Jump to text
On this page
Copy and source
Published source confirmed at last check

Source changed 2026-11-22 15:54:17 UTC · snapshot created 2026-11-22 15:55:11 UTC · last check 2026-11-22 16:10:10 UTC

Engineering.Reference:26.4.1.3 - Obtain a separate reason to credit a call

If the work needs a positive assessment of B’s first call, decide what the assessment means and obtain the evidence it needs. The table gives three alternatives.

Available groundConstruction and first usable resultWhat remains to establish
The domain requires obtaining the product n*m, and the observed return can be checked against it.Compare each actual return with n*m. The call-subresult labels are (1, 1, 0, 0). Use these as a separately qualified call-only target.Obtaining the product establishes neither the final answer nor the necessity of using a calculator. A legitimate different solution may not need this subresult.
An independently qualified next-call annotation is available at the decision history.Construct (history, permitted tool definitions, gold call set). For this exercise the declared method supplies calculator(expression="17*19"). Both A and B match; C and D do not.Establish that the annotation fits the chosen method and available input, including legitimate alternative calls. A final reward does not supply this gold set.
A qualified transition and continuation model is available.Evaluate each first call with the old continuation under MMP.8.SD:4.2 through MMP.8.SD:4.4. Here multiplication’s success probability is 0.5 and addition’s is 0; the old mixture is worth 0.25. Subtracting that baseline yields call credits (+0.25, +0.25, -0.25, -0.25).These values require the supplied deterministic law, all available action combinations and the old second-choice probabilities. Sparse stochastic logs alone do not establish them. A changed continuation requires reevaluation.

The first two rows assess a call against a checked subresult or a qualified annotation. The third compares its expected consequence under a specified continuation. None overwrites B’s recorded final failure. The baseline subtraction and the choice to update only the first logit are receiving choices. Starting again from the zero logits, a call-only mean of probability-ratio terms using those continuation credits has first gradient 0.125 for z1 and zero for z2. At rate 0.3 it gives z1=0.0375; the answer decision still needs its own learning signal.

ToolRL, NeurIPS 2025, §§2.1–2.4 and Appendix C supplies an annotated-call branch. It turns multi-step records into next-step tasks with preceding history, compares predicted and gold call sets using maximum matching over tool names, parameter keys and values, and normalizes the attainable score before scaling correctness to −3 through +3 and adding format reward. Group-relative advantages feed clipped policy optimization. Thus the first engineering input is a history with qualified call annotations. Use those annotated histories with the source’s scoring rule and trainer. If only the four terminal rewards are available, either construct and qualify the additional annotations or continue with the outcome branch; silently calling those rewards gold calls would change the evidence. Call agreement still needs a separate check of the resulting effect in the receiving work.