2-TLP: Contexts and Applications

• 2-TLP is a polysemous term used in dynamic traffic assignment, DeFi pricing, temporal link prediction, and tensor program tuning with distinct definitions in each domain.
• In dynamic traffic assignment, 2-TLP formulates a constrained optimization of toll rates that reduces local travel time by 21.6% while maintaining network density near a critical threshold.
• In finance and network prediction, 2-TLP quantifies liquidity premiums and forecasts multi-step graph links, and in tensor tuning it integrates paired cost models to enhance latency prediction.

Searching arXiv for papers using the term "2-TLP" and nearby variants to ground the article. “2-TLP” is a context-dependent abbreviation rather than a single standardized technical term. In the arXiv literature represented here, it appears in at least four distinct senses: the second toll-level problem in simulation-based dynamic traffic assignment (Gu et al., 2019), the two-period Liquidity-of-Time Premium in intertemporal pricing of time-bound stablecoins (Borjigin et al., 7 Oct 2025), 2-step Temporal Link Prediction in temporal networks (Xiong et al., 28 Feb 2025), and a shorthand for the paired use of TLP and MTL-TLP in tensor program tuning (Zhai et al., 2022). This distribution suggests that the expression is intrinsically polysemous and must be interpreted from disciplinary context rather than from the string “2-TLP” alone.

1. Terminological scope and disambiguation

The most explicit occurrences of “2-TLP” in the supplied literature are organized below.

Domain	Meaning of “2-TLP”	Source
Dynamic traffic assignment	Second toll-level problem	(Gu et al., 2019)
DeFi and asset pricing	Two-period Liquidity-of-Time Premium	(Borjigin et al., 7 Oct 2025)
Temporal networks	2-step Temporal Link Prediction	(Xiong et al., 28 Feb 2025)
Tensor program tuning	Combination of TLP and MTL-TLP	(Zhai et al., 2022)

Among these, the traffic paper uses “2-TLP” as a formally posed optimization problem with explicit constraints and solution machinery (Gu et al., 2019). The stablecoin paper uses “2-TLP” as a closed-form, two-period specialization of a term-structure model for the Liquidity-of-Time Premium (Borjigin et al., 7 Oct 2025). The temporal-network survey uses “2-step TLP” as a forecasting horizon within the broader Temporal Link Prediction taxonomy (Xiong et al., 28 Feb 2025). The tensor-program paper does not define “2-TLP” in the title or abstract, but the supplied integrated description uses it to denote the combined system formed by TLP and MTL-TLP (Zhai et al., 2022).

A common misconception is to treat “2-TLP” as a universal acronym. The broader literature represented here instead assigns “TLP” to unrelated constructs such as Ticket-Level Prediction, Task-Level Perturbations, Temporal Lift Pooling, Two Level Perceptron, Transboundary Loewner Property, Total Loss of Power, and Transient Lunar Phenomena (Prova et al., 17 Jun 2025, Yin et al., 14 Apr 2025, Hu et al., 2022, Jamet et al., 2024, Hakobyan et al., 2019, Prieto et al., 2013, AF, 2013).

2. 2-TLP as the second toll-level problem in dynamic congestion pricing

In dynamic traffic assignment, 2-TLP denotes the second toll-level problem, formulated for a cordoned toll subnetwork $G_p$ with $m=8$ successive tolling intervals and decision vector

$${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$

where $\tau^d_h$ is the distance toll rate and $\tau^t_h$ is the delay toll rate in interval $h$ (Gu et al., 2019). The problem extends a first TLP by adding an explicit control on the heterogeneity of congestion distribution in the pricing zone.

The first objective is to maintain the average network density $K_h$ in $G_p$ near the critical density $K_{\rm cr}=25$ vpkmpl. Its loss is

$$J_1({\boldsymbol\tau}) = \frac{1}{m}\sum_{h=1}^m \bigl\lvert\,K_h({\boldsymbol\tau})-K_{\rm cr}\bigr\rvert.$$

The second criterion is built from the “spatial spread of density”

$m=8$0

its lower-envelope cubic $m=8$1, and the “deviation from spread”

$m=8$2

The original bi-objective problem is then reformulated as a constrained optimization in which the NFD-control objective is minimized subject to a heterogeneity-limit constraint, with $m=8$3 vpkmpl (Gu et al., 2019).

The reformulated 2-TLP is

$m=8$4

This formulation makes 2-TLP a constrained control problem over time-varying joint distance and delay tolls rather than a static toll-design exercise (Gu et al., 2019).

3. Surrogate optimization, heterogeneity control, and reported effects

The 2-TLP in the Melbourne study is solved by a surrogate-based method using regressing kriging (RK) with expected improvement (EI) sampling to approximate the expensive simulation input-output mapping (Gu et al., 2019). The design of experiments stratifies each of the $m=8$5 decision variables into $m=8$6 intervals and draws an initial Maximin Latin Hypercube of size $m=8$7. The expensive outputs are the objective $m=8$8 and the constraint $m=8$9, both evaluated by the traffic simulator.

For the scalar objective, the RK metamodel is

$${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$0

with Gaussian correlation

$${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$1

and nugget regularization through $${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$2. Infill points are chosen by maximizing constrained expected improvement,

$${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$3

with $${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$4 (Gu et al., 2019).

The reported numerical outcome for the second problem is that 2-TLP reduces the average travel time in the cordoned zone $${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$5 by 21.6 % and in the entire network by 2.5 %. By comparison, the first TLP achieves 29.5 % in $${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$6 and 1.4 % in the entire network, so 2-TLP exchanges part of the within-zone gain for a larger whole-network benefit (Gu et al., 2019). The same report states that the average deviation $${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$7 under the single-objective solution was $${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$8 vpkmpl, whereas 2-TLP drives it down to the allowable bound $${\boldsymbol\tau}=(\tau^d_1,\dots,\tau^d_m,\;\tau^t_1,\dots,\tau^t_m),$$9, yielding more uniform densities and reducing the size of the hysteresis loop in the NFD.

Operationally, both TLPs were solved in under 100 simulator calls, and the use of regressing kriging plus constrained EI sampling required roughly 30–40 % fewer simulations than DIRECT, corresponding to $\tau^d_h$0 hours saved on a typical workstation (Gu et al., 2019). This suggests that, in this literature, “2-TLP” is not merely a second objective but a specific constrained reformulation whose central function is to regularize congestion heterogeneity while retaining NFD-based density control.

4. 2-TLP as the two-period Liquidity-of-Time Premium

In the stablecoin literature, TLP denotes the Liquidity-of-Time Premium, defined as

$\tau^d_h$1

where $\tau^d_h$2 is the underlying asset’s official closing price and $\tau^d_h$3 is the time-bound stablecoin’s market value during the closed period (Borjigin et al., 7 Oct 2025). Economically, TLP measures the compensation for bearing risk and illiquidity when the primary market is closed.

The paper then derives a term structure for TLP using a no-arbitrage model and a Black–Scholes put-price representation. For an $\tau^d_h$4-period horizon, with risk-neutral drift zero and lognormal overnight return, the supplied derivation writes

$\tau^d_h$5

Setting $\tau^d_h$6 yields the two-period specialization called 2-TLP: $\tau^d_h$7 and

$\tau^d_h$8

In this formulation, higher volatility $\tau^d_h$9 increases 2-TLP, and a lower loan-to-value ratio reduces both default probability and TLP (Borjigin et al., 7 Oct 2025).

The same work proposes a dynamic risk-control policy that adjusts $\tau^t_h$0 to keep TLP within a target band: $\tau^t_h$1 Empirical proxies include ADR premiums, overseas index futures versus cash index divergence, and pre-market versus official close gaps. The reported backtest summary includes a nightly 1-day TLP mean of 0.23%, median 0.18%, and 95th/99th percentiles 0.9%/1.8%, with low-vol assets having 2-day 2-TLP $\tau^t_h$2 and high-vol assets having 2-day $\tau^t_h$3 (Borjigin et al., 7 Oct 2025). Here “2-TLP” is thus a pricing object in intertemporal liquidity engineering rather than an optimization problem.

5. 2-step TLP in temporal networks

In temporal-network research, TLP denotes Temporal Link Prediction, and “2-step TLP” refers the task of predicting both $\tau^t_h$4 and $\tau^t_h$5 from a history of graph snapshots (Xiong et al., 28 Feb 2025). In the discrete-time setting,

$\tau^t_h$6

with adjacency matrices $\tau^t_h$7. The two-step forecasting problem is defined over a representation function

$\tau^t_h$8

and an inference function

$\tau^t_h$9

for $h$0, followed by

$h$1

The survey organizes the method space through a representation–inference taxonomy. On the representation side, it identifies snapshot-based models, feature-extraction methods such as CN, KI, and AA with temporal weighting, matrix/tensor-factorization models, random-walk embeddings, and discrete- or continuous-time GNNs (Xiong et al., 28 Feb 2025). On the inference side, it distinguishes direct matrix/tensor extension, RNN-based forecasting, and attention-based decoders. For 2-step prediction, examples include TSVD plus AR on latent factors, DynNode2Vec plus RNN, and attention-based GNN decoders such as DySAT-style temporal self-attention (Xiong et al., 28 Feb 2025).

The survey also emphasizes under-explored combinations, including continuous-time latent representations with matrix-factorisation inference for $h$2, neighbour-sequence representations with direct AR, and hybrid MF plus attention. This suggests that “2-TLP” in this literature denotes a multi-horizon forecasting regime within a general predictive taxonomy rather than a single canonical model (Xiong et al., 28 Feb 2025).

6. 2-TLP as a paired tensor-program tuning framework, and the broader problem of acronym collision

In the tensor-program tuning literature, the supplied integrated description uses “2-TLP” to denote the combination of TLP and MTL-TLP (Zhai et al., 2022). TLP is a deep learning-based cost model that treats schedule primitives as tensor languages and converts latency prediction into an NLP regression task. The regression target is written as

$h$3

with training by mean squared error over $h$4 pairs, or on normalized latency $h$5 (Zhai et al., 2022).

The input is a padded or cropped schedule-primitive sequence $h$6, where each primitive yields an embedding by concatenating primitive-type encoding, learned name-token embeddings, and scaled numerical features. The resulting matrix $h$7 is passed through input embedding, positional encoding, a contextual backbone consisting of a single Transformer block or alternatively an LSTM layer, and a regression head with average pooling and two residual linear layers (Zhai et al., 2022). MTL-TLP introduces one task per hardware target, shares the embedding and backbone, and assigns a task-specific final head $h$8.

When integrated into the Ansor framework on ResNet-50, MobileNet-V2, ResNext-50, BERT-tiny, and BERT-base, the reported results are: TLP achieves 9.1× average search-time reduction on CPU and 3.0× on GPU versus TenSet-MLP tuning 2,000 programs, while MTL-TLP with only 7 % of target-device data attains 4.7× on CPU and 2.9× on GPU (Zhai et al., 2022). In this usage, “2-TLP” is a convenient label for a paired methodology, not a formal mathematical object named in the original title.

Across the represented literature, the principal encyclopedic point is therefore negative but important: “2-TLP” has no stable cross-domain denotation. In one field it is a constrained toll-design problem (Gu et al., 2019); in another, a two-period liquidity premium (Borjigin et al., 7 Oct 2025); in another, a two-horizon graph-forecasting task (Xiong et al., 28 Feb 2025); and in another, a composite tensor-program tuning framework (Zhai et al., 2022). A plausible implication is that any technical reading of “2-TLP” should be treated as undefined until the surrounding domain vocabulary—traffic assignment, DeFi pricing, temporal networks, or tensor tuning—has been established explicitly.