DRACULA: Multidisciplinary Research Artifacts

Updated 10 July 2026

DRACULA is a term for distinct research artifacts spanning unsupervised ML, cryogenic detector testing, recursive data compression, and ACL2 pedagogy.
In deep research agents, the DRACULA dataset focuses on action-level feedback, isolating intermediate decisions with over 8,000 action preferences from expert annotations.
Applications range from reducing SN Ia spectra dimensionality with deep autoencoders to precise cryogenic testing of astrophysics detectors and structured feature extraction in sequential data.

DRACULA denotes multiple distinct research artifacts across machine learning, software engineering, and astrophysics. In current arXiv usage, the label appears as a dataset and study protocol for Scientific Deep Research agents, a cryogenic irradiation facility for sub-Kelvin detector testing, a Python toolbox for unsupervised analysis of Type Ia supernova spectra, a linear-programming framework for recursive compression of sequential data, and a classroom-oriented ACL2 environment built on DrRacket. In adjacent astronomical literature, the same query space also retrieves papers centered on the constellation Draco rather than on any DRACULA acronym expansion (Balepur et al., 26 Apr 2026, Sauvage et al., 25 Apr 2026, Sasdelli et al., 2015, Paskov et al., 2015, Eggensperger, 2013).

1. Terminology, capitalization, and disciplinary scope

The term is not monosemous. Its meaning depends on capitalization, expansion, and disciplinary context.

Form	Domain	Expansion or role
DRACULA	Deep research agents	Dataset and study protocol for intermediate-action feedback
DRACuLA	Cryogenic detector testing	Detector irRAdiation Cryogenic faciLity for Astrophysics
DRACULA	Astronomical ML	Dimensionality Reduction And Clustering for Unsupervised Learning in Astronomy
Dracula	Sequential data learning	Recursive dictionary-compression framework
DrACuLa	ACL2 pedagogy	DrRacket plugin / IDE for ACL2

This multiplicity is structurally important. The same string indexes a benchmark for long-horizon agents, a transportable dilution-refrigerator platform, an unsupervised astronomy package, a polyhedrally structured compression model, and a pedagogic programming environment. A plausible implication is that citation, expansion, and typography are necessary for disambiguation even within technical corpora.

2. DRACULA as action-level supervision for deep research agents

In the agent literature, DRACULA is the first dataset with user feedback on intermediate actions for Scientific Deep Research systems, rather than a benchmark that scores only the final report (Balepur et al., 26 Apr 2026). The central problem is that deep research agents make many sequential decisions—what papers to retrieve, how to organize sections, what to emphasize, and what to explain—while prior evaluation protocols largely collapse those decisions into a single final-report judgment. DRACULA isolates the bottleneck of action selection.

The study used a two-stage protocol. For each query, the system proposed 8 actions: four generic actions conditioned only on the query, and four “paper actions” conditioned on the query plus user-selected papers as context. The actions fell into four qualitative categories: Content, Style, Specificity, and Research Ideas. Users selected the actions they wanted, the DR system generated a report using the selected actions, and users then judged whether each selected action was actually executed well in the report. The interface presented actions as a checklist and highlighted where each action was reflected in the report.

The dataset was collected over five weeks from 19 expert CS researchers, with total annotation time of about 450 hours. It contains 8,103 action preferences, 5,230 execution judgments, 902 unique queries, 7,731 unique actions, and 13,333 query-action-judgment pairs. A follow-up relabeling study used 18 annotators, four queries each, five months later. Formally, the paper defines action prediction as a binary classification problem: given a query $q$ and an action $a$ , predict whether the user selected it, $\hat{l}_{select} \in \{0,1\}$ . The data are split chronologically into 80/10/10 train/dev/test, yielding 5982 / 1140 / 918 examples, and the main metric is Macro-F1 with bootstrap sampling $(n = 500)$ .

The empirical findings are highly specific. Few-shot prompting alone predicts action selection only modestly above random, especially for generic actions. The strongest improvement comes from the user’s full past selection history, which outperforms all-user history, self-reported intent or expertise, selected papers, inferred preference rules, rationale text alone, and long-context compression tricks. Model-inferred rules outperform user-written rules, but still do not match full history. The paper also shows that execution prediction is easier than action prediction, supporting the claim that the principal difficulty is not whether a selected action can be carried out, but whether it is the action the user wanted in the first place.

The follow-up analysis attributes part of the simulation ceiling to changing unstated goals. Users may alter their preferences because they now want “solutions, not new insights,” or because they no longer need a basic definition. This suggests that preference instability is not reducible to model error alone. In response, the paper studies an online intervention that generates actions from revealed preferences—past selections and judgments. The reported selection rates are 0.817 ± 0.030 for generic/query only, 0.682 ± 0.036 for research papers, 0.741 ± 0.034 for stated preferences, and 0.857 ± 0.027 for revealed preferences. Generic and revealed-preference actions overlap by only 12.1%, indicating substantive personalization rather than trivial reformulation.

3. DRACuLA as a cryogenic irradiation facility for astrophysics

In instrumentation, DRACuLA stands for Detector irRAdiation Cryogenic faciLity for Astrophysics, a mobile dilution-refrigerator irradiation platform developed at the Institut d'Astrophysique Spatiale to expose sub-Kelvin astrophysics detectors to particle beams while they operate at nominal cryogenic temperature (Sauvage et al., 25 Apr 2026). The facility was motivated by the need for pre-launch radiation qualification of increasingly sensitive superconducting detectors such as KIDs and TESs, and by the limitation of room-temperature irradiation tests, where radiation-induced defects can anneal before low-temperature characterization.

The design requirements were explicitly stringent: a base temperature as low as 10 mK, 500 $\mu$ W cooling power at 100 mK, compatibility with any accelerator beam line, and a sample space large enough for detector arrays or part of a focal plane. DRACuLA is built around a Bluefors LD400 dilution refrigerator with four successive cold stages at 50 K, 4 K, 1 K, and 100 mK. The 50 K and 4 K stages are provided by a Cryomech PT415-RM pulse tube cooler. The platform includes 24 RC/RF low-pass filtered readout lines using QDevil QFilter-II, four superconducting coaxial wires from room temperature to the cold plate, and four independent temperature measurement bridges. Mobility is engineered via a compact frame with wheels and micro-vibration attenuators. Beam access is enabled by four KF-50 ports through the vacuum can and inner shields at $0^\circ$ , $45^\circ$ , $90^\circ$ , and $180^\circ$ .

The September 2025 campaign at PARTREC in Groningen demonstrated beam-line integration on PRIMA/PRIMAger KID arrays developed by SRON. The detector samples were maintained at 120 mK throughout an approximately 12-hour proton irradiation run at 184 MeV, with beam flux at detector level of $\sim 2\times 10^6$ protons/cm $a$ 0/s. GEANT4 simulation tracked the residual beam energy from 184 MeV to approximately 150 MeV at detector level. Mechanical alignment used a supplementary support frame raising the cryostat to 1.5 m, a laser along the nominal beam axis, and an external collimator reducing the beam cross section to $a$ 1 mm at the detector plane. Before beam, the 120 mK stage stability was 120 mK $a$ 2 33 $a$ 3K; during beam it was 120 mK $a$ 4 695 $a$ 5K. After irradiation, about 15% of data points were flagged as glitches, compared to $a$ 6 before irradiation. This was attributed primarily to radio-activation of the gold-plated copper sample enclosure, with $a$ 7Cu identified as the dominant isotope and $a$ 8 h. Post-irradiation science measurements were therefore delayed by 24 hours.

A companion study places the facility in the broader context of cosmic-ray susceptibility of cryogenic detectors, using the Planck High Frequency Instrument experience as motivation (Besnard et al., 20 Apr 2026). After the Planck launch in 2009, HFI bolometers were considerably affected by cosmic rays; more than 30% of the data were affected, and several years of post-processing were required. IAS had already carried out three successful campaigns on bolometers, TES, and KID technologies. The paper reports two recent campaigns: a 2024 TES campaign at ALTO – IJCLab using 18 and 22 MeV protons on NIST TES prototypes for LiteBIRD MHFT, and a 2025 KID campaign at PARTREC using total dose proton irradiation of 6 krad on SRON KID arrays for PRIMA, intended to simulate 10 years of cosmic-ray exposure at L2.

The TES campaign recorded approximately 2800 events and identified three different glitch shapes, presumably corresponding to protons striking different regions of the detector chip. MCMC was used to extract the time constant, and denoising used Fourier-domain processing with DFT-based filtering of stochastic noise and a 10 kHz frequency cut. The KID campaign revealed substantial secondary-particle activity after irradiation, with about 130 glitches per second and about 15% of the dataset affected. Analysis remained in progress, with the remaining task of deconvoluting low-amplitude glitches from stochastic noise and quantifying performance degradation. The same DFT method, previously used only on individual glitches from the LiteBIRD campaign, was being extended to a larger dataset and planned for validation on a subset of Planck HFI data.

4. DRACULA as unsupervised learning infrastructure for Type Ia supernova spectra

In astronomical machine learning, DRACULA expands to Dimensionality Reduction And Clustering for Unsupervised Learning in Astronomy, a public Python toolbox introduced to study the spectroscopic diversity of Type Ia supernovae without slow, subjective, object-by-object visual classification (Sasdelli et al., 2015). The immediate problem is whether SN Ia spectra form genuinely distinct subclasses or a continuous sequence with a few extreme outliers. Traditional subtype labels such as High Velocity, 91T-like, and 91bg-like were historically defined through human inspection of a small number of spectra and empirical rules; DRACULA recasts that problem as dimensionality reduction plus clustering.

The workflow has three stages. First, the method uses transfer learning across epochs: rather than train only on spectra near maximum light, it builds a large matrix using all available SN Ia spectra across epochs, and only afterward selects spectra within $a$ 9 days of $\hat{l}_{select} \in \{0,1\}$ 0-band maximum for final clustering. Second, the data are preprocessed by Savitzky–Golay smoothing and by taking the derivative of the logarithm of flux, reducing calibration and distance-systematic effects while preserving line information. Third, the compressed representation is clustered, primarily with K-means, and interpreted using visualization tools such as Self-Organizing Maps and Isomap.

The central reduction method is a deep autoencoder. The key quantitative claim is that deep learning identifies the relevant structure in a 4 dimensional feature space, plus 1 for time evolution, while standard PCA barely achieves similar results using 15 principal components. Training on the full epoch range is crucial: with transfer learning the residual variance in reconstruction stabilizes with only 4 features, whereas without transfer learning even 10 features are insufficient to reach the same quality. The paper interprets this low-dimensionality as evidence that the progenitor system and explosion mechanism can be described by a small number of initial physical parameters.

Clustering in the reduced space yields a hierarchical group structure rather than a single rigid partition. With 2 clusters, the strongest separation is by line velocity; with 3 clusters, 91bg-like objects separate into their own group; with 4 clusters, 91T-like objects separate as well. The resulting structure is in close agreement with a previously suggested classification scheme, especially the Wang et al. organization of subtypes. The paper therefore treats line velocity as a first order effect in subtype determination, followed by 91bg-like events. At the same time, the latent-space geometry supports the view that SN Ia occupy a continuous manifold of spectral properties with transitional and outlier objects.

The package is presented as public infrastructure through COINtoolbox, with modular support for dimensionality reduction, clustering, plotting, parameter comparison, and SOM visualization. It is implemented with widely used libraries such as scikit-learn, while deep learning uses H2O. Its significance lies not only in reproducing familiar subtype labels, but in mapping the geometry of spectral diversity in a statistically coherent way for large survey regimes.

5. Dracula as recursive compression and feature learning for sequential data

In machine learning for sequential data, Dracula is a framework for unsupervised feature selection from sequences such as text, protein strings, or tagged sentences (Paskov et al., 2015). Its premise is that representation learning can be posed as lossless compression: frequently reused or structurally informative $\hat{l}_{select} \in \{0,1\}$ 1-grams should enter a corpus-specific dictionary. The distinctive step beyond shallow dictionary learning is recursive compression of the dictionary itself. In effect, Dracula is a deep extension of Compressive Feature Learning.

The formalism begins with an alphabet $\hat{l}_{select} \in \{0,1\}$ 2, a corpus $\hat{l}_{select} \in \{0,1\}$ 3, and the set $\hat{l}_{select} \in \{0,1\}$ 4 of all $\hat{l}_{select} \in \{0,1\}$ 5-grams appearing in the corpus. A pointer is defined as

$\hat{l}_{select} \in \{0,1\}$ 6

meaning that the string $\hat{l}_{select} \in \{0,1\}$ 7 is used at location $\hat{l}_{select} \in \{0,1\}$ 8 in string $\hat{l}_{select} \in \{0,1\}$ 9. In CFL, one learns a dictionary $(n = 500)$ 0 and a set of pointers reconstructing the documents. Dracula augments this with dictionary-reconstruction pointers $(n = 500)$ 1, so a full compression is $(n = 500)$ 2: documents are reconstructed using dictionary entries, and dictionary strings are themselves reconstructed using pointers to shorter strings. Unigrams serve as atomic symbols and play a special role in dictionary reconstruction.

The optimization is cast as a binary linear program with an LP relaxation. The exact problem is NP-complete in general, but the reconstruction subproblems are structurally tractable because the coverage matrices have interval structure and can be transformed into min-cost flow problems. Polyhedrally, the exact feasible set is $(n = 500)$ 3 and the relaxed outer approximation is $(n = 500)$ 4, with

$(n = 500)$ 5

The paper emphasizes that the solution path under continuous cost variation is well behaved: optimal solutions move across intersecting faces of the LP polyhedron rather than jumping arbitrarily.

A simple cost scheme controls the learned hierarchy. Dictionary costs are $(n = 500)$ 6, document pointer costs are $(n = 500)$ 7, and dictionary pointer costs are $(n = 500)$ 8 when the pointer uses a unigram and $(n = 500)$ 9 otherwise. Here $\mu$ 0 governs how difficult it is for an $\mu$ 1-gram to enter the dictionary, $\mu$ 2 governs the overall cost of reconstructing dictionary strings, and $\mu$ 3 directly controls depth by making characters cheaper or more expensive relative to higher-level pointers. For $\mu$ 4, all dictionary $\mu$ 5-grams of length at most $\mu$ 6 are built entirely from characters.

Every Dracula solution can be represented as a directed acyclic graph whose vertices are characters, dictionary strings, and documents, and whose edges are pointers. This is the paper’s notion of “deep”: hierarchical compression depth rather than neural-network depth. Features are extracted as bag-of- $\mu$ 7-grams-style counts from document reconstructions, and then extended by dictionary diffusion. If $\mu$ 8 is the top-level feature matrix and $\mu$ 9 collects dictionary-string feature vectors, the diffused representation is

$0^\circ$ 0

Because the dictionary graph is a DAG, the series converges and effectively yields $0^\circ$ 1. A subtle result follows: unregularized linear models are invariant to the dictionary reconstruction scheme, whereas regularized models are not. When regularization is present, the structure enters through $0^\circ$ 2, producing a graph-aware regularization effect analogous to flow conservation.

The reported experiments span protein-sequence classification, stylometry via POS-tag sequences, and sentiment prediction. Across these tasks, the paper presents Dracula as a method that compresses well, discovers hierarchical structure, and yields useful downstream features, with particular value when lower-order and higher-order sequential patterns must be balanced by interpretable parameters.

6. DrACuLa as an ACL2 teaching environment

In programming-language pedagogy, DrACuLa is a plugin for DrRacket that acts as an IDE for ACL2 and is primarily intended for classroom use (Eggensperger, 2013). It arose from the same motivation as ACL2 Sedan and later Proof Pad: the standard ACL2 installation is closely tied to Emacs, whereas many students accustomed to desktop IDEs react negatively to adapting to an unfamiliar interaction model. DrACuLa therefore built on DrRacket’s student-friendly interface and pedagogic reputation.

The system provided a familiar Lisp editor with parentheses matching and auto-indentation, features described as indispensable. It also extended ACL2 toward a modular programming style inspired by Scheme, with modules separated into interfaces and implementations and with mechanically verified contracts. Its teaching-oriented tooling included teachpacks, a QuickCheck-inspired property-testing library called DoubleCheck, and a REPL. DoubleCheck was significant not only for randomized testing, but as an accessible jumping-off point for students learning to think in terms of properties before proofs and to obtain counterexamples when proofs failed.

The same paper also documents the limitations that made DrACuLa the principal predecessor and comparison point for Proof Pad. Installation required acquiring and installing three separate software components, using the command line, and navigating potential pitfalls that were hard for beginners to recover from. Because DrACuLa executes ACL2 definitions in Scheme, only part of ACL2 is available; arrays and macros are not supported, and support would require “a lot of work.” The dual-implementation architecture also creates the possibility of divergence between DrACuLa behavior and real ACL2 behavior.

Questionnaire and classroom experience identified recurring user complaints: no runtime debugger or program stepper, weak runtime error messages, syntax errors that do not locate the problem well, slow startup, a hard-to-use multi-tab editor, poor documentation, and lack of autocomplete and argument display. Proof Pad is presented as building on the work of DrACuLa and ACL2s while changing the architectural choice: direct ACL2 integration rather than Scheme-based emulation, a simpler installation model, and workflow features such as the proof bar and clearer results handling. DrACuLa is therefore historically important as the system that demonstrated both the pedagogic viability and the architectural fragility of classroom-oriented ACL2 environments.

7. Draco-associated astronomical literature in the same query space

Some arXiv records associated with DRACULA queries concern Draco-region astronomy rather than a DRACULA acronym expansion. These papers are scientifically unrelated to the datasets, toolboxes, and facilities above, but they form part of the same retrieval landscape.

One cluster concerns the debated 3.5 keV X-ray line in the Draco dwarf spheroidal galaxy. A 2014 simulation study argued that Draco was an unusually clean and powerful test case for decaying dark matter, using Aquarius project haloes to predict flux ratios between the Galactic Center, M31, blank sky, and Draco (Lovell et al., 2014). The paper quoted an expected Draco line flux of $0^\circ$ 3, with $0^\circ$ 4 as the most probable value, and concluded that in 95 per cent of mock observations a 1.3 Msec-class XMM-Newton pointing could discover or rule out a 3.5 keV decay feature at the $0^\circ$ 5 level. A subsequent deep XMM analysis of Draco found no such line in either MOS or PN, with spectra consistent with an unfolded power law plus instrumental lines and no improvement upon adding a $0^\circ$ 6 keV Gaussian (Jeltema et al., 2015). The paper reported that the resulting upper limit rules out a dark matter decay origin for the previously claimed 3.5 keV line at greater than 99% confidence, and translated the constraint into sterile-neutrino parameter space as $0^\circ$ 7 and $0^\circ$ 8 at 95% confidence.

A second cluster concerns cataclysmic variables in Draco. One paper reports the discovery of DDE 32, identified with the ROSAT source 1RXS J161935.7+524630, as a new magnetic cataclysmic variable or polar (Denisenko et al., 2016). The ROSAT source had count rate $0^\circ$ 9 cnts/s and hardness ratios $45^\circ$ 0, $45^\circ$ 1. Photometry showed nearly 2 magnitudes of variability with a period refined to

$45^\circ$ 2

and the light curve changed from one dominant peak per cycle in 2005–2013 Catalina data to two peaks of nearly equal height in 2015. Together with prominent helium emission lines in the SDSS spectrum and He II 4686 Å effective width exceeding 30% of that of $45^\circ$ 3, the paper interprets the system as a polar whose accretion mode changed from one pole before 2014 to two poles in 2015. The paper also notes that it may possibly even be an asynchronous system like BY Cam.

Another Draco transient was established as a new SU UMa-type dwarf nova, GSC2.3 N152008120 / USNO-B1.0 1508-0249096 (0905.0809). Its first recorded superoutburst in October 2008 reached magnitude 14.9, had amplitude approximately 6 magnitudes, lasted at least 11 days, and was followed about 11 days later by a short-lived rebrightening by more than 2 magnitudes. Time-series photometry covering about 44 hours revealed clear superhumps with mean period 0.07117(1) d and amplitude 0.12 mag. The paper reports a distinct shortening of the superhump period around cycle 80, from $45^\circ$ 4 before the break to $45^\circ$ 5 after it, with only marginal evidence for an earlier period increase. These results place the object within the standard SU UMa phenomenology rather than the WZ Sge subclass.

Taken together, these Draco papers illustrate that the DRACULA query space spans not only acronymic technical systems, but also astrophysical targets and transients located in Draco. A plausible implication is that bibliographic disambiguation by expansion and domain is as important for astronomy retrieval as it is for software and machine-learning literature.