Hint-RFT
Hint-RFT is the proposed use of extra hints during rollout collection to obtain more correct training targets.
Project status: Research proposal/design in the inspected project sources; no implementation or completed local efficacy run identified. This entry describes the source audit of 14 September 2026; historical measurements retain their original dates.
Mechanism
Add a task hint to the generation prompt, verify resulting answers, and distill selected answers into a student that should work without the hint. The lever is easier generation during data collection rather than a changed preference loss.
Implementation and controls
Historical RFT literature notes and the next-arms ranking suggest Track B short prompts as a candidate; HumanEval already includes informative docstrings, so the proposed marginal benefit there is uncertain. A proper arm separates hint content, target selection, and whether the hint is present during training and evaluation.
Evidence and evaluation
The cited project document records this candidate and its intended experiment. It does not provide a completed local result for this method. Published-paper results mentioned by that document are background, not Trainfer measurements.
Limitations and interpretation
Hints can leak answers or teach a behavior dependent on the hint. Compare no-hint heldout performance and account for hint-generation cost. The project record contains a proposed adaptation, not a measured Hint-RFT result.
Sources
- agi: autoresearch/RFT_LITERATURE_NOTES.md — historical revision
3842fd8875ca. - agi: lile/docs/research/NEXT_ARMS_jun02.md — historical revision
3842fd8875ca.