Generating HTML from natural-language specifications is an important application of code large language models (LLMs). Despite producing renderable pages, current models often fail to ensure functionality and visual quality during actual interaction. Existing methods typically target task-specific webpage supervision, leaving demanding interactive artifact optimization underexplored. Using mini-games as a challenging testbed, we expose limitations in interaction experience, reward signal, and credit assignment fidelity that are obscured by shallow webpage interactions. We introduce WebCatalyst, a browser-interaction reinforcement learning framework with three complementary components: a game-level GUI assessor for sustained interaction and evaluation, Visual Anchor for mitigating VLM reward hacking, and Action-Trace Credit for grounding token-level credit assignment in runtime execution traces. Trained on fewer than 1K requests, predominantly for games, WebCatalyst raises the Cookie-Bench Mini Games score from 68.8 to 75.3 and also the WebGen-Bench accuracy from 48.8 to 56.9 on Qwen3.6-35B-A3B, outperforming Claude Opus 4.7 in average rank across evaluated metrics. Further analysis shows that Visual Anchor prevents aesthetic degradation, while runtime-grounded credit assignment accelerates functional improvement. Together, our results show that improving feedback fidelity is the key to substantially advancing interactive HTML generation.
Key result. With 640 training requests, WebCatalyst improves Qwen3.6-35B-A3B’s average rank from 14.0 to 4.0 within the evaluated 20-model pool, with performance comparable to Claude Opus 4.7.
Execute sustained interactions in a real browser. Evaluate state transitions, then aggregate action outcomes and visual judgments into a page-level reward.
§3.1 · Ground judgments in actual execution.02
Visual Anchor
Reward signal fidelity
Close the loophole where easier interactions mask visual degradation. Penalize visual regression relative to a high-quality reference; apply no penalty to comparable or better visuals.
§3.2 · Calibrate the reward against a reference.03
Action-Trace Credit
Credit assignment fidelity
Trace successful actions back to executed JavaScript. Map call stacks to generated tokens, then redistribute the advantage by call depth while preserving its mean.
§3.3 · Assign credit using runtime evidence.
A closer look at sign-aware credit assignment+
Sign-aware, mean-preserving credit assignment
For a positive page advantage, tokens on successful execution paths receive stronger positive updates. For a negative advantage, they receive weaker negative updates relative to untraced tokens.
Normalized token advantages enter the clipped GRPO objective. Successful behavior changes where the signal goes, while the sample-level scale stays intact.
Eq. (7) · Sign-aware modulation and mean-preserving normalization.
Main Results
Training on 640 requests improves both webpage generation and sustained interactive experiences.
Table 1 · Full comparison
Complete benchmark results
20 models / 2 benchmarks
Scroll horizontally to view all metrics →
Table 1. Complete main results: model date, size, all WebGen-Bench and Cookie-Bench Mini Games metrics, average rank, and changes from the base model.
Model
Date
Size
WebGen-Bench
Cookie-Bench Mini Games
Avg. Rank ↓
Start Fail. Rate ↓
Acc. ↑
Content Pres. ↑
User Inter. ↑
Data Mgmt. ↑
Func. Test ↑
Data Display ↑
Design Val. ↑
Aesth. ↑
Func. ↑
Total ↑
Bold marks group-wise best results, with WebCatalyst included in Other LLMs. The shaded rows identify the base model and WebCatalyst. Arrows indicate preferred directions.
Date is YY/MM. Size gives total/active parameters; undisclosed sizes are marked accordingly. Acc., Aesth., and Func. denote accuracy, aesthetics, and functionality. Average rank averages ranks across evaluation dimensions. Δ gives changes from Qwen3.6-35B-A3B as reported in Table 1 ↗.
Analysis
From reward design to interaction quality: what changes with high-fidelity feedback?
01 Training dynamics §4.4.1
Why each reward component matters
Functional gains can hide visual degradation.
With GUI-only rewards, functionality improves while aesthetics declines. The full method improves both, preserving the visual signal that task success alone can miss.
GUI-only · Mini Games aesthetics
81.5→71.3
Initial generation → training step 60
§4.4.1 · WebGen accuracy, Mini Games functionality, and Mini Games aesthetics.
The quality of the visual reference matters.
Self-anchoring helps early, but those gains erode later in training. A stronger GPT-5.5 reference better retains late-stage performance, despite checkpoint fluctuations.
Self Anchor · WebGen accuracy
56.4%→46.7%
Training step 20 → step 70
§4.4.1 · Reference quality and late-stage training performance.
Runtime credit reaches the same accuracy earlier.
Action-Trace Credit reaches 56.8% WebGen accuracy at step 30, compared with step 70 using the same strong anchor without ATC.
Steps to 56.8% WebGen accuracy
70→30
Without ATC → with ATC
§4.4.1 · Observed training checkpoints; this is a training-step comparison, not a wall-clock speedup.
02 Experience diagnostics §4.4.2
Better interactions, across dimensions
All six experience dimensions improve, alongside screenshot aesthetics. Gains also span every measured action type, indicating broader improvements in interaction quality.
Action success rate+19.4pp
48.0% →67.4%
Qwen3.6-35B-A3BWebCatalyst
Experience dimensions
All dimensions improve ↗
Action transitions
More successful interaction ↗
Figure 5 · Experience-dimension and action-transition scores before and after training.
View the numerical results +
Reward components
Dimension
Qwen
WebCatalyst
Visual stability
88.3
94.9
Screenshot aesthetic
75.5
85.1
Feedback clarity
48.7
64.9
Interaction reliability
56.0
69.7
State legibility
76.9
88.9
Prompt alignment
54.0
68.1
Task progress
26.1
42.1
Action transitions
Dimension
Qwen
WebCatalyst
Combo
83.6
85.6
Mouse move / hover
84.1
88.7
Press
76.3
84.2
Terminate
63.2
76.0
Type
93.0
96.0
Wait
74.3
84.5
Success rate (%)
48.0
67.4
03 Iterative refinement §4.4.2
Single-shot gains persist under refinement
WebCatalyst’s initial generation is already comparable to the base model after two refinement rounds. It maintains a higher Mini Games Total at every corresponding round, reaching 76.9 at round 3.
WebCatalystInitial generation
75.3
≈Mini Games Total
75.2
Qwen3.6-35B-A3BAfter 2 refinement rounds
Qwen3.6-35B-A3BWebCatalyst
Table 3. Mini Games performance under iterative refinement.
Table 3. All aesthetics, functionality, and total scores for refinement rounds 0 to 3.
Round
Qwen3.6-35B-A3B
WebCatalyst
Aes.
Func.
Total
Aes.
Func.
Total
0
81.5
56.1
68.8
87.5
63.2
75.3
1
85.1
57.9
71.5
88.1
63.4
75.7
2
87.7
62.8
75.2
89.5
61.8
75.7
3
86.2
62.1
74.2
91.1
62.7
76.9
Round 0 is the initial generation; rounds 1–3 revise the artifact using feedback from the GUI Assessor. Bold marks the better result for each metric within each round.
§4.4.2 · Table 3 · Mini Games performance with GUI-Assessor-guided refinement.
Interactive Demos
Compare original generated outputs for the same prompt. Watch recorded interactions or open the artifacts to try them directly.
3D flight / Prompt 17
Sky Runner
View the original prompt+
Loading the comparison gallery…
Recorded browser interaction · Original generated HTMLSelected examples illustrate behavior; benchmark results provide aggregate evidence.
Explore the examples
Full demo collection
50 prompts × 6 models · 300 generated artifacts
BibTeX
If you find this work useful, please consider citing our paper.
BibTeX
@misc{wang2026catalyzinginteractivehtmlgeneration,
title={Catalyzing Interactive HTML Generation Capabilities via High-Fidelity Reinforcement Learning},
author={Xujia Wang and Yixuan Liu and Shaohan Huang and Furu Wei and Lei Hou and Bin Xu and Juanzi Li},
year={2026},
eprint={2611.05151},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2611.05151},
}