LiGen: Disambiguated Research Systems

• LiGen in NLP is a logical grammar embedding model that unsupervisedly infers lexical categories and syntactic rules from raw text, enabling transparent grammatical composition.
• LiGen in drug discovery is a structure-based virtual screening application that optimizes throughput and quality on HPC systems using methods like EMaliboo and PAMaliboo.
• LiGen in indoor localization uses ambient-light spectral fingerprints with GAN-based augmentation to accurately predict 2D coordinates while reducing error compared to Wi‑Fi.

Searching arXiv for papers using the term "LiGen" to ground the article and disambiguate the name across domains. Using arXiv search for exact-title/name matches and nearby variants: LiGen, LGE, and related acronym collisions. LiGen is a reused research name rather than a single technical artifact. In the arXiv literature represented here, it denotes at least three unrelated systems: a logical grammar embedding model for unsupervised grammar induction from raw text (Deyo et al., 2023), a structure-based virtual screening application used in the EXSCALATE drug-discovery platform (Guindani et al., 2024), and an indoor localization system based on ambient-light spectral fingerprinting with GAN-based data augmentation (Lin et al., 5 Aug 2025). The label therefore requires domain-specific disambiguation before any technical claim, benchmark result, or implementation detail can be interpreted correctly.

1. Terminological scope

The name LiGen has been attached to conceptually independent research programs. In natural language processing, LiGen is also referred to as LGE (“logical grammar embedding”) and is presented as a model for learning lexical categories and syntactic composition rules from text. In computational drug discovery, LiGen is a real-world virtual screening software whose performance depends strongly on HPC deployment conditions. In indoor positioning, LiGen is a coordinate-regression pipeline trained on real and synthesized spectral fingerprints.

Usage of the name	Domain	Core function
LiGen / LGE	NLP	Unsupervised grammar induction from raw text
LiGen	Drug discovery / HPC	Structure-based virtual screening
LiGen	Indoor localization	Spectral-fingerprint coordinate prediction

The term is also easy to confuse with nearby names. LimGen is a dataset and benchmark for Suggestive Limitation Generation of research papers (Faizullah et al., 2024). LIGER is a training-free framework for long-horizon visual instruction generation (Suo et al., 12 Mar 2025). LOGen is a diffusion-based model for LiDAR object generation (Kirby et al., 2024). LaGen is an autoregressive LiDAR scene generation framework, and that paper explicitly states that “LiGen” is not the method being proposed there (Zhou et al., 26 Nov 2025).

2. LiGen as logical grammar embedding

In "A logical word embedding for learning grammar" (Deyo et al., 2023), LiGen is the logical grammar embedding (LGE), a model inspired by pregroup grammars and categorial grammars. Its central aim is to learn directly from a corpus both lexical categories for words and syntactic composition rules that explain how those categories combine into grammatical sentences. The model is framed as a logical, structure-sensitive alternative to ordinary word embeddings: rather than learning vectors mainly for similarity or prediction, it learns a representation intended to mirror the formal composition rules of grammar.

The underlying grammatical picture is categorial. Each word is associated with a category; categories can have syntactic valence; grammaticality corresponds to the reduction of a type sequence to a distinguished sentence type; and learning means recovering these categories and reduction patterns from corpora. The paper emphasizes unsupervised inference: no gold part-of-speech tags, dependency trees, or phrase-structure trees are required. The system instead searches for latent grammatical structure that best explains the observed sentences.

Several properties are treated as defining. The abstract states that LGE performs unsupervised inference of lexical categories and syntactic rules, produces comprehensible output summarizing its inferences, has a completely transparent process for producing novel sentences, and can learn from very small corpora, even on the order of a hundred sentences. The transparency claim is especially distinctive: sentence generation is described not as sampling from an opaque latent space, but as explicit grammatical composition in which one can inspect the chosen lexical categories, their combination, and the resulting derivation. This makes the learned grammar human-comprehensible in a symbolic sense rather than only effective as an internal representation (Deyo et al., 2023).

3. LiGen as structure-based virtual screening software

In "Efficient Parameter Tuning for a Structure-Based Virtual Screening HPC Application" (Guindani et al., 2024), LiGen is a real-world, highly parameterized structure-based virtual screening application used in the EXSCALATE drug-discovery platform. Its objective is to evaluate very large chemical libraries and rank ligands by how likely they are to bind to a target protein. The paper describes this as a throughput-versus-quality problem: greater search thoroughness can improve output quality but increases execution time, and the best trade-off depends on the specific HPC system.

The application implements an asynchronous pipeline with stages such as reader, parser, bucketizer, virtual screening / docking, and writer. The screening stage is the bottleneck and can offload computation to GPUs through a CUDA implementation. The paper identifies knobs that affect both throughput and quality—align_split, optimize_split, repetitions, num_restarts, clipping, and sim_thresh—and knobs that affect throughput only—cuda_threads and buffer_size. Some knobs are compile-time parameters, so exploring them requires rebuilding LiGen from source.

Quality and throughput are measured separately. Throughput is evaluated on 200 MEDIATE molecules replicated 10,000 times, giving a 2 million-molecule workload. Quality is measured on the PDBbind refined set, containing about 4,000 complexes with known co-crystallized poses, using the 75th percentile of RMSD,

$R(x)=\text{75th percentile of RMSD values},$

with feasibility threshold

$R(x)\le R_{\max}, \qquad R_{\max}=2.1.$

The paper defines the joint optimization target as

$f(x)=R^3(x)\,T(x),$

where $T(x)$ is execution time and lower is better.

The technical contribution of the autotuning study is not a new virtual screening algorithm but two parallel constrained-optimization methods for tuning LiGen on distributed HPC systems: EMaliboo, a distributed ensemble of independent MALIBOO agents, and PAMaliboo, a centralized asynchronous extension using a Kriging Believer strategy and periodic retraining of the constraint model. In the prototype environment with real LiGen executions, both methods significantly outperform OpenTuner: EMaliboo achieves about 40% improvement over OpenTuner, PAMaliboo achieves about 42% improvement, and relative to the expert default configuration EMaliboo is 35% better while PAMaliboo is 33% better. The reported final incumbents are 1279.1 for EMaliboo, 1304.9 for PAMaliboo, 2204.9 for OpenTuner, and 1943.9 for the default configuration (Guindani et al., 2024).

4. LiGen as GAN-augmented spectral fingerprinting for indoor localization

In "LiGen: GAN-Augmented Spectral Fingerprinting for Indoor Positioning" (Lin et al., 5 Aug 2025), LiGen is an indoor positioning system that combines ambient-light spectral fingerprinting with GAN-based data augmentation. The problem setting is standard fingerprint localization: given a fingerprint vector collected at an unknown position, predict the 2D coordinate in the room. The paper argues that spectral light fingerprints are a more stable sensing modality than Wi‑Fi RSSI and therefore better suited to fine-grained localization.

The sensing stack uses the AS7341 spectral sensor on a Raspberry Pi 4B, capturing light intensity over multiple wavelength channels from roughly 350 nm to 1000 nm. The reported room is 7 m × 7 m, divided into an 8 × 8 grid with 100 cm spacing, giving 64 reference points. At each point, the system collects 32 samples at 1-second intervals, for a total of 2048 samples. The temporal-stability argument is supported by the reported normalized average standard deviation at a fixed location: 0.00049 for spectral data and 0.03254 for Wi‑Fi.

LiGen has three stages: fingerprint acquisition, GAN-based augmentation, and localization model training. The final estimator is a deep MLP for coordinate regression. Two augmentation strategies are introduced. PointGAN is a conditional GAN that generates fingerprints conditioned on coordinates, so its samples are directly labeled. FreeGAN is an unconditional GAN whose outputs are pseudo-labeled by a weak localization model; the paper notes that there are effectively two identical localization MLPs in this pipeline, one used as the weak labeler and another trained on the augmented dataset.

The empirical claim is that both the sensing modality and the augmentation strategy matter. The paper reports that spectral fingerprints reduce localization error by approximately 50% compared with Wi‑Fi in most cases. In the light-signal-only setting, the DataLoc+ baseline has average error about 3.2 m, the plain MLP about 1.5 m, PointGAN / FreeGAN before augmentation achieve best-case around 0.5 m, and after augmentation the error improves to about 0.09 m. The paper also states that the best-case error in clean versus cluttered conditions differs by less than 0.2 m, which it interprets as evidence of robustness in cluttered environments (Lin et al., 5 Aug 2025).

5. Cross-domain comparison

Taken together, these uses show that LiGen is not a single research lineage but a reused label spanning symbolic-neural NLP, HPC-enabled drug discovery, and sensor-based indoor localization. The shared name masks major differences in input modality, optimization target, and interpretability regime.

LiGen variant	Primary input	Primary output
Logical grammar embedding	Corpus of sentences	Lexical categories, syntactic rules, transparent derivations
Virtual screening application	Large ligand libraries and target proteins	Ranked ligands under quality-throughput constraints
Spectral fingerprint localization	Ambient-light spectral fingerprints	2D room coordinates

The first LiGen is centered on formal composition rules of grammar and unsupervised inference from raw text. The second is an asynchronous pipeline whose practical issue is constrained autotuning on HPC systems. The third is a coordinate-regression system whose core innovation is the combination of spectral fingerprints with PointGAN and FreeGAN augmentation. This suggests that references to “LiGen” are meaningful only when accompanied by field, paper title, or arXiv identifier (Deyo et al., 2023, Guindani et al., 2024, Lin et al., 5 Aug 2025).

6. Disambiguation in adjacent arXiv literature

Several nearby names intensify the ambiguity. LimGen denotes a dataset of 4068 research papers from the ACL Anthology for the task of Suggestive Limitation Generation (SLG) and is unrelated to any of the three LiGen systems (Faizullah et al., 2024). LIGER refers to Long-horizon Instruction GEneration with logic and attribute self-Reflection, a training-free framework for generating coherent image sequences for multi-step tasks (Suo et al., 12 Mar 2025). LOGen means LiDAR Object Generator, a conditional DDPM for point-cloud objects in nuScenes (Kirby et al., 2024). LaGen is a long-horizon autoregressive LiDAR scene generator, and its paper explicitly warns that “LiGen” is not the method being proposed (Zhou et al., 26 Nov 2025).

A common misconception is therefore to assume that “LiGen” names a stable family of related generative models. The arXiv record summarized here does not support that reading. Instead, the exact label must be resolved against paper-specific context: LiGen/LGE for logical grammar induction, LiGen for EXSCALATE virtual screening, and LiGen for spectral indoor positioning. In bibliographic, software, and systems contexts, this distinction is not cosmetic; it determines the domain assumptions, datasets, metrics, and implementation stack to which any claim belongs.