rationale: Explaining the cognitive process behind the action.
  - Structured labels: Categorical selections for Team Goal, Perceived Partner Intent, and Self-Reasoning, along with an alignment status indicator.
Data Scale: The dataset comprises 2,987 collaboration actions from 50 participants across 25 dyadic sessions. Actions are standardized into discrete grid representations to facilitate LLM processing without relying on image comprehension.

Benchmark Experiments

The authors evaluated six LLMs (four open-source/proprietary: Qwen3.6-35B, Llama 3.3 70B, GPT-5.5, Claude 4.6 Sonnet; and two fine-tuned: Qwen3-4B, Qwen3-30B-A3B) on two tasks:

Next-Turn Behavior Prediction: Predicting the next action (message, draw, erase, etc.) given interaction history and persona.
Mental Model Prediction: Predicting the participant's mental state (team goal, partner intent, self-reasoning, rationale) for the next turn.

Experiments compared standard prompting, fine-tuning on ALMANAC, and the inclusion of ground-truth mental model annotations as input signals.

Key Contributions

ALMANAC Dataset: The first human collaboration dataset pairing authentic collaborative behaviors with theory-informed, action-level mental model annotations.
Annotation Framework: A novel design combining in-session checkpoints with post-session retrospective labeling to capture dynamic mental models and free-form rationales at the granularity of individual actions.
Empirical Benchmarking: A comprehensive evaluation of six LLMs demonstrating that while mental model annotations provide useful signals for predicting collaborative behavior, current models struggle to infer private reasoning processes.

Results

Next-Turn Behavior Prediction

Role Asymmetry: Predicting the Guide's next action (limited to messages) was trivial (100% accuracy), whereas predicting the Follower's action (involving drawing, erasing, and interpreting) was significantly harder.
Visibility Impact: Models generally performed better in the $C_{not\_visible}$ condition than in $C_{visible}$ . The authors attribute this to the higher behavioral variability and reliance on visually grounded corrections in the visible condition, which are harder to infer from text history alone.
Impact of Mental Models: Adding ground-truth mental model annotations as input generally improved performance in predicting Followers' actions, particularly for drawing accuracy in the $C_{not\_visible}$ condition. This suggests Followers' behaviors are more grounded in latent reasoning states than Guides' actions.
Fine-Tuning: Smaller models fine-tuned on ALMANAC (Qwen3-4B) achieved performance comparable to larger proprietary models, indicating that targeted supervision on collaboration-specific data can close the gap with larger models.

Mental Model Prediction

Predictability Hierarchy: Team Goal and Partner Intent were significantly easier to predict than Self-Reasoning. This aligns with the observation that shared components are often reflected in dialogue and progress, whereas self-reasoning captures private motivations not explicitly expressed.
Role Asymmetry: Follower mental models were easier to predict than Guide mental models. Guides' reasoning involves complex spatial planning and partner monitoring that is difficult to infer from interaction history alone.
Fine-Tuning Efficacy: Fine-tuned models (specifically Qwen3-4B) showed superior performance in predicting Followers' mental models and rationales compared to prompt-based models, suggesting ALMANAC provides effective supervision for learning collaboration-relevant reasoning.
Limitations: No single model consistently outperformed others across all conditions. The low accuracy in self-reasoning prediction highlights the challenge of inferring private mental models.

Significance and Claims

The paper claims that ALMANAC provides a foundational resource for developing LLM agents that function as genuine collaborative partners rather than mere task solvers.

Diagnostic Value: The results indicate that while mental model annotations offer signals beyond interaction history, current LLMs do not reliably leverage them to infer private reasoning. The gap between predicting shared states (team goal) and private states (self-reasoning) highlights a fundamental limitation in current models' ability to simulate the cognitive processes driving effective collaboration.
Process-Level Competence: By grounding evaluation in authentic human data with theory-informed annotations, ALMANAC opens a pathway toward agents that can maintain and align mental models throughout a collaboration, addressing a critical deficit in current agent benchmarks that focus solely on task completion.
Modest Scope: The authors acknowledge limitations, including the reliance on retrospective reporting (susceptible to recall bias), the modest dataset size compared to large-scale NLP benchmarks, and the restriction to a single task domain (Map Task). They note that future work is needed to extend these findings to diverse, real-world domains and to explore multimodal approaches for better spatial understanding.