What is HAI-Co²?
HAI-Co² (Human-AI Co-Construction) is a cross-domain reference architecture for cooperative problem solving between a domain expert and an LLM agent. Drawing on Dutta et al. (2025), it reframes generative AI as an equal partner that co-edits a set of persistent, structured artifacts with the user, rather than a one-shot autocompletion engine.
The “²” in the name marks the dual nature of co-construction: both the solution and the objective are shaped together. The collaborators don’t just decide how to reach a fixed goal, they also learn together what the goal should be.
A third mode, between automation and assistance
Two framings dominate today’s use of generative AI. In full automation, the model is expected to solve complex problems on its own. In pure assistance (or human-in-the-loop), the AI is a tool that the human directs and corrects. Both fall short for hard, expert-domain problems: current models are not reliable enough to act alone, and using them only as assistants underuses their strengths while overloading the expert with evaluation work.
HAI-Co² stakes out a third position: equal-partner co-construction. The human brings domain context, requirements, judgement and aesthetic intuition. The agent brings vast recall, fast exploration of alternatives, tool use and consistency. Neither party can solve the problem well alone, and the architecture is designed for exactly that case.
Why co-construction?
Complex problems in expert domains rarely fit a single prompt. They are solved by iterative refinement across multiple levels of abstraction, under preferences that evolve as the collaborators understand the problem better. Requirements get revised; trade-offs become visible only once a draft exists.
HAI-Co² makes this explicit. A persistent set of structured artifacts (preferences, todos, documents and, where relevant, domain-specific structures) sits between the user and the agent. Both parties read and modify them; the chat is one of several ways to interact alongside direct edits, rankings and categorical choices.
The result is closer to how human teams actually work: a search over shared drafts, with feedback flowing in both directions and partial work treated as a normal state rather than an error.
Four defining characteristics
- Complex problems that need active human participation. Solutions are non-unique, modular and lack a fully automatic verifier (writing software, building an ML pipeline, drafting a related-work section, triaging vulnerabilities).
- Co-construction of the solution. The artifact emerges from joint, iterative search across hierarchical levels of abstraction, rather than from a single generation step.
- Co-construction of the objective. Preferences are elicited and revised through interaction. The utility function over candidate solutions is itself learned, not assumed.
- Natural language as the main medium. Dialogue is what makes equal partnership possible at every level of abstraction, complemented by direct edits, rankings and categorical choices.
How it differs from related framings
| Paradigm | Role of the human | Treatment of the objective |
|---|---|---|
| Full automation | Specifier; absent during solving | Fixed up front |
| Human-in-the-loop / assistance | Oracle, labeler, corrector | Fixed; the human guards it |
| RLHF / assistance games | Source of preference labels | Static but unknown; inferred |
| HAI-Co² | Equal partner with complementary strengths | Co-constructed and revisable |
Ethics by partnership
Letting the AI participate in shaping the objective creates a real risk: the agent could steer preferences toward outcomes it finds easier to satisfy. HAI-Co² takes this seriously. Every preference change is explicit and visible; the user’s preferences prevail when they conflict with the agent’s; and the architecture is designed to allow independent monitor agents to flag suspicious shifts.
The flip side is that co-construction is itself part of the answer to alignment. Continuous human involvement keeps decisions in the open, the human understands the artifacts’ internal structure, and course corrections are possible at any level of abstraction.
Application domains
Related-Work Generation
Iteratively co-construct a related-work section grounded in a curated set of papers and a semantic graph of their contributions.
arXiv:2508.07955Chart Generation
Co-construct a chart specification (encodings, palette, annotations) with hard validity constraints over the underlying data.
CVE Retrieval & Triage
Cooperate with the analyst to retrieve, rank and summarise CVEs relevant to a given system or context.
Implementation
The HAI-Co² framework was originally introduced by Dutta et al. (2025). This reference implementation, the open-source codebase and the Co-construction Studio are developed and maintained by Petros Raptopoulos, PhD researcher at TU Darmstadt’s UKP Lab.
Collaborators & supervisors
HAI-Co² builds on conversations with, and the support of, the following people.
Reference
Dutta, S. et al. (2025). Problem Solving Through Human-AI Preference-based Cooperation. Computational Linguistics, Volume 51, Number 4.