This is a joint work by @aardauzunoglu @ZAlvin39105, and supervised by @DanielKhashabi. We also greatly appreciate Tianjian Li and Jack Zhang, as well as the JHU DSAI community for their valuable discussion and feedback!

Paper: https://arxiv.org/abs/2606.01000

Finally, we make data selection risk-controlled.

In practice, we often don’t know how noisy the target pool is. With a small calibration set, we can choose a trust threshold with statistical guidance, targeting a desired noise tolerance among selected weak labels.

Our L2T framework trains a trust function on labeled source data (importantly, not on the target data) to predict the reliability of weak labels from the teacher’s hidden states.

We then use it to score weak labels, keep high-trust examples, and train the strong student.

Weak-to-strong generalization trains a strong student on labels from a weaker teacher when reliable labels are scarce. But weak labels aren’t all equal.

We frame W2S as a data selection problem and ask, "Which weak labels should the student trust?"

L2T framework yields near-lossless W2S generalization.

Across world knowledge, quantitative reasoning, and strategy games, students trained on high-trust weak labels often match (and sometimes surpass) students trained on ground-truth labels.

More results in the paper!

The gains can also compound.

After training a student on high-trust weak labels, we can reuse it as the next teacher and leverage our L2T framework again.

This weak-to-strong chain produces stronger teachers over multiple rounds, leading to snowballing improvements.

Why do trust functions help?

Trust-based selection tends to:

• select easier examples • recover strong alternatives when ground truth is suboptimal • produce more aligned gradients

Together, these make weak supervision much more useful for strong students.