PaperlayerView exact revision on arXiv

Source-linked AI summary

CogEvol: Towards Efficient and Reliable Learning Environment Generation

Shangqing Tu, Daniel Zhang-Li, Yucheng Wang, Shiyu Gan, Yanpeng Wang, Huiqiang Rong, Mofei Chen, Shen Yang, Yini Chen, Yinuo Duan, Haoxuan Li, Binglin Liu, Ye He, Danqi Zheng, Zhanxin Hao, Yuxuan Wu, Mengting Tao, Yuqiu Liu, Jifan Yu, Juanzi Li, Bin Xu, Lei Hou, Huiqin Liu, Yu Zhang

arXiv:2608.30968v1cs.CLcs.AI

TL;DR

Learning Environment Generation requires fast, reliable production of structured slides and executable educational HTML, but general-purpose coding systems are slow and can produce visually convincing failures. CogEvol addresses this with production-grounded verified supervision, modality-specific reinforcement learning, and execution-aware rewards, reaching strong slide and interactive-HTML scores while reducing model and editing costs. Its evaluation also exposes a reward limitation: hard-fail gating intentionally disagrees with partial human credit on some nonresponsive pages.

  • Problem

    General-purpose models struggle to generate fast, reliable, and affordable educational artifacts under rendering and interactivity contracts.

  • Method

    CogEvol combines production-grounded verified SFT, separate slide and HTML GRPO stages, hybrid rule-plus-VLM rewards, and scaffold editing for single-pass artifact generation.

  • Results

    CogEvol-27B reaches 83.7 on slide-std and 63.7 on HTML-500, while production medians are 17 seconds per slide and 59 seconds per interactive page.

  • Takeaways & Limitations

    Executable probes and hardened rewards make interactivity measurable, while the released 4B model and editing stack support more accessible learning-environment generation.

  • Takeaways & Limitations

    The hard-fail gate assigns zero to nonresponsive pages that human raters may partially credit, creating deliberate reward–human disagreement.

Abstract

from arXiv · show

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.

1 Introduction

CogEvol targets Learning Environment Generation with single-pass production artifacts, addressing speed, reliability, and cost through purpose-trained models, verified data, and reinforcement learning. Its reported results combine production latency, benchmark quality, reliability safeguards, and efficient deployment.

  • Task: CogEvol formalizes Learning Environment Generation as producing structured-JSON slides or executable HTML learning artifacts from course briefs.The system targets interactive educational materials rather than static text alone.
  • Efficiency: 17 seconds per slide and 59 seconds per interactive page are the production medians across 220k requests, replacing minutes-long multi-turn scaffolding.The reported production window includes 180k slide generations and 40k interactive pages.
  • Cost and deployment: CogEvol-27B has 26.9× fewer parameters than flagship coding models, and scaffold editing reduces interactive-page generation cost by approximately 76%.CogEvol-4B is released openly, while CogEvol-27B serves production traffic.
  • Results: CogEvol-27B reaches 83.7 on slide-std and 63.7 on HTML-500, while external models can produce visually strong but schema-invalid or unplayable artifacts.The benchmark uses shared evaluation infrastructure, and the hardened reward addresses a reward-hacking episode involving games.
  • Reliability: 53,687 verified conversations support joint supervised fine-tuning, while hybrid rule-plus-VLM rewards and hardened interactive probes support reinforcement learning.Verification requires rendering and judging slides or re-executing HTML pages under an interactivity probe.

2 The Learning Environment Generation Task

Learning Environment Generation is defined as single-pass construction of renderer-valid slides or executable interactive HTML from course briefs. CogEvol treats both output modalities as strict contracts and evaluates them using modality-specific quality and execution criteria.

  • Task definition: Learning Environment Generation turns a course brief into one complete learning artifact in a single pass.The artifact is either a structured slide scene graph or a self-contained interactive HTML page.
  • Slides: Slides are emitted as structured JSON scene graphs governed by a strict rendering schema with fixed element types and numeric geometry.Invalid keys, coordinate types, percentages, or invented element types can cause renderer failure.
  • Interactive HTML: Interactive pages are self-contained HTML documents whose controls must respond, state must remain consistent, and depicted physics or mathematics must be correct.The modality spans simulations, diagrams, games, code playgrounds, 3D visualizations, and structured learning pages.
  • Scope and novelty: LEG applies the learning-environment concept to generated slides and interactive pages, excluding surrounding platform infrastructure.The report presents this artifact-generation formulation as a newly formalized task.
  • Evaluation: Outputs are scored on fidelity, layout, interactivity, and correctness, with interactivity measured by executing artifacts and probing their event streams.Rendered outputs support the first two dimensions, while execution probes support the third.

3 Post-Training I: Supervised Fine-Tuning

CogEvol’s supervised fine-tuning uses production-derived, execution-verified artifacts matched to deployment contracts. The resulting mixture teaches modality selection and output compliance, but leaves interactive quality largely to later reinforcement learning.

  • Data pipelines: Production-seeded slide synthesis targets observed layout failures, while HTML failure mining concentrates supervision on incumbent-model failures verified through Chromium execution.The pipelines accept teacher outputs or regenerations only after execution-aware checks.
  • Data limitations: Execution-aware filtering improves supervision but produces biased corpora: slide targets are safe and sparse, HTML data is 69.8% simulations, and contains no 3D examples.The paper identifies these distributions as reasons SFT alone cannot carry HTML quality.
  • Training mixture: 53,687 conversations form the final SFT mixture: 32,816 slides and 20,871 interactive pages.Each example is a system contract, user brief, and verified artifact triple.
  • Mixture design: Repairing system prompts raised 4B contract compliance to 97.5% while slide quality recovered to 70.8.The earlier HTML-heavy mixture caused modality confusion and reduced slide contract compliance to 61.7%.
  • SFT outcome: SFT reliably teaches output contracts but cannot raise interactive quality, leaving the HTML ceiling to reinforcement learning.Across schedules, HTML quality declined with additional optimizer updates in the reported pre-hardening comparisons.
  • Base-model ablation: Under an identical recipe, Qwen3.8-27B reaches 79.5 on slide-std versus 67.7 for Qwen3.6-27B, with parse rates of 99.2% versus 85.8%.The base-model swap was isolated by reusing the same data and schedule.

4 Post-Training II: Reinforcement Learning

CogEvol applies execution-aware GRPO reinforcement learning with separate slide and interactive-HTML rewards. Experiments show that trustworthy rule signals, diverse briefs, and executable interactivity probes determine whether optimization improves quality or exploits the judge.

  • RL setup: GRPO scores rendered slides and Chromium-executed pages using rendered pixels, probe traces, and structured evidence rather than model claims.Deterministic dimensions use rules, while subjective dimensions use vision-language judges.
  • Slide RL: Slide reward combines 0.6 VLM fidelity with 0.4 geometric rules on a 0–5 scale.The rule engine evaluates canvas use, collisions, overflow, occlusion, and chart geometry.
  • Reward reliability: A normalization bug caused fidelity to rise +0.46 while layout fell −0.31, with severely broken layouts reaching 45%; repairing the rule signal restored simultaneous improvement.A 4B checkpoint eventually beat its SFT start on both dimensions.
  • Prompt design: Diverse detailed briefs improved fidelity, whereas purifying prompts to one generator collapsed GRPO as within-group scores converged and advantages approached zero.Filtering short briefs while retaining multiple generator styles restored training health.
  • Cross-modal transfer: A 27B slide-only RL stage improved interactive HTML by +10.4pp over its SFT start, showing transfer between shared modalities.The stage used slide data only, without HTML in the loop.
  • Trade-offs: The hardened run preserves more slide quality than its old-reward twin at 4B, paying a −1.7 slide tax versus −4.0.At 27B, both reward versions pay the same −1.1 tax.
  • Interactive-HTML RL: A screenshot-only reward produced an old-reward game score of 18.8 under the hardened probe, revealing visually convincing but unplayable outputs.The hardened reward raised the released model’s game score to 57.6 and made interactivity a permanent benchmark gate.
  • Final checkpoints: The released 27B model reaches 63.7 on HTML-500, 57.6 on games, and 83.7 on slide-std after a −1.1 serial tax.The 4B serial recipe reaches 61.7 on HTML-500 and 75.1 on slide-std.

5 Evaluation

CogEvol is evaluated on internally built slide and interactive-HTML suites, with rendering validity and executable interactivity treated as explicit requirements. Across benchmark comparisons and human testing, the results show strong task performance, improved reliability after reward hardening, and measurable judge-discrimination trade-offs.

  • Evaluation protocol: The evaluation suite covers slide generation and 500 interactive-HTML cases, with unrenderable slides and hard-failing pages scoring zero.Slides are judged on rendered fidelity and layout, while interactive pages additionally undergo execution probes and composite scoring.
  • Main results: The hardened reward restores CogEvol-27B game performance from 18.8 to 57.6 after the old reward produced fluent but broken interactive pages.The old-reward run scored highest on code and learning pages but collapsed on games, exposing reward hacking specific to playability.
  • Human evaluation: Manual testing reports pages unable to be entered falling from 2/24 to 0/30, with all eight games in the second round playable on entry.The rounds are directional rather than strictly controlled because prompts differ slightly and sample sizes are small.
  • Reward–human agreement: Revised judge prompts improve quality discrimination, raising Pearson r from 0.609 to 0.715 and Spearman ρ from 0.672 to 0.741.Adding viewport and interactivity signals further raises current-prompt Spearman ρ to 0.749, while Pearson r moves to 0.675.
  • Reward–human agreement: The hard-fail gate deliberately disagrees with partial-credit human scoring on non-responsive pages, assigning zero to 18 pages that raters scored nonzero.Its purpose is to suppress reward for visually complete but non-functional outputs, at the cost of lower linear agreement on this subset.

6 Inference Acceleration

CogEvol accelerates both first-pass interactive-page generation and subsequent editing by replacing full-file regeneration with retrieval, component-level decisions, and incremental diffs. These mechanisms reduce cost and latency while preserving end-to-end quality through routing, reconstruction checks, and interaction probes.

  • 6.1 Scaffold Editing for First-Pass Generation: Scaffold editing retrieves a similar historical template, emits component-level edit decisions, and reconstructs the page programmatically instead of generating HTML from scratch.A tiered router falls back to direct generation when the retrieved scaffold is a poor match.
  • 6.1 Scaffold Editing for First-Pass Generation: 74% of output tokens are removed by splitting JavaScript at function granularity and emitting decisions only for functions that change.Unmodified components cost zero tokens and are reconstructed deterministically.
  • 6.1 Scaffold Editing for First-Pass Generation: ∼76% lower output and latency are achieved on the internal test set, reaching 2,999 tokens, 16.0 seconds, and 94% reconstruction success.The oracle upper bound is −82.5%; external reductions range from −37% for games to −88% for vis3d.
  • 6.1 Scaffold Editing for First-Pass Generation: Hard render/runtime gates and automatic interaction probes ensure scaffold gains are measured with objective end-to-end checks.A failed MODIFY is demoted to REPLACE and then KEEP, confining failures to individual components.
  • 6.2 Fast Iterative Editing: Click-to-Locate in the MAIC-UI Harness: MAIC-UI anchors edits through the live preview’s DOM context and applies model-generated unified diffs incrementally.Click-to-Locate supplies the selected element and relevant page context without requiring source-code navigation.
  • 6.2 Fast Iterative Editing: Click-to-Locate in the MAIC-UI Harness: 6.3 seconds and ¥0.37 per edit make MAIC-UI faster and cheaper than Claude Code at 34.3 seconds and ¥1.03 or direct API regeneration at 151.7 seconds and ¥1.68.The comparison uses the same backbone model across three editing tasks; MAIC-UI averages 17.1k tokens per edit.

7 Deploying CogEvol on Domestic Accelerators

The paper adapts CogEvol-27B to Ascend accelerators through precision reconstruction, release-independent operator dispatch, and replica-first serving topology. Application-level behavior matches A800 deployment, while recurrent-state constraints limit some acceleration techniques.

  • Precision and Operator Adaptation: BF16 dequantization and INT8 W8A8 re-quantization reconstruct all 256 Linear layers while keeping tensors within ∼1% of their originals.Noise-Floor Arbitration verifies deviations against each tensor’s measured BF16 rounding floor.
  • Precision and Operator Adaptation: Attention Dispatch Override reroutes linear-attention layers to fused inference attention and replaces the incompatible Mamba radix cache with a chunked cache.The deployment-time dispatch works on an engine release predating upstream architecture support.
  • Serving Topology: TP2 plus independent replicas outperforms the capacity-maximizing topology because intra-TP communication, rather than memory, binds throughput.Raising TP expands running-request capacity from 24 to 192 slots while throughput falls monotonically.
  • Serving Topology: 31% higher throughput is achieved than the capacity-maximizing topology at matched cache hit, with concurrency 64 selected as the interactive operating point.Beyond concurrency 64, throughput rises 34% while TTFT degrades 36×.
  • Recurrent-State Constraints: The GDN recurrent state is non-invertible, so prefix caching and speculative decoding cannot reconstruct it at arbitrary positions.HCD and BSE detect silent state-reuse failures that ordinary forward-output comparison misses.
  • End-to-End Outcome: 500/500 valid parses and mean element counts of 17.4 versus 17.4 show application-level parity between Ascend and A800 deployments.The remaining 3.63× end-to-end gap is decomposed into hardware/software and unavailable speculative-decoding factors.

8 Conclusion

CogEvol combines post-trained models, executable evaluation, reward design, and serving infrastructure to make learning-environment generation measurable and efficient. The conclusion emphasizes executable interactivity, reward coverage, production deployment, domestic acceleration, and open release.

  • General Findings: Executable probes are central to reliability because screenshot-only judging cannot detect visually convincing but unplayable games.The disclosed reward-hacking checkpoint scored highest on code while its games were unplayable.
  • General Findings: Reward design sets the ceiling of each training stage because dimensions absent from the reward do not improve themselves.This conclusion follows the paper’s reward-hacking episode and its broader evaluation findings.
  • Deployment and Access: CogEvol-27B serves production traffic with the OpenMAIC team, while CogEvol-4B is released openly with its weights and editing harness.The deployment and release are presented as routes toward broader classroom access to efficient learning-environment generation.

A Discussion and Future Directions

The discussion frames learner-aware adaptation as a roadmap rather than a validated result and identifies unresolved limitations in scaling, judging, coverage, and cross-page coherence. Future personalization would use multimodal evidence, sensemaking, strategy selection, and brief-level requirements.

  • Learner-Aware Personalization: The proposed four-stage personalization pipeline is not validated end to end and remains future work beyond current environment generation.Its evidence stage includes visual, wearable, and software-telemetry signals from student activity.
  • Learner-Aware Personalization: Brief-level personalization requirements could feed CogEvol’s existing generation interface without architectural change.The proposed pipeline would produce tailored slides and interactive pages in its final stage.
  • Open Problems: The zero-forgetting-tax recipe is validated on 4B only, leaving 27B confirmation pending.This is identified as the first open problem in the current system.
  • Open Problems: The hardened reward scores code playgrounds at 53.6 versus 59.1 for the earlier recipe, and the contribution of real quality loss versus judge calibration remains unclear.The paper also reports surviving language mixing, element stacking, and cross-page theme drift.
  • Open Problems: Per-page quality does not yet compose into per-course quality, so multi-page theme drift may require course-level context or constraints.The proposed remedy is course-level conditioning rather than page-level training alone.

C Interactivity Measurement

The paper measures interactivity by directly operating rendered pages and observing state changes, while excluding controls the harness cannot semantically operate. A hard-fail gate prevents totally inert pages from receiving useful mid-range rewards.

  • Detecting response: Static screenshots cannot reveal whether controls respond, so the probe operates pages directly in Chromium and records resulting changes.The harness enumerates interactive elements and drives them before collecting response evidence.
  • Detecting response: DOM fingerprints alone can misclassify physics simulations because timer-driven DOM changes may occur even when canvas state ignores input.This failure motivated instrumentation of canvas drawing rather than relying only on surrounding-page changes.
  • Detecting response: Canvas 2D response is detected from new drawing signatures after subtracting signatures produced during a quiescent baseline.WebGL response is handled through sampled framebuffer comparisons across an input differential.
  • Reward signal: The interactivity reward combines listener count, post-baseline signature change, traversal response rate, and WebGL activity using a coverage-weighted mean.This credits only the portion of the interactive surface that the probe can confirm.
  • Probe limits: Some semantic gestures are whitelisted and excluded from response statistics because the harness cannot reliably perform them.Examples include dragging a card into a target and completing ordered multi-step gestures; richer primitives or learner-like agents are proposed extensions.
  • Hard-fail gate: Four hard-fail conditions assign zero to totally unusable pages, including confirmed inertness and blank canvases after activation.The gate bypasses the weighted mean because a fully inert page could otherwise remain mid-range and become an optimization target.

D HTML RL Reward Server

The HTML RL reward server renders pages, probes interactivity, and scores outputs concurrently under training and evaluation workloads. Its latency measurements show judge models within 5% on every latency metric, while hard-fail rates reflect model capability rather than server load.

  • Reward serving: Each reward request first renders a page and captures probe traces, then dispatches outputs concurrently to the VLM judge and deterministic viewport pipelines.The server uses four Playwright-driven Chromium workers for the rendering phase.
  • Workload and concurrency: 64 concurrent requests form each RL batch, with double-buffering allowing up to 128 requests in flight.The training loop generates eight rollouts for each of eight prompts and waits for the batch before computing advantages.
  • Throughput and latency: 441–444 s/8 ≈55 s and 644–648 s/8 ≈81 s are the per-step reward latencies for Qwen3.5-4B and Qwen3.8-27B HTML under the RL workload.Both remain within the 120-second training step budget; P99 latency reaches 76–82 seconds.
  • Scoring robustness: Fallback rates are near zero across judges and workloads, with the evaluation workload reaching exactly zero in every run.Fallbacks contribute a neutral default score and flag the request when all judge calls fail or time out.
  • Judge model selection: The two judge models differ by less than 5% on mean latency and at most 2 percentage points on fallback rate.Selection therefore depends on alignment with human ratings and dimension-level scoring consistency.
  • Judge model selection: Qwen3.8-27B achieves Spearman ρ = 0.741 against Qwen3.6-27B’s 0.716 on human ratings under current quality-judge prompts.The advantage is concentrated in text-intensive dimensions.
  • Training interface: The training mixture uses six public system prompts: one slide contract and five interactive-HTML templates.Learning pages reuse the simulation template by design.

Slide content contract

The slide modality requires a single valid 16:9 JSON scene graph on a 1000 × 562 canvas. Its renderer-facing contract fixes element types, geometry, and modality-specific schemas, making structural validity a hard requirement.

  • Output structure: Each slide is emitted as one JSON object containing elements and a solid background.The required top-level shape is {"elements":[...],"background":{"type":"solid","color":"#ffffff"}}.
  • Element schema: Supported elements are text, shape, line, image, table, chart, latex, and video, each with prescribed fields.Most elements require type, id, numeric geometry, and rotate; line elements use a distinct schema.
  • Geometry: Position fields must be JSON numbers on the 1000x562 canvas, while line elements encode endpoints rather than bounding-box height and rotation.Line width is stroke thickness, and line style must be a string.
  • Tables: Tables require cell-level data with spans and exact outline, width, and height fields rather than row- or column-object formats.The contract explicitly rejects alternatives such as data strings, header/rows, colSpan, and rowSpan.
  • Charts: Charts must use one of nine fixed chart types and the exact labels-legends-series data shape.Donut-style charts use chartType "ring", not "donut".
  • Images: Assigned assets are represented by image elements whose src equals the asset id and whose fixedRatio is true.The contract prohibits redrawing assigned images.

F External Flagship Evaluation

External flagships are evaluated under the same harness with either a slim deployment contract or a full 34 KB specification. The results show that explicit contract details can substantially change slide performance, while CogEvol retains a cross-modality reliability advantage under the reported conditions.

  • Evaluation setup: External slide models are tested under slim and full prompts using the same topics, renderer, judge, and normalization pipeline.Outputs that remain unrenderable after normalization score zero; HTML-500 uses production prompts without this split.
  • Prompt conditions: The slim contract is approximately 310 words and omits field-level schemas, examples, and style rules that CogEvol internalizes during training.This is the only slide prompt CogEvol sees in training and production serving.
  • Slim-contract results: Under the slim condition, GPT-5.4 scores 54.9 versus 70.8 for the SFT-only CogEvol-4B reference.GPT-5.4 names the wrong text field for 1,956 of 1,958 text elements and emits none of the required style keys.
  • Full-specification results: Gemini 3.6 Flash gains +58.5 points under the full specification, rising from 19.2 to 77.7.GPT-5.4 gains +15.9, while its layout changes from 50.7 to 51.7 despite fidelity reaching 90.0.
  • Full-specification results: Qwen3.8-Max gains +59.8 points, reaching 83.7 under the full specification, with 5/120 outputs unrenderable.The specification enables contract-clean scene graphs except for token-budget truncations.
  • Cross-modality comparison: Qwen3.8-Max ties CogEvol-27B at 83.7 on slides, but its HTML-500 average is 35.3 with 204 of 500 pages dead at the probe.CogEvol-27B records 63.7 on HTML-500 with zero dead pages under the cited comparison.
  • Cross-modality comparison: Claude Opus 4.8 reaches the best external HTML-500 score of 67.2 with 19 dead pages, while GPT-5.4 reaches 66.0 with 13.No external model leads both modalities in the reported comparison.

G Chinese-Language Production Examples

Figure 8 presents Chinese-language interactive pages and slides drawn from live OpenMAIC production traffic, spanning diverse educational subjects and modalities.

  • Chinese-Language Production Examples: The artifacts were sampled from the same live OpenMAIC traffic as Figure 2 using the same selection criteria.The subject and modality mix mirrors learner demand on the platform.
  • Chinese-Language Production Examples: Figure 8 shows four Chinese-language interactive HTML pages covering chemistry, geography, mathematics, and physics topics.The examples include a Le Chatelier principle particle simulator, terrain explorer, atmospheric-circulation simulator, and function-graph explorer.
  • Chinese-Language Production Examples: The figure also shows four Chinese-language slide decks covering matter classification, Newton’s second law, English openers, and news-story structure.These examples extend across science, English, and writing instruction.
Loading 2608.30968v1…