Adaptive Natural-Language Curricula for Specialist Generators

Why People Wanted It

The problem is reachability. A specialist generator can be useful within the part of design space it already knows, yet a narrow scientific goal may lie so far away that its samples almost never reach it. In that situation, asking directly for the final goal can produce no useful distinction among current candidates. So the real question is not simply how to describe the destination; it is how to choose a next step the generator can actually learn from, without losing the connection to that destination.

The Core Move

Let a frozen language model inspect the generator’s current samples and their observable properties, propose an attainable sub-goal connected to the final goal, and tighten that sub-goal as the generator improves. The paper states this method as an online sequence of goal regions, $\Omega_{g_0} \rightarrow \Omega_{g_1} \rightarrow \cdots \rightarrow \Omega_g$: each region is a stage on the way toward the final goal.

Why the Goal Depends on the Current Generator

Here is the first deeper layer: a sub-goal is useful only relative to what the current generator can produce. At a round, the generator supplies a batch of designs; observables for those designs give the teacher a view of the region the generator currently covers. The teacher uses that view to propose a scientifically connected, attainable objective, rather than applying the final goal blindly.

The paper gives the key failure condition as negligible probability mass on the final discovery region, written $p_{\theta_0}(\Omega_g) \approx 0$. In words, the initial generator almost never samples from the region that satisfies the goal. Under that condition, direct judgments against the final goal may not distinguish the current samples well enough to guide adaptation. The curriculum therefore begins with a broad, useful relaxation and tightens it as the student changes.

Why Observables Are a Capability Check

Now the more technical point: the teacher does not infer capability from the goal description alone. It receives the current batch and its observables, which are presented in a domain-appropriate way to make the generator’s currently generable region legible. The batch is evidence about what the student can reach now; the observables are the interface through which the teacher assesses it.

A proposed sub-goal is accepted only when it yields enough positive and negative candidates in that batch to support discrimination. That makes attainability operational: the goal should be connected to the final objective, but it also needs to separate examples the current generator actually produced. The exact procedure for proposing and validating goals is an implementation choice, not a required fixed schedule.

Why Progression Is Online Rather Than Prewritten

Here is the final layer of the mechanism: the curriculum is revised in response to the generator’s changing samples, not set once as a timetable. The protocol retains a current sub-goal while the generator improves under it, then tightens toward the final goal. In the paper’s implementation description, the teacher first checks whether the final goal is informative on the current batch; if not, it selects a relaxation that separates the samples.

That is especially relevant when the desired region is poorly covered or out of distribution for the initial generator. The progression provides intermediate objectives instead of demanding an immediate jump to that region. It is not a promise that every proposed rung is scientifically necessary; the stated goal is plausible progress while preserving a connection to the final target. What should happen if the generator cannot make progress on the current rung?

Minimal Causal Code

This plain loop shows the curriculum state changing in view of the current generator. choose_subgoal stands for the teacher’s proposal and validation; the update itself is left abstract because this module is about goal selection, not the preference-distillation objective.

# Conceptual PyTorch-style loop; the state update is intentionally visible.
current_goal = final_goal

for round_idx in range(num_rounds):
    samples = generator.sample(batch_size)
    observed = observe(samples)

    # Select a goal that fits the current batch and remains connected
    # to the immutable final goal.
    current_goal = teacher.choose_subgoal(
        final_goal=final_goal,
        previous_goal=current_goal,
        samples=samples,
        observables=observed,
    )

    labels = teacher.assess(samples, goal=current_goal)
    generator = update_from_feedback(generator, samples, labels)

What You Gain, and What You Lose

You gain a goal-selection process that responds to the generator’s present capabilities and can create intermediate learning signals when the final target is too distant. The sub-goal is not arbitrary: it must remain scientifically connected to the final goal and constitute plausible progress toward it.

You also add dependence on the teacher’s assessment of samples and observables. If the observables do not make the current design region understandable, or a proposed goal does not separate the batch, the curriculum cannot provide the intended useful distinction. And the progression is adaptive rather than a guaranteed route: the method describes how to select and tighten objectives, not a guarantee that every specialist generator will reach every final goal.