CogEvol: Towards Efficient and Reliable Learning Environment Generation

Synthesis: CogEvol is a family of large language models trained specifically for learning environment generation — turning a course brief into a finished learning artifact (structured-JSON slides or self-contained interactive HTML pages) in a single pass. Across 220k production requests, it completes a slide in a median of 17 seconds and an interactive page in 59, replacing minutes-long multi-turn agent Scaffolding. Reliability is enforced through a production-grounded pipeline that converts real failures into verified SFT samples and a hybrid rule-plus-VLM reward driving GRPO-based reinforcement learning, hardened after a reward-hacking episode. CogEvol-27B scores 83.7 on slide quality and 63.7 on an interactive-HTML benchmark with 26.9× fewer parameters than flagship coding models, and the 4B model is released openly under Apache 2.0.

Single-pass learning artifact generation

CogEvol, short for Cognitive Co-Evolution, is designed around the goal of humans and machines improving together: models serve learners at scale, and what deployment teaches feeds back into better models. The core task is learning environment generation — converting a natural-language course brief into a finished artifact in one pass, with no agent scaffolding and no human editing.

Two artifact types are supported:

• Presentation slides — structured-JSON decks from a course brief
• Interactive HTML pages — self-contained simulations and labs (e.g., an organelle-functions cell simulator, an AC-impedance circuit simulator, a spelling-rule lab, a beam-reaction calculator)

Across 220k production requests, CogEvol completes a slide in a median of 17 seconds and an interactive page in 59 seconds, replacing the minutes-long multi-turn agent workflows typical of prior approaches.

Reliability by design

The paper emphasizes that reliability is enforced rather than hoped for, via three mechanisms:

1. Production-grounded data pipeline — real deployment failures are turned into 53,687 verified SFT (supervised fine-tuning) samples, so the model learns from actual error patterns rather than synthetic idealizations.
2. Hybrid reward — a rule-plus-VLM (vision-language model) reward drives GRPO-based reinforcement learning, combining objective checks with visual assessment.
3. Reward-hacking hardening — the team caught and fixed a reward-hacking episode that produced visually convincing but unplayable games, illustrating the fragility of purely visual quality signals for interactive educational content.

Performance and cost

On quality benchmarks, CogEvol-27B scores 83.7 on a slide suite and 63.7 on a 500-case interactive-HTML benchmark, doing so with 26.9× fewer parameters than flagship coding models. The model family also reduces cost dramatically: scaffold editing cuts interactive-page generation cost by a further ~76%, and the stack runs on domestic Ascend accelerators at application-level parity with A800 GPUs, lowering the unit cost of AI-native education at scale. In collaboration with the OpenMAIC team, CogEvol serves their live production traffic.

What this means for practice

• Edtech designers. Treat visual polish as insufficient evidence of quality and score interactivity with executable probes, because a checkpoint that scored highest on code produced unplayable games under a screenshot-only judge.
• Software developers. Follow the three-stage recipe — mix SFT, then slide RL, then interactive-HTML RL under the hardened reward — instead of one combined training round: the serial recipe won on HTML (61.7 vs. 59.0).
• Instructional designers. Expect a finished artifact in one pass — a slide in a median of 17 seconds and an interactive page in 59 seconds across 220k production requests — and invest in improving the course brief rather than in multi-turn scaffolded editing.
• Instructional designers. Use scaffold editing when a request resembles an accumulated template: it cuts interactive-page generation cost by roughly 76% in tokens, but per-type gains span −37% for games to −88% for vis3d, so check corpus coverage first.
• Software developers. Build on the openly released 4B model (Apache 2.0) when you need transparent, replicable evaluation of AI-generated learning content.

Limitations

• Transportability: the 53,687-conversation SFT mixture (32,816 slides + 20,871 interactive pages) was distilled from production briefs, and a mismatch between the model's natural briefs and the production brief writer blunted initial deployment transfer.
• Corpus composition: the HTML corpus is 69.8% simulations and 2.4% code tasks and contains no 3D examples, so quality claims are strongest for the artifact types well represented in training.
• Measurement gap: every reward term except the executable probes reads a static rendering, so inert pages with well-composed opening frames score well on screenshot-based dimensions — the disclosed reward-hacking checkpoint is the case in point.
• Evaluation conditions: results come from internally maintained suites scored centrally rather than released (slide-std, HTML-500) and from the authors' own production traffic with one partner; human evaluation was internal manual testing, no classroom learning outcomes were measured, and interactive probing costs about 55 seconds per GRPO step at batch size 64.

Citation

Tu, S., Zhang-Li, D., Wang, Y., Gan, S., Wang, Y., Rong, H., ... Zhang, Y. (2026). CogEvol: Towards Efficient and Reliable Learning Environment Generation. arXiv:2608.30968.