Adaptation and Evidence-grounding of Generative Interventional Systems (AEGIS) Causality and Uncertainty in Artificial Intelligence for Decision Making and Policy Awareness 

about aegis

Abstract

AEGIS advances generative AI that is fit for decision‑making by grounding models in interventions, counterfactuals, and calibrated uncertainty. The workshop, held during ACM KDD 2026 conference, convenes researchers and practitioners in causality, LLMs, and prescriptive analytics to address when and how generative systems should recommend actions in healthcare and public policy. We invite methods that couple structural causal reasoning with LLMs, diffusion and sequence models; techniques for off‑policy evaluation, dynamic treatment regimes, and feedback‑aware learning; semi‑synthetic benchmarks and governance practices; and measures beyond predictive fidelity, including policy regret, counterfactual calibration, and safety constraints. The program will feature a keynote talk and peer‑reviewed presentations. By aligning with KDD’s emphasis on trustworthy, scalable AI, AEGIS endeavors to establish shared evaluation protocols and artifacts that make prescriptive models reliable under distribution shift, so recommendations remain robust and accountable from development to deployment across settings and populations.

Motivation and rationale

Generative models, including foundation models and large language models (LLMs), are increasingly deployed for prescriptive purposes: recommending treatments and policies, simulating patient trajectories, and planning long‑horizon strategies. However, models that excel at observational prediction or distributional simulation can fail when repurposed to compare or recommend actions—especially in settings with feedback loops, confounding, selection effects, and evolving practice patterns. In healthcare and public policy, these failures translate into unsafe, biased, or unequal outcomes, as recommendations may not be causally valid even when predictive metrics look strong. Addressing this gap requires methods that (i) are valid from a cause-effect standpoint, i.e., explicitly modeling interventions and counterfactuals, (ii) quantify and calibrate uncertainty to support safe decision‑making, and (iii) evaluate models by decision validity (e.g., policy regret, counterfactual calibration), not just predictive accuracy.

Focus and topics

AEGIS examines generative AI for prescriptive analytics with a particular emphasis on healthcare and public policy as stress‑test domains, while remaining relevant to other KDD areas (platforms/marketplaces, finance, mobility). We focus on methods that integrate causal reasoning with generative architectures (including LLMs, diffusion models, etc.) to produce intervention‑ready and counterfactual‑aware recommendations under real‑world data complexities.

Representative themes include: causality with and within generative models and LLMs; decision‑making under feedback, confounding, and selection; uncertainty quantification, calibration, and safety for recommendations; evaluation beyond predictive fidelity, including counterfactual and decision‑validity metrics; fairness, accountability, and model‑level governance for prescriptive AI; applied case studies using heterogeneous healthcare and policy data.

Target audience

The workshop targets researchers and practitioners working on generative modeling and LLMs, causal inference with an emphasis on interventions and counterfactuals, and uncertainty quantification and evaluation for decision‑making. It also serves applied scientists and engineers in healthcare, public health, and policy (primary audiences), as well as those in platforms and marketplaces, finance, and industrial organizational psychology who deploy prescriptive models under feedback, bias, and distribution shift. Finally, it welcomes students and early‑career scholars seeking mentoring and hands‑on exposure.

A global stage

AEGIS will be held as a workshop for participants of the annual ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), August 9-13, 2026, Jeju, South Korea. KDD’s 2026 themes put modern generative AI and trustworthy data science front‑and‑center, and with rapid real‑world adoption of LLM‑driven decision systems, this is the timely venue to consolidate community standards in a way that only the SIGKDD community can deliver at scale. The workshop is tightly coupled to the KDD 2026 Research Track scope and themes, particularly Foundations of Knowledge Discovery and Data Science, Modern AI & Big Data, Trustworthy & Responsible Data Science, Systems for Data Science & Scalable AI, Data Science Applications, and the new AI‑for‑Sciences.

Agenda

• Introduction by Dr. Beth Virnig, University of Florida, USA, and welcome by the organizers (15m)
• Keynote speech (45m) with Q&A (10m)
• Coffee break (30m)
• Selected talks from accepted papers (4x15m)

call for papers

AEGIS will accept short paper (1500-2500 words excluding references) submissions. There are no specific formatting requirements, but authors should follow ACM guidelines about authorship and usage of AI (see ACM authorship policy and ACM FAQs). Submissions will be peer reviewed (single blind) by the AEGIS program committee and possibly selected for oral presentation during the workshop.

Accepted papers will not be included in the main conference proceedings, and the authors retain rights to publish the work elsewhere. However, the AEGIS program committee encourages authors to share the contribution to the scientific community though open repositories such as biorxiv in the format they prefer.

The AEGIS organizers are currently negotiating a post-workshop publication opportunity for accepted papers in a special issue of Springer’s Journal of Healthcare Informatics Research. Authors will be required to submit an extended version of their paper, undergo additional peer review, and pay applicable article processing charges.

Paper topics include but are not limited to:

• Causality in/with generative models and LLMs (e.g., structural‑causal‑model–constrained generation, causal‑graph grounding for prompts/decoding, counterfactual generation)
• Interventional and counterfactual methods (e.g., sequential treatment policies from observational data, identification with proxies or instruments, sensitivity and bounding analyses)
• Feedback‑aware decision‑making and dynamics (e.g., dynamic treatment regimes, performative effects and decision‑induced shifts, intervention‑aware representations)
• Uncertainty, robustness, and safety (e.g., calibration for interventions, distributional robustness, safety constraints as causal or temporal invariants)
• Evaluation and governance beyond predictive fidelity (e.g., semi‑synthetic or synthetic‑twin benchmarks, policy regret and off‑policy evaluation, fairness auditing and auditable decision traces)
• Data and systems at scale (e.g., electronic health records, claims and multimodal or graph data; reproducible pipelines, logging and lineage, human‑in‑the‑loop what‑if tools)
• Knowledge and mechanistic integration (e.g., program‑guided or neuro‑symbolic generation, knowledge‑graph grounding, incorporation of domain ontologies and mechanistic priors)
• Cross‑cutting themes (e.g., reproducibility and artifacts, privacy‑preserving learning and data governance, negative results and failure analyses, ethical and societal impacts)

Attendance and registration

Participation in AEGIS 2026 requires registration to the main conference, KDD 2026. For accepted papers, at least one author must register at the conference.