PaperlayerView exact revision on arXiv

Source-linked AI summary

Coverage You Can Steer: Online Conformal Calibration for RL-Driven Hardware-Aware NAS

Pedro Brandimarte, Nerea Aranjuelo, Marcos Nieto, Oihana Otaegui

arXiv:2610.03127v1cs.LG

TL;DR

Hardware-aware NAS is costly because candidate rewards require training, while RL-driven proposal shifts violate the exchangeability assumption behind static conformal filters. The paper uses online conformal feedback control and its variants for filtering and acquisition; across three families and both search formulations, coverage tracks requested levels while pruning evaluations, and the bound also beats random search on one constrained testbed.

  • Problem

    Static conformal filtering assumes exchangeability between calibration and test candidates, but RL policy improvement and sequential depth shifts make the search stream non-exchangeable.

    Source: §1 · p. 1, §3.3.3 · p. 12

  • Method

    The method replaces static calibration with online adaptive conformal feedback for reward upper bounds, extending it with tuning-free, locally-adaptive, and group-conditional variants.

    Source: §3.3.1 · p. 11, §3.6 · p. 17

  • Results

    Adaptive filtering tracks requested coverage within ∼10^-3 across three architecture families and both search formulations while pruning 25–50% of evaluations at no measured solution-quality cost.

    Source: §5 · p. 28

  • Takeaways & Limitations

    The calibrated optimistic bound can also guide proposals, with greedy acquisition beating random search on one constrained testbed.

    Source: §5 · pp. 27–28

    Per-depth Mondrian coverage depends on enough observations in each group, and some structurally difficult groups remain under-covered even when well-sampled.

    Source: §5.1 · p. 29

Abstract

from arXiv · show

Hardware-aware neural architecture search (NAS) is dominated by evaluation cost: every architecture must be trained before its reward is known. Conformal-prediction filters cut this cost by pruning candidates whose predicted-reward upper bound misses a threshold, with a distribution-free guarantee that at most a fraction $\delta$ are wrongly discarded. That guarantee assumes exchangeability between calibration and test candidates, which the surrounding reinforcement-learning (RL) loop violates: the policy's proposals improve as search proceeds and, in layer-by-layer construction, shift within every episode. We replace one-shot quantile estimation with online feedback control (Adaptive Conformal Inference, with tuning-free, locally-adaptive, and group-conditional variants), restoring steerable coverage: dialing the target delivers it, monotonically and reproducibly, for arbitrary sequences. Across three neural-network architecture families and both single-step and sequential search (three seeds), it tracks every requested level to within ${\sim}10^{-3}$ while pruning 25-50% of evaluations at no measured accuracy cost, whereas static calibration loses control of its coverage and a Gaussian-process baseline stays conservative regardless of the request. Finally, used as an acquisition function on one constrained testbed, the same optimistic bound beats random search, a gain that fixed optimism already carries and online calibration sharpens. The source code is available at https://github.com/Vicomtech/rl-hw-nas.

1 Introduction

Hardware-aware NAS must balance accuracy and device constraints, but candidate evaluation is costly and reward is only observed after training. Static conformal filtering loses its exchangeability premise as RL proposals shift, motivating online calibration that both steers pruning and can direct search.

1 Introduction

  • Every candidate must be trained before its reward is known, making evaluation the dominant NAS cost in combinatorial hardware-aware design spaces.The objectives include accuracy, memory, latency, and quantization precision.Source: §1 · pp. 1–2
  • RL policy improvement shifts candidate distributions away from the warm-up calibration distribution, while sequential construction adds within-episode depth shifts.Static coverage can drift from the requested level, causing wasted evaluations or false discards of strong candidates.Source: §1 · pp. 2–3
  • Online Adaptive Conformal Inference replaces one-shot quantile estimation with feedback updates, with tuning-free, locally-adaptive, and group-conditional extensions.The group-conditional construction targets coverage at each layer depth rather than only across the stream.Source: §1 · p. 3
  • The adaptive filter tracks requested coverage within ∼10^-3 across three architecture families and both search formulations, pruning 25–50% of evaluations at no measured solution-quality cost.The aggregated filter needs no step-size tuning on these streams.Source: §1 · p. 3
  • The calibrated optimistic bound also serves as an acquisition function, and greedy acquisition beats random search on one constrained testbed.Fixed optimism carries most of the gain, online calibration adds an edge, and tree search provides only an early-budget head start.Source: §1 · pp. 3–4
  • The reproduced RL controller does not outperform random search in these spaces; the paper locates its method’s value in calibrated evaluation-budget control and acquisition.The Gaussian-process upper confidence bound remains near 0.99 realized coverage regardless of the requested level, so it cannot be steered.Source: §1 · p. 4

2 Related Work

Related work spans hardware-aware search, evaluation-cost reduction, learned controllers, Bayesian optimization, tree search, and conformal prediction under shift. This paper focuses on online calibration for RL-induced drift and uses the resulting bound for both filtering and search.

2.1 Hardware-aware neural architecture search

2.2 Reducing the evaluation cost

  • Evaluation-cost methods include partial training, weight sharing, learned predictors, zero-cost proxies, and multi-fidelity allocation, while conformal filtering decides whether to evaluate at all.The filter wraps a predictor in a conformal envelope; better predictors can tighten bounds without changing the guarantee argument.Source: §2.2 · p. 5

2.3 Reinforcement-learning controllers, and how strong random search is

  • Random search can match or beat learned NAS controllers on moderate spaces, and the reproduced MARCO controller optimizes expected reward while search rewards the argmax.This motivates locating the contribution in controller-agnostic budget management and argmax-aligned acquisition.Source: §2.3 · pp. 5–6

2.4 Bayesian optimization and optimism-based search

  • GP-UCB supplies optimism under correct model specification, but its model-based uncertainty can miscalibrate under shift; here its realized coverage stays near 0.99 at requested levels.The paper distinguishes RL policy improvement and within-episode depth shift from acquisition-induced feedback shift, and uses online feedback control.Source: §2.4 · p. 6

2.5 Tree search for architecture construction

  • Prior NAS tree-search methods use learned predictors or space partitions, whereas conformal-optimistic MCTS values simulated completions with calibrated bounds and trains only committed candidates.The construction retains an AlphaZero-style structure with controller priors and PUCT value backups.Source: §2.5 · pp. 6–7

2.6 Conformal prediction on non-exchangeable data

  • Conformal prediction methods offer marginal, locally-adaptive, or group-conditional coverage, while ACI controls long-run coverage on arbitrary sequences through online updates.The paper applies online conformal inference to NAS, measures and corrects depth-conditional coverage, and uses calibrated bounds as an acquisition function.Source: §2.6 · p. 7

3 Methodology

The methodology couples a MARCO-style reinforcement-learning search controller with a conformal filter that prunes candidates before training. It replaces static calibration with adaptive coverage control and also uses the calibrated bound to steer search.

3 Methodology

  • The target problem is hardware-aware edge NAS, where candidate training or partial training is the dominant evaluation cost.The controller follows the MARCO style, and the conformal filter prunes unpromising candidates before training.Source: §3 · p. 7
  • Adaptive conformal inference replaces the static split-conformal filter whose exchangeability assumption RL policy improvement violates, with tuning-free, locally-adaptive, and group-conditional variants.The approach maintains a distribution-free coverage guarantee on the non-stationary candidate stream.Source: §3 · p. 7
  • The calibrated machinery also turns the pruning bound into an acquisition function that directly steers search.The methodology is organized around the objective and search spaces, controllers, filters, acquisition, and evaluation protocol.Source: §3 · p. 8

3.1 Problem formulation

The problem formulation combines task accuracy with hardware penalties over discrete architecture spaces, evaluated through partial-training proxies and an analytical cost model. It covers both complete-architecture decisions and layer-by-layer construction.

3.1.1 Hardware-aware objective

  • The objective maximizes a hardware-aware reward over a family’s discrete architecture space under device latency and memory budgets.The reward combines accuracy with soft hardware penalties; the memory penalty is a continuous barrier that remains informative near the budget.Source: §3.1.1 · p. 8

3.1.2 Search spaces

  • The study uses macro families for single-step bandit search and a per-layer family for sequential MDP search, exposing the filter to policy drift and within-episode depth shift.The macro families include LeNet-like, ResNet-like, and MobileNet-like CNNs; the sequential space grows partial networks layer by layer.Source: §3.1.2 · pp. 8–9
  • Macro spaces vary depth, width, kernel, skip connections, pooling, and per-block bit-width; the per-layer family assigns kernel, channels, and bit-width at each layer.The sequential search has roughly 20 choices per layer and depths up to 20, yielding on the order of 2^20 terminal architectures.Source: §3.1.2 · pp. 8–9

3.1.3 Evaluation proxy and quantization

  • Candidates use a partial-training proxy: five epochs on a fixed random 10% training subset by default, with a stronger 15-epoch, 30% proxy for search-quality comparisons.The stronger proxy restores a climbable reward landscape in the ResNet-like family; calibration tracks bound-versus-outcome performance and is described as largely proxy-agnostic.Source: §3.1.3 · p. 9
  • Fake post-training quantization applies each block’s chosen bit-width before scoring, making lower bit-widths potentially reduce accuracy as well as cost.The quantization step prevents low bit-width choices from receiving cost-model credit without an accuracy effect.Source: §3.1.3 · p. 9

3.1.4 Analytical hardware cost model

  • A deterministic MCU-style oracle estimates latency and memory from per-layer parameter and multiply-accumulate counts, bit-widths, weight storage, and peak activations.The model is not calibrated to a specific device but is monotone in design knobs; truncated sequential configurations use the same spatial schedule as construction.Source: §3.1.4 · p. 9

3.2 Search controllers

The search controllers generate the drifting, non-exchangeable candidate streams that the filter must handle, while the filter interacts with them only through proposals and observed rewards. The section describes single-step and sequential multi-agent PPO formulations and evaluation-matched search baselines.

3.2 Search controllers

  • The filter is policy-agnostic: each controller generates the drifting candidate stream, and the filter interacts through proposed candidates and observed rewards.The controllers are described as the source of the difficulty the filter addresses, rather than as contributions of this work.Source: §3.2 · pp. 9–10

3.2.1 Multi-agent PPO: single-step formulation

  • In the single-step formulation, the HCA selects macro-architecture knobs and the QA selects per-block bit-widths autoregressively, emitting one complete architecture per decision.Both actors use a shared normalized advantage and minimize their own PPO clipped surrogate; this horizon-1 setup isolates policy improvement as the drift source.Source: §3.2.1 · p. 10

3.2.2 Sequential per-layer MDP

  • The sequential controller builds layers from partial-architecture states, selecting layer specifications and bit-widths before terminating at a sampled depth.Per-step proxy rewards support TD credit assignment; a terminal-only variant evaluates only completed networks to match complete-network evaluation budgets.Source: §3.2.2 · pp. 10–11

3.2.3 Search baselines

  • Random search samples architectures uniformly, while regularized evolution mutates tournament winners; both baselines are matched to RL by real evaluations consumed.Random search is identified as a strong NAS baseline.Source: §3.2.3 · p. 11

3.3 Conformal evaluation filtering

The filtering loop uses a warm-up calibration set, then gates training with a reward upper bound and feeds measured rewards back to the filter. The surrogate and static split-conformal baselines are specified alongside the exchangeability condition that RL search violates.

3.3.1 The filtering loop

  • After W warm-up candidates, a proposal is evaluated only if U(x) ≥ τ; real evaluations feed back measured rewards, and coverage is measured on the post-warm-up evaluated stream.Pruned candidates are not trained and receive the fixed rejection reward rrej = −1; the sequential no-prune audit mode makes depth-conditional coverage observable.Source: §3.3.1 · p. 11

3.3.2 Reward surrogate

  • The reward surrogate is a three-layer, width-64 ReLU MLP trained by mean-squared error, with family-specific architecture features and a held-out split for conformal scores.The split avoids using the surrogate's training residuals, which would understate generalization error.Source: §3.3.2 · pp. 11–12

3.3.3 Static split-conformal baselines

  • Static split conformal guarantees coverage only under exchangeability, which RL proposals violate as the policy improves; periodic recalibration every P = 50 evaluations tracks drift with lag and offers no between-update guarantee.The calibrate-once variant fixes α after warm-up; the P = 50 recalibration period is an inferred setting because MARCO specifies no period.Source: §3.3.3 · p. 12

3.3.4 Adaptive conformal inference

Adaptive Conformal Inference updates an additive offset after each evaluation, targeting the requested miscoverage rate without requiring exchangeability or stationarity. AgACI aggregates several step-size experts to remove the need to tune a single step size.

3.3.4 Adaptive conformal inference

  • ACI raises its offset after miscoverage and lowers it otherwise, with a fixed point at miscoverage δ; bounded rewards yield long-run coverage converging to 1 − δ for arbitrary sequences.The offset starts from split calibration after warm-up, and the constant step size allows tracking persistent shift.Source: §3.3.4 · pp. 12–13

3.3.5 AgACI: removing the step size

  • AgACI aggregates ACI offsets from five fixed step sizes, γk ∈ {0.01, 0.02, 0.05, 0.1, 0.2}, using exponential weighting, with its grid and schedule fixed a priori.The construction removes the single step-size hyperparameter; it aggregates offsets rather than the experts’ interval bounds.Source: §3.3.5 · p. 13

3.3.6 Normalized conformal: a per-candidate width

Normalized conformal scaling makes bound widths vary by candidate and uses that width as an acquisition-search exploration signal. Mondrian offsets target group-conditional coverage, while the GP comparator has model-based uncertainty without a distribution-free guarantee on drifting streams.

3.3.6 Normalized conformal: a per-candidate width

  • A global offset gives every candidate the same additive adjustment and no candidate-specific uncertainty; normalized scores instead scale the adaptive multiplier by a candidate difficulty estimate.A second network h(x) supplies the mean-absolute-deviation scale, making width q u(x) vary by candidate and serve as the acquisition search’s exploration signal.Source: §3.3.6 · pp. 13–14

3.3.7 Mondrian conformal: group-conditional coverage

  • Mondrian conformal maintains an ACI offset per group, initialized from that group’s residual quantile when it has at least 10 calibration points; smaller or unseen groups use the global offset.Sufficiently observed groups inherit the coverage guarantee within the group, with slower adaptation because each offset learns only from its own stream.Source: §3.3.7 · p. 14

3.3.8 Gaussian-process baseline

  • The GP baseline uses posterior mean and standard deviation under an RBF-plus-white-noise kernel, but its nominal upper bound has no distribution-free guarantee on a drifting, non-exchangeable stream.Its uncertainty depends on the kernel’s stationarity and Gaussian residual assumptions, unlike the adaptive filters’ arbitrary-stream coverage guarantee.Source: §3.3.8 · p. 14

3.4 Conformal acquisition: from pruning to directed search

The calibrated optimistic bound can serve as an acquisition function, steering search rather than only filtering externally proposed candidates. Greedy acquisition, optimism-shaped reinforcement learning, and conformal-optimistic tree search provide three ways to use it.

Conformal acquisition: from pruning to directed search

  • The calibrated bound U(x) = g(x)+q u(x) is used as a distribution-free analogue of GP-UCB to steer search.As an acquisition function, its optimism bonus inherits the coverage control described for the conformal bound.Source: §3.4 · p. 14

Greedy acquisition

  • Greedy acquisition selects the bound-argmax from a uniformly sampled pool of 64 candidates, with no policy or tree search.It is a myopic one-step optimization, but the uncertainty term q u(x) rewards exploration.Source: §3.4 · p. 15

Optimism-shaped reinforcement learning

  • Optimism-shaped reinforcement learning combines incumbent improvement with an intrinsic optimism bonus in its per-step training reward.The shaping does not reliably improve on the unshaped controller and can destabilize it, so it is included as a tested negative.Source: §3.4 · p. 15

Conformal-optimistic tree search

  • Conformal-optimistic tree search treats partial networks as nodes and uses an AlphaZero-style planner with the PUCT rule to select proposals.Leaves are valued by the calibrated bound rather than training; after 200 simulations, the visit-count-greedy leaf receives the step’s single real evaluation, and that evaluation updates the surrogate and ACI state.Source: §3.4 · p. 15

Evaluation comparison

  • The acquisition experiments compare greedy and tree-search proposers with random, evolution, and PPO baselines at matched real evaluations; shaped and unshaped controllers are matched by episodes.The shaped controller is compared separately using sequential per-step currency.Source: §3.4 · p. 15

3.5 Evaluation protocol

The evaluation protocol measures coverage and its control across requested levels, alongside evaluation savings or search quality. It also tracks coverage reproducibility across seeds, with separate reporting for depth-conditional audits.

Coverage

  • Coverage is the post-warm-up fraction of evaluated candidates whose reward is at most U(x), reported marginally and by layer depth in the sequential MDP.Depth-conditional coverage is measured in no-prune audit mode to avoid censoring.Source: §3.5 · p. 16

Calibration error and steerability

  • Steerability is assessed by the error from target coverage across δ values 0.05, 0.1, 0.2, and 0.3, and by whether realized coverage moves monotonically with the requested level.The δ dial is treated as an operational property for practitioners.Source: §3.5 · p. 16

Evaluation saving

  • Evaluation saving is measured as the pruned fraction in filtered search or best reward at matched evaluation budgets in directed search..Source: §3.5 · p. 16

Solution quality

  • Solution quality is assessed by comparing the best accuracy or reward discovered across filters at fixed seed, exposing harm from falsely discarding top candidates.Source: §3.5 · p. 16

Reproducibility

  • Coverage reproducibility is measured by across-seed standard deviation, using three seeds for calibration sweeps and eight for depth-conditional audits.Per-depth estimates are noisier because each depth group contains only a fraction of the stream.Source: §3.5 · p. 16

3.6 Summary of coverage guarantees

The paper distinguishes guarantees by their assumptions: split conformal relies on exchangeability, while ACI supports long-run coverage under arbitrary sequences. Mondrian ACI adds group-level coverage, and the tuning-free aggregate and GP baseline have distinct guarantees.

Split conformal

  • Split conformal provides finite-sample marginal coverage under exchangeability, an assumption violated by construction in RL-driven search.Periodic recalibration mitigates the violation but lags drift and gives no guarantee between recalibrations.Source: §3.6 · p. 16

ACI and normalized ACI

  • ACI and normalized ACI provide distribution-free long-run marginal coverage for arbitrary sequences, including policy-improvement drift and within-episode depth shifts.The guarantee holds for a single online offset, global for ACI or locally scaled for normalized ACI.Source: §3.6 · p. 17

AgACI

  • AgACI’s weighted-average offset does not inherit the exact ACI coverage guarantee, but its expert aggregation is competitive with the best fixed γ in hindsight and empirically tracks the target within ∼0.004 mean error.Source: §3.6 · p. 17

Mondrian ACI

  • Mondrian ACI provides the ACI guarantee per group, degrading gracefully to marginal coverage when a group lacks data.Source: §3.6 · p. 17

GP bound and acquisition

  • The GP bound has no distribution-free coverage statement, while conformal acquisition inherits the arbitrary-sequence guarantee for its exploration signal.The optimism bonus is continuously driven toward 1 − δ rather than interpreted as posterior credence.Source: §3.6 · p. 17

4 Experiments

The experiments assess online coverage control under RL-driven proposal drift, the effect of pruning on solution quality, and sequential depth-conditional calibration across architecture families. They also examine directed search on a 256 kB testbed, including acquisition, planning, and controller-reward choices.

4.1 Experimental setup

  • The experiments span LeNet-like/MNIST, ResNet-like/CIFAR-10, and MobileNet-like/CIFAR-100, with 4- or 8-bit per-block fake post-training quantization.Bandit runs use 1,000 proposed candidates, including 100 warm-up candidates; sequential experiments run 150 construction episodes. Acceptance thresholds are calibrated per family to its reward distribution.Source: §4.1 · pp. 17–18

4.2 Calibration, steerability and reproducibility

  • ACI tracks requested coverage within ∼10^-3 and monotonically across families; tuning-free AgACI closely matches it, while static conformal filters lose steerability and GP-UCB remains near 0.99.Across LeNet-like, ResNet-like, and MobileNet-like families, ACI mean absolute calibration errors are 0.001, ∼10^-3, and 0.002; AgACI errors are 0.003, within 0.004 of ACI, and 0.004, respectively. Figure 1 shows the cross-family comparison; Table 2 reports the ResNet-like δ-sweep.Source: Figure 1 · p. 18, §4.2 · pp. 18–19

4.3 Solution quality: pruning is free, miscalibration is not

  • Adaptive and AgACI filters prune 25–50% of evaluations without measured quality loss; split-recal falsely discards the ResNet-like optimum, lowering best accuracy to 0.481 versus 0.505 for other arms.The false discard occurs on every seed; its penalty matters most on the isolated ResNet-like peak, unlike families with clusters of near-optimal architectures.Source: §4.3 · p. 21

4.4 Sequential search: depth-conditional coverage

  • At target coverage 0.90, depth-conditional Mondrian adaptation is near-exact when groups are well-sampled, but data-starved deep ResNet-like groups fall back toward the global offset.On MobileNet-like, coverage is 0.893–0.902 across all twelve depths and mean conditional error is 0.003 versus 0.039 for adaptive global ACI; on strong-proxy ResNet-like, Mondrian overall error is 0.052 versus 0.113 for adaptive and 0.140 for split-recal. Binning starved deep groups halves their error from 0.160 to 0.079.Source: Figure 3 · p. 23, §4.4 · pp. 24–25

4.5 Does the controller matter? RL vs uninformed search

  • With identical 1,000-evaluation budgets and filtering disabled, PPO offers no search-efficiency advantage over random search or regularized evolution and is overtaken by both by evaluation 200.PPO’s mean proposal reward rises 0.25 →0.40 (+58%), but it concentrates on a good-mean macro region and misses the optimum’s tail, which random search finds.Source: §4.5 · pp. 25–26

4.6 Directed search with the calibrated acquisition

  • The acquisition study tests greedy conformal-bound search and conformal-optimistic MCTS against random search and terminal-only PPO on one constrained sequential testbed, at matched budgets up to 1,000 evaluations.The ResNet-like strong-proxy testbed uses a 256 kB memory budget; an attribution ablation compares online-calibrated optimism with fixed, uncalibrated optimism.Source: §4.6 · p. 26
  • On this harder, different 256 kB landscape, terminal PPO ties random within noise (0.659 vs 0.656), unlike its result on the standard-budget landscape.Absolute rewards in this testbed are lower and are not comparable across the two tables.Source: §4.6 · p. 26
  • At 1,000 evaluations, greedy conformal acquisition finishes +0.040 reward/+0.061 accuracy above random and +0.036/+0.048 above PPO, with seed deviation 0.001.Seed deviations were 0.021 for random and 0.032 for PPO.Source: §4.6 · pp. 26–27
  • Reward ranks calibrated > uncalibrated > mean (0.695 > 0.691 > 0.682); fixed optimism beats random by 0.036, and online updates add 0.004.The mean-only arm gains exactly zero over its last 500 evaluations, while both optimistic arms continue improving; online calibration also reduces variance about 6× versus the mean arm.Source: §4.6 · p. 27
  • Every MCTS variant beats random at 100–200 evaluations but plateaus by 500 and finishes at or below random; cpuct sweeps keep best reward within ∼0.01 at every budget.The passage attributes the early-only advantage to too few real evaluations reaching optimistic simulated leaves to correct them; cheap pool-argmax acquisition, rather than planning depth, sustains the advantage.Source: §4.6 · p. 27
  • Optimism-shaped PPO does not reliably beat unshaped PPO: median reward is 0.63 vs 0.67 and discovered accuracy is 0.72 vs 0.70, with two shaped seeds diverging.The matched-episode comparison used eight seeds and per-step currency; divergence occurred as the optimism bonus drove the controller into the infeasible region.Source: §4.6 · p. 27

5 Conclusion

Online feedback calibration restores steerable coverage in RL-driven search, while static calibration loses control; the paper also extends the calibrated bound to conditional coverage and acquisition. Its conclusions are bounded by the hardware model, proxy rewards, data availability, and evaluation scope.

5 Conclusion

  • Static split-conformal calibration loses control under policy drift, while per-depth adaptation addresses depth-dependent miscoverage and optimistic acquisition beats random search on one constrained testbed.On data-starved depths, conditional coverage may fall back toward marginal coverage; the acquisition evidence is limited to one family and testbed.Source: §5 · pp. 27–28, §5.1 · p. 29
  • Coverage tracks every requested level to within ∼10^-3 across three architecture families and both search formulations, while pruning 25–50% of evaluations at no measured solution-quality cost.The result is monotonic and reproducible; seed deviation is ∼0.002, and fixed-grid ACI experts match the fixed-γ filter within 0.004.Source: §5 · p. 28

5.1 Limitations

  • The analytical hardware oracle is not physically calibrated, so hardware conclusions concern orderings and budget-binding regimes, not absolute latency or memory.Hardware generality is demonstrated only across parameterizations of one MCU-style model; validation against measured GPU, NPU, and physical-edge costs remains future work.Source: §5.1 · p. 28
  • Mondrian adaptation closes depth-conditional gaps only where groups have enough data; data-starved or structurally harder groups can remain miscalibrated.The deepest ResNet-like depths are data-starved, while the LeNet-like terminal depth under-covers even when well-sampled; such harder groups are family-specific.Source: §5.1 · p. 29
  • Coverage in filtered runs is measured only on accepted candidates, so full-population coverage requires no-prune audits or a statistical selection-bias correction.Pruned candidates are never trained, leaving their rewards unobserved; the adaptive filter’s measured coverage is τ-invariant across the reported operating range, except near the reward ceiling.Source: §5.1 · p. 29
  • Evidence comes from from-scratch CNN classification on MCU-scale spaces; transfer to transformers, other tasks, pretrained-model fine-tuning, or deployed toolchains is asserted by design, not tested.The conformal-acquisition evidence is narrower still: one family, one constrained testbed, and three seeds.Source: §5.1 · p. 29
  • Three calibration-sweep seeds and eight depth-audit seeds support headline effects but leave wide intervals for sparse deep-depth estimates and small-budget search comparisons.All rewards use partial-training proxies, and whether their rankings survive full training is untested; proxy strength also shapes measurable search quality and requires per-family recalibration of τ.Source: §5.1 · pp. 28–29

5.2 Outlook

  • The outlook identifies calibrated acquisition, conditional validity on structured spaces, and device-measured objectives as consequential next steps.It also proposes using continuously verified bounds for exploration, resource rationing, and stopping rules.Source: §5.2 · p. 30
Loading 2610.03127v1…