Human-AI Co-discovery (HACO)

• HACO is a human-AI co-discovery system that embeds active AI agents within the scientific method to generate and test hypotheses.
• It leverages structured workflows, including literature synthesis and code generation, to streamline experimental design and validation.
• Empirical systems like Denario and MaskGXT demonstrate HACO’s potential to advance science while ensuring human interpretability and control.

Human-AI Co-discovery System (HACO) denotes a class of research systems in which humans and agentic AI participate jointly in scientific discovery, with AI embedded within the scientific method rather than deployed as a passive prompt-response tool. In this formulation, AI systems perform literature synthesis, code generation, data analysis, hypothesis proposal, model criticism, and experimental or simulation planning, while humans retain leadership in problem framing, interpretation, ethical judgment, and institutional accountability. The term is used explicitly for scientific algorithm discovery in crystal structure prediction, where a HACO system identified and refined a masked generative approach that became MaskGXT; related work frames the same paradigm more generally as multi-agent “AI scientists,” co-evolving epistemic partnerships, and human-centered collaboration architectures (Seong et al., 22 Jun 2026, Jimenez et al., 22 Jun 2026, Lin, 6 May 2025, Gao et al., 28 May 2025).

1. Conceptual foundations and epistemic status

A HACO system is grounded in the view that “AI scientists” are multi-agent, agentic AI systems that are “embedded within the scientific method” and act as epistemic actors rather than as generic assistants. In this conception, AI is not restricted to answering prompts; it uses tools, maintains state, plans, and executes sequences of actions across “literature synthesis, code generation, data analysis, hypothesis proposal, and model criticism.” The scientific cycle is correspondingly reorganized into a mixed workflow: human-led question framing, AI-led literature and data landscape mapping, multi-agent hypothesis and model proposal, AI-coordinated experiment or simulation design, model criticism and replication attempts, and finally human-led interpretation and communication. A common misconception is that HACO implies replacement of the scientist. The central formulation is the opposite: AI expands the hypothesis-generating and verification capacity of science, while humans remain responsible for meaning, direction, and ethics (Jimenez et al., 22 Jun 2026).

A second conceptual strand treats HACO as a co-evolving epistemic partnership rather than a static tool configuration. In the Cognitio Emergens framework, authority can appear in Directed, Contributory, and Partnership modes, and collaborations oscillate among them rather than moving linearly from tool to teammate. The same framework organizes capability along six “Epistemic Dimensions”: Divergent Intelligence, Interpretive Intelligence, Connective Intelligence, Synthesis Intelligence, Anticipatory Intelligence, and Axiological Intelligence. This reframes HACO as a system whose quality depends not only on task performance but also on how authority is distributed, how interpretability is maintained, and how risks such as “epistemic alienation” and “epistemic closure” are controlled (Lin, 6 May 2025).

A third line of work describes long-horizon human-AI interaction as co-learning, defined as the ability of human and AI to “interact and learn from/with, and grow with their collaborator.” Its three core characteristics are mutual understanding, mutual benefits, and mutual growth. In HACO terms, this means that both participants update their mental models over time: the human learns where the AI is strong or brittle, and the AI learns from corrections, annotations, and strategic guidance. This view is especially important for discovery settings in which the object of collaboration is not a single answer but an evolving research program (Huang et al., 2019).

2. Architectural organization and formal models

The canonical HACO architecture is a multi-agent “algorithmic research group.” One agent proposes models or interpretations; a critic agent searches for inconsistencies, overfitting, or violated constraints; a verification agent instantiates models and tests predictions against data; and a controller or reasoning layer orchestrates the workflow, selects tools, and maintains the state of the research program. These agents are scaffolded by a shared tool layer that can include literature search APIs, code execution environments, dataset interfaces, and visualization libraries. Denario is the best-developed prototype of this pattern: it “integrates automated literature exploration, code generation and execution, dataset analysis, and iterative hypothesis testing,” with cooperating agents communicating through structured prompts while a central reasoning layer coordinates the workflow (Jimenez et al., 22 Jun 2026).

A more formal account casts HACO as search in a construction space $\mathcal{X}$ of candidate artifacts, hypotheses, or workflows. In HAICo2, this space is organized hierarchically as $$\mathcal{X}_0, \mathcal{X}_1, \ldots, \mathcal{X}_J$$, with surjective abstraction maps

$f_j : \mathcal{X}_j \rightarrow \mathcal{X}_{j-1},$

and refinements

$$f_j^{(-1)}(x') = \{ x \in \mathcal{X}_{j} \mid f_j(x) = x' \}.$$

Candidate quality is represented by a latent multi-dimensional utility

$u(x) = (u_1(x), \ldots , u_K(x))^\top$

and an aggregated scalar utility $U(x)$. This formalization makes explicit that a HACO system is not merely generating outputs; it is conducting iterative, preference-guided search over structured scientific objects at multiple abstraction levels (Dutta et al., 2024).

An even more explicitly inferential perspective appears in co-creative learning via symbol emergence. There, each participant holds a local belief $q_t^m(s)$ over shared symbols $s$, and the collective belief is

$q_t(s) \propto \prod_{m \in \mathcal{A}} q_t^m(s).$

The associated collective free energy is

$$\mathcal{F}_t = - \log p(X^{\text{AI}}, X^{\text{Human}}) + \mathrm{KL}\!\left(q_t(s) \,\big\|\, p(s \mid X^{\text{AI}}, X^{\text{Human}})\right).$$

This suggests a formal reading of HACO as decentralized inference over shared external representations: the human and the AI do not merely exchange answers, but progressively align on a jointly useful hypothesis system (Okumura et al., 18 Jun 2025).

3. Human roles, communication, and co-learning

Across the literature, HACO is explicitly human-centered. Human scientists retain problem selection and framing, interpretation and theory building, and ethical and epistemic oversight. One influential distinction is between AI’s ability to explore “the adjacent possible” and the human capacity to “create new possibility spaces,” “reframe a problem conceptually,” and make “imaginative leaps.” The risk, on this view, is not chiefly malicious AI but “generating papers without understanding them, outsourcing judgment without noticing it is being outsourced.” HACO therefore places human checkpoints around high-consequence actions, acceptance of robust findings, and publication (Jimenez et al., 22 Jun 2026).

Human-centered collaboration work sharpens this requirement into the principle of human-led ultimate control. In the HCHAC framework, humans hold final decision rights and strategic oversight, while AI is designed to empower humans through complementary capabilities. The architecture is modeled as a joint cognitive system with a value-alignment channel between human and AI mental models, shared responsibilities in decision-making and planning, and explicit control-transfer mechanisms such as handover, takeover, and override. For HACO, the implication is that adjustable autonomy is not a peripheral safety feature; it is part of the basic scientific workflow (Gao et al., 28 May 2025).

Communication design strongly shapes whether AI is experienced as a tool or a partner. A study of 38 participants comparing a baseline co-creative system with a communicating variant found that the communicating AI produced a significantly higher collaboration score ($$\mathcal{X}_0, \mathcal{X}_1, \ldots, \mathcal{X}_J$$0), and 30 out of 38 participants said that the experience felt more like collaborating with a partner. The difference came not from changed capability—both systems had the same underlying ability—but from AI-to-human communication through speech, text, feedback requests, and an affective avatar. For HACO, this is highly consequential: two-way communication can increase perceived reliability, intelligence, and personalness, but it can also inflate trust and encourage over-reliance if interface cues outrun actual competence (Rezwana et al., 2022).

The co-learning literature adds that collaboration quality depends on shared mental models. Mutual understanding means that each participant can anticipate the other’s strengths and failures; mutual benefit means that their asymmetries become productive rather than destabilizing; and mutual growth means that both update strategies through repeated interaction. A HACO system that does not support this longitudinal alignment is likely to remain either brittle or superficial, regardless of its raw generation capability (Huang et al., 2019).

4. Representations, interaction media, and discovery workflows

One recurring design pattern in HACO-adjacent systems is the replacement of disposable prompts with persistent, human-readable structures. TombWriter’s “story instrument” is hierarchical, persistent, and executable: premise, characters, scenes, beats, and prose are maintained as explicit objects rather than transient prompts. Although developed for co-writing, its deeper relevance lies in the separation of structure from surface realization and the use of iterative, beat-level refinement to discover latent possibilities. This suggests that HACO systems benefit from persistent external representations of hypotheses, designs, and intermediate reasoning states, rather than relying on ephemeral chat histories (Andersson et al., 19 May 2026).

A second pattern is small-batch, expert-guided exploration. In the PFAS replacement workbench, subject matter experts iteratively request small batches of candidate molecules from a generative model, inspect them, select promising candidates as new targets, and thereby steer exploration through a branching graph of discovery. The system is designed so that experts need not understand model internals; their tacit knowledge is instead captured through target selection, branch continuation, and branch abandonment. This suggests a general HACO principle: when exploration spaces are vast, expert steering is often more effective when expressed as repeated local selection than as one-shot global specification (Ferreira et al., 2023).

A third pattern is schema discovery and schema application. Schemex organizes examples by structural similarity, extracts dimensions and attributes, and then uses contrastive refinement to compare schema-guided outputs against trusted examples. SchemaBuilder extends this into workflow specification, turning discovered schemas into explicit human-AI co-creative procedures. For HACO, this implies that knowledge representations should often be structural rather than merely semantic: dimensions, constraints, and relations may be more useful than undifferentiated retrieved text when the task is to generate and test new scientific procedures (Wang, 7 Aug 2025).

These patterns converge on a common claim: HACO works best when hypotheses, plans, and intermediate artifacts are made inspectable, revisable, and branchable. This suggests a shift from prompt engineering toward explicit scientific workspaces with persistent state, contrastive views, and navigable exploration histories.

5. Verification, provenance, credit, and safety

The institutional core of HACO is verification. The strongest formulations insist that AI scientists must be deployed inside institutions “redesigned for verification, accountability, interpretability, and dual-use safety.” This includes “structured, machine-readable provenance” records, “executable pipelines” that allow independent re-running of workflows, automated checks for code execution, data integrity, statistical claims, and citation accuracy, and capability gating for high-risk tools in biology, chemistry, cyber, or autonomous experimentation. Logging, provenance, capability control, and evaluation are not external compliance layers; they are part of the architecture itself (Jimenez et al., 22 Jun 2026).

Questions of authorship and credit follow directly from this architecture. The current consensus is that “AI tools cannot be assigned authorship,” because “moral, legal, and intellectual responsibility for scientific publications remains with humans.” Yet AI contributions must still be made visible. HACO therefore requires fine-grained attribution of which agent produced which hypothesis, analysis, or draft, combined with human accountability for endorsement, refinement, and release. A useful technical analogue comes from fine-grained machine-generated text detection under coauthoring: on the HACo-Det benchmark, metric-based methods achieved only a 0.462 average F1 score for word-level detection, whereas a fine-tuned DeBERTa model reached 0.831 at the word level and 0.966 at the sentence level, with document-level AI-rate prediction errors of 1.84% at word level and 1.78% at sentence level. These results do not solve attribution, but they show that token- and sentence-level contribution tracing is technically plausible and materially superior to binary document-level judgments (Su et al., 3 Jun 2025).

Safety concerns extend beyond publication integrity to the physical and dual-use domain. Capability gating, pre-deployment evaluation, audit logs, anomaly monitoring, and explicit logged human authorization are repeatedly proposed for “AI-initiated physical experiment[s], irreversible intervention[s], or release of a self-modifying agent.” The same literature warns about publication flooding, hallucinated references, methodological homogenization, unsafe self-improvement loops, and “knowledge collapse,” the erosion of deep human expertise as AI becomes dominant. Cognitio Emergens adds the more specific risks of epistemic alienation, where researchers endorse results they cannot explain, and epistemic closure, where reinforcing human-AI feedback loops narrow the space of considered theories. In HACO, safety is therefore epistemic as well as operational (Lin, 6 May 2025).

6. Representative systems and empirical results

The HACO literature includes both full scientific systems and adjacent implementations that isolate particular design principles. Together they show that co-discovery is not a single technique but a family of architectures, protocols, and evaluation regimes (Jimenez et al., 22 Jun 2026, Seong et al., 22 Jun 2026, Okumura et al., 18 Jun 2025, Lin et al., 3 Nov 2025).

System	Domain	Reported result
Denario (Jimenez et al., 22 Jun 2026)	Modified gravity (DHORST theories)	Identified the hidden symmetries that keep certain modified-gravity theories “healthy” (ghost-free)
HACO → MaskGXT (Seong et al., 22 Jun 2026)	Crystal structure prediction	79.06% METRe on the MP-20 polymorph split, versus 70.87% for the strongest evaluated baseline
MH-based co-creative learning (Okumura et al., 18 Jun 2025)	Joint attention naming game	In 69 participants, human-AI pairs with an MH-based agent improved categorization accuracy and achieved stronger convergence toward a shared sign system
APEX (Lin et al., 3 Nov 2025)	Cleanroom flexible electronics fabrication	About 24% tool-recognition gain over general multimodal LLMs, 53% higher step-tracking accuracy in RIE, and expertise transfer to beginners

These systems cover distinct layers of the HACO stack. Denario exemplifies the multi-agent “algorithmic research group” in a largely computational science setting. MaskGXT demonstrates scientific algorithm discovery through cross-domain search plus sparse human steering, with state-of-the-art performance on CSP benchmarks. The MH-based naming-game work provides a formal and empirical template for shared-symbol emergence under partial observability. APEX shows how co-discovery architecture can extend into physical experimentation through mixed reality, step tracking, and real-time corrective guidance.

A plausible implication is that HACO becomes especially effective where validation is cheap, fast, and well aligned with the scientific objective, as in the MaskGXT case. Where experimentation is embodied and safety-critical, as in APEX, the same general paradigm persists but requires stronger gating, richer perception, and more explicit human authorization.

7. Open problems and ecosystem trajectory

Several open problems recur across the literature. One is evaluation: how to benchmark the quality and novelty of AI-generated science systematically, especially when systems act at scale or in partnership modes that blur contribution boundaries. Another is methodological diversity: shared models and corpora may induce convergence, so funders and institutions are urged to support heterodox methods, independent re-implementations, and adversarial replication studies. A third is preserving human expertise: as AI handles more routine exploration, institutions must avoid “knowledge collapse” and maintain scientists’ capacity for deep interpretation, ambiguity tolerance, and original thought (Jimenez et al., 22 Jun 2026).

At the ecosystem level, recent work shifts from single HACO instances toward infrastructures of cooperative agents. “Agentic Discovery” proposes role-specialized agents for Objective, Knowledge, Prediction, Service, Analysis, Publish, Exploration, Planning, and Enforcement, organized around a closed-loop version of the scientific method. OmniScientist extends this ecosystemal direction by combining a structured knowledge system built upon citation networks and conceptual correlations, a collaborative research protocol called OSP, and an open evaluation platform, ScienceArena, based on blind pairwise user voting and Elo rankings. This suggests that HACO may evolve from a lab-local co-discovery tool into a broader social-technical substrate in which humans and AI collaborate not only on experiments and models but also on review, attribution, and collective epistemic selection (Pauloski et al., 15 Oct 2025, Shao et al., 21 Nov 2025).

The long-term trajectory is therefore not a single autonomous “AI scientist,” but a reconfigured research environment in which humans, specialized agents, scientific databases, execution platforms, and evaluation institutions are coupled tightly enough to support continuous co-discovery. Whether that trajectory remains scientifically productive will depend less on raw model capability than on verification, interpretability, pluralism, and the preservation of human judgment.