A Survey on Human-AI Collaboration with Large Foundation Models

Technical Summary: A Survey on Human-AI Collaboration with Large Foundation Models

Problem Statement
As Large Foundation Models (LFMs), including Large Language Models (LLMs) and Large Vision Models (LVMs), expand rapidly in capability, the paradigm of Human-AI (HAI) collaboration is shifting from purely automated systems to symbiotic partnerships. While LFMs offer unprecedented pattern recognition and predictive capabilities derived from vast datasets, their deployment introduces persistent challenges regarding safety, fairness, control, and accountability. The core problem addressed by this survey is the lack of a comprehensive framework that integrates human intellect with the raw power of LFMs to create reliable, trustworthy, and beneficial systems. The authors argue that successful HAI systems are not merely the automatic result of stronger models but require careful, human-centered design across the entire model lifecycle.

Methodology
This paper presents a structured survey of the literature on HAI collaboration, specifically analyzing the impact of LFMs. The authors conducted a search on Google Scholar covering the period from 2014 to 2025, utilizing keywords such as "Human-AI," "Human-AI collaboration," "large models," "foundation models," "large language models," and "large vision models." The scope was expanded to include targeted updates on 2025–early 2026 works concerning tool-use agents, multi-agent orchestration, and human supervision of long-horizon workflows.

The survey is organized around the lifecycle of HAI collaboration with LFMs, as illustrated in Figure 3 of the paper:

1. Build-time: Human-guided model development (data curation, objective shaping, evaluation).
2. Run-time: Design principles for collaboration (interaction, role allocation, control handoff).
3. Governance: Cross-cutting ethical, societal, and regulatory frameworks.
4. Applications: Domain-specific implementations.

The authors employed recurring collaboration patterns as a guiding lens to synthesize the literature, including human steering, AI augmentation, mixed-initiative collaboration, dynamic delegation, and emerging supervisory orchestration. Notably, the authors themselves utilized HAI collaboration (combining human literature review with AI-assisted language flow and citation placement) to prepare the manuscript.

Key Contributions and Structure

1. Human-Guided Model Development (Build-time)
The survey details how humans shape LFMs before deployment through three phases:

• Human-in-the-Loop (HITL) Data & Active Learning: Moving beyond mass annotation, humans act as targeted teachers during fine-tuning. The paper highlights active learning workflows where models identify uncertainty, and humans provide manual annotations to retrain the model. A key challenge identified is designing bias-aware active learning that addresses pre-trained knowledge skew.
• Human-Guided Objective Shaping: This section covers preference-based alignment methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO). These methods allow humans to reshape model objectives by providing rankings or critiques, steering models toward safer and more helpful behaviors. The survey notes extensions into multimodal feedback and robotics (e.g., InstructRL).
• Human Evaluation and Alignment Metrics: Evaluation is framed not as a static report card but as an iterative feedback loop. The paper emphasizes the need for human evaluators to stress-test systems, detect subtle failures, and refine rubrics, moving beyond automated benchmarks to assess trust, safety, and transparency.

2. Design Principles for Run-time Collaboration
Once deployed, LFMs act as partners requiring specific design principles:

• Interaction and Interface Design: The survey discusses the shift from rigid graphical interfaces to fluid, conversational interactions. It highlights the tension between natural language flexibility and the need for precision, advocating for hybrid interfaces that offer visible reasoning and structured controls.
• Collaboration Patterns and Role Allocation: The authors categorize collaboration into five distinct patterns (Table 2):
  • Human-Steers-AI: Humans provide instructions and corrections.
  • AI-Augments-Human: AI supports human decision-making (e.g., in healthcare).
  • Mixed-Initiative: Both parties propose and revise tasks.
  • Dynamic Delegation/Control Handoff: Control shifts based on situational awareness (critical in autonomous vehicles).
  • Emerging Supervisory Orchestration: A new pattern for agentic systems where humans monitor multi-step workflows, set goals, and intervene, rather than guiding every step.
• System Integration and Evaluation: The paper stresses the need for backward compatibility with legacy systems, managing the learning curve for users, and developing holistic metrics that measure team dynamics and user experience, not just task accuracy.

3. Ethical, Societal, and Governance Aspects
The survey examines cross-cutting issues:

• Fairness: Addressing how human oversight can paradoxically reinforce bias and the need for fair-by-design frameworks.
• Autonomy and Well-being: Balancing AI augmentation with the risk of human deskilling and loss of ownership.
• Labor Dynamics: Analyzing the reshaping of jobs, the need for reskilling, and the economic impacts of LLMs.
• Privacy and Security: Discussing threats like prompt injection and data poisoning, and solutions like federated learning and differential privacy.
• Trust and Accountability: The need for transparent decision trails (e.g., Agent-Deed-Consequence models) and the challenge of assigning liability in agentic workflows.
• Policy: Highlighting the gap between rapid LFM development and the creation of enforceable legal frameworks.

4. Applications Across Domains
The survey maps these patterns to specific high-stakes domains:

• Healthcare: Focuses on calibrated augmentation, expert deferral, and the critical need for trust and explainability in clinical settings.
• Autonomous Vehicles: Highlights dynamic control handoffs and the use of LFMs for situational awareness and scenario generation.
• Surveillance and Security: Discusses the balance between AI autonomy and human oversight in triaging alerts and detecting anomalies.
• Games: Explores LLMs as active partners and content creators, noting challenges with coherence and bias.
• Education: Examines personalized tutoring and the risks of parasocial attachments and opacity.
• Accessibility: Focuses on co-adaptive systems that learn from user feedback to assist individuals with disabilities.

Results and Open Challenges
The survey concludes that while LFMs have expanded the potential of HAI collaboration, significant open challenges remain:

• Scalable and Diverse Human Guidance: Ensuring feedback is representative and not limited to narrow groups.
• Fluid but Controllable Interaction: Designing interfaces for agentic systems that prevent silent error accumulation in long action chains while maintaining natural interaction.
• Accountability and Governance: Creating auditable records and enforceable standards for complex, multi-step AI workflows where responsibility is blurred.
• Contextual Grounding: Moving beyond one-size-fits-all models to domain-specific adaptations that respect professional norms and user needs.

Significance
The paper claims its significance lies in synthesizing the current state of HAI collaboration in the era of LFMs. It argues that the future of AI depends not on autonomy alone, but on the quality of partnerships built through careful design. By identifying key open challenges and structuring the literature around the model lifecycle, the survey aims to guide future research toward creating HAI systems that are reliable, trustworthy, and beneficial to society. The authors emphasize that realizing the potential of LFMs responsibly requires addressing safety, fairness, and control through human-centered design rather than relying solely on model scale.