Policy-Driven Quantization Control

• Policy-driven quantization control is a set of adaptive methods that dynamically selects quantization parameters based on system feedback and task quality.
• These techniques balance performance metrics such as accuracy, latency, and memory usage through multi-objective optimization and precise policy control.
• Applications span neural compression, adaptive beamforming, and quantized feedback control, enabling efficient resource allocation under varying system constraints.

Searching arXiv for the cited works to ground the article and confirm metadata. Policy-driven quantization control denotes a class of methods in which quantization parameters, feedback resolutions, or quantization-related resource allocations are selected by an explicit policy rather than fixed a priori. In neural compression, this typically means per-layer or per-module mixed precision; in communication and control, it means dynamic quantizer switching or event-driven feedback; in quantization-aware distillation, it can mean spatially adaptive allocation of distillation effort under low-bit constraints. Across these settings, the common structure is a control variable, a performance objective, and a policy that trades task quality against memory, BitOPs, latency, feedback cost, or control cost (Xu et al., 2018, Maity et al., 2020, Wang et al., 2023).

1. Conceptual scope

A central premise of the literature is that uniform quantization is often structurally mismatched to heterogeneous systems. In convolutional networks, different layers have different representational capacities and sensitivities to quantization, so a single bit-width forces overly conservative precision on robust layers and overly aggressive precision on sensitive layers. DNQ formalized this observation by learning a per-layer mixed-precision assignment with a bit-width controller and a distance-aware quantizer (Xu et al., 2018). DQMQ extended the same principle from architecture heterogeneity to input heterogeneity, arguing that the relative layer-wise quantization sensitivity changes under varying input data quality, and that bit-widths should therefore adapt to different data qualities online (Wang et al., 2023). SEAM, by contrast, treated the policy as a transferable per-layer bit-width allocation searched on a small proxy dataset, with transferability enforced through large-margin discriminability rather than dataset identity (Tang et al., 2023).

The same control logic appears outside conventional neural compression. In event-driven beamforming, the controlled variable is whether channel-state feedback is turned on or off, and finite-rate quantization changes the effective feedback price; the optimal policy is threshold-type in the CSI-accuracy variable $z$ and depends on channel power $g$ (0909.4983). In quantized-feedback LQG, the decision variable is the choice among multiple quantizers with different operating costs, and the optimal solution separates into certainty-equivalence controller synthesis and an MDP for quantizer switching (Maity et al., 2020). In spiking transformers and image restoration, policies also control candidate evaluation or loss allocation: AQ4SViT uses a search gating policy based on membrane potential drift, while SPARK uses a stochastic spatial weight map to modulate the KD loss during QAT (Putra et al., 14 Jun 2026, Rasool et al., 13 Jun 2026). This suggests that “policy-driven quantization control” is best understood as an umbrella for adaptive quantization decisions, not solely as layerwise bit allocation.

2. Formal objectives and controlled variables

Despite domain differences, the formulations are consistently multi-objective or constrained. In DNQ, the controller outputs per-layer bit-widths $b_\ell \in \{2,3,\dots,8\}$ and optimizes a reward

$R = Acc + \lambda \cdot r,$

where $Acc$ is validation accuracy and $r$ is the compression ratio computed from the selected bit-width sequence and centroid overhead (Xu et al., 2018). DQMQ uses a hybrid objective

$$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$

where the policy selects per-layer bit-widths conditioned on activations and Hessian-trace sensitivity, and $R_l(a_l)$ is the memory size reduction (Wang et al., 2023). SEAM searches a mixed-precision policy $P=\{(b_i^w,b_i^a)\}$ by minimizing

$$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$

combining class-level separation, intra-class compactness, and a BitOPs proxy (Tang et al., 2023).

AQ4SViT casts post-training quantization of SViTs as

$g$0

with $g$1 percentage points on ImageNet-1K, and supplements this with a gate-aware surrogate

$g$2

together with the hard gate $g$3 (Putra et al., 14 Jun 2026). SimQ-NAS formulates a joint search over architecture $g$4 and quantization policy $g$5, seeking Pareto-optimal $g$6 that maximize validation accuracy while minimizing latency or model size (Sridhar et al., 2023).

In communication and control, the objectives are written as explicit stochastic control criteria. Beamforming maximizes net throughput

$g$7

where $g$8 is the feedback decision, $g$9 is channel power, and $b_\ell \in \{2,3,\dots,8\}$0 is the alignment variable (0909.4983). QF-LQG minimizes

$b_\ell \in \{2,3,\dots,8\}$1

where $b_\ell \in \{2,3,\dots,8\}$2 selects a quantizer from a finite set and $b_\ell \in \{2,3,\dots,8\}$3 is the quantizer-cost vector (Maity et al., 2020).

Setting	Controlled quantity	Objective form
DNQ	Per-layer $b_\ell \in \{2,3,\dots,8\}$4	$b_\ell \in \{2,3,\dots,8\}$5
DQMQ	Per-layer dynamic $b_\ell \in \{2,3,\dots,8\}$6	$b_\ell \in \{2,3,\dots,8\}$7
SEAM	Per-layer $b_\ell \in \{2,3,\dots,8\}$8	$b_\ell \in \{2,3,\dots,8\}$9
AQ4SViT	Mixed-precision assignment $R = Acc + \lambda \cdot r,$0	Minimize memory under accuracy and MPD constraints
Beamforming / QF-LQG	Feedback or quantizer action	Net throughput or quadratic cost with feedback price

These formulations differ mainly in whether the trade-off is handled by a scalarized reward, a constrained search, or a dynamic programming objective. A plausible implication is that the control vocabulary of quantization has broadened from static compression to general resource allocation under task-dependent uncertainty.

3. Policy classes and optimization algorithms

The policy classes range from stochastic sequence models to threshold laws and predictor-guided search. DNQ models per-layer bit allocation as an MDP whose state is the partial bit-width sequence $R = Acc + \lambda \cdot r,$1, whose action is the current layer bit-width, and whose policy $R = Acc + \lambda \cdot r,$2 is parameterized by a BiLSTM. Training uses REINFORCE without a baseline, SGD with learning rate $R = Acc + \lambda \cdot r,$3, batch size $R = Acc + \lambda \cdot r,$4, and $R = Acc + \lambda \cdot r,$5 controller iterations, with Monte Carlo completions used to reduce variance (Xu et al., 2018).

DQMQ adopts a hybrid RL formulation. Its Policy Decision Agent is a lightweight, layer-specific four-layer CNN that takes $R = Acc + \lambda \cdot r,$6, where $R = Acc + \lambda \cdot r,$7 is the current activation and $R = Acc + \lambda \cdot r,$8 is the Hessian-trace sensitivity. Training proceeds in two stages: a soft differentiable stage that relaxes discrete bit-widths to continuous probabilities, followed by a hard stage with policy-gradient optimization over discrete actions (Wang et al., 2023). SEAM uses a different relaxation: each layer maintains logits $R = Acc + \lambda \cdot r,$9 and $Acc$0 over weight and activation bit candidates, the search is differentiable, and the final discrete policy is obtained by argmax after proxy-data optimization. The search runs for $Acc$1 epochs with SGD at learning rate $Acc$2, while the first and last layers remain fixed at $Acc$3 bits (Tang et al., 2023).

SimQ-NAS replaces direct policy gradients with multi-objective predictor-guided search. Candidate solutions are tuples $Acc$4 that jointly encode elastic architecture parameters and per-module INT8/FP32 choices for weights and activations. The search algorithm is LINAS, which iteratively samples candidates, evaluates them through static PTQ and validation, trains lightweight predictors such as ridge regression or SVR, and proposes new candidates likely to improve the Pareto front (Sridhar et al., 2023). AQ4SViT also uses explicit search policies, but its defining feature is the gating mechanism: Greedy search performs a hierarchical global $Acc$5 stage-wise $Acc$6 block-wise search, while Beam search performs breadth-limited exploration with beam width $Acc$7, repair passes, and gate-based pruning (Putra et al., 14 Jun 2026).

In control-theoretic formulations, the policy structure is often analytically characterized. In event-driven beamforming, the optimal stationary policy is threshold-type: there exists $Acc$8 such that $Acc$9 if $r$0 and $r$1 otherwise, and this remains true under quantized feedback (0909.4983). In QF-LQG with perfect measurement, quantizer selection is an MDP on the sufficient state $r$2, and the optimal policy partitions the state space into disjoint regions, one per quantizer (Maity et al., 2020). A structurally different, but still policy-driven, partitioning mechanism appears in critic-driven Voronoi state partitioning: the critic $r$3 is not added to the cloning loss, but it drives where the Voronoi quantizer introduces new cells by identifying low-value regions for splitting (Deproost et al., 14 May 2026).

4. Quantizers, surrogates, and state signals

Policy-driven control depends not only on the optimizer but also on the underlying quantizer model and the signal used to judge candidate quality. DNQ employs nonuniform, centroid-based weight sharing via per-layer $r$4-means with

$r$5

where $r$6 and one code is reserved to encode exact zero. Its quantizer then ranks weights by quantization distance

$r$7

clusters distances into $r$8 groups per layer, quantizes larger-distance groups earlier, and retrains only unquantized weights through a mask $r$9 (Xu et al., 2018). DQMQ instead uses uniform fake quantization with per-layer scales derived from tensor min/max and an explicit dequantizer after each layer so that quantization bias does not accumulate across layers or across changing bit-width decisions (Wang et al., 2023).

SEAM relies on standard QAT and STE during both search and fine-tuning, but its distinctive control signal is not a reconstruction metric; it is class-level discriminability. The search loss explicitly promotes inter-class separation and intra-class compactness under a Gaussian Mixture Distribution model, reflecting the claim that sub-optimal policies shrink class margins and blur decision boundaries (Tang et al., 2023). SimQ-NAS uses post-training static quantization via Intel Neural Compressor, but the paper does not specify whether quantization is per-channel or per-tensor, symmetric or asymmetric; the explicit policy choices are only INT8 and FP32 for both weights and activations (Sridhar et al., 2023).

AQ4SViT uses uniform symmetric simulated quantization for weights,

$$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$0

with per-tensor symmetric quantization, $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$1, $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$2, and $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$3. Its policy signal is membrane potential drift (MPD), computed by comparing membrane-potential sequences from FP32 and quantized models over a small calibration batch; the default hard-gating threshold is $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$4 (Putra et al., 14 Jun 2026). SPARK uses yet another surrogate family: four no-grad spatial difficulty maps—Laplacian variance, pixel variance, student reconstruction error, and teacher-student knowledge gap—are fed into a compact policy CNN that outputs a stochastic spatial weight map, which is then detached and used to modulate the KD loss during QAT (Rasool et al., 13 Jun 2026).

In robotics and autonomous driving, the control signal is distributional rather than structural. Quantization-Aware Imitation Learning combines behavior cloning on expert and FP-policy rollouts with a Quantization-Robust Behavior Cloning term that aligns quantized and full-precision action distributions through average $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$5 distance in the final logits; for long-horizon manipulation, a weighted variant increases the alignment weight on high-saliency states (Park et al., 2024). In beamforming, the relevant state variables are channel power $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$6 and alignment $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$7, and under isotropy and Markovity the controller state reduces to these two scalars (0909.4983). In QF-LQG, the selector uses quantized innovations $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$8 and, in the quantized-measurement case, covariance terms $$J(\theta) = E_{a \mid X,h}\!\left[L_{task}(\theta; a,h) - \alpha \sum_{l=1}^L R_l(a_l)\right],$$9 rather than raw states (Maity et al., 2020).

5. Empirical performance across domains

On canonical CNN benchmarks, policy-driven mixed precision typically improves the accuracy–compression trade-off relative to uniform quantization. DNQ reported, on CIFAR-10 with CIFAR-Net, that uniform 3-bit INQ achieved $R_l(a_l)$0 at $R_l(a_l)$1, DNQ Module 2 only achieved $R_l(a_l)$2 at $R_l(a_l)$3, and DNQ Module 1+2 achieved $R_l(a_l)$4 at $R_l(a_l)$5 with a learned assignment of conv1 = 5-bit and conv2/conv3 = 3-bit. On ImageNet, DNQ achieved Top-1 $R_l(a_l)$6 on AlexNet at $R_l(a_l)$7, compared with full precision $R_l(a_l)$8, and Top-1 $R_l(a_l)$9 on ResNet-18 at $P=\{(b_i^w,b_i^a)\}$0, compared with full precision $P=\{(b_i^w,b_i^a)\}$1 and INQ 3-bit $P=\{(b_i^w,b_i^a)\}$2 at $P=\{(b_i^w,b_i^a)\}$3 (Xu et al., 2018). DQMQ reported on ImageNet with ResNet-18 Top-1 $P=\{(b_i^w,b_i^a)\}$4 with $P=\{(b_i^w,b_i^a)\}$5 weight compression and model size $P=\{(b_i^w,b_i^a)\}$6 MB, exceeding the cited FP baseline of $P=\{(b_i^w,b_i^a)\}$7; on mixed-quality CIFAR-10 and SVHN, it reported $P=\{(b_i^w,b_i^a)\}$8 and $P=\{(b_i^w,b_i^a)\}$9, respectively (Wang et al., 2023). SEAM searched policies on proxy datasets of about $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$0 of ISLVRC-2012 scale and reported up to $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$1 times improvement in MPQ searching efficiency, including ResNet-18 4-bit mixed precision at $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$2 and $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$3G BitOPs in $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$4 GPU-hours using CIFAR-10 proxy data, versus FracBits at $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$5 and $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$6 GPU-hours (Tang et al., 2023).

The joint optimization of architecture and quantization policy produces a different empirical profile. SimQ-NAS reported, depending on the backbone, up to $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$7 latency improvement and $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$8 model-size improvement without degradation in accuracy compared to fully quantized INT8 baselines. Specific examples include BERT Base on SST-2 with up to $$\mathcal{L} = \mathcal{L}_{cls} + \lambda \mathcal{L}_{inc} + \gamma \mathcal{L}_{comp},$$9 speedup at the same accuracy, ViT Base on ImageNet-1K with $g$00 speedup without accuracy sacrifice and a top-accuracy point of $g$01 that remained $g$02 faster than baseline, and OFA ResNet152 with $g$03 latency improvement and $g$04 size reduction without accuracy loss (Sridhar et al., 2023). In SViTs, AQ4SViT-Greedy reported up to $g$05 faster search time and up to $g$06 memory saving compared to the state of the art, while AQ4SViT-Beam reported up to $g$07 memory reduction with about $g$08 longer search time, all while maintaining high accuracy within $g$09 from the original/non-quantized models on ImageNet; for SDTv2 specifically, Beam reached $g$10 MB and $g$11 Top-1 (Putra et al., 14 Jun 2026).

Quantization-aware policies also improve behavior or reconstruction under deployment constraints. In robot manipulation, the imitation-learning framework reported INT4 weight quantization on a real edge GPU with up to $g$12 speedup and $g$13 energy savings while preserving accuracy; for 4-bit weight and activation quantized self-driving models, it reported up to $g$14 speedup and $g$15 energy saving on a low-end GPU (Park et al., 2024). In image restoration, SPARK reported on LOLv1 with a self-quantized MIRNet student $g$16 dB PSNR, $g$17 SSIM, and $g$18 LPIPS, improving over PTQ by $g$19 dB PSNR, $g$20 SSIM, and $g$21 LPIPS; on SIDD, the same student reached $g$22 dB and outperformed PTQ by $g$23 dB (Rasool et al., 13 Jun 2026). In event-driven beamforming, optimal controlled feedback improved net throughput over periodic feedback by up to $g$24 bit/s/Hz without extra bandwidth or antennas (0909.4983). In QF-LQG, numerical experiments showed a Pareto curve with minimum control cost $g$25 at quantization cost $g$26, maximum control cost $g$27 at quantization cost $g$28, and an average bit-rate of $g$29 bits per sample in one comparison to perfect feedback (Maity et al., 2020).

6. Limitations and research directions

A recurrent misconception is that policy-driven quantization control is synonymous with static mixed-precision bit allocation. The literature is broader: it includes feedback triggering, dynamic quantizer switching, search gating, spatial KD weighting, and value-guided Voronoi refinement (0909.4983, Maity et al., 2020, Rasool et al., 13 Jun 2026, Deproost et al., 14 May 2026). The policy object may therefore be a bit-width vector, a feedback action, a gating threshold, a spatial mask, or a partition refinement rule.

The methods also inherit substantial optimization and deployment costs. DNQ is computationally intensive because each policy update requires repeated quantization and retraining cycles, and it does not report runtime, use a baseline, or use entropy regularization (Xu et al., 2018). DQMQ relies on activation statistics and Hessian traces as data-quality signals and introduces per-layer policy agents, while severe quality shifts may require server-side re-quantization in the deployment pattern described (Wang et al., 2023). SEAM requires labels to form class-level margins and can degrade when proxy similarity is too low or the constraint is extremely tight (Tang et al., 2023). SimQ-NAS restricts numerics to INT8/FP32 and requires latency predictors to be retrained and latencies re-measured for new hardware (Sridhar et al., 2023). AQ4SViT’s MPD is a strong but imperfect surrogate, Beam search is about $g$30 slower than Greedy, and hardware-aligned extensions such as per-channel quantization, integer activation/state quantization, and accumulator-width constraints remain future work (Putra et al., 14 Jun 2026).

The control-oriented papers make their own assumptions explicit. Beamforming threshold optimality relies on channel Markovity, isotropy, and a stochastic-dominance property for the $g$31 dynamics (0909.4983). QF-LQG assumes perfect, delay-free, distortion-free communication and faces a $g$32-dimensional continuous-state MDP in the perfect-measurement case, so exact solution becomes difficult in high dimensions (Maity et al., 2020). In robotic control, the quantized policy is static at runtime, there is no formal safety certification, and dynamic precision control is left to future work (Park et al., 2024). In critic-driven Voronoi distillation, high-dimensional state spaces, Euclidean-distance assignment, and discontinuities at Voronoi boundaries limit fidelity and stability, especially on harder systems such as BipedalWalker (Deproost et al., 14 May 2026).

The direction of travel is nonetheless clear. The surveyed works repeatedly move from fixed quantization schedules toward policies conditioned on state, input quality, architectural elasticity, class margins, spike dynamics, or deployment hardware. A plausible implication is that future quantization research will increasingly treat quantization not as a one-time compression step but as a control problem over discrete, structured, and often constrained actions.