MARGIN: Runtime Confidence Calibration for Multi-Agent Foundation Model Coordination

MARGIN. Runtime Confidence Calibration for Multi-Agent Foundation Model Coordination. By Joss Armstrong, from Ericsson. Published on the eighth of October. No code repository is linked to it.

Picture a coordinator that sends one question to several different foundation models and then has to decide whose answer to trust. Each model reports a confidence, and the obvious move is to lean on the most confident one. The trouble is that a stated confidence need not mean the same thing from one responder to the next, and its meaning can drift as the workload changes. A high score from one model may be worth far less than the same score from another, and may be worth less again once the kind of task shifts.

MARGIN, short for Multi-Agent Runtime Grading via Incremental Normalisation, calibrates those confidences while the system is running. It learns a correction specific to each model from the outcomes it actually observes, as answers turn out right or wrong. Two things it avoids: it does not retrain any of the models, and it does not need a held-out calibration set prepared in advance.

The correction works through confidence bands. Within each band, MARGIN keeps track of a model's recent accuracy alongside its stated confidence, and uses the ratio between the two to adjust what the model reports. Some bands will see few answers, so their corrections are blended toward an estimate made at the level of the whole model. The corrected scores then weight the candidate answers when the group reaches a collective decision.

Evaluation spans code generation, question answering and mathematics, with a pool of eighteen models and a nine-model subset for distribution-shift tests. One finding sets the scene. On BigCodeBench, a model's mean confidence is negatively related to its accuracy, and among pairs where one response is correct and the other is not, picking the more confident responder does worse than chance.

Against five online calibration baselines, each given identical feedback and keeping its learned state across every transition, MARGIN reaches lower expected calibration error after the shift than all five in two code-generation transitions. In a question-answering transition it beats four of them, and the remaining comparison is inconclusive. In separate coordination experiments on code generation, calibration improves how correct responses are ranked, and raises answer-selection accuracy by four point three percentage points on one benchmark and by fourteen points on another, relative to weighting by uncalibrated confidence. That gain appears on two of the three benchmarks.

The author frames the scope carefully. The results support model-specific calibration at runtime for coordination when workloads change, provided that feedback on correctness is available for the models taking part. Under that condition, a raw confidence score stops being taken at its word and starts being read against each model's own track record.