Amadeus in Science & Technology

• Amadeus is a polysemous designation spanning scientific writing pedagogy, particle physics experiments, acoustic neutrino systems, language models, music generation, and customer segmentation.
• Each domain applies a structured methodology—from a 9-step scientific writing recipe to advanced retrieval-augmented frameworks and diffusion-enhanced music models—demonstrating its versatility.
• Empirical results in physics and digital applications highlight how Amadeus frameworks optimize processes, from detector calibration in kaon interactions to personalized travel recommendations.

Amadeus is a polysemous designation in the arXiv literature. It names a scientific-writing strategy for non-native users of English; a set of experimental programs in low-energy kaon–nucleus physics and acoustic neutrino detection; a family of Brazilian Portuguese LLMs; retrieval-augmented role-playing agents; symbolic-music generation systems; and an industrial framework for customer segmentation in flight search recommendation [0611013, (Scordo et al., 2015, Collaboration et al., 2010, Cruz-Castañeda et al., 20 May 2025, Park et al., 4 Aug 2025, Daniel et al., 2019, Daniel et al., 2019, Su et al., 28 Aug 2025)].

1. AMADEUS as a scientific-writing strategy

In scientific-writing pedagogy, AMADEUS refers to the strategy introduced in "Developing strategies to produce better scientific papers: a Recipe for non-native users of English" [0611013]. The paper states that the AMADEUS strategy had been used to produce scientific writing tools for non-native users of English for 15 years, emphasizes a learn-by-doing approach, and provides a 9-step recipe for students to compile writing material according to a procedure described as efficient in scientific writing courses [0611013].

The stated target population is non-native users of English, especially students and novice writers. The emphasis on a "learn-by-doing" approach indicates that AMADEUS is framed as a procedural writing pedagogy rather than only a descriptive style guide. The accompanying description characterizes it as a structured, step-by-step methodology centered on compiling writing material and improving scientific writing through iterative practice; this suggests a focus on reusable procedures for producing publishable prose rather than on isolated grammar correction alone [0611013].

A plausible implication is that AMADEUS belongs to a family of method-driven scientific-writing interventions in which writing is decomposed into repeatable operations. Within the available record, however, the minimally secure claims are its long-term use, its orientation toward non-native writers, its learn-by-doing emphasis, and its 9-step recipe [0611013].

2. AMADEUS in low-energy antikaon and kaon–nucleus physics

In hadron and nuclear physics, AMADEUS denotes a dedicated program at DA$\Phi$NE and KLOE devoted to low-energy interactions of negatively charged kaons with nucleons and nuclei, and to related questions in non-perturbative QCD with strangeness (Piscicchia et al., 2013, Scordo et al., 2015, Marton et al., 2016, Curceanu et al., 2015, Marton et al., 2017, Piscicchia et al., 2013, Scordo et al., 2013). The acronym is given as "Antikaon Matter At DA$\Phi$NE: Experiments with Unravelling Spectroscopy" or closely related formulations, and the program exploits the nearly monochromatic low-momentum kaons from $\phi \to K^+K^-$ at DA$\Phi$NE, with kaon momentum about $127\,\mathrm{MeV}/c$ (Piscicchia et al., 2013, Scordo et al., 2015, Curceanu et al., 2015).

The central scientific goals are consistent across the papers. AMADEUS aims to study low-energy $K^-$–nucleon and $K^-$–nucleus interactions, clarify the nature of the $\Lambda(1405)$ in the nuclear medium, measure low-momentum $K^-$ cross sections on light nuclei, characterize single- and multi-nucleon absorption, and search for deeply bound kaonic nuclear states such as $K^-pp$, $\Phi$0, $\Phi$1, and $\Phi$2 (Piscicchia et al., 2013, Marton et al., 2016, Marton et al., 2017). The broader motivation is that the low-energy $\Phi$3 interaction is attractive and may affect the structure of dense baryonic matter, including scenarios relevant to compact stars (Scordo et al., 2015, Marton et al., 2016, Marton et al., 2017).

A defining methodological feature is the use of KLOE as both spectrometer and active target. The KLOE drift chamber gas is $\Phi$4, the entrance wall is carbon fiber plus aluminum, and the detector environment supplies effective targets including H, $\Phi$5, $\Phi$6, and $\Phi$7 (Scordo et al., 2015, Marton et al., 2017). AMADEUS "step 0" consisted of reanalysis of KLOE 2004–2005 data, and a later step added a dedicated pure carbon target in 2012 to obtain a high-statistics sample of at-rest $\Phi$8 interactions (Scordo et al., 2015, Curceanu et al., 2015, Marton et al., 2017). Future configurations described in the literature include dedicated gaseous and solid targets, notably cryogenic $\Phi$9 and $\phi \to K^+K^-$0 (Piscicchia et al., 2013, Marton et al., 2016).

The program has also generated detector R&D tailored to this physics. A GEM-based TPC prototype was developed as an inner tracker for AMADEUS, with requirements including $\phi \to K^+K^-$1, $\phi \to K^+K^-$2, low material budget, and continuous operation; prototype measurements reported up to $\phi \to K^+K^-$3 detection efficiency and spatial resolution around $\phi \to K^+K^-$4 in the drift direction under suitable conditions (Lener et al., 2013, Lener et al., 2013). A scintillating-fiber trigger read out by MPPCs was developed to tag charged kaon pairs, with a measured average efficiency for protons for a double layer of scintillating fibers of $\phi \to K^+K^-$5 (Scordo et al., 2013).

The physics output of the program is unusually granular for low-energy $\phi \to K^+K^-$6 capture. In the $\phi \to K^+K^-$7 channel, global fits to $\phi \to K^+K^-$8, $\phi \to K^+K^-$9, and momentum distributions separated 2NA-QF, 2NA-FSI, 3NA, and 4NA-like contributions, and showed that genuine quasi-free two-nucleon absorption is only a subdominant component in that channel (Scordo et al., 2015). In the same analysis, a hypothetical $\Phi$0 component with best fit around $\Phi$1 and $\Phi$2 improved the fit, but the F-test significance was only about $\Phi$3, insufficient for a discovery claim (Scordo et al., 2015). In $\Phi$4 final states from $\Phi$5 absorption in helium, the collaboration reported about 150 $\Phi$6 events, described as the highest-statistics sample to date for that channel (Marton et al., 2017).

AMADEUS has also pursued the spectroscopy of $\Phi$7 through neutral and charged $\Phi$8 channels. The $\Phi$9 mode is treated as a particularly clean probe because it is pure $127\,\mathrm{MeV}/c$0 and avoids $127\,\mathrm{MeV}/c$1 contamination (Piscicchia et al., 2013). Analyses of the $127\,\mathrm{MeV}/c$2 channel additionally investigated possible $127\,\mathrm{MeV}/c$3 internal conversion in nuclei (Scordo et al., 2013). Collectively, these studies place AMADEUS at the interface of hadron spectroscopy, few-body antikaon dynamics, hyperon final-state interactions, and dense-matter phenomenology (Marton et al., 2016, Marton et al., 2017).

3. AMADEUS as the acoustic test system of ANTARES

In astroparticle physics, AMADEUS stands for "ANTARES Modules for the Acoustic Detection Under the Sea" and denotes an acoustic subsystem integrated into the ANTARES deep-sea neutrino telescope in the Mediterranean Sea (Collaboration et al., 2010, Lahmann, 2011, Graf, 2010). Its purpose is to investigate techniques for acoustic detection of ultra-high-energy neutrinos in the deep sea and to assess the acoustic background relevant to such a detector (Collaboration et al., 2010, Lahmann, 2011).

The system was completed in May 2008 and consists of six acoustic clusters or storeys, each holding six acoustic sensors arranged at distances of roughly 1 m from each other, with inter-cluster spacings from 14.5 m to 340 m and depths between about 2050 m and 2300 m (Collaboration et al., 2010, Lahmann, 2011). The sensors employ piezo-electric elements for broad-band recording of signals up to 125 kHz, with typical sensitivity around $127\,\mathrm{MeV}/c$4 including preamplifier (Collaboration et al., 2010, Lahmann, 2011). The acoustic data are continuously acquired, sent to shore, processed by online filter algorithms, and reduced to a daily recorded data volume of about 10 GB (Collaboration et al., 2010). Reported recorded data volumes were 1.6 TB in 2008 and 3.2 TB in 2009 (Lahmann, 2011).

The physical motivation is the thermo-acoustic model of neutrino detection. A very-high-energy neutrino interaction deposits energy rapidly in water, creating a bipolar pressure pulse whose spectral density peaks around $127\,\mathrm{MeV}/c$5 and whose emission is concentrated in a thin disk perpendicular to the cascade axis (Graf, 2010). The paper describes a representative amplitude of order $127\,\mathrm{MeV}/c$6 per $127\,\mathrm{MeV}/c$7 cascade energy at a vertical distance of $127\,\mathrm{MeV}/c$8 from the cascade (Graf, 2010). Because acoustic attenuation lengths in water are much larger than optical attenuation lengths, the method is motivated as a possible route toward detectors on the order of $127\,\mathrm{MeV}/c$9 (Graf, 2010).

Within ANTARES, AMADEUS also functions as a technological and calibration testbed. The hybrid opto-acoustic environment supports studies of ambient noise, transient signal classification, signal correlations on multiple length scales, source localisation, and acoustic positioning of detector structures (Collaboration et al., 2010, Graf, 2010). The reported heading reconstruction agreement between compass and acoustic methods had an RMS deviation of about $K^-$0, corresponding to a positioning uncertainty of only a few centimeters (Graf, 2010). A plausible implication is that AMADEUS contributed not only to feasibility studies for acoustic neutrino detection, but also to the design logic of later hybrid optical–acoustic neutrino observatories.

4. Amadeus in language modeling and role-playing systems

In contemporary machine learning, the name appears in at least two unrelated systems: Amadeus-Verbo, a family of Brazilian Portuguese LLMs, and AMADEUS, a retrieval-augmented framework for consistent role-playing agents (Cruz-Castañeda et al., 20 May 2025, Park et al., 4 Aug 2025). Both are concerned with language modelling, but they address different problems and use different architectures.

Amadeus-Verbo is a family of models derived from Qwen2.5 and adapted specifically for Brazilian Portuguese (Cruz-Castañeda et al., 20 May 2025). The family includes base-tuned, merged, and instruction-tuned models in sizes of 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B parameters (Cruz-Castañeda et al., 20 May 2025). The reported objective is to show how easy it is to fine-tune foundation models to democratize the open-source development of Brazilian Portuguese LLMs when data and resources are available (Cruz-Castañeda et al., 20 May 2025). Training uses supervised fine-tuning on approximately 600k instruction examples, and evaluation uses a Portuguese adaptation of the LM evaluation harness across tasks including ENEM, OAB exams, semantic textual similarity, natural-language inference, hate speech, offensive language, and sentiment classification (Cruz-Castañeda et al., 20 May 2025). The models are released on Hugging Face under the AmadeusAI organization (Cruz-Castañeda et al., 20 May 2025).

The role-playing framework called AMADEUS is explicitly a RAG-based system for character-consistent role-playing agents (Park et al., 4 Aug 2025). It consists of three modules: Adaptive Context-aware Text Splitter (ACTS), Guided Selection (GS), and Attribute Extractor (AE) (Park et al., 4 Aug 2025). ACTS selects an optimal chunk length and hierarchical contexts for each character; GS filters retrieved chunks according to whether they support attribute inference for the query; and AE extracts general character attributes from the selected chunks to form the final context (Park et al., 4 Aug 2025). To support this line of work, the paper introduces CharacterRAG, with persona documents for 15 fictional characters totaling 976K written characters and 450 question–answer pairs (Park et al., 4 Aug 2025).

The reported evaluation emphasizes both factual grounding and persona consistency. On CharacterRAG, AMADEUS reaches $K^-$1 accuracy with GPT-4.1 and yields a lower hallucination score than the compared RAG baselines; on personality-style probing, it reports $K^-$2 MBTI accuracy and $K^-$3 SLOAN accuracy (Park et al., 4 Aug 2025). This indicates that the framework is intended not merely to retrieve facts about a character, but to stabilize beliefs, values, and psychological traits across out-of-knowledge queries. A plausible implication is that AMADEUS treats persona as a structured latent constraint rather than as a passive document store.

5. Amadeus in symbolic-music generation

In symbolic music generation, the name identifies two different systems from different periods: a 2019 algorithmic composition system based on LSTMs and reinforcement learning, and a 2025 framework combining autoregressive note modelling with bidirectional attribute diffusion (Daniel et al., 2019, Su et al., 28 Aug 2025).

The 2019 system, called Amadeus, is an algorithmic music composition system for polyphonic piano music (Daniel et al., 2019). It combines a multi-stream note representation, a deep LSTM recurrent neural network trained on classical piano music, and an RL agent that searches over a plan space of high-level controls (Daniel et al., 2019). The multi-stream representation divides polyphonic music into a small number of monophonic streams, thereby reducing combinatorial complexity while preserving the ability to model melodies, chords, and occasional contrapuntal sequences (Daniel et al., 2019). The RL agent tunes plan inputs and temperature parameters so that the resulting compositions satisfy musical criteria including pitch entropy, chord usage, rhythmic behavior, limited repetition, limited rest overuse, and reduced cross-correlation with training songs (Daniel et al., 2019). Reported comparisons show higher chord incidence and pitch entropy, and lower repeated identical note-sets, aggregate rest duration, and cross-correlation peaks when RL is used (Daniel et al., 2019).

The 2025 "Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music" reformulates the problem around a different architectural claim (Su et al., 28 Aug 2025). Its central observation is that note attributes are best treated as a concurrent and unordered set rather than as a sequence with fixed temporal dependency structure (Su et al., 28 Aug 2025). The architecture therefore uses a two-level design: an autoregressive model for note sequences and a bidirectional discrete diffusion model for note attributes (Su et al., 28 Aug 2025). To improve latent representation quality, it adds the Music Latent Space Discriminability Enhancement Strategy (MLSDES), which incorporates contrastive learning constraints, and the Conditional Information Enhancement Module (CIEM), which strengthens note latent vector representation via attention mechanisms (Su et al., 28 Aug 2025). The paper states that the model significantly outperforms state-of-the-art systems across multiple metrics while achieving at least $K^-$4 speed-up, and that it supports training-free, fine-grained note attribute control (Su et al., 28 Aug 2025).

The 2025 work is also paired with AMD, described as the largest open-source symbolic music dataset to date, intended for both pre-training and fine-tuning (Su et al., 28 Aug 2025). Taken together, the two music-generation systems show that the name Amadeus has been used both for sequence-model-based compositional planning and for a more recent note/attribute factorization strategy in symbolic music.

6. Amadeus in travel recommendation and customer segmentation

In travel recommendation, Amadeus refers to the industrial setting of Amadeus S.A.S. and to the customer search data used in a multi-objective consensus clustering framework for flight search recommendation (Chatterjee et al., 2020). The problem addressed is personalization of flight-search results by segmenting customers according to search behavior and then optimizing recommendation weights separately for each segment (Chatterjee et al., 2020).

The dataset is composed of Amadeus customer search queries and uses nine features: distance between two airports, geography, number of passengers, number of children, advance purchase, stay duration, day of the week of the departure, day of the week of the return, and day of the week of search (Chatterjee et al., 2020). Because the raw data contain millions of customers, the paper uses stratified sampling to construct datasets of 500, 1000, and 1500 customers while preserving important empirical distributions (Chatterjee et al., 2020).

The clustering framework is ensemble-based and multi-objective. It combines multiple clustering results, each from a different algorithmic configuration, into a consensus solution and uses NSGA-II to optimize two objectives: maximize the average Adjusted Rand Index between the consensus and the base clusterings, and minimize the standard deviation of those similarities (Chatterjee et al., 2020). A weighted co-association matrix is introduced to incorporate co-occurrence, confidence, and base-solution quality; the method also automatically determines an appropriate number of clusters without requiring user input (Chatterjee et al., 2020). This matters operationally because customer segmentation is not assumed to have a fixed, business-defined granularity.

The external validation is an Amadeus business metric based on relative improvement in booking probability. The framework is integrated into the Amadeus flight recommendation optimizer, and segment-specific weighting of recommendation price and convenience criteria is used to increase the probability that returned itineraries are booked (Chatterjee et al., 2020). The paper reports that the prior business-knowledge segmentation yielded a $K^-$5 improvement, while the proposed model achieved $K^-$6 without relying on predefined segment assumptions (Chatterjee et al., 2020). In this usage, "Amadeus" therefore denotes both the enterprise environment and the applied recommendation context in which consensus clustering becomes a production-facing personalization mechanism.