---

license: apache-2.0 base_model:

• Logics-MLLM/Logics-PPT-Qwen3.6-27B-SFT tags:
• presentation-generation
• html-generation
• slide-generation
• code-generation
• reinforcement-learning
• grpo
• verifiable-rewards
• agentic-post-training
• multimodal-layout
• openclaw

---

Logics-PPT-Qwen3.6-27B-RL

Model card for Logics-PPT-Qwen3.6-27B-RL. The technical report is under review and a public link will be added when available.

📰 News

• [2026.09.20] Released Logics-PPT-Qwen3.6-27B-RL under the Apache-2.0 license.
• This model is the reinforcement-learning checkpoint built on top of the SFT presentation-generation model.
• The technical report is currently under review. A citation link will be added after public release.

🔎 Overview

Logics-PPT-Qwen3.6-27B-RL is a 27B-parameter model for automatic presentation generation. It is initialized from Logics-PPT-Qwen3.6-27B-SFT and further optimized with GRPO-based reinforcement learning using verifiable aesthetic rewards computed on fully rendered HTML slides.

The model targets single-page HTML presentation generation in the OpenClaw framework. Instead of relying on an LLM judge for visual refinement, this RL stage uses deterministic, rule-based rewards defined on post-render outputs.

RL objective

The RL stage refines slide aesthetics along five verifiable dimensions:

1. aspect-ratio compliance
2. whitespace adequacy
3. element-collision avoidance
4. visual balance
5. font readability

These rewards are computed on rendered slides using a browser-based rendering pipeline with headless Chromium / Playwright. The reward design is intended to provide a transparent, auditable signal for aesthetic refinement beyond text-only supervision.

🧠 Training Recipe

Initialization

The paper selects the RL initialization checkpoint by choosing the SFT model that:

• preserves performance on the GDPVal PPT subset, while
• achieving a high score on PresentBench.

For the 27B dense backbone, the initialization is the selected SFT checkpoint released as:

• Logics-PPT-Qwen3.6-27B-SFT

RL data selection

The paper studies two RL data-selection strategies:

• task-hard
• aesthetic-hard

These are designed to expose the policy to:

• globally difficult presentation-generation tasks, and
• model-specific visual failure cases.

RL optimization

The policy is optimized with GRPO under long-horizon agentic rollouts.

Reported shared RL settings include:

• 8 prompt instances per step
• group size 8
• 64 rollouts per update
• rollout temperature 1.0
• up to 128 actions per trajectory
• total sequence budget 131,072 tokens
• one training epoch

The paper also studies runs with and without an explicit KL anchor to the SFT initialization.

📊 Experimental Results

Verifiable aesthetic scores

The strongest reported 27B RL variant in the paper reaches:

This demonstrates that RL improves several post-render visual properties relative to the no-train baseline and often relative to SFT on the rule-based reward dimensions.

🎯 Intended Use

This model is intended for research use in:

• reinforcement learning for presentation generation
• verifiable post-render reward design
• HTML-based layout generation
• agentic slide generation
• aesthetic refinement of long-horizon LLM outputs

This model is especially suitable for work studying:

• rule-based visual rewards,
• RL refinement after SFT,
• post-render evaluation and optimization.

⚠️ Limitations

• This RL model is optimized for verifiable aesthetic metrics, not solely for benchmark-level semantic quality.
• As reported in the paper, the RL stage primarily improves the five rule-based visual dimensions; users interested in maximum overall benchmark score may prefer the SFT checkpoint depending on use case.
• The model remains optimized for single-page HTML slide generation.
• Real usage still depends on the surrounding agent scaffold, rendering pipeline, and prompt format.
• The paper’s future-work discussion explicitly notes that content-quality reward modeling remains open.

🚀 Quickstart

Transformers: text-generation smoke test

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Logics-MLLM/Logics-PPT-Qwen3.6-27B-RL"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

messages = [{
    "role": "user",
    "content": (
        "Generate a visually clean 5-slide presentation as separate HTML slides. "
        "Keep a 16:9 canvas, avoid overlapping elements, reduce excessive whitespace, "
        "and maintain readable fonts."
    ),
}]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=2048,
        do_sample=True,
        temperature=1.0,
        top_p=0.95,
        top_k=20,
    )

completion = output[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(completion, skip_special_tokens=True))

This is only a minimal text-generation smoke test.
Actual presentation generation in our setup additionally uses:

• the OpenClaw-style agent framework,
• task instruction files and reference materials,
• browser rendering

📚 Citation

The technical report is currently under review.
Citation information will be added after public release.

🙏 Acknowledgements

We thank the Qwen, GLM, OpenClaw, PresentBench, and GDPVal teams and the broader research communities working on agentic generation, presentation intelligence, and reinforcement learning for LLMs.

License

The model weights are released under the Apache License 2.0.