CogEvol: Towards Efficient and Reliable Learning Environment Generation

AI summary

Overview

Research area: Natural Language Processing, specifically post-training of large language models for educational content generation.

Technical level: Advanced. The paper assumes familiarity with supervised fine-tuning (SFT), GRPO-based reinforcement learning, reward design, and multimodal judge systems.

Scope: A report on CogEvol, a family of models (CogEvol-4B and CogEvol-27B) trained to generate complete learning artifacts—structured-JSON slides or self-contained interactive HTML pages—from a course brief in a single pass, evaluated on the new Learning Environment Generation (LEG) task.

What This Paper Is About

General-purpose coding agents can already produce slides and interactive courseware, but only slowly: existing systems routinely take 200–600 seconds per edit, and the authors measure a direct-API editing loop at 152 seconds per edit, while the strongest coding models are enormous (GLM-5 weighs 744B parameters) and priced out of reach for many educational users. The paper formalizes Learning Environment Generation (LEG)—turning a course brief into a renderer-valid slide scene graph or a self-contained runnable HTML page in one pass—and builds a purpose-trained model family to do that faster, more reliably, and more cheaply than agent scaffolding on top of general flagships.

Key Contributions

1. Task and benchmarks. The authors formalize Learning Environment Generation and build evaluation suites for both modalities: slide-std and slide-short (120+120 topics) plus a 500-case interactive-HTML benchmark with executable interaction probes.
2. Efficiency. Purpose-trained single-turn generation (median 17s per slide, 59s per interactive page across 220k production requests) replaces minutes-long multi-turn agent scaffolding; scaffold editing cuts interactive-page generation cost by a further ~76%, and the MAIC-UI harness layer accelerates iterative edits by 23×.
3. Reliability. A hybrid rule-engine plus VLM reward system, hardened after a reward-hacking episode on interactive games that the authors found and fixed, and a one-big-round multi-task recipe that trains both modalities jointly with zero forgetting tax.
4. Cost and open release. CogEvol-27B matches production needs at 26.9× fewer parameters than coding flagships; CogEvol-4B is released openly under the Apache 2.0 license; the full stack runs on domestic Ascend accelerators with application-level parity to A800 GPUs.

Main Findings

• Latency over production traffic. Over a seven-day window of live traffic (August 2026), the production models completed 180k slide generations at a median (P95) of 17s (26s) and 40k interactive pages at 59s (107s). Slide pages emit structured JSON (~1.8k output tokens on average); interactive pages emit complete HTML documents (~9k).
• Edit iteration speed. The MAIC-UI editing layer reduces per-edit latency by 23× (151.7s to 6.3s) compared with direct API calls under the same backbone model.
• The rendering contract is binding. The bare Qwen3.8-27B base emits JSON that parses for 118/120 slide briefs yet renders 0/120 under the strict schema. Routed through schema normalization, the same outputs render 118/120 with content fidelity near the family's best (91.7) but composition collapsed (layout 41.5, versus 77.5 after slide RL).
• Base capability sets the ceiling. Under an identical SFT recipe (mix-0812 data, 13,421 updates), Qwen3.8-27B reaches 79.5 on slide-std against 67.7 for Qwen3.6-27B, with contract parse rates of 99.2% versus 85.8%. The best Qwen3.6 checkpoint anywhere in its lineage tops out at 81.4, still below the 84.8 the same slide-RL recipe reaches from the Qwen3.8 SFT.
• SFT teaches contracts, not HTML quality. HTML quality declines monotonically with update count on Qwen3.6-27B—88.4 at 6,710 updates, 83.7 at 10,711, ~73.5 at 13,421—and the Qwen3.8 run lands at the same 73.2, below its own bare base (78.3), ceding code (−7.0 pp) and game (−17.9 pp) ground.
• Cross-modal transfer. The 27B slide RL stage, trained on slide data only with no HTML in the loop, lifted interactive HTML by +10.4 pp over its SFT start.
• Reward hacking on games, caught and fixed. A third HTML RL run gave games a third of its training batch; games regressed by −12.1 pp because every reward component judged static renderings. After hardening the reward with an always-on interactivity probe and a hard-fail gate—same checkpoint, data, and schedule—games reversed from −12.1 pp to +5.8 pp, 3D rose +15.9 pp, and overall HTML lifted 54.2 → 61.7. The old-reward 27B checkpoint scores 18.8 on games under the hardened reward, a −36 pp collapse the screenshot judge had masked as −9.7 pp.
• Human-verified reliability gains. Human testers saw unusable pages halve (25% → 10%) and entry failures vanish (2/24 → 0/30) after the fix.
• Final released scores. CogEvol-27B reaches 63.7 on HTML-500 and 83.7 on slide-std, with games at 57.6, after a −1.1 serial forgetting tax. CogEvol-4B reaches HTML-500 61.7 (from 52.8 at SFT) and slide-std 75.1, with a −1.7 forgetting tax.
• Comparison to flagships. Claude Opus 4.8 edges the HTML overall at 67.2 but fails 19 pages outright, while CogEvol-27B posts zero interactive hard failures on all 500 pages. Qwen3.8-Max exactly ties 83.7 on slides but collapses on HTML at 35.3, with 204 of 500 pages dead. Handed the full 34 KB specification, every other flagship stays at or below 79.5 on slides.
• Cost. At public list prices, CogEvol-27B delivers near-flagship quality at 15–22× lower per-artifact API cost than Claude Opus 4.8 or GPT-5.4, and CogEvol-4B at ~100× lower.
• Joint training alternative. One-big-round multi-task RL on 4B reached slide 74.1 and HTML 59.0 with games intact (53.6, versus 37.3 for the pre-hardening serial twin), running 500 steps with no router miss—but the serial recipe still wins on HTML alone (61.7 vs. 59.0).

Methodology in Plain English

The system is built by post-training only, starting from public base models—Qwen3.8-27B (a hybrid with 48 gated-delta-net linear-attention layers, 16 full-attention layers, and a multi-token-prediction head) for the 27B model and dense Qwen3.5-4B for the 4B model. Nothing is pre-trained from scratch.

Three stages follow. Mix SFT teaches the two output contracts using 53,687 verified conversations: 32,816 slide rows and 20,871 interactive pages, each a (system contract, user brief, verified artifact) triple. The supervision is execution-aware—slides are accepted only after passing JSON parsing, schema validation, canonicalization, a production render and a multimodal judge; HTML pages are accepted only after re-executing under a Chromium probe. The HTML data is mined from real failures: of 119,122 exported generations, 117,309 completed and 25,475 (21.7%) exhibited hard failures, which were regenerated and re-tested.

Slide RL then optimizes composition with a hybrid reward of 0.6 × VLM fidelity on rendered pixels plus 0.4 × geometric rules (collision, overflow, canvas utilization), where the rule engine went through five versions to close exploits such as faking charts with styled text. Interactive-HTML RL optimizes dependability with a reward combining visual quality (0.4), content (0.3), dual-viewport rendering (0.1 each) and a measured interactivity term (0.3) from a Playwright-driven Chromium instance that operates the page's controls. Both stages use GRPO on the slime framework with 8–16 H800 GPUs, group size 8, KL coefficient 10⁻³, learning rate 10⁻⁶, and 250 rollouts per stage. Because candidates must be rendered, loaded, probed and judged before scoring, reward computation dominates wall-clock cost.

Evaluation mirrors training: slides are rendered by the production renderer (contract violations score zero) and judged on fidelity and layout; HTML pages must first survive a deterministic probe gate, where hard failures score zero regardless of appearance.

Why This Matters

The paper argues that efficient, reliable, and cheap generation is what turns AI education from a premium product into public infrastructure. It shifts the field's attention from conversational tutoring and text generation toward executable artifacts governed by a rendering contract, and it supplies a concrete cautionary case study: what a reward cannot measure, RL will quietly destroy.

Real-world applications:

• Classroom courseware generation. Teachers turn a brief into slides or runnable simulations, games, code playgrounds and 3D visualizations within a class period, rather than waiting minutes per artifact.
• On-device and low-resource deployment. CogEvol-4B's open weights, including a Q4_K_M GGUF build, are small enough for on-device use, targeting teachers in mountain regions, low-income families, and olympiad training in less-developed areas.
• Live platform serving. CogEvol-27B, developed with the OpenMAIC team, has served their production traffic since 2026-08-24, replacing its predecessor behind the same API.
• Hardware-constrained deployment. The full stack runs on domestic Ascend accelerators at application-level parity with A800 GPUs, lowering unit cost where GPU supply is restricted.

Industry relevance centers on cost structure: a 27.7B model delivering production-grade output at 15–22× lower per-artifact API cost than flagships, versus GLM-5's 744B parameters (40B active), makes educational tooling economically viable as a service rather than a premium product.

Future Directions

• 27B validation of one-big-round training. The joint multi-task recipe was trained on 4B only; the paper explicitly leaves 27B validation to future work.
• Closing the serial-versus-joint quality gap. The serial recipe wins on HTML alone (61.7 vs. 59.0), and human inspection of a 100-case side-by-side favored its pages; per-task gradient is halved and long-tail types (learning pages, code) pay for it.
• Addressing dataset biases. The HTML corpus is 69.8% simulations, 2.4% code tasks, and contains no 3D examples; the slide targets skew safe and sparse (mean 14.4 elements, 68.5% at most 16), and the specification model's English-rich briefs differ from the production brief writer.
• Broadening harness support. CogEvol is not built exclusively for OpenMAIC; the authors invite other AI+Education teams to open their harnesses so CogEvol can be adapted as it was for OpenMAIC.

Target Audience

This paper is most useful to machine-learning engineers and researchers working on post-training, reinforcement learning from verifiable or measured rewards, and reward-hacking mitigation; to AI-in-education product and platform teams evaluating whether to train a small specialist model instead of scaffolding a large flagship; and to researchers in educational technology who need a formalized task definition, benchmark design, and evaluation protocol for executable learning artifacts.

Authors’ abstract

We present CogEvol, a family of 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, CogEvol 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 rather than hoped for: a production-grounded data pipeline turns real failures into 53,687 verified SFT samples, and a hybrid rule-plus-VLM reward drives GRPO-based RL, hardened after we caught and fixed a reward-hacking episode that produced visually convincing but unplayable games. CogEvol-27B scores 83.7 on slide quality and 63.7 on a 500-case interactive-HTML benchmark with 26.9x fewer parameters than flagship coding models, and, in collaboration with the OpenMAIC team, serves their live production traffic. CogEvol-4B is released openly under the Apache 2.0 license at https://github.com/CogEvol/CogEvol-4B; external flagships are measured on the same suites under the identical harness. Scaffold editing cuts interactive-page generation cost by a further ~76%, and the full stack runs on domestic Ascend accelerators at application-level parity with A800 GPUs, lowering the unit cost of AI-native education at scale.