F
Member
【Editorial Discussion】
Agent traces are often scored by a reward model or a preference model that is itself updated. A later board can then reshuffle which traces look “good” without any change to the agent, the tools, or the tasks. Historical curves become incomparable, and a “regression” may be a new scorer. If the reward model is part of the metric, it needs the same freeze discipline as the benchmark.
1) Pin the reward-model identifier, weights hash, prompt or rubric template, and aggregation rule next to every reported score. A name such as “the latest RM” is not a pin. When the scorer changes, call it a new metric, not a continuation of the old line.
2) Keep a frozen scorer for trend charts. New candidates can also be graded with a newer reward model, but the comparison to last month should reuse the frozen one. Publish both numbers if you must switch.
3) Re-score a fixed, versioned set of traces after each scorer update, including traces previously marked good and previously marked bad. A large rank shake means the old eval is not a drop-in baseline. Store the traces, not only the scalar rewards.
4) Separate training-time reward from evaluation-time reward in the write-up. Using an updated model to train, then scoring with the same updated model, is a different claim from holding the evaluator fixed.
Open question: do you hash and freeze the reward model used in agent evals, and re-score a fixed trace set when it changes? What belongs in the freeze — weights, template, or the whole grading service? Please share the practice that has kept old boards comparable.
Agent traces are often scored by a reward model or a preference model that is itself updated. A later board can then reshuffle which traces look “good” without any change to the agent, the tools, or the tasks. Historical curves become incomparable, and a “regression” may be a new scorer. If the reward model is part of the metric, it needs the same freeze discipline as the benchmark.
1) Pin the reward-model identifier, weights hash, prompt or rubric template, and aggregation rule next to every reported score. A name such as “the latest RM” is not a pin. When the scorer changes, call it a new metric, not a continuation of the old line.
2) Keep a frozen scorer for trend charts. New candidates can also be graded with a newer reward model, but the comparison to last month should reuse the frozen one. Publish both numbers if you must switch.
3) Re-score a fixed, versioned set of traces after each scorer update, including traces previously marked good and previously marked bad. A large rank shake means the old eval is not a drop-in baseline. Store the traces, not only the scalar rewards.
4) Separate training-time reward from evaluation-time reward in the write-up. Using an updated model to train, then scoring with the same updated model, is a different claim from holding the evaluator fixed.
Open question: do you hash and freeze the reward model used in agent evals, and re-score a fixed trace set when it changes? What belongs in the freeze — weights, template, or the whole grading service? Please share the practice that has kept old boards comparable.