Build1 publisher2 min readPublished
One-word answers leak a fine-tuned model's coding skill to a student
Researchers from five institutions lifted a student model 5.34 points over a control on HumanEval+ using 5,664 one-word answers from a code-tuned teacher. None of those answers contained code, so a filter that screens outputs for task content would have passed every one.
The Engineer · Build desk
What happened
- Researchers asked the coding-tuned teacher to pick one word on ordinary prompts where the original model split roughly 50/50, such as jacket versus tie.
- The student was a fresh copy of the original model and never saw the teacher's coding examples, weights, probabilities or any long answer.
- Comparison students trained on the same words reassigned to different prompts, and the gain's 95% confidence interval ran from 1.22 to 9.60 points.
- Evidence of the effect also appeared in science, commonsense reasoning and reading comprehension, and on other Qwen generations, sizes and Llama.
- Code-trained and science-trained teachers each produced their largest gains on the matching kind of task.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- exposure Teams that post-train an open-weights base and accept outside queries match this setup, since anyone holding the same base can find the undecided prompts.
- cost Collecting the teacher's side of the dataset takes a few thousand one-word completions, far less output than the worked answers conventional distillation trains on.
- precedent Earlier subliminal-learning work moved behavioral traits; a coding gain scored by executed programs adds capability to what unrelated outputs have been shown to carry, in one small-model study.
A prompt balanced that finely is sensitive to small changes in the model. If post-training nudged the teacher even slightly, its answer might flip [5]. Each query returns one ordinary word [6]. It is an odd way to learn to code. The researchers call what leaks through this channel the "behavioral shadow" of post-training [3].
The control is the part I would defend at review. Had the student been compared with an untouched model, ordinary fine-tuning could explain the gain. With shuffled pairings, both groups saw essentially the same ingredients, and the only difference was whether each word stayed attached to the prompt where the teacher chose it [15]. HumanEval+ also runs the generated programs, so the score counts code that works [9].
The interval is wide. It spans 8.38 points, more than the 5.34-point estimate it brackets [10][1]. The main run used Qwen2.5-1.5B-Instruct, a 1.5-billion-parameter model [7]. For the gain to carry over to a production fine-tune, it would have to hold on models many times that size [7].
The constraint that limits who can run this sits in prompt selection. Finding prompts where the base model is split near 50/50 means running the base model [2]. In this setup the base was public and only the post-training was private [4].
The Neuron's account of the paper, by researchers at Peking University, Georgia Tech, ShanghaiTech University, Tsinghua University and Lovart AI, conditions its reading on the result holding up [1][14]. On that condition, it says models may reveal and potentially transfer more of what they learned than their responses let on [14]. The evidence supports a narrower claim. The researchers picked the prompts for the purpose and sent them straight to the teacher [5]. Every result in the summary comes from that curated setup, not from answers to ordinary user traffic [5][11].
What to watch
- Results where teacher and student start from different base models, which would show whether the probe works without a shared starting point.
- Replications on models well above 1.5 billion parameters, and whether the interval's lower bound moves clear of about one point.
- Tests that draw probe prompts from ordinary traffic in place of prompts curated for base-model indecision.