🫧 Qwen3.5 4B Trace Inverter

A compact experimental model for trace inversion. Given a problem or context, a known final answer, and one or more compressed reasoning bubbles, it expands those clues into a detailed, plausible synthetic reasoning trace.

The task follows Jackrong/Trace-Inverter-4B. It uses the same two trace-inversion datasets as amkkk/Trace-Inverter-4B-NoBubble, but this model is bubble-conditioned: reasoning bubbles are part of the expected input. The base model is xtools-at/Qwen3.5-4B-MultiDistill-3xMerge-UNCENSORED-heretic; this checkpoint should not be described as a fine-tune of Jackrong's model.

💡 What it does

Trace inversion uses the visible reasoning summary as a scaffold. The model aims to preserve the broad order and intent of the supplied bubbles while filling in intermediate steps, producing output in a <think>...</think> block. It is intended for research and synthetic-data workflows rather than as a general-purpose answer model.

The generated trace is not the original hidden reasoning of Claude or any other source model. It is a model-generated reconstruction, and should be labelled as synthetic.

📚 Training data

The model uses 11100 training- and 555 evaluation-entries from the following datasets (also used by the NoBubble reference model):

• Jackrong/Claude-opus-4.6-TraceInversion-9000x — about 8.67k rows.
• Jackrong/Claude-opus-4.7-TraceInversion-5000x — about 4.76k rows.

The key fields have been mapped and converted to this dataset: xtools-at/trace-inversion-opus-synthetic

The reasoning bubble is retained in the prompt at inference time.

🛠️ Usage

Include all three ingredients—the problem, the answer, and the reasoning bubble(s)—in the prompt. For example:

# System:

You are a trace inversion 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:
- Treat the reasoning bubbles as the skeleton of the trace.
- Expand each bubble 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}

Reasoning Bubbles:
{reasoning_bubbles}

Reconstruct the full reasoning trace.

Expected output shape:

<think>
{synthetic reconstructed 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

• Expanding compressed reasoning summaries into candidate synthetic traces.
• Research into trace inversion and reasoning supervision.
• Creating draft examples for downstream supervised fine-tuning, after careful filtering and verification.

⚠️ Limitations & responsible use

• Synthetic, not recovered: outputs are reconstructions, not verified private reasoning from a source model.
• Answer-conditioned: the model treats the supplied final answer as a constraint and may produce a convincing explanation for an incorrect answer.
• Bubbles matter: this model is designed to receive reasoning bubble(s); omitting them is outside its intended input setup and may reduce quality.
• Verify before reuse: check calculations, run code where relevant, and fact-check claims before using 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

• Jackrong/Trace-Inverter-4B — bubble-conditioned trace inversion reference.
• amkkk/Trace-Inverter-4B-NoBubble — uses the same upstream datasets but removes bubbles from the inference prompt.
• 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}
}