The six long-chain Stage 3 workflows average 0.321 versus 0.559 across all other Stage 3 tasks and are the lowest-scoring family for 15 of 17 models¶

Claim · All claims
Evidence · 1 study · design r? · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? In Stage 3, long-chain workflows score far below other tasks (0.321 versus 0.559) and are the lowest-scoring family for 15 of 17 models, consistent with composition pressure across evidence, actions, artifacts, persistent state, and communication. → Zixin Chen 2026

Evidence¶

Zixin Chen 2026¶

Zixin Chen, Peng Liu, Rui Sheng, Haobo Li, Jianhong Tu, Xiaodong Deng, Kashun Shum, Dayiheng Liu, Huamin Qu. (2026). Are Agents Ready to Teach? A Multi-Stage Benchmark for Real-World Teaching Workflows. Preprint. https://arxiv.org/abs/2605.14322 doi:10.48550/arxiv.2605.14322

q2 · i? · design · r?

Descriptive contrast from the Stage 3 results of the 17-model evaluation. The article reports the averages for the six long-chain workflows versus all other Stage 3 tasks and states the association is consistent with composition pressure without identifying chain length or any interface as the cause.

"In Stage 3, the six long-chain workflows average 0.321 versus 0.559 across all other Stage 3 tasks and are the lowest-scoring family for 15 of 17 models."

Discussion¶

Related Claims¶

• The strategic surface learner is the lowest-scoring Stage 2 profile for 15 of 17 models, averaging 0.702 versus 0.757 across the other profiles — related
• Across 17 frontier models, Stage 1 pedagogical-judgment scores are high (mean 0.893) while rank correlations with Stage 2 (ρ=0.24) and Stage 3 (ρ=0.21) are weak, with workflow scores topping out at 0.704 — related
• Claude Opus 4.8 has the highest overall score (0.803) and is the only model ranking in the top three on all three stages, yet does not saturate situated tutoring (0.773) or workflows (0.704) — related