On the live single-node deployment, short-answer rounds take ≈32 s (≈$0.015/round) while multiple-choice rounds are ≈4× faster (≈7.6 s, ≈$0.005/round) because grading is deterministic¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Measured on the live backend, short-answer rounds cost ≈32 s and ≈$0.015 each, while multiple-choice rounds take ≈7.6 s and ≈$0.005 because only next-item generation calls the LLM. → Kailai He 2026

Evidence¶

Kailai He 2026¶

Kailai He, Zhihao Wu, Linhai Zhang, Runcong Zhao, Yulan He, Jiazheng Li. (2026). CoLearn: An Agentic Tutor that Learns its Learner in a Human–AI Co-Learning Loop. https://arxiv.org/abs/2609.21154 doi:10.48550/arxiv.2609.21154

q2 · i? · design · r2

Latency and cost measured on the live single-node deployment (AWS Bedrock, Claude Sonnet 4.5); the article reports "Short-answer ≈32 s/round (grade+feedback dominate); MCQ ≈7.6 s (determin- istic answer)" with cost estimated from token usage.

"Short-answer ≈32 s/round (grade+feedback dominate); MCQ ≈7.6 s (determin- istic answer). Cost is estimated from token usage."

Discussion¶

Related Claims¶

• Taking initial multiple-choice tests without feedback can lead students to later produce the incorrect lure answers they selected, even when an overall retrieval practice benefit occurs — related
• The prototype maintains total end-to-end latency of 3-5 seconds between query completion and RAG-generated response delivery — related
• The AI feedback system returns responses in under a second for MCQs and about six seconds for open-ended questions, and its reusable prompt is roughly 2.77 times more time-efficient to prepare than equivalent educator text — related
• LLM-generated retrieval practice was practical to deploy and well received, but high correctness rates (about 70% of items at p=1.0) restricted score variance — related
• LearnLens's modular pipeline outperforms direct-prompt baselines on scoring quality while matching latency and cost on 100 authentic student answers — related
• Prebuilt quiz banks of 435 questions across 65 modules decouple practice from live generation, with warm practice-request medians of 4.8 and 5.8 ms — related
• Larger dense generators passed smoke probes but all exceeded the +30% classroom latency limit in bake-off round 1 — related
• System reliability and latency acted as a prerequisite bottleneck: 56.4% of started solves completed, with a 32.0-second average latency and 68.5-second p90 — related