AdvChar: Diverse Implementations Across Domains

• AdvChar is an overloaded label with distinct meanings across narrative modeling, semantic parsing, adversarial NLP, and formal mathematical character theories.
• In narrative modeling, AdvChar systems use a two-stage LLM-based generation and validation process to capture rich, evolving character attributes with high precision.
• In semantic parsing and adversarial NLP, AdvChar leverages character-level feature integration to boost model accuracy and ensure minimal, interpretation-preserving perturbations.

AdvChar is an overloaded technical label rather than a single standardized concept. In the provided arXiv literature, it denotes at least four distinct families of objects: a character-centric narrative representation workflow built with CHIRON; character-level augmentations to semantic parsers; a black-box, interpretation-preserving adversarial attack on interpretable NLP systems; and several mathematically specialized character theories, including admissible characters in rational conformal field theory, D-characters in operator algebras, characters associated with finite-order Cartan automorphisms of extended affine Lie algebras, and cohomological algebraic characters for Harish-Chandra modules. The common lexical element is the word “character,” but the underlying objects range from narrative entities and orthographic sequences to modular functions, homomorphisms, and Euler characteristics (Gurung et al., 2024).

1. Terminological scope

The label appears in multiple, non-equivalent senses. In long-form narrative modeling, the CHIRON paper frames practical guidance as “Building an AdvChar system with CHIRON,” where the target is rich, source-grounded representations of fictional characters over story time (Gurung et al., 2024). In semantic parsing, “AdvChar” is the paper’s name for two character-level integration strategies in sequence-to-sequence DRS parsing: a single-encoder variant that concatenates char-CNN features to token embeddings, and a two-encoder variant with a separate character encoder and dual attention (Noord et al., 2020). In adversarial NLP, AdvChar is a black-box attack on interpretable NLP pipelines that flips the classifier’s prediction while preserving the explanation and keeping edits minimal (Abdukhamidov et al., 22 Jul 2025).

The mathematical uses are again different. One source explicitly presents “admissible characters (‘AdvChar’) and quasi-characters in rational CFT,” specialized to the three-character case (Mukhi et al., 2020). Another presents “Advanced character theory via D-characters in operator algebras (AdvChar),” where scalar characters are replaced by completely contractive homomorphisms into a C$^*$-subalgebra (Blecher et al., 2019). Related sources study characters attached to finite-order Cartan automorphisms of extended affine Lie algebras and cohomological algebraic characters for Harish-Chandra modules (Azam, 2024, Januszewski, 2012).

A central misconception is therefore terminological: AdvChar is not a field-wide canonical acronym. It is a reused label whose meaning must be recovered from the surrounding domain, object class, and formalism.

2. AdvChar in character-centric narrative modeling

Within the CHIRON framework, the practical “AdvChar system” is a structured character-sheet representation intended for long-form narratives. Its goal is to capture “rich, multi-faceted, and evolving aspects of characters” without reducing them to graph nodes, atomic attributes, or a single brief summary. The representation is defined as

$$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$

where each category contains atomic, source-grounded, single-sentence facts about a character, and validated facts are retained only when the entailment model assigns label $5$:

$$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$

Aggregation across story time is

$R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$

with deduplication by TF-IDF cosine similarity $> 0.9$ (Gurung et al., 2024).

Construction proceeds in two stages. The Generation Module prompts an LLM with category-specific open-ended questions for each snippet–character pair, enforces “short, simple sentences with no dependent clauses or transition words,” and then simplifies outputs into atomic statements with pronouns resolved. The Validation Module combines prompted ambiguity and informativeness checks, chain-of-thought comparison to relevant spans, and a domain-specific entailment model: Mistral 7B Instruct v0.2 fine-tuned with LoRA $(r=16, \alpha=32, \text{dropout}=0.1)$ to predict labels in $\{1,\dots,5\}$. On the annotated entailment task, “Trained All (ICL + CoT shown)” achieves precision $0.930$, recall $0.746$, and accuracy $$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$0 at the strict threshold $$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$1, while “Zero-shot None” attains precision $$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$2, recall $$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$3, and accuracy $$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$4. The strict threshold is used because “high precision is prioritized to ensure that kept statements are correct.”

The downstream evaluation is masked-character prediction. For unseen STORIUM scenes with the three most frequent characters masked as $$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$5, the model predicts the character-to-mask assignment

$$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$6

and, in the category-wise aggregation setting,

$$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$7

On 147 data points from 49 finished stories, the best result is “CHIRON Agreed-Subset (OnlyRole on, Knowledge+Physical): 58.5%,” compared with the best summary baseline at $$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$8. The paper also defines a density metric for character-centricity,

$$R = \{F_c : c \in C\}, \qquad C = \{\text{Dialogue, Physical/Personality, Knowledge, Goals}\},$$9

and reports Pearson $5$0 $5$1 between automated density and summed human character-centricity judgments. The same study notes that $5$2 of generated statements are unverifiable or false before validation, making validation structurally central rather than auxiliary.

3. AdvChar as character-level augmentation in semantic parsing

In DRS-based semantic parsing, AdvChar denotes explicit character-level representations added to a seq2seq parser in the presence of strong contextual LLMs. The base model is a bi-LSTM encoder with attention and an LSTM decoder. The single-encoder AdvChar computes character features per token using a char-CNN and concatenates them to the token representation:

$5$3

The two-encoder AdvChar instead runs separate bi-LSTM encoders over tokens and characters, initializes the decoder with both context vectors,

$5$4

and uses dual attention,

$5$5

where $5$6 is dot-product attention (Noord et al., 2020).

The empirical claim of the paper is that character-level information still matters “even in the age of BERT.” On PMB 3.0.0 English dev, BERT-only reaches $5$7 F1, while BERT + char reaches $5$8 in both the single-encoder and two-encoder setups, a $5$9 absolute gain that the paper marks as significant. On PMB 3.0.0 English test, BERT-only reaches $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$0, BERT+char $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$1 in the single-encoder setting, $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$2 in the two-encoder setting, and the best model, characters plus semantic tags, reaches $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$3. Similar improvements are reported across ELMo, BERT-large, RoBERTa-base, and RoBERTa-large, as well as across German, Italian, and Dutch.

The analysis is phenomenon-specific. Using semantic tags to select subsets, the paper reports that characters help especially with numerals, named entities, pronouns, attributes, logical operators, and some forms of modality. On combined PMB 3.0.0 English dev+test, the gains for numerals are $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$4, $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$5, and $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$6 depending on the character configuration. The paper also reports reduced ill-formed outputs and more perfect DRS matches. A distinctive implementation result is that freezing LM weights is essential: fine-tuning degrades performance substantially, with F1 around $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$7–$$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$8. In this usage, AdvChar is therefore not an adversarial method and not a narrative character representation; it is an input-side character-sequence enhancement to semantic parsers.

4. AdvChar as an interpretation-preserving adversarial attack

In a separate NLP literature, AdvChar is a black-box adversarial attack on Interpretable Natural Language Processing Systems. The attacked pipeline consists of a classifier $$keep(f) = 1 \text{ if } \hat y(f)=5, \text{ else } 0.$$9 and an interpreter $R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$0, where $R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$1 yields a token-level attribution vector $R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$2. The attack seeks a perturbed text $R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$3 such that the prediction flips,

$R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$4

the interpretation remains similar,

$R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$5

and the edit cost stays minimal, with

$R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$6

Token importance is normalized to $R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$7, high-importance tokens are traversed greedily, and candidate edits are single-character substitutions chosen to be visually subtle or typographically similar (Abdukhamidov et al., 22 Jul 2025).

The threat model is explicitly black-box. The attacker can query the classifier for labels and the interpreter for attribution vectors, but does not need gradients, model parameters, or architecture access. The reported interpreters are SHAP, Saliency Maps, and LIME. The evaluation covers seven classifiers—GPT-2, BERT, DistilBERT, Electra, CANINE, FNet, and XLM-R—on SST-2, AG News, and Yahoo Answers. Metrics include Attack Success Rate, Query Count, Misclassification Confidence, interpretation similarity via IoU, and the average number of adversarial character changes.

The paper’s headline empirical claim is that AdvChar can “significantly reduce the prediction accuracy of current deep learning models by altering just two characters on average in input samples.” Representative results include SST-2 with SM+BERT at ASR approximately $R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$8; AG News with CANINE+LIME at ASR approximately $R_t = \bigcup_{\tau \le t} \{\hat f_i(\tau)\},$9, QC approximately $> 0.9$0, and approximately $> 0.9$1 characters changed; and Yahoo Answers with LIME+BERT at ASR approximately $> 0.9$2, QC approximately $> 0.9$3, and $> 0.9$4 characters changed. IoU values are often high under LIME, with SST-2 commonly around $> 0.9$5–$> 0.9$6, AG News peaking around $> 0.9$7, and Yahoo Answers often at or above $> 0.9$8. Compared with TextBugger, the paper reports fewer queries, far fewer character changes, higher ASR, higher MC, and substantially higher interpretation IoU. The attack therefore targets trust in transparency: the classifier becomes wrong while the explanation remains deceptively stable.

5. Mathematical AdvChar: admissible characters, D-characters, and algebraic character formalisms

One mathematical use is explicit: the RCFT source presents “admissible characters (‘AdvChar’) and quasi-characters in rational CFT,” specialized to the three-character case. There, vector-valued characters have $> 0.9$9-expansions

$(r=16, \alpha=32, \text{dropout}=0.1)$0

quasi-characters have integral but not necessarily positive coefficients, and admissible characters require non-negative integer coefficients after normalization. The three-character case is organized through MLDEs, the Wronskian index $(r=16, \alpha=32, \text{dropout}=0.1)$1, and the sum rule

$(r=16, \alpha=32, \text{dropout}=0.1)$2

The paper conjectures several infinite families, shows that linear combinations can generate admissible characters with arbitrarily large Wronskian index, and emphasizes that the structure is “completely different from the order two case” (Mukhi et al., 2020).

A second explicit mathematical usage is “Advanced character theory via D-characters in operator algebras (AdvChar).” Here a D-character is a contractive unital homomorphism $(r=16, \alpha=32, \text{dropout}=0.1)$3 that restricts to the identity on the C$(r=16, \alpha=32, \text{dropout}=0.1)$4-subalgebra $(r=16, \alpha=32, \text{dropout}=0.1)$5. Such maps are automatically completely contractive. The paper develops Jensen-type inequalities, a Bishop–Ito–Schreiber type characterization, and Gleason parts for D-characters. The algebraic core includes the ball-Jensen condition $(r=16, \alpha=32, \text{dropout}=0.1)$6 under $(r=16, \alpha=32, \text{dropout}=0.1)$7 and $(r=16, \alpha=32, \text{dropout}=0.1)$8, the theorem that this condition is equivalent to $(r=16, \alpha=32, \text{dropout}=0.1)$9 being a $\{1,\dots,5\}$0-preserving homomorphism under the stated hypotheses, and the equivalence of several Gleason-part criteria such as $\{1,\dots,5\}$1 and $\{1,\dots,5\}$2 (Blecher et al., 2019).

The EALA source studies characters in a different sense. For almost all extended affine Lie algebras, any finite-order Cartan automorphism is diagonal, and the associated combinatorial map

$\{1,\dots,5\}$3

is a character, multiplicative on summable roots. The synthesis theorem states that if $\{1,\dots,5\}$4 and $\{1,\dots,5\}$5, then finite-order Cartan automorphisms modulo the stated conjugacy relation are in bijection with finite-order characters $\{1,\dots,5\}$6; in index-zero cases these extend to root-lattice characters on $\{1,\dots,5\}$7 (Azam, 2024).

The Harish-Chandra source does not brand its formalism as AdvChar, but it is a closely related advanced character theory in the provided corpus. It defines the algebraic character of a $\{1,\dots,5\}$8-module $\{1,\dots,5\}$9 by

$0.930$0

establishes additivity, multiplicativity, duality, transitivity, and restriction functoriality, and uses the resulting localized Grothendieck-group formalism to attack branching and discretely decomposable restriction problems (Januszewski, 2012). In these mathematical settings, “character” no longer refers to textual units or narrative agents; it refers to modular objects, homomorphisms, or cohomological invariants.

6. Comparative themes and persistent confusions

The most persistent confusion is categorical rather than technical. In the NLP papers, “character” may refer either to narrative entities tracked over time or to orthographic character sequences used as subtoken signals. In the adversarial paper, the same label names an attack whose perturbation primitive is also character-level, but whose object of interest is the stability of explanations under misclassification. In the mathematical papers, “character” denotes neither story agents nor orthographic units, but modular functions, multiplicative maps, or Euler-characteristic constructions. The same lexical surface therefore spans radically different ontologies (Gurung et al., 2024, Noord et al., 2020, Abdukhamidov et al., 22 Jul 2025).

A second recurring theme is structural control. CHIRON enforces atomicity, source-grounding, unambiguity, and strict entailment filtering. Semantic-parsing AdvChar uses explicit architectural factorization, either by concatenating char-CNN features or by separating token and character encoders with dual attention. The adversarial AdvChar enforces an interpretation-similarity bound and a minimal edit budget. The mathematical formalisms likewise impose strong structure: RCFT admissible characters satisfy modular covariance and positivity constraints; D-characters are unital completely contractive homomorphisms fixing $0.930$1; EALA characters arise from diagonal finite-order Cartan automorphisms; Harish-Chandra algebraic characters are localized Euler characteristics (Mukhi et al., 2020, Blecher et al., 2019, Azam, 2024, Januszewski, 2012).

A plausible implication is that “AdvChar” functions less as a stable term of art than as a recurring naming pattern for methods that refine coarse character notions into formally constrained objects. In narrative modeling this means validated, category-specific sheets; in semantic parsing it means orthographic features beyond token-level contextual embeddings; in adversarial NLP it means minimal perturbations that preserve explanation geometry; and in mathematics it means character theories with explicit modular, algebraic, or cohomological control. Any technically precise use of the term therefore requires immediate disambiguation by domain, formal object, and evaluation criterion.