Karl: Astronomy, Quantum, and AI Innovations

• Karl is a multifaceted term referring to significant figures like Rakos and Popper as well as advanced systems in named entity recognition and student learning.
• In astronomy and quantum mechanics, Karl denotes pioneering contributions—from Rakos’s digital innovations in Vienna to Popper’s realist challenges to Copenhagen orthodoxy.
• In AI, KARL embodies knowledge-aware augmentation for NER and content-aware flashcard scheduling, yielding measurable improvements in model performance and student retention.

to=arxiv_search.search 北京赛车群 ＿老司机_json {"query":"id:(Odell, 2012) OR id:(Santo, 2018) OR id:(Chawla et al., 2021) OR id:(Santo, 2017) OR id:(Shu et al., 2024)","max_results":10,"sort_by":"relevance"} to=arxiv_search.search ’wini 久赢_json {"query":"(Odell, 2012)","max_results":5,"sort_by":"relevance"} to=arxiv_search.search 天天爱彩票是json {"query":"Karl Rakos Obituary", "max_results": 5, "sort_by": "relevance"} to=arxiv_search.search 乐盈_json {"query":"Karl Popper quantum debate (Santo, 2018)", "max_results": 10, "sort_by": "relevance"} to=arxiv_search.search 天天中彩票不能_json {"query":"KARL-Trans-NER Knowledge Aware Representation Learning for Named Entity Recognition using Transformers", "max_results": 5, "sort_by": "relevance"} In the arXiv literature represented here, “Karl” designates both named individuals and acronymic research systems across astronomy, foundations of quantum mechanics, named entity recognition, and educational technology. The principal referents are Karl Dragutin Rakos, a Croatian-Austrian astronomer and co-founder of modern astrophysical research at the University of Vienna (Odell, 2012); Karl Popper, treated as a central yet neglected participant in postwar and 1980s debates on the foundations of quantum mechanics (Santo, 2018, Santo, 2017); KARL-Trans-NER, where KARL denotes Knowledge Aware Representation Learning for transformer-based named entity recognition (Chawla et al., 2021); and KARL, where the acronym expands to Knowledge-Aware Retrieval and Representations aid Retention and Learning in Students, a content-aware flashcard scheduling framework (Shu et al., 2024).

1. Principal referents of “Karl”

The term appears in two distinct modes: as a personal name and as an acronym.

Referent	Domain	Core characterization
Karl Dragutin Rakos	Astronomy	Research astronomer, teacher, and institutional modernizer in Vienna
Karl Popper	Philosophy of physics / FQM	Realist critic of Copenhagen orthodoxy and participant in quantum-foundations debates
KARL-Trans-NER	NLP	Knowledge-aware augmentation mechanism for NER
KARL	Educational technology	Content-aware student model and flashcard scheduler

As a personal name, “Karl” denotes figures whose work is embedded in institutional and conceptual change. Rakos is described as a “valued researcher, university teacher and co-founder of modern astrophysical research at the Institut für Astronomie of the University of Vienna” (Odell, 2012). Popper is reconstructed not merely as a philosopher with views on quantum theory, but as a “quantum dissident” operating between philosophical and physical communities and later as someone who “effectively became part of the physics community” in the 1980s (Santo, 2018, Santo, 2017).

As an acronym, “KARL” identifies two unrelated technical systems. In NER, it names a Knowledge Aware Representation Learning network that augments contextual encoders with knowledge-graph-derived global representations (Chawla et al., 2021). In flashcard scheduling, it names a Knowledge-Aware Retrieval and Representations aid Retention and Learning in Students model combining deep knowledge tracing, retrieval, and BERT to predict recall and drive scheduling (Shu et al., 2024).

A plausible implication is that “Karl” functions here as a cross-domain label for both intellectual agents and systems concerned with modernization, realism, or explicit knowledge integration, but the underlying subjects remain historically and technically distinct.

2. Karl Dragutin Rakos in astronomy

Karl Dragutin Rakos was born on November 1, 1925, in Stefanje, Croatia, and died on October 31, 2011, one day before his 86th birthday (Odell, 2012). He is presented as a Croatian-Austrian astronomer, university teacher, and major figure in the modernization of astronomy in Vienna. The obituary states that with his death, “the Vienna astronomical community lost a valued researcher, university teacher and co-founder of modern astrophysical research at the Institut für Astronomie of the University of Vienna” (Odell, 2012).

His early life is described as marked by an early interest in science and technology. He became aware of the stars on Christmas Eve when he was nine, read books in physics and chemistry, experimented with batteries in the kitchen, and by fourth grade was building radios (Odell, 2012). He became captivated by astronomy on New Year’s Day of 1938 after finding a book on astronomy in an attic box of books. He studied science at the Universities of Zagreb and Belgrade, but wartime and postwar political conditions in Yugoslavia disrupted his education; after an adventurous flight in 1950, he completed his studies in Graz (Odell, 2012).

In 1956, he graduated with work titled “Development and Construction of a Photoelectric Star Photometer”. He then served as an assistant at the University Observatory in Graz and subsequently as an assistant at the University of Vienna Observatory (Odell, 2012). Beginning in 1960, with a Fulbright Scholarship and later an NSF grant, he spent four years at Lowell Observatory in Flagstaff, Arizona, where he met Dr. Alois Purgathofer. During this period, while Harold Johnson was developing the UBV photometric system, Rakos used it to observe magnetic variable stars (Odell, 2012).

His technical and institutional contributions in Vienna were extensive. He established highly advanced methods of digital electronics, helped develop a photon counting photometer, introduced digital detectors—first a Reticon, later CCDs—and was instrumental in acquiring PDP12, PDP11, and later a DEC-VAX, thereby introducing computerization to the Vienna Institute of Astronomy (Odell, 2012). He also developed the Area Scanner for precision astrometric and photometric measurements of double stars; this work led, among other things, to the first usable photometric data on the white dwarf companion of Sirius in 1974 (Odell, 2012).

His research program evolved from stellar photometry and chemically peculiar stars to extragalactic astronomy. Earlier work focused on magnetic variable stars, Ap stars, double and multiple stars, Sirius B, close visual binaries, UV resonance lines in hot Ap stars, and photoelectric and narrow-band photometry (Odell, 2012). Later, his interests shifted “from questions about stars to the evolution of galaxies.” In that phase he used Strömgren filter wavelengths to break the degeneracy of the UBV system for determining the age and metallicity of cluster galaxies, and he used filters shifted in wavelength to match cluster redshift, reducing large k corrections (Odell, 2012).

Rakos also exercised leadership and sustained educational influence. He served as Head of the Institute of Astronomy from 1979 to 1981, contributed to the 1976 report “A plan for astronomical research in Austria,” and supervised “countless students and dissertations” (Odell, 2012). After retiring in 1989, he reconnected with Croatia, became a regular lecturer in Zagreb, founded a chair of astronomy at the university, and arranged for a 36-inch telescope from the Vienna institute to be transferred to the Island of Hvar (Odell, 2012). His memberships included the International Astronomical Union, the American Astronomical Society, the European Astronomical Society, the Astronomische Gesellschaft, the European Science Foundation, the Croatian Astronomical Society, and external membership in the Croatian Academy of Sciences and Arts (Odell, 2012).

3. Karl Popper and the foundations of quantum mechanics in the 1950s and 1960s

Karl Popper is presented as a major but long underrecognized participant in the postwar debate on the foundations of quantum mechanics. The central historiographical claim is that he became a genuine “quantum dissident”: a philosopher engaged in a sustained effort to challenge the orthodox, Copenhagen-style interpretation and to keep realist alternatives intellectually active (Santo, 2018). What made his role “central yet neglected” was his position at the boundary between the philosophical and physics Denkkollektiv, together with the fact that much of his influence circulated through correspondence, conference proceedings, and unpublished materials rather than mainstream physics journals (Santo, 2018).

At the level of argument, Popper’s best-known intervention is the critique of “the observer” in quantum mechanics. In Quantum Mechanics without the Observer (1967), he wrote that he wanted to “exorcise the ghost called ‘consciousness’ or ‘the observer’ from quantum mechanics” (Santo, 2018). He rejected instrumentalism and argued that theories should aim at objective truth or nearness to truth rather than merely functioning as predictive instruments. Closely connected to this was his criticism of subjective or epistemic readings of probability. Against them he developed the propensity interpretation, according to which probabilities are “physically real” and should be understood as “physical propensities,” permitting objective probabilities for single events rather than only ensembles (Santo, 2018).

He also challenged common readings of the uncertainty principle. He did not deny the formal validity of Heisenberg’s relations, but he rejected the standard epistemological interpretation that casts them as limits on what can be known about individual systems. Instead, he proposed a statistical or ensemble-based reading: the uncertainty relation constrains probability distributions, not the actual value of a measured quantity (Santo, 2018). In the same vein, he argued that wave-packet reduction is, at least partly, a trivial feature of probability theory: when one conditions on additional information, the relevant probability distribution changes. The paper summarizes this as Popper’s attempt to dissolve part of the “great quantum muddle” produced when statistical measures are treated as physical properties of individual systems (Santo, 2018).

The archival reconstruction stresses both influence and limitation. Popper’s long relationships with Alfred Landé, David Bohm, Hermann Bondi, and Henry Margenau demonstrate that his ideas did gain traction among physicists skeptical of Copenhagen orthodoxy (Santo, 2018). Landé eventually described himself as Popper’s “most staunch adherent among the physicists,” and Bohm regarded Popper’s remarks on propensities as “a genuine contribution to clarifying the issues” (Santo, 2018). At the same time, the paper is explicit that Popper’s propensity interpretation did not solve the distinctively quantum problem of interference of amplitudes. Popper himself eventually admitted that the peculiarity of quantum theory lies in the “principle of superposition of wave amplitudes,” which has no classical parallel (Santo, 2018). Thus his framework is presented as a powerful realist resistance to observer-centered orthodoxy, but not as a complete solution to the measurement problem.

4. Popper’s EPR-like experiment and the 1980s physics community

The later reconstruction of Popper’s role in the 1980s sharpens the claim that he was not merely commenting from outside physics. The paper argues that he “effectively became part of the physics community” through direct engagement with foundations researchers, especially Jean-Pierre Vigier and Franco Selleri (Santo, 2017). His publications, conference participation, correspondence, and the circulation of drafts and criticisms placed him within what the paper calls a “quantum subculture” dissatisfied with Copenhagen orthodoxy (Santo, 2017).

The central episode is the genesis of Popper’s EPR-like thought experiment. An unpublished letter dated 09/06/1980 to Vigier shows that Popper had formulated the experiment two years before its first publication in 1982 (Santo, 2017). In that letter he proposed an arrangement with a source $S$, detectors and coincidence counters, and screens with variable parallel slits used to “measure” the $x$-position of particles $A$ and $B$. The operational core was explicit: vary slit widths, obtain coincidence statistics, and compare scatter distributions before and after widening the slit at $B$ while localizing $A$ (Santo, 2017). The proposal was intended as a crucial test between the Copenhagen interpretation and a realist/statistical interpretation.

Historically, the paper traces a sequence from Popper’s mistaken 1934 Gedankenexperiment against Heisenberg’s uncertainty relations, through renewed contact with Vigier, to the 1980 letter, the 1981 co-authored papers by Garuccio, Popper and Vigier, the 1982 publications, the 1983 Bari workshop, debate across 1984–1987, and eventual implementation in 1999 (Santo, 2017). The unpublished letter is presented as decisive evidence of priority and of the experiment’s emergence from dialogue with Vigier after a meeting that Popper described as “two happier afternoons” (Santo, 2017).

Reception was mixed and technically focused. The proposal was debated by referees and by figures including Shimony, Bell, Aspect, Mandel, Sudbery, Krips, and Collett and Loudon (Santo, 2017). The disputed questions were whether the experiment was feasible and, if feasible, what it actually tested. Some critics argued that the required source conditions made realization impractical or impossible; others argued that the setup did not uniquely test the Copenhagen interpretation (Santo, 2017). Popper nevertheless continued to defend it as a legitimate crucial test.

The 1983 Bari conference marked a turning point in Popper’s public presence within the field. There he presented “Realism in Quantum Mechanics and a New Version of the EPR Experiment,” and the discussion involved Marcello Cini, Francesco De Martini, Karl Kraus, Trevor W. Marshall, Helmut Rauch, C. Robinson, Franco Selleri, J. Six, Gino Tarozzi, and Jean-Pierre Vigier (Santo, 2017). The broader significance lies in the paper’s portrayal of Popper as a controversial but recognized participant whose views were sought even after Aspect’s 1981 Bell-test experiments, especially on the distinction between experiments concerning nonlocal action and those concerning the Copenhagen interpretation (Santo, 2017).

5. KARL-Trans-NER: knowledge-aware augmentation for sequence labeling

In NLP, KARL stands for Knowledge Aware Representation Learning, and the full model is named KARL-Trans-NER (Chawla et al., 2021). It is a named entity recognition system that augments standard neural sequence labeling with knowledge graph-derived global representations. The motivation is that encoders such as BERT, ELMo, and Flair model surface contextual information effectively, but do not explicitly inject structured world knowledge from external knowledge bases (Chawla et al., 2021). The paper argues that this matters because named entities and relations are polysemous in knowledge graphs, prior knowledge-graph embeddings are often static and non-contextualized, some NER-oriented methods drop the tail entity of a triplet, recurrent approaches for KG encoding are computationally expensive, and sentence-level augmentation is too coarse for entity-level sequence labeling (Chawla et al., 2021).

World knowledge is represented as fact triplets

$KG = \{<s,r,o> \mid (s,o) \in E,\ r \in R\},$

with graph contexts

$$S = \{(s \rightarrow r \rightarrow o) \mid (s,o) \in E,\ r \in R\}.$$

To train the knowledge-graph embedding module, the object entity is masked: $T = \{(<s,r,[MASK]>, o)\}.$ This yields a masked entity prediction task inspired by masked language modeling and CoKE (Chawla et al., 2021).

The KGE encoder is Transformer-based and differs from CoKE by adding character-level encoding for entities and relations, word-level embeddings, and relative positional embeddings rather than sinusoidal positions (Chawla et al., 2021). For an input triplet sequence $T=(t_1,t_2,t_3)$, the token representation is formed as

$x$0

followed by $x$1 Transformer layers

$x$2

The masked object prediction uses the final hidden state $x$3: $x$4 with cross-entropy loss

$x$5

These details are important because the paper treats the KG not as a static embedding lookup but as a contextualized triplet-encoding problem (Chawla et al., 2021).

For NER, the model does not consume all triplets. It performs entity shortlisting via n-gram matching and ranks ambiguous matches by subject frequency, retaining the top $x$6 entities for each word $x$7 in a candidate set $x$8 (Chawla et al., 2021). It then performs relation shortlisting, selecting the top $x$9 most frequent relations to construct the graph context $A$0. Each shortlisted entity-relation pair is appended with [MASK], encoded by the pretrained KGE Transformer, and converted into contextualized triplet embeddings. If the final-layer hidden states are $A$1, they are concatenated into

$A$2

Using the BERT contextualized embedding $A$3 of the target word as query, attention is computed with

$A$4

and the global knowledge-aware representation is

$A$5

The resulting token representation combines six feature types: word-level GloVe embeddings, character-level IntNet CNN encoding, context-level BERT embeddings, sentence-level label-attention sentence embedding, document-level key-value memory network features, and global-level KARL knowledge-aware embeddings $A$6 (Chawla et al., 2021).

Experimentally, the model is evaluated on CoNLL 2003, CoNLL++, and OntoNotes v5, using Wikidata filtered to around 10 million relevant triplets from roughly 400 million total, and reporting average span-level F1 and standard deviation over three runs (Chawla et al., 2021). On CoNLL 2003, performance rises from 90.12 for the word-level baseline to 93.74 with + Global/KARL; on CoNLL++, the word-level baseline is 90.90 and + Global/KARL reaches 94.52; on OntoNotes v5, the word-level baseline is 86.97 and + Global/KARL reaches 91.41 (Chawla et al., 2021). The paper also states that KARL outperforms KAWR by about 0.65 F1 on CoNLL 2003 under the same context+knowledge configuration, and that the augmentation yields gains of about 0.35–0.5 F1 across the datasets (Chawla et al., 2021).

A common misconception the paper explicitly rejects is that KARL replaces BERT. It does not: KARL is not a replacement for BERT; it is a feature augmentation mechanism on top of standard contextual and non-contextual features (Chawla et al., 2021). The stated limitations are that the KGE module is trained separately from the NER model and that entity shortlisting does not use semantic similarity (Chawla et al., 2021).

6. KARL for student modeling and flashcard scheduling

In educational technology, KARL expands to Knowledge-Aware Retrieval and Representations aid Retention and Learning in Students (Shu et al., 2024). The system is introduced in the context of flashcard scheduling, where a scheduler relies on a student model to predict recall and a teaching policy to decide which cards to show next. The paper’s central claim is that prior student models are largely content-blind: they rely on study history features such as counts of correct and incorrect responses, time since last review, spacing, and interval variables, but ignore the semantic content of the cards themselves (Shu et al., 2024).

KARL is described as a DKT-inspired student model combining retrieval over the student’s history, BERT embeddings of flashcard text, flashcard-level behavioral features, and a classifier that predicts whether the student will answer the current flashcard correctly (Shu et al., 2024). Given a current flashcard $A$7, prior study history $A$8, and hand-crafted features $A$9, the model predicts $B$0 (Shu et al., 2024). Instead of embedding the full history, it retrieves the top-$B$1 most semantically similar prior flashcards using BERT encodings

$B$2

with similarity

$B$3

and Maximum Inner-Product Search (MIPS) implemented with FAISS (Shu et al., 2024). This is presented as improving efficiency and robustness to topic shifts.

The representation stage combines BERT embeddings and flashcard-level features. The appendix lists 20 features, including number of cards studied by the student, number of correct and incorrect responses, time since last review, session-level counts, previous delta and spacing features, and Leitner and SM-2 variables (Shu et al., 2024). The classifier architecture is specified as BERT model, Dropout, Linear layer, GELU, LayerNorm, Dropout, Linear layer, with hidden state $B$4 (Shu et al., 2024). The paper gives the loss as

$B$5

and clarifies in the appendix that training uses binary cross entropy with mini-batch gradient descent; the optimizer is Adam, the learning rate is $B$6, batch size is 64, epochs are 10, retrieval size is $B$7 in one experiment and $B$8 elsewhere, and data are split 75\% train / 25\% test by chronological order (Shu et al., 2024).

A major contribution is the dataset. The authors construct flashcards from QANTA trivia questions spanning 11 topics and obtain 23,918 unique flashcards (Shu et al., 2024). They built a web/mobile flashcard app and collected user interactions over four months, producing 123,143 study records / logs from 543 users (Shu et al., 2024). Additional statistics include 44.02% of studies on new flashcards, average user study of 226.78 cards, and 71.80% correctness on shown flashcards (Shu et al., 2024).

Offline evaluation emphasizes AUC, ECE, and accuracy conditioned on correct and incorrect outcomes (Shu et al., 2024). On seen cards, KAR achieves AUC: 0.7922, ECE: 0.1006, Acc Correct: 0.9213, and Acc Incorrect: 0.5419 (Shu et al., 2024). On unseen cards, KAR reaches AUC: 0.7294, ECE: 0.1206, Acc Correct: 0.6010, and Acc Incorrect: 0.6103, while LM-KT attains AUC 0.6838, ECE 0.2343, Acc Correct 0.6665, and Acc Incorrect 0.5899 (Shu et al., 2024). The ablation study reports that removing BERT produces the largest degradation, with Seen AUC: 0.692, Unseen AUC: 0.612, and Unseen ECE: 0.205, while removing $B$9 also harms performance, yielding Seen AUC: 0.680 and Unseen AUC: 0.620 (Shu et al., 2024).

The paper also introduces a delta-based teaching policy intended to improve on threshold-based policies. Instead of selecting cards near a fixed retention threshold, it estimates the expected future gain from studying a card now: $B$0 where

$B$1

The scheduler then chooses the top-$B$2 cards with highest expected gain (Shu et al., 2024).

Online evaluation is deliberately modest. In a within-subject study with 27 users and 32 completed study trajectories, the comparison is between FSRS and KAR + delta policy over five days of study and a day-6 post-test (Shu et al., 2024). Average accuracy is reported as FSRS: pre-test 0.44, post-test 0.91 and KAR: pre-test 0.45, post-test 0.90 (Shu et al., 2024). The paper interprets this as evidence that content-aware scheduling is practical and promising. A necessary qualification is that the online results are comparable to FSRS rather than decisively superior, while the stronger advantage appears in offline prediction quality and handling of unseen cards (Shu et al., 2024).