Double-Layer Optimization and DLO Ambiguity

• Double-Layer Optimization is an umbrella term describing layered design strategies that vary by context, from hardware architectures to iterative algorithmic frameworks.
• Its applications range from THz beamforming using double-layer TTD schemes and surrogate density models in black-box optimization to dynamic layer operations in LLMs and robotics for manipulating deformable objects.
• The layered approaches emphasize distinct trade-offs such as reduced hardware costs and enhanced computational efficiency, illustrating practical challenges in system design and optimization.

Searching arXiv for exact usage of “double-layer optimization” and closely related “DLO” terminology to ground the article in current literature. In the arXiv literature considered here, “Double-Layer Optimization (DLO)” is not a standardized technical term. Instead, the acronym DLO is used for several unrelated constructs: “Deterministic Langevin Optimization” in expensive black-box global optimization (Sullivan et al., 2023), “Dynamic Layer Operation” in transformer depth scaling (Tan et al., 2024), and deformable linear objects in robotics (Li et al., 22 Oct 2025, Zhu et al., 16 Jun 2026, Cao et al., 2 Jun 2026). The closest instance of an explicitly double-layer design appears in THz communications, where the contribution is a double-layer true-time-delay (TTD) scheme with a layered alternating optimization procedure, not a formally named bilevel method called “Double-Layer Optimization” (Sun et al., 2023). Accordingly, the term is best treated as an ambiguous label whose meaning depends entirely on citation context.

1. Terminological status and scope

Within this corpus, no paper introduces a canonical framework named Double-Layer Optimization. The main source of confusion is acronym collision: the same three letters denote unrelated methods, architectures, or object classes across optimization, LLMs, communications, and robotics.

arXiv id	Meaning of “DLO”	Research area
(Sullivan et al., 2023)	Deterministic Langevin Optimization	Expensive black-box global optimization
(Tan et al., 2024)	Dynamic Layer Operation	Vertical scaling of LLMs
(Sun et al., 2023)	Double-layer TTD scheme	THz beamforming with distributed RISs
(Li et al., 22 Oct 2025)	Deformable linear objects	Long-horizon robotic routing
(Zhu et al., 16 Jun 2026)	Deformable linear objects	Industrial manipulation benchmark
(Cao et al., 2 Jun 2026)	Deformable linear objects	Differentiable physics and benchmark

The communications paper is the only one in this set whose title contains an explicit double-layer construction, but it states that the paper does not explicitly formulate a bilevel optimization problem called “Double-Layer Optimization”; rather, “double-layer” refers primarily to a two-layer TTD hardware architecture at the base station (Sun et al., 2023). This establishes an important terminological boundary: “double-layer” in the available literature is often architectural rather than an optimization taxonomy.

A plausible implication is that references to “DLO” without an accompanying paper identifier are intrinsically under-specified. For technical writing, the acronym therefore requires disambiguation by title or domain.

2. Closest technical analogue: double-layer TTD with layered alternating optimization

The most direct approximation to what might be informally called double-layer optimization is the THz beamforming framework of "Beamforming Design for the Distributed RISs-aided THz Communications with Double-Layer True Time Delays" (Sun et al., 2023). Its novelty is a hierarchical TTD network in the analog beamforming hardware: the second layer contains $K_{\rm H}$ TTDs with relatively larger delay range, each feeding $K_{\rm L}$ first-layer TTDs with smaller delay range, and each first-layer TTD then connects via phase shifters to $P$ antennas, giving

$N = K_{\rm H}K_{\rm L}P.$

The motivation is explicitly hardware-oriented: compared with a conventional single-layer TTD architecture, the double-layer arrangement reduces the number of large-range delay devices, which are associated with insertion loss, power consumption, and amplification overhead.

The optimization target is an achievable-rate maximization problem over hybrid analog/digital beamforming, TTD delays, and RIS reflection coefficients under discrete hardware constraints. The paper formulates

$$R_{\rm sum} = \sum_{k=1}^{K}\sum_{m=1}^{M}\log_2(1+\gamma_{m,k}),$$

and then maximizes $R_{\rm sum}$ subject to transmit-power, finite-resolution TTD, finite-resolution phase-shifter, and finite-resolution RIS-phase constraints. The resulting problem is nonconvex and is described as NP-hard (Sun et al., 2023).

Algorithmically, the method is layered, but not bilevel in the formal mathematical-programming sense. The analog beamformer is designed first using geometry and phase-compensation principles; thereafter, the digital beamformer is updated via an MMSE/WMMSE reformulation, and RIS coefficients are updated by coordinate search. The paper therefore exemplifies a hierarchical design plus alternating optimization / block-coordinate framework, rather than an upper-level problem constrained by lower-level optimality conditions (Sun et al., 2023).

The paper’s explicit performance tradeoff reinforces that interpretation. For $N=128$, $U=32$, $K_{\rm H}=8$, and $K_{\rm L}=4$, the single-layer design uses total bits $K_{\rm L}$0, whereas the double-layer design uses

$K_{\rm L}$1

yielding a hardware-cost ratio

$K_{\rm L}$2

In the reported comparison, rate degradation is only $K_{\rm L}$3 relative to the single-layer TTD baseline (Sun et al., 2023). This suggests that, in this literature, a “double-layer” idea is most naturally understood as hierarchical variable partitioning aligned to hardware structure, followed by staged optimization.

3. Deterministic Langevin Optimization and the surrogate–density decomposition

In "Deterministic Langevin Unconstrained Optimization with Normalizing Flows" (Sullivan et al., 2023), DLO means Deterministic Langevin Optimization, not Double-Layer Optimization. The method addresses expensive, black-box, unconstrained, global optimization, with maximization posed as

$K_{\rm L}$4

Its central acquisition combines a surrogate prediction with a density penalty: $K_{\rm L}$5 Here $K_{\rm L}$6 is a surrogate for the annealed objective, and $K_{\rm L}$7 is a normalizing-flow estimate of the density of previously evaluated samples. The method is motivated by deterministic Langevin dynamics and the Fokker–Planck equation, with stationary density

$K_{\rm L}$8

The resulting exploration mechanism differs qualitatively from standard GP-based Bayesian optimization. Instead of using predictive variance, the method explores regions where the sample density is low. The implementation has two learned components at each iteration: an objective surrogate $K_{\rm L}$9 and a density model $P$0. The paper is explicit, however, that this is not a nested “double-layer” optimization method in the bilevel sense (Sullivan et al., 2023).

That clarification is central. The paper describes an outer loop of sequential expensive evaluations and inner computations for fitting a surrogate and a flow, but it states that DLO is not presented as double-layer optimization, bilevel optimization, or nested optimization in that formal sense. The two learned components may explain occasional informal misremembering, yet the correct name remains Deterministic Langevin Optimization (Sullivan et al., 2023).

4. Dynamic Layer Operation in LLM scaling

In "DLO: Dynamic Layer Operation for Efficient Vertical Scaling of LLMs" (Tan et al., 2024), DLO again does not denote Double-Layer Optimization. It denotes Dynamic Layer Operation, a dynamic depth-scaling framework for transformer-based LLMs. The method performs group-based layer expansion, attaches per-layer routers, and conditionally executes or skips the MLP submodule at token level based on layerwise feature similarity.

The paper’s formulation is architectural and routing-centric. If an original model has $P$1 layers, grouped as $P$2, each group is expanded to

$P$3

yielding final depth

$P$4

Routing supervision is induced heuristically from cosine similarity between the attention output and the post-MLP output. Low similarity indicates that the MLP changes the representation substantially, and that token-layer pair is therefore more valuable to execute.

Training is integrated into supervised fine-tuning (SFT) rather than continual pre-training (CPT), with total loss

$P$5

The method is explicit that it targets vertical scaling, not width scaling as in Mixture-of-Experts. It also does not define a double-layer optimization formalism. Any resemblance to a “two-layer” structure comes from the coexistence of expansion and routing mechanisms, not from a recognized optimization category (Tan et al., 2024).

A plausible implication is that searches for “DLO” in LLM literature can easily retrieve dynamic-depth routing work rather than optimization work. In bibliographic contexts, title-level disambiguation is therefore mandatory.

5. DLO as deformable linear objects in robotics

A third major usage is entirely non-optimization: in robotics, DLO commonly means deformable linear objects, such as cables, ropes, wires, and rubber bands (Li et al., 22 Oct 2025, Zhu et al., 16 Jun 2026, Cao et al., 2 Jun 2026). This usage spans hierarchical planning, benchmarks, differentiable simulation, and sim-to-real transfer.

"Hierarchical DLO Routing with Reinforcement Learning and In-Context Vision-LLMs" studies long-horizon routing of deformable linear objects through multiple clips from language instructions. Its architecture combines a high-level vision-language planner with low-level skills—Insert, Pull, and Flatten—and reports an overall success rate of 92.5% across long-horizon routing scenarios (Li et al., 22 Oct 2025). Here, DLO is an object class rather than a method family.

"WireCraft: A Simulation Benchmark for Industrial DLO Manipulation" likewise uses DLO to denote industrial wire and cable manipulation. It introduces three task families—connector insertion, clip routing, and channel seating—and concludes that privileged state-based RL solves representative settings with over $P$6 success, whereas current vision-based learning remains limited, especially in the contact-rich alignment phase of connector insertion (Zhu et al., 16 Jun 2026).

"DLO-Lab: Benchmarking Deformable Linear Object Manipulations with Differentiable Physics" extends this robotics usage through a differentiable simulator supporting extensibility, elasticity, bending plasticity, loop topology, self-contact, and coupling with rigid and soft materials. It evaluates PPO, SAC, SHAC, SAPO, GD, and CMA-ES, reporting an average fixed-horizon success rate of 86.6% for CMA-ES across eight tasks (Cao et al., 2 Jun 2026). None of these papers uses DLO as an optimization acronym.

This accumulation of robotics usage makes acronym ambiguity especially acute. In contemporary arXiv indexing, DLO is at least as likely to refer to deformable linear objects as to any optimization-related construct.

6. Relation to bilevel, hierarchical, and layered optimization

The cited literature makes a consistent distinction between double-layer architectures, two-component models, and formal bilevel optimization. The THz beamforming paper has the strongest “double-layer” content, but its own characterization is a double-layer TTD hardware architecture combined with analog design plus alternating optimization (Sun et al., 2023). The black-box optimization paper has a repeated loop with a surrogate model and a density model, but explicitly states that it is not a double-layer or bilevel optimization framework (Sullivan et al., 2023). The LLM paper involves expansion, activation, and skipping, yet those are dynamic routing operations rather than a named double-layer optimization theory (Tan et al., 2024).

This suggests a useful editorial distinction. If “Double-Layer Optimization” is used at all in relation to these works, it should be treated as an informal umbrella phrase for layered computation or hierarchical design, not as the title of an established method class. The phrase becomes technically precise only when replaced by the actual construct named in the cited source: Deterministic Langevin Optimization, Dynamic Layer Operation, double-layer TTD scheme, or deformable linear objects.

A common misconception is therefore to read the acronym DLO as semantically stable across fields. The available arXiv evidence points in the opposite direction. The term is best handled as a citation-dependent abbreviation whose meaning must be reconstructed from the paper title, application domain, and explicit definitions provided by the authors (Sullivan et al., 2023, Tan et al., 2024, Sun et al., 2023).