The Battle Between Raw Traces and Compressed Efficiency in AI Reasoning Models

In the area of AI reasoning models, the debate over raw trace accuracy versus compressed efficiency is heating up. Two prominent models, Qwen3.5-397B-A17B and gpt-oss-120B, have generated a staggering 283,000 correct reasoning traces each. The question is, at what cost?

Compressing the Complexity

These models aren't just churning out verbose outputs. they're creating long chain-of-thought traces that are expensive to distill. Enter the saviors: instruction-tuned models capable of compressing these traces down to a mere 8.6-21.0% of their original character length. That's a significant trim, but does it compromise the accuracy?

Across a strong 48-run main grid and additional Qwen-teacher truncation tests, the compression was shown to reduce training tokens to just 12-30% of their raw counterparts. This means speedier training, up to 7.6 times faster in some cases, and much shorter inference outputs. But, of course, the smaller the model, the less dramatic the reductions.

Raw Versus Compressed: A Trade-Off

Here's the kicker. Despite the efficiency gains, raw traces still hold the crown for downstream accuracy. For every model size and teacher, they outperform compressed versions. So, is compression just a numbers game? Not quite. When matched for length, model-compressed traces often outperform naive truncation, especially for smaller models, without the verbosity.

So what's the real takeaway here? Reasoning-trace compression isn't a magic bullet. It's an accuracy-efficiency trade-off. Students retain up to 96% of raw-trace accuracy while offering up to 18 times higher per-token efficiency. That's a substantial figure, but as the 0.8B scale under LoRA suggests, compressed traces can't quite close the gap with raw traces, let alone surpass them.

The Future of AI Reasoning

What does this mean for the future of AI reasoning models? It's a clear reminder that slapping a model on a GPU rental isn't a convergence thesis. Developers must weigh the benefits of efficiency against the cost of accuracy. As models grow, the stakes will only get higher. If the AI can hold a wallet, who writes the risk model?

In an industry obsessed with faster, cheaper, and better, this trade-off will define the trajectory of reasoning models. Show me the inference costs. Then we'll talk about what's truly worth pursuing.