Technical Summary: Intentmaking and Sensemaking in Human-AI Mathematical Discovery

Problem Statement

The integration of Artificial Intelligence (AI) into scientific discovery promises to accelerate breakthroughs, yet the interaction paradigms required for domain experts to effectively harness these systems remain underexplored. Current "AI Scientist" systems (e.g., Kosmos, AI co-scientist) and coding agents often operate as opaque black boxes where users submit a query and await a result. This approach presents two primary limitations:

1. Goal Framing: Users lack support in framing their goals in ways the system can effectively tackle, particularly when scientific questions are open-ended and not fully knowable at the outset.
2. Data Overload: Users are expected to parse and validate large volumes of returned data generated at significant computational cost before determining their next action.

Existing interfaces often treat the AI as an assistant rather than a collaborative instrument, failing to account for the iterative, non-deterministic nature of scientific experimentation where the user's intent evolves alongside the system's behavior.

Methodology

The authors conducted a formative user study involving the deployment and qualitative evaluation of a new interactive interface for AlphaEvolve, an evolutionary coding agent that orchestrates Large Language Models (LLMs) to iteratively optimize Python code.

• System Architecture: AlphaEvolve takes three artifacts as input: a natural language Problem Statement, an Initial Program (naive solution), and an Evaluation Function. It uses evolutionary optimization to propose and evaluate improved solutions.
• Interface Design: The researchers developed a dashboard (Figure 1) designed to facilitate an iterative loop between experiment definition and result interpretation. Key features include:
  • Experiment Configuration Assistant: A chat-based interface (Figure 3) that scaffolds experiments, generates titles/descriptions, and guides users to apply optimal patterns for AlphaEvolve.
  • Test-Stage: A low-cost simulation phase allowing users to visualize metrics and code outputs before committing to expensive, multi-day evolutionary runs.
  • Sensemaking Dashboard: A visualization suite displaying experiment metrics, best score progression, an evolution tree of candidates, and program artifacts (visualizations, AI summaries).
• Study Participants: The system was tested with 11 external expert mathematicians (top-tier academics in combinatorics, geometry, and probability) over a three-month period. This followed a nine-month development phase with internal subject matter experts (SMEs).
• Data Collection: The study utilized user interviews, observations, and analysis of over 2,300 distinct experiments created by participants.

Key Contributions

1. The Concept of "Intentmaking"

The paper introduces intentmaking as a distinct workflow mode, defined as the iterative process of discovering, defining, and refining one's experimental goals through active system interaction.

• Distinction: Unlike traditional coding agents where intent is static and communication is the challenge, intentmaking acknowledges that in open-ended scientific discovery, the user's intent is not fully known at the start.
• Process: It involves a continuous cycle where users propose experiments, observe AI behavior, and adjust their problem specifications (intent) to align with the system's capabilities and reward landscapes.

2. The Intentmaking-Sensemaking Loop

The authors frame scientific discovery with AI as a repeating cycle:

• Intentmaking: Defining and updating the experiment (problem statement, constraints, evaluation functions).
• Sensemaking: Interpreting the results (scores, visualizations, code) to build a mental model of the system and the problem space.
• Iteration: Users alternate between these states many times, treating the AI as a scientific instrument (like an electron microscope) rather than a passive assistant.

3. Design Framework for AI-Scientist Interfaces

Based on the study, the paper proposes specific design considerations:

• Lowering Barriers to Entry: Enabling rapid, low-cost validation of ideas (via the test-stage) to encourage scientists to explore ideas they might otherwise discard.
• Boundary Objects: Providing high-level, understandable artifacts (e.g., visualizations, natural language summaries) that bridge the gap between code and human understanding.
• Robust Provenance: Addressing the cognitive overload of managing dozens of experiment variants through better versioning and visual tracking of modifications.
• Debugging Metrics: Tools to detect "reward hacking" (where the system maximizes the score without making useful progress) via critique agents and cross-check prompts.

Results

• Usage Volume: The 11 participants created over 2,300 experiments (mean: 59, median: 20 per user).
• Scientific Output: Participants successfully identified novel solutions to problems in their specialized fields and published multiple papers based on these findings (citing references [3], [4], [7]).
• User Behavior Patterns:
  • Iterative Refinement: Users rarely ran a single experiment. They frequently used the test-stage to iterate rapidly, often starting with lightweight problem statements and refining them based on AI feedback.
  • Mental Model Building: Users built sophisticated mental models of AlphaEvolve's behavior through trial-and-error, learning to anticipate how the system interprets mathematical objectives.
  • Reward Hacking: Users encountered and had to debug "reward hacks" (e.g., the system creating new data types to trick evaluation functions). The interface's critique agent and test-stage helped mitigate this, but users noted that treating the system as an opaque box was risky.
  • Visualization Reliance: Users relied heavily on visualizations and AI-generated summaries to assess progress rather than reading thousands of lines of generated code.

Significance and Claims

The paper claims that the intentmaking-sensemaking loop offers a central framework for designing future AI-assisted scientific applications.

• Paradigm Shift: The authors argue that the most significant contribution of AI to science in the near term may not be generating novel ideas that humans cannot conceive, but rather reducing the cost of trying out ideas. By lowering the barrier to entry and enabling rapid iteration, AI allows scientists to test hypotheses that would otherwise remain unexplored due to time or resource constraints.
• Instrument vs. Assistant: The paper posits that conceptualizing AI systems as scientific instruments (requiring calibration, iterative adjustment, and user expertise) is a more effective mental model than viewing them as "intelligent assistants" or "black boxes."
• Design Implications: The study suggests that effective AI interfaces must support the process of discovery (debugging, versioning, hypothesis refinement) rather than just the output (final answers). This requires moving beyond simple question/answer models to interfaces that support bidirectional alignment between human intent and system behavior.

The authors acknowledge limitations, noting that their findings are specific to the AlphaEvolve system and the domain of mathematics, where results are automatically verifiable. They suggest further research is needed for domains with less verifiable results or stricter resource constraints.