LAGER: Multi-Domain Research Applications

• LAGER is a polysemous term representing distinct constructs in astronomy, cephalometric algorithms, document entity recognition, LLM evaluation, and brewing systems biology.
• In astronomy, LAGER uses a custom NB964 filter on DECam to detect z≈7 Lyα-emitters, achieving key performance metrics like 90% peak transmission and precise field coverage.
• Across other domains, LAGER advances cephalometric planning, improves few-shot layout-aware token modeling, optimizes LLM alignment, and elucidates lager yeast proteostasis for enhanced brewing.

LAGER is a polysemous research term used for several unrelated constructs across astronomy, clinical geometry, document intelligence, LLM evaluation, and brewing systems biology. In observational cosmology, it most prominently denotes Lyman Alpha Galaxies in the Epoch of Reionization, a DECam narrowband survey designed to discover Ly$\alpha$ emitters at $z\sim 7$ and use them to probe cosmic reionization. In craniomaxillofacial planning it denotes the LAndmark GEometric Routine for estimating a premorbid midsagittal plane. In document AI it denotes layout-aware graph-based entity recognition, a graph-augmented few-shot entity-recognition model for document images. In LLM evaluation it denotes a framework for improving LLM-as-a-Judge alignment with human scoring via internal representations. In brewing-oriented systems biology, the term appears in its ordinary sense, referring to lager fermentation and lager yeast physiology (Zheng et al., 2019, Jajoo et al., 2018, Krishnan et al., 2023, Lai et al., 5 Aug 2025, Telini et al., 2019).

1. Nomenclature and research domains

The term has no single cross-disciplinary definition. Its astronomical usage is the most extensively developed programmatically: a wide/deep narrowband survey using DECam and the custom NB964 filter to isolate $z\sim 7$ Ly$\alpha$-emitting galaxies and infer the ionization state of the intergalactic medium during the epoch of reionization. Other usages are acronymic and domain-specific: a landmark-based geometric routine in CMF surgical planning, a graph-enhanced document entity recognizer, and a layer-aggregated scoring framework for LLM-as-a-judge alignment. The brewing literature represented here uses “lager” in the context of industrial beer fermentation rather than as an acronym.

Usage	Expansion or sense	Representative work
LAGER	Lyman Alpha Galaxies in the Epoch of Reionization	(Zheng et al., 2019)
LAGER	LAndmark GEometric Routine	(Jajoo et al., 2018)
LAGER	layout-aware graph-based entity recognition	(Krishnan et al., 2023)
LAGER	LLM-as-a-Judge alignment with human via internal representations	(Lai et al., 5 Aug 2025)
lager	lager yeast and lager beer fermentation context	(Telini et al., 2019)

This multiplicity is not merely terminological. Each usage encodes a distinct methodological agenda: survey cosmology, cephalometric optimization, graph-structured document representation learning, internal-state aggregation in LLM evaluation, and inter-organellar stress biology.

2. LAGER as an astronomical survey infrastructure

In astronomy, LAGER is a DECam narrowband imaging program built to detect Ly$\alpha$ emitters at $z\sim 7$, where Ly$\alpha$ transmission is strongly modulated by neutral hydrogen in the reionizing IGM. The enabling hardware is the custom NB964 filter, designed for the Dark Energy Camera on the Blanco 4 m telescope at CTIO. Its measured operating-band parameters are a central wavelength of $964.2$ nm and FWHM $9.2$ nm, very close to the design optimum near $W_c \simeq 964\pm 1\,\mathrm{nm}$ and $z\sim 7$0 derived from sky transmission, OH background, DECam CCD quantum efficiency, and LAE-yield optimization (Zheng et al., 2019). The basic redshift relation is

$z\sim 7$1

so NB964 targets Ly$z\sim 7$2 near $z\sim 7$3, with an approximate half-power interval $z\sim 7$4 to $z\sim 7$5 (Zheng et al., 2019).

The filter paper also established the instrumental constraints that shaped the survey. Because DECam operates in an $z\sim 7$6 converging beam, the passband had to be pre-compensated for angle-of-incidence shifts across a $z\sim 7$7 mm clear aperture. The delivered filter has peak transmission about 90% on average, spatial variation in peak transmission of 3.83%, and stringent out-of-band blocking of $z\sim 7$8 from 300–1050 nm and $z\sim 7$9 from 1050–1200 nm (Zheng et al., 2019). On-sky, NB964/$z\sim 7$0 sky count-rate ratios of $z\sim 7$1 matched design expectations, validating the sky-noise optimization (Zheng et al., 2019).

The first LAGER results from COSMOS described the program as the largest narrowband survey for $z\sim 7$2 galaxies then available, with a design goal of 12 deg$z\sim 7$3 over four fields and total comoving volume of about $z\sim 7$4. The initial COSMOS data set comprised 34 hr of NB964 imaging over a 3 deg$z\sim 7$5 field, reaching a $z\sim 7$6 limiting magnitude of 25.6 AB in a $z\sim 7$7 diameter aperture (Zheng et al., 2017). Later spectroscopic reporting described LAGER as a 24 deg$z\sim 7$8 survey using deep DECam imaging in multiple fields, which suggests that the program footprint expanded over time (Harish et al., 2021).

Although conceived for $z\sim 7$9 Ly$\alpha$0 cosmology, the same NB964 infrastructure also yields large foreground emission-line samples. In a single $\alpha$1 COSMOS DECam pointing, the survey identified 1577 $\alpha$2 H$\alpha$3 emitters, 3933 $\alpha$4 [OIII] emitters, and 5367 $\alpha$5 [OII] emitters, enabling luminosity-function constraints and contaminant characterization directly relevant to LAE selection (Khostovan et al., 2020).

3. Photometric selection, spectroscopic confirmation, and reionization inferences

LAGER’s LAE selection is field-dependent but conceptually uniform: a strong NB964 excess plus non-detection in bluer veto bands. The earliest COSMOS paper used $\alpha$6, identifying 23 candidate $\alpha$7 LAEs in the central 2 deg$\alpha$8 region, all with $\alpha$9 and $\alpha$0 (Zheng et al., 2017). The CDFS spectroscopic follow-up paper selected candidates with

$\alpha$1

together with non-detections in $\alpha$2, while the later COSMOS and WIDE-12 follow-up used SNR $\alpha$3 in NB964 and $\alpha$4 (Yang et al., 2019, Harish et al., 2021).

The first spectroscopic LAGER paper observed 12 of the 23 COSMOS candidates with IMACS on Magellan and reported Ly$\alpha$5 detections in 6 sources. Because only 9 had spectra deep enough to test the selection fairly, the effective confirmation rate was $\alpha$6. Three of the confirmed objects, LAE-1, LAE-2, and LAE-3, were described as luminous LAEs with $\alpha$7, then the highest-Ly$\alpha$8-luminosity spectroscopically confirmed galaxies known at $\alpha$9 (Hu et al., 2017). Two of the lines, in LAE-1 and LAE-3, had strongly asymmetric red-wing profiles with weighted skewness $z\sim 7$0 and $z\sim 7$1, reinforcing the Ly$z\sim 7$2 identification (Hu et al., 2017).

The 2019 CDFS follow-up confirmed two narrowband-selected galaxies as Ly$z\sim 7$3 emitters at $z\sim 7$4 and $z\sim 7$5 from IMACS spectroscopy. Their Ly$z\sim 7$6 luminosities were $z\sim 7$7 and $z\sim 7$8, with intrinsic Ly$z\sim 7$9 FWHM of 229 km s$\alpha$0 and 318 km s$\alpha$1 from IMACS. For CDFS-LAE1, deep FIRE spectroscopy yielded non-detections of high-ionization UV nebular lines and upper limits such as $\alpha$2 and $\alpha$3, which the authors interpreted as evidence that the ionizing emission is dominated by normal star formation rather than AGN activity (Yang et al., 2019).

A larger Keck/LRIS campaign later observed 21 LAGER candidates in COSMOS and WIDE-12 and detected emission lines in 17. Of those, 15 were interpreted as $\alpha$4 Ly$\alpha$5 emitters, while 2 were low-redshift [OIII] interlopers at $\alpha$6. Including two earlier Magellan confirmations among the LRIS non-detections, the effective success rate became 17/21, reported as $\alpha$7 or $\alpha$8. With this work, the total number of spectroscopically confirmed LAGER LAEs reached 33, and the stacked spectrum yielded a $\alpha$9 upper limit of

$964.2$0

again favoring star-formation-dominated ionization (Harish et al., 2021).

On the luminosity-function side, the first COSMOS-only analysis found 23 candidate LAEs and a bright-end excess driven by 4 galaxies with $964.2$1. It reported a fourfold reduction in Ly$964.2$2 luminosity density from $964.2$3 to $964.2$4 and inferred a factor of $964.2$5 suppression of Ly$964.2$6 transmission by the $964.2$7 IGM, implying $964.2$8–0.6 for Ly$964.2$9 velocity offsets of $9.2$0–$9.2$1 (Zheng et al., 2017). The subsequent two-field “tale of two LAGER fields” analysis expanded the sample to 79 LAEs—49 in COSMOS and 30 in CDFS—confirmed the bright end bump in COSMOS with 6 LAEs above $9.2$2, showed that the bump was absent in CDFS, and derived an average neutral fraction $9.2$3–0.4 at $9.2$4 (Hu et al., 2019). The later four-field synthesis used 174 LAE candidates over $9.2$5 Mpc$9.2$6, found that the Ly$9.2$7 luminosity density declined at the same rate as the UV continuum luminosity density from $9.2$8 to $9.2$9, and concluded that the data were consistent with a fully or largely ionized $W_c \simeq 964\pm 1\,\mathrm{nm}$0 IGM, with a conservative $W_c \simeq 964\pm 1\,\mathrm{nm}$1 limit $W_c \simeq 964\pm 1\,\mathrm{nm}$2 (Wold et al., 2021).

Taken together, these results show that LAGER evolved from a photometric candidate survey into a spectroscopically anchored reionization program. A plausible implication is that increasing field coverage, explicit completeness modeling, and direct treatment of field-to-field variance materially changed the inferred global neutral fraction.

4. Overdensities, protoclusters, and the topology of reionization

One of the most consequential LAGER findings is the protocluster LAGER-z7OD1 at $W_c \simeq 964\pm 1\,\mathrm{nm}$3. The discovery paper identified an overdensity of 21 member LAEs, of which 16 were spectroscopically confirmed, with a galaxy overdensity

$W_c \simeq 964\pm 1\,\mathrm{nm}$4

The structure is elongated, consists of two sub-protoclusters, and is expected to evolve into a present-day cluster of mass

$W_c \simeq 964\pm 1\,\mathrm{nm}$5

Its surveyed volume was given as $W_c \simeq 964\pm 1\,\mathrm{nm}$6, or $W_c \simeq 964\pm 1\,\mathrm{nm}$7, and the summed ionized bubble volume inferred for the 21 members was $W_c \simeq 964\pm 1\,\mathrm{nm}$8, slightly larger than the protocluster volume itself (Hu et al., 2021).

That early ionization-budget argument was later sharpened with new JWST/NIRSpec, JWST/NIRCam, and Keck/LRIS observations of nine LAEs in LAGER-z7OD1. Using systemic redshifts from Balmer and [OIII] lines together with Ly$W_c \simeq 964\pm 1\,\mathrm{nm}$9 velocity offsets and transmission estimates, the 2025 study reconstructed a three-dimensional ionized topology. In the most conservative no-ISM-correction picture, the inferred distances to the first neutral patch were

$z\sim 7$00

while a dust-free ISM correction gave at least

$z\sim 7$01

The nine LAEs were found to occupy three distinct sub-clusters rather than a single monolithic bubble, and the conservative ISM-corrected topology implied ionized pockets of about $z\sim 7$02 cMpc, $z\sim 7$03 cMpc, and $z\sim 7$04 cMpc. The authors concluded that “five of the nine LAEs are plausibly the primary ionizing agent of the $z\sim 7$05 bubbles” and that the system is consistent with accelerated reionization in the protocluster environment (Martin et al., 15 Oct 2025).

LAGER has also motivated survey-statistical work on the Void Probability Function of LAEs. Using simulated LAGER-like geometries and LAE samples, the VPF study argued that even one DECam field could distinguish a mostly neutral from a mostly ionized universe, that four fields could distinguish 30%, 50%, and 95% ionized scenarios, and that eight fields could distinguish 30%, 50%, 73%, and 95% ionized scenarios, although 73% vs 95% remained difficult (Perez et al., 2022). This provides a clustering-based complement to luminosity-function analyses and supports the broader interpretation that LAGER fields are sampling a patchy, spatially inhomogeneous reionization process.

The protocluster work therefore changes the meaning of LAGER from a survey that counts bright LAEs to a program that maps ionized environments. This suggests that the survey’s most powerful legacy lies in linking luminous LAEs, overdensities, and ionized-bubble overlap within a single empirical framework.

5. Clinical and computational acronymic reuses

Outside astronomy, LAGER is also the name of a cephalometric algorithm for estimating a premorbid midsagittal plane in subjects with craniomaxillofacial deformity. The LAndmark GEometric Routine is based on three steps: quantifying landmark asymmetry by Euclidean Distance Matrix Analysis, recursively dropping outlier landmarks, and optimizing a midsagittal plane from the retained landmarks. If $z\sim 7$06 and $z\sim 7$07 are right- and left-side distance vectors, the asymmetry vector is

$z\sim 7$08

and the final plane minimizes weighted point-to-plane distances for unpaired landmarks and paired midpoints together with an angular term enforcing perpendicularity of left-right landmark vectors to the plane. In validation on 20 synthetic models derived from 5 normal subject skull CT models, all LAGER-generated midsagittal planes satisfied the clinical criteria $z\sim 7$09, $z\sim 7$10 mm, $z\sim 7$11 mm, and $z\sim 7$12 mm (Jajoo et al., 2018).

In document intelligence, LAGER denotes layout-aware graph-based entity recognition, a few-shot document-image entity recognizer built by placing Graph Attention Network layers on top of layout-aware transformer embeddings. Tokens are treated as graph nodes, and edges are formed either by $z\sim 7$13-nearest bounding boxes or by directional nearest-neighbor heuristics over multiple angles. With backbone embeddings $z\sim 7$14, the graph-enhanced representation is summarized as

$z\sim 7$15

or, in the multi-angle version,

$z\sim 7$16

Using LayoutLMv2 and LayoutLMv3 backbones, the method improved few-shot F1 on both FUNSD and CORD and showed smaller performance drops under shifting, scaling, and rotation; for example, with a LayoutLMv3 backbone in 5-shot FUNSD, vanilla performance was 72.07, versus 74.30 for LAGER$z\sim 7$17 and 75.22 for LAGER$z\sim 7$18 (Krishnan et al., 2023).

In LLM evaluation, LAGER denotes a lightweight framework for improving LLM-as-a-Judge alignment with human scoring via internal representations. Instead of relying only on the final-layer score token distribution, it aggregates score-token logits across layers of a frozen decoder-only LLM: $z\sim 7$19 followed by a softmax over candidate scores and the expected-score prediction

$z\sim 7$20

The tuned version learns only $z\sim 7$21 scalar layer weights. On FLASK, HelpSteer, and BIGGen, the paper reported improvements of up to 7.5% over the best baseline, and on downstream data selection the highest-LAGER-with-tuning subset achieved an AlpacaEval 2.0 length-controlled win rate of 12.65, above 11.82 for longest responses and 10.69 for SuperFiltering (Lai et al., 5 Aug 2025).

These three uses share a family resemblance despite their disciplinary separation: all treat internal structure as more informative than a naïve surface representation. In one case the hidden structure is bilateral craniofacial symmetry, in another it is token-topology under document transformations, and in the third it is layerwise representation geometry inside an LLM.

6. Lager fermentation and inter-organellar proteostasis

In the brewing paper represented here, “lager” is not an acronym but the fermentation context of an industrial Saccharomyces pastorianus system. The study is a Hypothesis and Theory manuscript centered on the claim that yeast performance in lager beer fermentation is governed by an integrated, inter-organellar proteostasis network spanning the endoplasmic reticulum, mitochondria, and cytosol. The authors refer to this as a CORE network of inter-organellar/cross-organellar communication and response, particularly relevant under high-gravity (HG) and very-high-gravity (VHG) brewing, where VHG wort is about 24 $z\sim 7$22P (Telini et al., 2019).

The empirical basis is a re-analysis of public DNA microarray data for the proprietary lager strain CB11 from Coors Brewing Limited. Using GSE9423, GSE10205, and GSE16376, and filtering differentially expressed genes by mean $z\sim 7$23, logFC SD $z\sim 7$24, and FDR $z\sim 7$25, the study identified 36 proteostasis-associated Pan-DEGs and 54 chaperone-associated Pan-DEGs (Telini et al., 2019). Upregulated genes implicated ER proteostasis, ERAD, the GET pathway, mitochondrial import and AAA proteases, cytosolic HSP systems, ubiquitin-proteasome machinery, and probable links to processing bodies and stress granules.

The physiological argument is that ethanol accumulation creates a heat-shock-like proteotoxic environment. The paper states that moderate ethanol, around 5% v/v, alters enzyme kinetics and affects metabolism and the cell cycle, whereas ethanol above 5% v/v more dramatically alters protein structure and promotes unfolded proteins and amorphous aggregates (Telini et al., 2019). On this basis, the manuscript proposes that viability, vitality, attenuation, flavor consistency, and serial repitching performance depend on coordinated proteostasis rather than on isolated compartment-specific stress responses. It also identifies methodological limits: the evidence is transcript-based rather than proteomic or functional, and the Affymetrix Yeast Genome 2.0 Array was not designed for the hybrid S. pastorianus genome (Telini et al., 2019).

Within the scope represented here, the lager-brewing usage of the term is therefore conceptually distinct from the acronymic usages but structurally similar in one respect: it also organizes a systems-level interpretation around hidden coordination among subsystems. In this case, the subsystems are organelles rather than survey fields, landmarks, tokens, or transformer layers.