Engineering.Reference:26.4.6.2 - Turn selected guidance into a qualified learning target
Choose the recurring operation to be learned. The two guide files a and b can supply the same interpretation, so their separate scores depend on which other guide is present. Retain one while developing that interpretation; keep the current range read and arithmetic execution. Recover a useful action at a history where the intended learner has the inputs that warrant it. Qualify the action against the unit definition and required report. A guide may help a teacher produce that target; its presence does not qualify every resulting action.
SIRI v1, §3 offers one selective construction: induce guidance from interaction experience, compare assisted and unassisted group returns, and use positively credited guided actions as targets for the policy without that guidance. Its group benefit covers the retrieved bundle. Give the student the history that led to the decision and the observations available then. Skill0.5 v1, §3 offers a hybrid: retain particular guidance while using task difficulty to choose guided demonstrations or outcome learning. An empty set of successful demonstrations supplies no imitation target. Choose the loss and probability domain through 26.4.5; a truncated distribution or implementation-specific reward needs its own meaning.
The finite fixture uses milliseconds and minutes equally often. For milliseconds, output 1 selects division by 1,000 and output 0 leaves the amount unchanged. For minutes, output 1 leaves the amount unchanged and output 0 selects multiplication by 60. Qualified targets are therefore (1,0). For these positive inputs, the two available actions give different reported values. Selecting the specified action is therefore necessary to return the correct report. The amount remains an input to the retained arithmetic operation. This representation fixes which interpretation can change; new unit conventions need a new applicable definition or another representation.
Let p_j = sigmoid(theta_j) be the probability of output 1 in input class j. With equally weighted qualified targets y=(1,0), minimize
L = -[log(p_0) + log(1-p_1)]/2, with dL/dtheta_j = (p_j-y_j)/2.
Starting at theta=(-ln 4, ln 4), both interpretations have probability 0.20 of the correct action. Four updates with step size eta=4 give:
| Update | theta_0 | theta_1 | Correct completion without either guide |
|---|---|---|---|
| 0 | −1.386294 | 1.386294 | 0.200000 |
| 1 | 0.213706 | −0.213706 | 0.553224 |
| 2 | 1.107258 | −1.107258 | 0.751617 |
| 3 | 1.604023 | −1.604023 | 0.832580 |
| 4 | 1.938863 | −1.938863 | 0.874227 |
The table gives exact report accuracy for the stated positive durations with the remaining operations reliable. If a different commission includes zero duration, both conversions return zero there. Compute results under its actual mix. For a zero fraction z in each unit class and the same action probabilities, report accuracy becomes z + (1-z) * success(theta). To assess selection of the prescribed operation itself, use the action criterion even when the reported numbers coincide. The fourth update meets the stipulated 0.85 floor without a or b. It remains below the unchanged aided performer’s supplied 0.90 rate. Preserve both arrangements for the whole-result comparison; satisfying a floor does not settle whether a loss of accuracy is worth a saving in support. The trained performer with a guide is another joint arrangement to assess.
The target also determines which trace positions enter the loss. Consider a trace containing a user’s request, the assistant’s current-rule call, the returned rule and the assistant’s report:
| Trace position | Action-imitation output target | Whole-sequence prediction target |
|---|---|---|
| User request | Context | Predict the recorded request |
| Assistant call | Selected generated action | Predict the recorded call |
| Observed rule | Context for the next action | Predict the recorded return |
| Assistant report | Selected generated action | Predict the recorded report |
daVinci-Dev, ICML 2026, §§2–3 uses full-sequence mid-training followed by action-masked fine-tuning. Such a target switch changes what the fit means. A failed attempt can supply a recorded sequence without supplying actions to imitate as successful conduct. Retain whether its context was observed during the attempt or reconstructed with hindsight. The report needs evidence of the rule in force for this request. Predicting the recorded rule in training cannot supply that evidence. The finite update here fits the two selected interpretation actions.