TraceForge Distills Reasoning in the Student Model’s Own Style

Introduction

Distillation often follows a simple recipe: a teacher model supplies answers and reasoning, while a smaller student learns from both. That recipe becomes harder to use when a closed model returns only a final answer or a short explanation rather than a detailed reasoning trace. TraceForge takes a different route. Instead of attempting to recover the teacher’s hidden chain of thought, it asks the student model family to generate a plausible trace of its own.

How TraceForge works

The system receives the task context and the teacher’s answer. A generator then writes a reasoning trace in the style of the target student family. Training has two stages:

• Supervised learning: the generator learns to reconstruct reasoning traces associated with its own model family from context and answer.
• Adversarial learning: a critic learns to distinguish realistic family-style traces from less convincing candidates, providing a signal for improving the generator.
• Candidate selection: several traces can be produced for the same example, after which the critic selects the one that appears most authentic.

This is not an attempt to reproduce the teacher’s private computation step by step. The teacher contributes the answer, while the reasoning text is synthesized by a model trained in the student’s ecosystem. The distinction matters: the approach is designed to make training data more compatible with the student, not to establish that the generated explanation was the teacher’s actual reasoning.

What the experiments show

The paper evaluates the method on five benchmarks while holding other training conditions as constant as possible. In an authenticity assessment, TraceForge traces received an average score of 0.68, compared with 0.66 for real traces. Selecting from four generated candidates with the critic raised the score to 0.82. These numbers measure how genuine a trace appears; they are not accuracy rates and should not be read as proof that the reasoning is correct.

The clearest gains came from coding experiments. Using the same 5,000 questions and the same teacher answers, training with TraceForge traces improved HumanEval+ by 3.7 points and MBPP+ by 7.1 points at the final checkpoint. The MBPP+ advantage was reported at every checkpoint. The authors suggest that traces written in the student family’s native style may be easier for that student to learn from than traces copied from another model’s reasoning conventions.

Why it matters—and what remains unresolved

TraceForge reframes reasoning distillation. The goal is no longer to obtain a teacher’s hidden thoughts, but to construct supervision that a particular student can absorb efficiently. This could offer a practical path for training from closed-model answers without relying on methods that attempt to recover protected internal traces.

There are important caveats. A generated trace can be a post-hoc explanation rather than the process that produced the answer. A high authenticity score only indicates stylistic or distributional similarity, not logical validity. The reported results also center on the supplied benchmarks, especially coding tasks. More work is needed to test whether the approach transfers to mathematics, factual question answering, and longer multi-step problems, and to determine when student-style reasoning improves learning rather than merely improving its appearance.

Source: Hugging Face Daily Papers