Logics-PPT-Qwen3.6-35B-A3B-SFT

Logics-PPT-Qwen3.6-35B-A3B-SFT

Model card for Logics-PPT-Qwen3.6-35B-A3B-SFT. The technical report is under review and a public link will be added when available. Note: This model was previously named Logics-PPT-Qwen-3.6-35B-A3B-SFT and has been renamed to Logics-PPT-Qwen3.6-35B-A3B-SFT on 2026-09-20 for naming consistency.

📰 News

• [2026.09.20] Released Logics-PPT-Qwen3.6-35B-A3B-SFT under the Apache-2.0 license.
• This model is the supervised fine-tuning (SFT) checkpoint for the MoE presentation-generation backbone in our presentation-intelligence pipeline.
• The technical report is currently under review. A citation link will be added after public release.

🔎 Overview

Logics-PPT-Qwen3.6-35B-A3B-SFT is a presentation-generation model built on top of Qwen3.6-35B-A3B, an MoE backbone, for single-page HTML slide generation.

The model is trained for end-to-end presentation generation through a slide-wise HTML generation paradigm within the OpenClaw framework. In our study, Single-HTML is adopted as the final generation setting because it offers stronger robustness and optimization efficiency than multi-page HTML and Python-based generation.

This release corresponds to the SFT stage of our unified training pipeline for presentation intelligence:

1. Cross-validated data synthesis
2. Guided supervised fine-tuning
3. Reinforcement learning with verifiable rewards

This checkpoint is the final pre-RL SFT model for the 35B-A3B MoE backbone.

Why this model?

PPT generation is a long-horizon task that jointly requires:

• document understanding,
• information acquisition,
• content organization,
• layout planning,
• and visually coherent slide construction.

This model is designed to improve presentation generation beyond a generic base model through:

• cross-validated synthetic task and rubric construction
• teacher-enhanced trajectory distillation
• mixed-teacher supervision
• lightweight filtering of low-quality trajectories

The goal of this SFT stage is to substantially improve both:

• benchmark-level presentation quality, and
• basic slide aesthetics before RL refinement.

🧠 Training Recipe

Stage 1: Cross-validated data synthesis

The paper constructs a synthetic PPT-generation training corpus through a five-stage pipeline:

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

Stage 2: Guided SFT with mixed teachers

The model is trained from Qwen3.6-35B-A3B using supervised fine-tuning on distilled presentation-generation trajectories.

Teacher models:

• Qwen-3.8-Max: strong general coding and agentic reasoning
• GLM-5.2: presentation-specialized teacher

Teacher guidance is strengthened through progressively richer prompts:

• Base
• Pro
• Max

Training corpora are combined under:

• Single
• Mixed
• Mixed+Filter

The released model corresponds to the selected mixed+filter checkpoint used as the RL initialization for the MoE backbone.

Selected setting for this release:

• Target backbone: Qwen-3.6-35B-A3B
• Training regime: mixed + filter
• Selected teacher composition: qwen + glm_pro (filter)

📊 Experimental Results

Main SFT results

Reported overall benchmark scores from the paper:

Compared with the base model, this SFT checkpoint improves:

• PresentBench: 64.93 → 74.28 (+9.35)
• GDPVal (PPT subset): 74.01 → 78.17 (+4.16)

These results show that the SFT pipeline produces substantial gains on both the in-domain benchmark and the external PPT evaluation subset.

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
• multimodal layout generation
• long-horizon agentic generation
• supervised post-training for deck synthesis
• MoE backbones for structured generation

It is especially suitable for research on systems that:

• read source materials,
• plan a multi-slide presentation,
• generate one HTML file per slide,
• and render outputs for downstream evaluation or refinement.

⚠️ Limitations

• This is an SFT checkpoint, not the final RL-refined model.
• Although it significantly improves overall presentation quality, post-render visual issues may still remain, including:
  • excessive whitespace,
  • imperfect balance,
  • element overlap,
  • and other fine-grained layout defects.
• The model is optimized for single-slide HTML generation rather than native .pptx output.
• Real deployment still requires the broader agent and rendering stack used in the paper.
• Performance depends on task prompting, reference-material access, and rendering setup.

🚀 Quickstart

Transformers: text-generation smoke test

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Logics-MLLM/Logics-PPT-Qwen3.6-35B-A3B-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 an 8-slide presentation as separate HTML slides for a technical topic. "
        "Use a consistent design system, clear hierarchy, and readable visual layout."
    ),
}]

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.