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

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

Model card for Logics-PPT-Qwen3.6-35B-A3B-RL. 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-RL 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-RL under the Apache-2.0 license.
• This model is the reinforcement-learning checkpoint for the 35B-A3B MoE presentation-generation backbone.
• The technical report is currently under review. A citation link will be added after public release.

🔎 Overview

Logics-PPT-Qwen3.6-35B-A3B-RL is a reinforcement-learning-refined presentation-generation model built on top of Logics-PPT-Qwen3.6-35B-A3B-SFT.

The model targets single-page HTML presentation generation and is optimized with GRPO-based reinforcement learning using verifiable post-render visual rewards.

Rather than using an opaque learned critic or VLM judge for training-time reward assignment, this model is refined using deterministic, rule-based reward functions computed on fully rendered slides.

Verifiable aesthetic rewards

The RL stage optimizes five visual dimensions:

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

These rewards are computed after rendering, using a browser-based pipeline with headless Chromium / Playwright and semantic bounding-box extraction.

This design is intended to make slide-aesthetic refinement:

• transparent,
• stable,
• auditable,
• and directly tied to rendered outputs.

🧠 Training Recipe

Initialization

The RL policy is initialized from the selected MoE SFT checkpoint:

• Logics-PPT-Qwen3.6-35B-A3B-SFT

The paper chooses RL initialization checkpoints by favoring strong in-domain performance while preserving out-of-domain generalization on the GDPVal PPT subset.

RL data selection

Two RL data-selection strategies are studied:

• task-hard
• aesthetic-hard

These strategies expose the model to:

• globally difficult presentation-generation tasks, and
• backbone-specific visual failure modes after SFT.

RL optimization details

The paper trains the model with GRPO using long-horizon rollouts.

Reported shared settings include:

• 8 prompts per step
• 8 rollouts per prompt
• 64 rollouts per update
• rollout temperature 1.0
• top-k = 50
• total sequence length 131,072
• per-turn response cap 8,192 tokens
• up to 128 actions per trajectory
• one RL epoch

The study also compares runs:

• with KL anchor
• without KL anchor

For the MoE backbone, the paper reports especially strong gains in visual-aesthetic metrics.

📊 Experimental Results

Main benchmark results

Verifiable aesthetic scores

The strongest reported 35B-A3B RL variant reaches:

These are the strongest overall reported aesthetic totals for the MoE backbone in the paper.

Human evaluation note

The paper additionally reports a blinded human evaluation comparing post-RL models against their pre-RL counterparts. For the MoE backbone, the post-RL model achieves:

• 82.4% win rate in visual quality
• 71.8% win rate in overall quality

on the reported sample and annotation protocol.

🎯 Intended Use

This model is intended for research use in:

• reinforcement learning for presentation generation
• verifiable reward design for slide aesthetics
• post-render optimization of HTML slides
• MoE-based long-horizon agentic generation
• presentation-intelligence systems

This checkpoint is especially useful for studies focused on:

• visual refinement after SFT,
• deterministic reward shaping,
• and RL for structured multimodal outputs.

⚠️ Limitations

• This model is optimized for visual-aesthetic refinement, not only for benchmark-level semantic score.
• As discussed in the paper, content-quality reward modeling remains an open direction.
• The model is optimized for single-page HTML slide generation, not native PowerPoint editing.
• Real use still requires an external framework for prompting, tool invocation, rendering, and file handling.
• Reported behavior depends on the exact rollout and rendering setup used in the study.

🚀 Quickstart

Transformers: text-generation smoke test

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Logics-MLLM/Logics-PPT-Qwen3.6-35B-A3B-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 polished slide deck as separate HTML slides. "
        "Keep the canvas at 16:9, avoid overlap, maintain visual balance, "
        "use readable fonts, and reduce excessive whitespace."
    ),
}]

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.