On recurrence-heavy INSECTS, the CJSD gate held 13 experts, compared with 18 to 52 for exchange-based decisions; on Covertype, it maintained one to two. The decision layer treats reuse, spawning a new expert and waiting for more evidence as separate choices on changing data streams.
The evaluation covered a four-regime synthetic stream with ground-truth mapping IDs, Electricity, Covertype, and abrupt and incremental-reoccurring INSECTS streams with documented change points.
A gate that can wait
At the center is conditional Jensen–Shannon discrepancy, or CJSD, a statistical measure used here to compare tasks along two axes. A comparability check is followed by one-sided tests for reuse and spawning; in the indifference zone from tau to three times tau, neither exit is triggered and the system defers.
For each expert, betting-style e-processes—running evidence scores—update from point-by-point loss differences. The discriminators used to score those differences are frozen before use, and the paper claims finite-time anytime validity for the observable surrogate discrepancy generated by that predictable sequence.
For population CJSD, the paper describes a one-sided transfer with discriminator-specific excess-risk slack. Exact zero-slack conservativeness is conditional on threshold correction with a valid excess-risk bound or on the stated downward-bias regularity.
The benchmark results varied by stream
With routing equalized, accuracy across adaptive policies on the synthetic stream ranged from 0.939 to 0.945. The CJSD gate alone recorded zero false spawns and zero missed concepts at the ideal expert count; under gradual drift, its spawn rate was 0.10, compared with 0.16 to 0.22 for loss and CPD policies and 1.0 for a policy that always spawned.
The default streaming-native configuration reported 0.85 accuracy on the synthetic stream, versus 0.91 for the batch gate. False-spawn and false-reuse rates were both 0.00 at one expert, while accuracy on recurrence-heavy INSECTS was 0.675.
In an instance-accounted restarted-bank comparison, post-switch mis-reuse and false-spawn rates were both 0.00 on the synthetic stream. Accuracy there was 0.846, versus 0.856 for the windowed heuristic; reported accuracy changed from 0.616 to 0.675 on INSECTS-reoccurring and from 0.770 to 0.790 on Covertype.
In a three-seed synthetic ablation, reported sensitivity was localized to the threshold tau rather than window length. At tau = 0.01, combined false-spawn plus false-reuse was 0.01 with 1.3 experts; at tau of 0.03 or higher, combined error was 0.50. Window length was inert across four to 32 chunks, and no uncertainty intervals were reported for the grid.
At a controlled drift boundary near the threshold, repeated confidence-interval peeking committed 64% of runs to near-coin-flip decisions, while valid monitoring schemes deferred in 90% to 97% of runs. On clear cases, reported delay factors were 1.8 to 2.6, and the e-process false-alarm rate was 0.00 when delays were 20% above the naive monitor.
The design has a defined boundary
The paper frames expert pools around discrete recurring concepts and marks continuous drift as a boundary of applicability. On continuous-drift Electricity specifically, all reported accuracies were 0.74 to 0.78, and the paper reported no benefit from an expert pool.
For the restarted bank, the paper states that family-level anytime validity is preserved with memory that grows logarithmically with time, written as O(log t). It also reports an alpha-spending scheme for familywise error control across arbitrarily many expert-creation events without a capacity cap.
The paper reports discriminator computation at about three times the cost of loss-trigger baselines. In a benchmark comparison on 33-dimensional INSECTS, predictable top-k shortlisting was reported at a 5.9-times speedup with no measured decision change; switching lagged by one mini-batch rather than one chunk.
The supplied document is an arXiv version 1 dated 20 Aug 2026. Full benchmark sample sizes, chunk counts, run counts and uncertainty intervals are not reported; the numerical findings are limited to the named streams, regimes, policies and reported runs. The paper leaves sharper-than-union-bound multiplicity control and finite-sample theory for the underlying estimator open.
Paper data and sources
Original title: Evidence Before Expansion: Reuse, Spawn, or Defer in Lifelong Expert Pools
Authors: Kentaro Oda
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text