AI Enters the Experiential Era: WAIC 2026 Shows AI‑Native Learning Labs in Action

The WAIC exhibition featured the iLoveStudy AI learning agent, which transforms traditional "answer‑first" tools into an interactive tutoring system. In a geometry problem demo, the agent first asks the student to explain the reasoning behind the answer, cues the student when the response is vague, and only proceeds once the underlying knowledge point is clarified. This "search‑for‑reasoning" workflow replaces the usual "show‑the‑solution" approach.

During an English‑writing lesson, the agent treats the student's sentence as a living draft, prompting for adjectives, adverbs, prepositional phrases, and non‑finite clauses. Errors are highlighted, suggestions are offered, and the dynamic board updates in real time, mirroring a human teacher’s feedback loop.

iLoveStudy’s core is a self‑trained large model that acts as a teaching decision hub: it first identifies the student’s state, decides the next pedagogical action, and then executes it through language, board rendering, and interaction pacing. A dual evaluation system ensures both instruction compliance (over 99% compliance) and response quality (natural, concise, logical).

To bridge the gap between prototype and product, iLoveStudy employs a two‑stage data flywheel. Offline, synthetic classroom data are used for safe failure testing and policy iteration. Online, real student interactions generate reinforcement signals that are validated in a sandbox before being fed back into the model, continuously expanding its teaching skill set.

On the multimodal side, the company introduced AnyAvatar, a 3D Gaussian‑based head avatar that self‑calibrates camera pose and facial geometry without studio‑grade equipment, dramatically lowering production cost while achieving high‑fidelity rendering. MoGaFace provides end‑to‑end voice‑print separation and noise reduction, achieving 97% recall, 98% precision, >15 dB SNR improvement, and a 40 dB reduction in background noise, with registration time cut from 20 seconds to 1 second.

Engineering optimizations target sub‑second first‑sentence latency: large‑model request scheduling, KV‑Cache and Prefix‑Caching, and token‑to‑wave separation reduce end‑to‑end latency to ~1.5 seconds even under massive concurrency. The system is designed for re‑entrancy, preserving each learner’s knowledge state, goals, and interaction history so that a student can leave and return without losing context.

Safety is enforced through staged rollouts, gray‑scale testing, and sandbox verification before any new teaching strategy reaches production, preventing accidental leaks or inappropriate feedback.

Overall, iLoveStudy combines a self‑trained teaching‑decision model, a multimodal body (speech, vision, digital human), and a robust engineering foundation to deliver a seamless, interactive AI‑driven classroom experience.