Paper Explained: Post-Training Leaves Behavioral Shadows on Unrelated Decisions — capability transfer through one unrelated word

Ask a code-tuned model to pick between 'jacket' and 'tie' a few thousand times, and another model can inherit part of its coding gain. A from-scratch walkthrough of Active Taskless Distillation (ATD): the math, the controls, and the limits.

Post-Training Leaves Behavioral Shadows on Unrelated Decisions

What this paper claims

The paper is "Post-Training Leaves Behavioral Shadows on Unrelated Decisions" (Zhang et al., September 2026, arXiv:2609.29233).

Here is the abstract in plain terms. Language models can pass capabilities along through text that has nothing to do with the task. Post-training — the extra training done after pretraining — normally improves a model with task-specific data. Earlier work on "subliminal learning" showed that information about such updates can leak through unrelated generations, but it mostly studied traits or preferences and relied on large amounts of teacher output. The authors introduce Active Taskless Distillation (ATD), which transfers capability using a single word from the teacher per prompt. ATD picks prompts where the public model that teacher and student both descend from is almost exactly torn between two ordinary words, and records which one the teacher picks. In the main coding experiment with Qwen2.5-1.5B, 5,664 one-word teacher responses lift HumanEval+ by 5.34 points over a control that scrambles the prompt–response pairing. Transfer also shows up in science knowledge, commonsense reasoning, and reading comprehension, with positive mean gains on other model generations, sizes, and families.

The student never sees a line of code and still gets better at writing it. Let's work out how that can happen.

An analogy: asking the undecided question

Picture identical twins who grew up with the same tastes. One of them spends a month at a coding bootcamp. The other knows nothing about what was taught there.

The stay-at-home twin isn't allowed to ask "teach me to code." All they can do is ask throwaway questions — "jacket or tie?" — and get a one-word answer.

The trick is to only ask questions that the pre-bootcamp twin (that is, yourself) would find a coin flip. On questions nobody hesitates over ("ramen or a rubber tire?"), the slightly changed twin answers the same way, so you learn nothing. But when the scale is perfectly balanced, even a small shift in weight tips the answer. Collect thousands of those tips, imitate them, and part of what the bootcamp changed rubs off. That is ATD in one paragraph.

In the widget below, the balanced state is two bars at nearly the same height. Notice how little the scores need to move before the ranking flips.

The paper calls these changes in behavior on task-unrelated inputs the behavioral shadow of post-training (§1).

The setup: public ancestor, private teacher, same-ancestor student

There are three players (§3):

• Public model $M_0$ (the ancestor), with parameters $\theta_0$. Anyone can inspect it.
• Private teacher $M_T$, obtained by post-training $M_0$ on a target task. Its parameters are $\theta_T=\theta_0+\delta$, and the update $\delta$ is unknown.
• Student, initialized from the same $\theta_0$.

The teacher is a black box that returns exactly one next token per query, chosen greedily over the full vocabulary. The student never sees target-task data. And the set of queries must be fixed before a single teacher response is observed.

How it works: why probing near a tie pays off

Put two candidate words $a_i$ and $b_i$ after a prompt $x_i$ and take the difference of their logits — the raw scores before softmax (§3):

In words, $m_i$ is the gap between the logits of the two candidate words: positive means the model leans toward $a_i$, negative toward $b_i$, and zero means a dead heat.

Think of $m_i$ as the tilt of a scale: positive leans toward $a_i$, negative toward $b_i$, zero is a dead heat. $z_\theta(w\mid x_i)$ is the logit a model with parameters $\theta$ assigns to word $w$.

A first-order expansion tells us how the tilt moves under the teacher's parameters $\theta_0+\delta$:

In words, the teacher's tilt is roughly the ancestor's tilt plus the push the update $\delta$ gives this particular scale. The term $g_i^\top\delta$ measures how well the update lines up with the direction that would tip the scale toward $a_i$.