Q-DEQ: Discrete Solving and Quantization for Deep Equilibrium Models in Time Series Forecasting under Edge Deployment Coding Constraints
Ruotong Yang, Hongdong Zhu, Qi Gao, Yin Ma, Hai Wei, Kai Wen
cs.LG
Quick verdict
Q-DEQ 将 DEQ 的 Anderson 更新重构为 QUBO + W8A8 以用于边缘预测；准确性保持不变（−1.16% 至 +2.90% MSE），但缺乏硬件、运行时或成本证据。
Summary
该论文提出 Q-DEQ，这是对用于多变量时间序列预测的深度平衡模型 (DEQ) 的一种改进，用离散优化问题取代了 Anderson 加速的连续系数搜索。候选更新方向基于当前残差和迭代历史构建；局部二次残差模型、m=8 系数的 8 位二进制编码（64 个二进制变量）以及向 Ising 形式的转换产生了一个可通过模拟退火或 Kaiwu 相干 Ising 机求解的 QUBO。在不动点求解后，W8A8 伪量化重前向传播对共享的 iTransformer 层进行量化。在 Weather、ETTh1、ETTh2、ECL 和 Traffic 数据集（输入长度 96，地平线 96，三个种子）上，与显式多层基线相比，Q-DEQ-INT8 显示出 −1.16% 至 +2.90% 的 MSE 差距；参数量下降了 1.80 倍至 3.82 倍，静态权重存储下降了 4.30 倍至 12.80 倍。FP32 消融实验发现，离散求解器在五个数据集中的四个上略优于 Anderson。单次 CIM 和 SA 执行给出了几乎相同的预测，Willow 图匹配附录检查了架构迁移。局限性：没有真实的硬件、延迟、功耗或内存测量；求解器成本（每次迭代额外的 m 次前向传播加上 SA）未报告；CIM 验证 MSE 使用了无法解释的尺度；更长的地平线（单次运行）在所有五个数据集上都有所退化；且未提供显著性检验或代码。
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Local DEQ update coefficients can be discretized into a 64-variable QUBO and solved by simulated annealing with accuracy comparable to Anderson acceleration, slightly better on four of five datasets in FP32.
- Combining DEQ parameter sharing (1.80x-3.82x) with W8A8 fake quantization cuts static weight storage 4.30x-12.80x versus the explicit multi-layer baseline while MSE stays within -1.16% to +2.90% at horizon 96 (three seeds).
- The same local QUBO problems executed on the Kaiwu CIM physical backend and CPU SA produce closely matching downstream forecasts (single-run comparison, Table 6).
- The discrete solver is architecture-agnostic, transferring to a weight-tied graph-matching Transformer with near-equal accuracy (0.974-0.980 on Willow, single run).
Strengths
Novel, completely derived solver-level bridge from DEQ to QUBO/Ising backends
Section 3.2 gives the full chain — finite-difference directional derivatives (Eq. 3), local quadratic residual energy (Eq. 4), 8-bit binary encoding into a 64-variable QUBO (Eq. 5) and Ising Hamiltonian (Eq. 6) — and Section 5.4 executes it on both CPU SA and the physical Kaiwu CIM with near-identical downstream forecasts (Table 6).
Clean four-configuration ablation design
Table 1/3 isolates network structure (A vs B), solver (B vs C) and quantization (C vs D) on one shared EncoderLayer across five datasets x 3 seeds; Q-DEQ-FP32 beats Anderson on 4/5 datasets (e.g., Weather 0.1736 vs 0.1773, Traffic 0.3978 vs 0.3994), showing the discrete solver carries no systematic accuracy penalty.
Decomposed storage accounting
Candid, well-hedged reporting
Architecture-agnosticity backed by a second task
Weaknesses
No deployment-relevant measurements despite the framing
Solver compute cost never measured
Unexplained scale mismatch in the CIM validation
Thin statistics for near-parity claims
Narrow baselines and undiagnosed failure modes
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Q-DEQ (arXiv 2609.24042, cs.LG) proposes to make deep equilibrium models (DEQs) cheaper to store for multivariate time-series forecasting by intervening at two stages. First, the forward solver: instead of Anderson acceleration's continuous mixing-coefficient search (Eq. 2), Q-DEQ builds m=8 candidate directions from the current residual and iteration history (Rademacher-padded when history is short), estimates directional derivatives by finite differences (Eq. 3), fits a local quadratic residual model (Eq. 4), encodes each of the m coefficients as an offset plus an 8-bit binary vector — 64 binary variables in total — and converts this into a QUBO and Ising Hamiltonian (Eqs. 5-6) solved by simulated annealing. Second, after the approximate fixed point is reached, a single re-forward pass applies W8A8 fake quantization to the shared iTransformer EncoderLayer (per-channel weight, per-tensor activation, straight-through estimator; Section 3.3). Training keeps implicit DEQ differentiation and never backprops through the solver (Algorithm 1, Section 3.4). The stated contributions are: a local discrete solver, its integration with a shared layer plus W8A8, portability across combinatorial backends (CPU SA and the Kaiwu CIM), and an architecture-agnostic PyTorch plugin.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Q-DEQ in one sentence?
Q-DEQ is a variant of deep equilibrium models (DEQ) for time-series forecasting that replaces Anderson acceleration's continuous update-coefficient search with a discrete QUBO optimization problem (64 binary variables, solved by simulated annealing or a coherent Ising machine) and then applies W8A8 fake quantization to the shared layer in a re-forward pass.
How much does Q-DEQ compress the model?
Weight tying inside the DEQ reduces parameter counts 1.80x-3.82x versus an explicit multi-layer iTransformer; adding W8A8 fake quantization brings static weight storage to 4.30x-12.80x smaller (for example, Traffic drops from 24.46 to 1.91 MiB). These are storage-arithmetic figures, not on-device measurements.
Does Q-DEQ lose forecasting accuracy?
At a 96-step horizon across five datasets with three seeds, MSE differs from the explicit baseline by -1.16% to +2.90% (better on Weather and ETTh2, worse on ETTh1, ECL and Traffic). At a 336-step horizon (single run) it is worse on all five datasets by about 2.1% on average.
Is code or real-hardware deployment data available for Q-DEQ?
The paper announces no code release, and it contains no latency, power, or memory measurements on real edge hardware: all quantization is simulated (fake quantization), and the CIM experiment uses the cloud-based Kaiwu backend in a single run. This is stated as future work in the paper itself.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-7 are referenced but not analyzable from the extracted text; Figure 7's per-category Willow accuracies and the horizon-192/336 curves cannot be independently verified.
- Hyperparameter values (finite-difference step eta, coefficient offsets, SA schedule, stopping tolerances, training details) are absent from the extracted text.
- Table 6's differing MSE scale may reflect an evaluation-setup detail not stated in the extracted text; it is flagged as an inconsistency rather than adjudicated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.