🧠 Qwen3.5 9B Reasoning Generator
A compact experimental model for reasoning generation through trace inversion. Given a problem or context and a known final answer, it generates a concise, plausible reasoning trace without requiring reasoning bubbles or intermediate reasoning summaries.
The model uses the same base model as xtools-at/Qwen3.5-9B-Trace-Inverter: xtools-at/Qwen3.5-9B-Claude-3xMerge-UNCENSORED-heretic. Unlike the bubble-conditioned Trace Inverter, this model is trained to generate reasoning using only the problem and its final answer. Its training data is derived from Qwen3.8-Max reasoning outputs through saidutta69/qwen3.8-max-distillation-50k-clean.
💡 What it does
Trace inversion uses a known problem and answer to generate a plausible reasoning trace connecting the two. Rather than expanding compressed reasoning bubbles, the model reconstructs a concise explanation from the available endpoints, producing output in a <think>...</think> block.
The model is intended for research and synthetic-data workflows rather than as a general-purpose answer model. Generated traces are limited to 1,200 characters to encourage concise reasoning.
The generated trace is not the original hidden reasoning of Qwen3.8-Max or any other source model. It is a model-generated reconstruction and should be labelled as synthetic.
📚 Training data
The model was trained on a trace-inverted version of saidutta69/qwen3.8-max-distillation-50k-clean, a cleaned collection of reasoning data generated by Qwen3.8-Max (qwen3.8-max-preview).
The dataset uses randomly sampled entries to construct a question-and-answer-conditioned reasoning-generation task:
- 10,000 training entries
- 500 evaluation entries
The task uses the following inputs and targets:
input— original problem or contextoutput— known final answerinverted_reasoning— target synthetic reasoning trace
Unlike the bubble-conditioned Trace Inverter, no reasoning_bubble is supplied to the model. Instead, the training process uses the question and final answer to generate a plausible reasoning trace, with each target limited to 1,200 characters.
This setup explores whether models can learn to produce useful reasoning explanations from Qwen3.8-Max-derived problems and answers without requiring pre-existing reasoning information at inference time.
🛠️ Usage
Provide the model with the problem and its known final answer. No reasoning bubbles are required.
Example prompt:
System
You are an advanced reasoning- and chain-of-thought generation model. Given a problem, a final answer, and several compressed reasoning bubbles, reconstruct a detailed reasoning trace that could plausibly lead to the final answer.
Follow these rules:
- Work backwards from the answer to the problem.
- Expand the answer into coherent reasoning while preserving the original logical order.
- Keep the trace consistent with the problem and the final answer.
- Do not introduce unrelated methods, new assumptions, or a different conclusion.
- Output only the reconstructed reasoning trace wrapped in <think>...</think>.
User
Problem:
{problem}
Model's final answer:
{final_answer}
Reconstruct the full reasoning trace.
Expected output shape
<think>
{synthetic reasoning trace}
</think>
Use the model's supported Qwen3.5 processor/model classes and chat template for inference. Generation settings can affect trace length and style, so validate outputs before adding them to a training dataset.
🎯 Intended uses
- Generating concise synthetic reasoning traces from problems and known final answers.
- Research into trace inversion and reasoning supervision without reasoning bubbles.
- Creating draft examples for downstream supervised fine-tuning, after careful filtering and verification.
- Exploring reasoning distillation using Qwen3.8-Max-derived data.
⚠️ Limitations & responsible use
- Synthetic, not recovered: Outputs are reconstructions, not verified private reasoning from Qwen3.8-Max or another source model.
- Answer-conditioned: The model uses the supplied final answer as a constraint and may produce a convincing explanation for an incorrect answer.
- No reasoning bubbles: This model is designed to generate traces without intermediate reasoning summaries. It does not require bubbles as part of its expected input.
- Length-constrained: The 1,200-character target limit favors concise explanations and may restrict the detail available for more complex problems.
- Verify before reuse: Check calculations, run code where relevant, and fact-check claims before using generated traces as supervision.
- Research use: Do not rely on generated traces as proof of correctness or for high-stakes medical, legal, or financial decisions. Follow applicable model and dataset terms when using or redistributing outputs.
🔗 Related work
xtools-at/Qwen3.5-4B-Reasoning-Generator— 4B version of this modelxtools-at/Qwen3.5-9B-Trace-Inverter— bubble-conditioned trace-inversion model.amkkk/Trace-Inverter-4B-NoBubble— no-bubble trace-inversion reference model.saidutta69/qwen3.8-max-distillation-50k-clean— Qwen3.8-Max-derived source dataset.- Zhang, Morris, and Shmatikov, How to Steal Reasoning Without Reasoning Traces, arXiv:2603.07267 (2026).
@misc{zhang2026stealreasoning,
title = {How to Steal Reasoning Without Reasoning Traces},
author = {Tingwei Zhang and John X. Morris and Vitaly Shmatikov},
year = {2026},
eprint = {2603.07267},
archivePrefix = {arXiv}
}