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

Model card for Logics-PPT-Qwen3.6-27B-SFT. 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-SFT under the Apache-2.0 license.
• This model is the supervised fine-tuning (SFT) checkpoint for HTML-based presentation generation in our presentation-intelligence training pipeline.
• The technical report is currently under review. A citation link will be added after public release.

🔎 Overview

Logics-PPT-Qwen3.6-27B-SFT is a 27B-parameter model for automatic presentation generation. It is built on top of Qwen3.6-27B and trained for single-page HTML slide generation within the OpenClaw agent framework.

Our work targets end-to-end presentation generation rather than next-slide completion. The model is trained to generate one self-contained HTML file per slide, which is then rendered into final visual outputs. We adopt Single-HTML as the generation paradigm because it is more robust and easier to optimize than multi-page HTML and Python-based alternatives in our study.

The training pipeline behind this model contains three stages:

1. Cross-validated data synthesis for presentation-generation tasks and instance-level rubrics.
2. Guided supervised fine-tuning (SFT) with mixed-teacher trajectory distillation.
3. Reinforcement learning (RL) with verifiable visual rewards.

This release corresponds to Stage 2, i.e., the final SFT checkpoint before RL.

What makes this SFT model different?

The model is trained with a supervision pipeline designed specifically for long-horizon PPT generation:

• Cross-validated synthetic data with explicit control over difficulty and validity.
• Mixed-teacher supervision combining:
  • Qwen-3.8-Max for broad coding and agentic reasoning capability.
  • GLM-5.2 for presentation-specific structuring priors.
• Guided teacher enhancement with progressively stronger prompts.
• Lightweight outcome-driven filtering to remove low-quality trajectories such as rendering failures, layout overflow, and content overlap.

This model is intended to improve both:

• content quality of generated presentations, and
• basic visual organization before reward-based aesthetic refinement.

🧠 Training Recipe

Stage 1: Cross-validated data synthesis

We synthesize PPT-generation training instances using a five-stage pipeline:

• source profiling
• task construction
• rubric generation
• principle audit
• rubric tightening

Stage 2: Guided SFT with mixed teachers

For SFT, we distill presentation-generation trajectories from:

• Qwen-3.8-Max
• GLM-5.2

We use guided prompts with increasing strength:

• Base
• Pro
• Max

and combine teacher corpora under different mixing strategies. The released 27B SFT model corresponds to the selected filtered mixed-teacher checkpoint used to initialize RL in the paper.

Selected setting for this release:

• Target backbone: Qwen-3.6-27B
• Training regime: mixed + filter
• Selected teacher composition: qwen + glm_pro + glm_max (filter)

📊 Experimental Results

Main SFT results

Reported overall benchmark scores from the paper:

Compared with the base model, this SFT checkpoint improves:

• PresentBench: 70.45 → 77.57 (+7.12)
• GDPVal (PPT subset): 74.49 → 79.36 (+4.87)

The paper also reports that another mixed-teacher setting reaches the highest PresentBench score of 77.90, while this released checkpoint is the one selected for subsequent RL because it preserves stronger external generalization.

Aesthetic metric snapshot

The paper also reports rule-based aesthetic scores for the selected SFT checkpoint:

🎯 Intended Use

This model is intended for research use in:

• automatic presentation generation
• HTML-based slide generation
• agentic document-to-deck generation
• long-horizon multimodal/code generation
• supervised post-training for presentation intelligence

Typical usage involves a system that:

• reads task instructions and reference materials,
• plans a slide deck,
• generates one HTML file per slide,
• renders the result in a browser-based pipeline.

⚠️ Limitations

• This is an SFT model, not the final aesthetically refined RL checkpoint.
• It improves presentation generation quality substantially, but some post-render visual defects may remain, such as:
  • excessive whitespace,
  • visual imbalance,
  • element overlap.
• The model is optimized for single-page HTML slide generation, not native .pptx authoring.
• Reliable deployment still requires the surrounding agent stack, rendering loop, and task prompt format.
• This release is intended for research and evaluation, not as a guaranteed production presentation system.

🚀 Quickstart

Transformers: text-generation smoke test

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

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

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

messages = [{
    "role": "user",
    "content": (
        "Create a 6-slide academic presentation as separate single-slide HTML files. "
        "The deck should summarize a research topic, maintain a consistent 16:9 layout, "
        "and keep each slide visually balanced and readable."
    ),
}]

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.