Introduction
Most modern AI systems are notoriously "data-hungry," requiring millions of annotated examples to master a task. However, in real-world scenarios involving personal preferences, health restrictions, or subjective tastes, such massive datasets are rarely available. This paper addresses the challenge of Human-Agent Interaction (HAI) for subjective tasks, specifically focusing on ingredient substitution. The goal is to create an agent that learns incrementally from a human teacher, much like a student, rather than training on a static, black-box dataset.
The Human-Agent Teaching Framework
The researchers proposed a three-pronged architecture designed to solve the "cold-start" learning problem, where an agent must become useful after only a handful of interactions. This framework is structured around three core research questions (RQs):
- Domain Knowledge (RQ1): How should the agent be "seeded"? The study compared sub-symbolic embeddings (like FoodBert) against symbolic knowledge from the FoodOn ontology.
- Learning Method (RQ2): How does the agent process new examples? The authors moved away from standard "Prototypical Networks"—which average example vectors and cause "semantic dilution"—in favor of an Accumulative approach. By adding vectors together, the model preserves the magnitude of signals, naturally weighting habitual or popular substitutions.
- Tutoring Policy (RQ3): In what order should a human provide examples? The study compared a Random order against a Balanced policy that uses a bucketing system to ensure a mix of common patterns and rare, diverse scenarios.
Case Study: Ingredient Substitution
To validate this framework, the authors utilized the Recipe1MSubs dataset, containing approximately 70,520 real-world ingredient substitutions. The task is highly contextual: if a recipe calls for salt but the user is out, the "correct" substitute depends on the specific dish and user needs.
In this pipeline, the human teacher provides a substitution example (e.g., replacing ingredient $s$ with $t$ in recipe $R$). The agent perceives the query, integrates it with its existing domain knowledge, and updates its learning method to suggest better substitutions in the future.
Key Results and Insights
The experimental results revealed several critical insights into efficient AI instruction:
- Symbolic Knowledge Generalizes Better: Interestingly, the structured hierarchy of the FoodOn ontology (e.g., knowing "Sea Salt" is a type of "Salt") proved more useful for generalization than pre-trained text embeddings.
- Strategic Teaching is a Force Multiplier: The order of instruction is as vital as the content. Using the Balanced tutoring policy, the agent reached half of its peak performance after only 100 examples—representing just 0.2% of the training data. At this 100-example mark, the Balanced policy achieved a 10.99% Hit@1, nearly tripling the 3.82% achieved by random instruction.
- Data Redundancy: The findings suggest that when tutored strategically, nearly 80% of a large dataset is redundant. An agent can reach "complete dataset performance" much earlier if the human teacher selects examples wisely.
Conclusion and Future Work
This research demonstrates that combining symbolic domain knowledge with strategic tutoring allows AI agents to learn subjective tasks with extreme efficiency. While this study used a simulated teacher, future work will explore Active Learning, where the agent identifies its own blind spots and asks the human for specific guidance. This architecture paves the way for more intuitive, personalized AI assistants in fields ranging from culinary arts to healthcare.
Summary updated