LASiR: Cross-Domain Applications

• In neuroimaging, LASIR implements a covariate-dependent mixture of spatially regularized image-on-scalar regressions to identify latent subgroups with distinct brain–covariate associations.
• For smart-contract security, LASiR integrates static taint analysis, LLM-assisted semantic checks, and symbolic execution to detect signature replay vulnerabilities in Solidity contracts.
• In indoor optical wireless networking, LASiR denotes a laser-based system using an intelligent reflecting surface to improve coverage and achieve significant sum-rate enhancements.

LASiR denotes distinct constructs in multiple technical literatures rather than a single standardized object. In neuroimaging, LASIR—also written LASiR—stands for Latent Subgroup Image-on-Scalar Regression, a model-based framework for identifying latent subpopulations with heterogeneous brain–covariate associations in large multi-site imaging studies (Lin et al., 2023). In smart-contract security, LASiR is a hybrid analysis system for detecting Signature Replay Vulnerabilities by combining static taint analysis, LLM-assisted semantic analysis, and symbolic execution (Wang et al., 12 Nov 2025). In indoor optical wireless networking, “LASiR” (Editor’s term) can denote a laser-based OWC architecture augmented with an intelligent reflecting surface, specifically a VCSEL-based angle-diversity transmitter coupled to a passive mirror array for blockage mitigation and sum-rate improvement (Hamad et al., 2024).

1. Scope and nomenclature

The term is used in at least three technically unrelated senses.

Research area	Meaning	arXiv id
Neuroimaging	Latent Subgroup Image-on-Scalar Regression	(Lin et al., 2023)
Smart-contract security	Detection system for Signature Replay Vulnerabilities	(Wang et al., 12 Nov 2025)
Optical wireless networking	“LASiR” (Editor’s term): laser-based OWC with an intelligent reflecting surface	(Hamad et al., 2024)

The neuroimaging usage is a formal acronym. The smart-contract usage is the name of a detection framework. The optical-networking usage is not presented as a formal acronym in the source paper; rather, the paper is described as “a first, relatively compact realization” of a laser-plus-IRS system of the kind designated here as LASiR. This distinction matters because the three literatures share neither methodology nor objective: one concerns mixture modeling for image-valued responses, one concerns program analysis for Solidity contracts, and one concerns Gaussian-beam propagation and specular reflection in indoor OWC.

2. LASIR in neuroimaging: latent subgroup image-on-scalar regression

In neuroimaging, LASIR addresses image-on-scalar regression in settings where the image is the outcome, the predictors are scalar covariates, and the associations are heterogeneous across subjects (Lin et al., 2023). The motivating setting includes large multi-site studies such as the ABCD study, where brain activation patterns and brain–behavior associations can vary substantially across individuals, while subgroup allocation may itself depend on sociodemographic characteristics.

A classical spatially varying coefficient model would write

$$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_j(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$

with a single homogeneous set of coefficient images for the entire population. LASIR replaces this with latent subgroup structure:

$$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$

where

$$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$

Here, $\delta_{ik}=1$ indicates membership of subject $i$ in subgroup $k$, and $\alpha_{kj}(v)$ is the subgroup-specific coefficient function for exposure $j$. The result is an individual-specific coefficient surface that equals the subgroup-specific effect image for the subject’s latent class.

Subgroup allocation is covariate-dependent rather than purely unsupervised. LASIR uses a multinomial logit model

$$\Pr(\delta_{ik}=1 \mid \mathbf{z}_i) = \frac{\exp(\mathbf{w}_k^\top \mathbf{z}_i)} {\sum_{c=1}^K \exp(\mathbf{w}_c^\top \mathbf{z}_i)}, \quad \mathbf{w}_K \equiv \mathbf{0}.$$

This design distinguishes LASIR from clustering on images: the subgroups are intended to be homogeneous with respect to brain–covariate associations, heterogeneous across groups, and predictable from individual characteristics.

Because direct voxelwise modeling is infeasible when $d \sim 10^4\!-\!10^5$, LASIR introduces a basis expansion using orthonormal basis functions $$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$0. Subgroup effects, site effects, and control-covariate effects are expanded in this basis, and the spatial noise covariance is represented as

$$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$1

After projection into basis space, the model becomes a finite mixture of multivariate regressions with diagonal covariance $$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$2. This makes estimation tractable while retaining a GP-motivated spatial structure.

A common misconception is to treat LASIR as a clustering method on raw images. The model is explicitly designed not to cluster on image similarity alone. The latent classes are tied to subgroup-specific regression surfaces and to a covariate-dependent allocation mechanism, which is a materially different target.

3. Estimation, model selection, and empirical behavior of the neuroimaging LASIR model

LASIR is estimated with an efficient stochastic expectation maximization algorithm (Lin et al., 2023). The E-step computes posterior subgroup probabilities. The S-step samples hard assignments from those posterior probabilities. The M-step updates shared coefficients, subgroup-specific coefficients, and multinomial-logit parameters. The stochastic step is emphasized because, unlike deterministic EM, it reduces the tendency to get stuck in local modes of a multimodal likelihood.

Model selection is based on BIC,

$$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$3

with

$$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$4

under the paper’s assumptions. The basis dimension $$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$5 is chosen from the GP eigenvalue spectrum so that a fixed fraction of variation is captured; the paper uses $$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$6.

The simulation program is extensive. In synthetic cube images, LASIR achieves NMI of approximately $$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$7–$$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$8, while KMLR is approximately $$y_i(v) = \sum_{j=0}^p x_{ij}\,\beta_{ij}(v) + \sum_{r=1}^q z_{ir}\,\eta_r(v) + \sum_{s=1}^S u_{is}\,\gamma_s(v) + \epsilon_i(v),$$9; MSEs for group-specific and individual-specific coefficients are an order of magnitude smaller with LASIR than with KMLR and much smaller than with a single SVCM (Lin et al., 2023). In a brain-shaped simulation, LASIR attains NMI of approximately $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$0, versus approximately $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$1–$$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$2 for KMLR, and shows much better type I error control, often below $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$3, with reasonable power. When no heterogeneity is present, BIC correctly selects $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$4 in all 50 replicates.

The principal application uses ABCD baseline working-memory task fMRI with 2-Back vs 0-Back contrast maps. After preprocessing and quality control, $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$5 participants remain; the exposure of interest is the cognitive $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$6-factor, the control vector has $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$7 variables, and there are $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$8 site indicators. With a modified squared-exponential kernel and basis dimension $$\beta_{ij}(v) = \sum_{k=1}^K \delta_{ik}\,\alpha_{kj}(v).$$9, BIC selects $\delta_{ik}=1$0, and the best run converges in 11 iterations; the reported computation time is about 13 minutes for 20 runs on an 8-core Xeon iMac (2021) (Lin et al., 2023).

The four identified subgroups are labeled MH, ML, PL, and NL. “Contrast” refers to the group intercept $\delta_{ik}=1$1, and “association” refers to $\delta_{ik}=1$2, the slope of the $\delta_{ik}=1$3-factor. In MH and PL, positive association between $\delta_{ik}=1$4 and activation appears in the frontoparietal task control network, default mode network, salience network, and sensory/somatomotor regions. In ML, essentially only the DMN shows significant association with $\delta_{ik}=1$5. In NL, most voxels show no association, with a small subset in the somatomotor network showing negative association and some DMN voxels showing positive association. The held-out prediction analysis reports the smallest MSE for within-subgroup modeling, larger MSE for a no-subgroup model, and the worst MSE for shuffled subgroup labels, supporting the claim that the identified subgroups capture predictive structure rather than arbitrary partitioning.

The paper also notes an important inferential limitation: voxelwise inference treats subgroup labels as known and therefore does not propagate membership uncertainty. This suggests that finite-sample uncertainty may be optimistic even when asymptotically acceptable.

4. LASiR in smart-contract security: signature replay vulnerability detection

In smart-contract security, LASiR is a system for detecting Signature Replay Vulnerabilities (SRVs) in Solidity contracts on EVM chains (Wang et al., 12 Nov 2025). An SRV occurs when a single valid signature can be accepted more than once or in unintended contexts because the contract does not bind the signature tightly enough to context, does not track or enforce one-time use, or accepts multiple cryptographically equivalent signatures.

The paper derives a five-part taxonomy from 108 SRV cases identified in 1,419 audit reports collected from 37 blockchain security companies and bug bounty platforms. The five types are Cross-chain Replay Attack (X-CRA), Cross-project Replay Attack (X-PRA), Contract Account Signature Replay (CASR), Signature State Management Issue (SSMI), and Signature Malleability Attack (SMA). X-CRA concerns the omission of block.chainid; X-PRA concerns omission of address(this); CASR concerns missing identity binding in EIP-1271-style flows; SSMI concerns nonce, timestamp, and usage-tracking failures; SMA concerns failure to enforce canonical ECDSA constraints such as $\delta_{ik}=1$6 and lower-half-$\delta_{ik}=1$7 rules.

LASiR operates in three phases: Slicing with LLM Analysis, Inspection of Signature Verification, and Path Reachability Verification. The pipeline builds an inter-contract program dependency graph, locates ecrecover() and EIP-1271 validation logic, uses Prompt_A to identify signature-related key variables, performs function-level slicing, uses Prompt_B to identify sanitized variables and expected checks, runs static taint analysis over the slice, uses Prompt_C to infer function-call sequences or business flows, and finally applies symbolic execution to check whether the flagged paths are feasible.

The implementation combines SlithIR / Slither for Solidity parsing and IR generation, an I-PDG derived from AST and control flow, Rattle for CFG reconstruction and symbolic execution, and the DeepSeek-V3 API with 128K context length, Temperature $\delta_{ik}=1$8, and a 3-request redundancy mechanism (Wang et al., 12 Nov 2025). Taint sources include msg.sender, msg.data, msg.value, block attributes such as block.chainid and block.timestamp, external-call return values, and user-supplied parameters. Sinks include ecrecover(hash, v, r, s), isValidSignature(bytes32,bytes), and related verification routines.

A central design feature is that the LLM is used as a semantic oracle, not a standalone detector. Static analysis verifies dataflow and syntactic conditions, and symbolic execution checks reachability. This is intended to address the weaknesses of pure LLM reasoning, particularly hallucination and poor path sensitivity.

5. Empirical study and practical implications of the smart-contract LASiR system

The large-scale evaluation starts from 918,964 contracts with source code across Ethereum, BSC, Polygon, and Arbitrum, then filters to 15,383 contracts that explicitly use ecrecover() (Wang et al., 12 Nov 2025). The reported counts are 4,513 on Ethereum, 5,590 on BSC, 4,140 on Polygon, and 1,140 on Arbitrum.

Chain	Signature-using contracts	% contracts with any SRV
Ethereum	4,513	19.63%
BSC	5,590	9.29%
Polygon	4,140	7.11%
Arbitrum	1,140	5.94%

These results are paired with an asset-risk analysis. Among SRV-flagged contracts with non-zero balances, 258 contracts remain. Manual exploitation analysis confirms 31 contracts with exploitable signature replay behaviors affecting \$4.76M in active assets. For 24 contracts, the validation uses Tenderly simulations; for 7 contracts, it uses manual proof-of-concept exploit construction.

Detection accuracy is evaluated on a manually labeled set of 500 contracts, with 72 positives and 428 negatives. The reported confusion counts are TP $\delta_{ik}=1$9, FP $i$0, FN $i$1, and TN $i$2, yielding Precision $i$3, Recall $i$4, and F1-score $i$5 (Wang et al., 12 Nov 2025). The abstract reports an F1-score of 87.90%, which is close to the detailed DB2 calculation. Comparative experiments show that GPT-4o, DeepSeek-R1, DeepSeek-V3 used as an LLM-only baseline, Slither4SRV, Siguard, and GPTScan4SRV all underperform LASiR on the reported benchmark. The ablation results are especially strong: full LASiR reaches F1 $i$6, whereas No_LLM reaches F1 $i$7.

The false-positive analysis identifies two recurrent causes. First, unrelated business-state restrictions can already enforce one-time use, such as a state reset after withdrawal even when signature handling is formally weak. Second, implementation errors can make a verification path non-functional, so LASiR flags a cryptographic weakness even though exploitation is infeasible. The false-negative analysis attributes the main misses to assembly-based custom ecrecover implementations, where the LLM and the static analysis fail to reconstruct low-level semantics accurately.

The practical coding recommendations follow directly from the five-type taxonomy: include block.chainid for X-CRA, include address(this) for X-PRA, bind the contract-account identity for CASR, incorporate nonces and deadlines plus explicit usage tracking for SSMI, and enforce canonical ECDSA or use OpenZeppelin’s ECDSA for SMA. The paper frames these not as stylistic choices but as requirements for correct authorization semantics.

6. “LASiR” in indoor optical wireless communication

In indoor optical wireless networking, “LASiR” (Editor’s term) denotes a laser-based OWC system assisted by an intelligent reflecting surface, instantiated in the source paper as a multiuser downlink indoor OWC network with an IRS on one wall and a laser-based access point on the ceiling (Hamad et al., 2024). The room size is $i$8. The AP is mounted at the center of the ceiling. The IRS is a mirror array mounted on one wall. Users are uniformly distributed on the communication floor. Each user receives a superposition of a direct LoS component and a single-bounce NLoS component reflected by one IRS mirror.

The AP employs an angle diversity transmitter with 5 branches, one at the center and four placed symmetrically around it. Each branch is a $i$9 VCSEL array at wavelength $k$0, with bandwidth $k$1 and beam waist $k$2. The beam is modeled as a Gaussian beam, not Lambertian. Each user has an angle diversity receiver with photodiodes of area $k$3, responsivity $k$4, and field-of-view $k$5; the ADR branches have azimuth angles $k$6 and elevation $k$7.

The IRS is modeled as a passive mirror array of $k$8 rotational mirrors, with examples $k$9 and $\alpha_{kj}(v)$0, reflectivity $\alpha_{kj}(v)$1, and mirror area $\alpha_{kj}(v)$2. This is a specular-reflection model with physical mirror orientation control, not a wavelength-scale metasurface with explicit phase control. That distinction is fundamental: the control variable is mirror orientation, and the channel is governed by geometric optics and Gaussian-beam coupling rather than by coherent phase engineering.

The propagation model uses

$\alpha_{kj}(v)$3

for the beam waist at distance $\alpha_{kj}(v)$4, and

$\alpha_{kj}(v)$5

for the transverse intensity distribution. The received power through a circular aperture of radius $\alpha_{kj}(v)$6 is

$\alpha_{kj}(v)$7

The effective channel gain for user $\alpha_{kj}(v)$8 is modeled as

$\alpha_{kj}(v)$9

and the IM/DD-style signal model is

$j$0

The SINR and achievable rate are

$j$1

The simulation results show that IRS deployment improves sum rate for a fixed transmit SNR, and increasing mirror density from $j$2 to $j$3 yields further gains (Hamad et al., 2024). For a given SNR, a $j$4 mirror array achieves approximately 26% higher sum rate than a $j$5 mirror array and approximately 71% higher sum rate than the system without IRS. Sum rate also remains consistently higher than the non-IRS system as the number of users increases. The paper interprets the IRS benefit mainly through improved coverage and alternative NLoS paths rather than through explicit moving-blocker simulations.

A frequent misconception is to equate this optical IRS with RF metasurface IRS models. The paper instead describes a practical mirror array with rotational or switchable elements, passive operation, and specular redirection of narrow Gaussian beams. Relative to conventional non-IRS optical wireless systems, this suggests a way to preserve narrow-beam, high-SNR VCSEL transmission while partially relaxing the usual rate–coverage trade-off by redirecting beams into shadowed or off-axis regions.

7. Comparative perspective

The three usages of LASiR are linked only by nomenclature, but each addresses a form of latent or hidden structure. In neuroimaging, the latent object is subgroup membership governing heterogeneous coefficient images. In smart-contract security, the hidden object is replay-enabling program semantics that are not captured by simple pattern matching. In optical wireless networking, the critical hidden structure is the alternative NLoS geometry created by a reconfigurable passive surface.

Their methodological cores are correspondingly different. Neuroimaging LASIR is a covariate-dependent mixture of spatially regularized image-on-scalar regressions with GP-inspired basis reduction and SEM estimation (Lin et al., 2023). Smart-contract LASiR is a hybrid analysis pipeline combining I-PDG-based taint propagation, LLM semantics, and symbolic execution to detect five SRV classes (Wang et al., 12 Nov 2025). Optical “LASiR” is a Gaussian-beam and specular-reflection system model for IRS-assisted VCSEL-based indoor OWC (Hamad et al., 2024).

The term therefore functions as a cross-domain homograph. Accurate interpretation depends entirely on disciplinary context: in neuroimaging it names a latent-subgroup regression model; in blockchain security it names a detector for signature replay; in optical communications it designates, in editor’s terminology, a laser-plus-IRS architecture.