CoLL: Multi-domain Frameworks

• CoLL is a multi-domain framework encompassing arithmetic logic, graph-based AI, and collider analyses that unifies principles for collecting, synthesizing, and bounding complex information.
• Its logical instantiation refines witness collection in arithmetic, bridging weak theories and full induction through uniform boundedness of existential proofs.
• Its AI and HEP applications employ LLMs and GNNs to enhance anomaly detection and automate collider analyses, yielding significant improvements in performance and interpretability.

CoLL refers to a family of concepts and systems across several domains, notably including logic (the Collection scheme “Coll”), AI frameworks for graph anomaly detection (“Court of LLMs”), and engineering toolkits for collider analyses. In each context, CoLL or COLL denotes a foundational mechanism for collecting, synthesizing, or augmenting information—be it logical witnesses in arithmetic, textual/structural evidence for anomaly detection, or code/feature pipelines for scientific analysis. The following account rigorously documents major CoLL instantiations in logic (mathematical logic and model theory), machine learning (graph AI), and high energy physics (AI-accelerated detector studies), precisely as described in key arXiv literature.

1. The Collection Scheme “Coll” in Arithmetical Logic

The full Collection Scheme, denoted Coll, is a pivotal family of first-order axiom schemas formulated in the language of arithmetic $\mathcal{L}_{\text{arith}} = \{0,1,+,\times,<\}$. Coll is stated for each formula $\varphi(x,y, \bar z)$ as the schema: $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$ This principle asserts that the existence of witnesses $y$ for all $x<v$ can always be tightened to a uniform global bound $w$, independent of $x$. Coll occupies a position strictly stronger than basic bounding schemes (like $\Delta_0$-Collection) but is strictly weaker than the full induction schema; added to weak base theories, it guarantees induction for bounded-existential formulas (E$_1$-induction), and when combined with I$\Delta_0$ yields full Peano Arithmetic (PA). However, in isolation (e.g., with IOpen), Coll does not secure arbitrary induction (Enayat et al., 20 Aug 2025).

2. Restricted Complexity and Completions in Weak Arithmetical Theories

A first-order theory $\varphi(x,y, \bar z)$0 is of restricted complexity if there exists $\varphi(x,y, \bar z)$1 and a possibly infinite set $\varphi(x,y, \bar z)$2 of $\varphi(x,y, \bar z)$3-sentences such that $\varphi(x,y, \bar z)$4 can be axiomatized by $\varphi(x,y, \bar z)$5, i.e., all axioms are of bounded formula-complexity (quantifier alternation depth). This property is sensitive to the logical strength of $\varphi(x,y, \bar z)$6: any consistent extension of I$\varphi(x,y, \bar z)$7+Exp admits no restricted-complexity completion, while weaker fragments may. The central theorem of Enayat–Łełyk–Visser (Enayat et al., 20 Aug 2025) establishes that:

• There exists a model $\varphi(x,y, \bar z)$8 of IOpen+Coll such that $\varphi(x,y, \bar z)$9 (the complete theory of $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$0) is of restricted complexity, i.e., $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$1, with $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$2 a single first-order sentence.
• The construction uses iterated Shepherdson real-closed fraction field extensions—successively adjoining infinitely many transcendentals indexed by $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$3. This yields a model $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$4 whose elements’ supports (occurring variable indices) are first-order definable, and whose diagram can be coded in $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$5.
• Through an interpretability-retract plus partial-isomorphism (Ehrenfeucht–Fraïssé) framework, it is demonstrated that $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$6 may always be given a single-sentence (modulo $$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$7-completeness) axiomatization.

This result demarcates the precise “incompleteness barrier” at I$$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$8+Exp, and positions IOpen+Coll (and PA$$\forall v \left[ \left( \forall x < v\ \exists y\ \varphi(x,y,\bar{z}) \right) \rightarrow \exists w\ \forall x < v\ \exists y < w\ \varphi(x,y,\bar{z}) \right]\,.$$9+Coll) as the first natural weak arithmetical theories that admit restricted-complexity completions, providing a new taxonomy below the full induction threshold (Enayat et al., 20 Aug 2025).

3. CoLL in Graph Anomaly Detection: Court of LLMs

CoLL (“Court of LLMs”) (Xu et al., 1 Aug 2025) is an architectural framework for unsupervised anomaly detection on text-attributed graphs (TAGs), integrating LLMs and graph neural networks (GNNs) for evidence-augmented detection:

• Stage I (Evidence-Augmented Generation): Two specialized LLM “prosecutors” analyze, respectively, node-local text (contextual prosecutor) and node-plus-neighbors text (structural prosecutor), each emitting explanatory textual evidence. A higher-capacity LLM “judge” synthesizes these prosecutor outputs and the node text into a final human-readable “verdict.”
• Stage II (High-Order Structure Completion): Node-level features are formed by encoding both raw text and judge verdict via a frozen text encoder; an LSTM-style gating mechanism fuses these signals, which are then passed through a GNN. Training uses a node-subgraph contrastive objective (positive if the node embedding matches its neighborhood summary, negative for random subgraphs). No additional regularization or alignment losses are needed.

Quantitatively, CoLL surpasses 11 established baselines on four large TAG benchmarks, yielding a mean Average Precision (AP) increase of +13.37% and AUC gain of +2.39%. Ablation establishes both the utility of dual-prosecutor evidence (+3–5% AUC from a judge) and the advantage of gating over mean/simple fusion (+1–2% AUC). Despite its multi-LLM workload in evidence generation, the full pipeline is $y$09 times faster than prior deep GAD models (Xu et al., 1 Aug 2025). The collaboration of LLM-based evidence and GNN high-order consistency is a distinctive technical advance in explainable, scalable anomaly detection.

4. CoLLM: LLM+AutoML Toolbox for Collider Analyses

CoLLM (“CoLL”) (Esmail et al., 6 Feb 2026) is an applied AI engineering platform designed to automate end-to-end high-energy physics (HEP) collider analyses:

• Domain-driven Code Synthesis: Users specify event selection criteria, histogram outputs, and result structure in plain language; an LLM system prompt embeds HEP conventions, including LHCO format, four-momentum formulae, and Pythonic parsing modules.
• Automated Deep Learning: The system compiles the user spec into validated Python code for event selection and feature extraction. A separate AutoML module (MLP, GNN, Transformer) accepts these features and orchestrates full deep-learning pipelines, supporting metric logging and model checkpointing.
• User Interfaces: Both command-line and graphical (Streamlit) interfaces allow interactive inspection, prompt editing, training run configuration, and monitoring of results (loss, AUC, confusion matrices).
• Code Robustness: The PyFixer module iteratively debugs generated scripts using LLM-assigned patches derived from Python tracebacks, achieving high reproducibility (9/10 runs deterministic output, with prompt ambiguity as the main source of residual variation).
• Performance: For typical LHC-style analyses (e.g., $y$1, $y$2), code generation and validation is completed rapidly. Deep models typically achieve AUC scores in the $y$3–$y$4 range, with transformers outperforming GNNs and MLPs in the test set. Training times are 2–10 minutes for $y$5-event datasets.

Notable limitations include: support for LHCO-only input (no ROOT support), lack of built-in hyperparameter optimization (manual YAML/GUI), and handling only trivial GNN edge features. Planned extensions include richer graph construction, improved prompt clarification, and expansion of data format compatibility (Esmail et al., 6 Feb 2026).

5. Related Logical, Physical, and Statistical Uses of “coll”/f_{coll}

In cosmological simulation and reionization modeling, the “collapse fraction field” $y$6 denotes the per-cell fraction of mass in collapsed dark-matter halos: $y$7 A recent GPR-based emulator (Pundir et al., 2024) stochastically models the conditional distribution $y$8, given density contrast $y$9. This enables highly accurate ($x<v$0 error) and efficient ($x<v$1 speedup over N-body) production of HI/HII fields for reionization studies, with the approach providing a substantial upgrade over analytic mass function prescriptions. This usage is conceptually distinct from the logical or AI-centric CoLL frameworks, but shares the “collection/collapse” nomenclature in denoting aggregate witness, mass, or feature collection.

6. Significance and Outlook

The various instantiations of CoLL across mathematical logic, machine learning, and physical sciences share a programmatic focus: formalizing the collection, synthesis, or bounding of complex information (witnesses, evidence, features) through either logical schema, AI collaboration, or domain-specific automation. In logic, Coll precisely characterizes the point at which weak theories admit completions of minimal complexity, delineating the border between effective and ineffective axiomatizability. In AI, both Court of LLMs and CoLLM frameworks demonstrate the synergy of human-readable reasoning and machine feature integration for anomaly detection and automated science, with substantial empirical gains in reproducibility, interpretability, and speed. In physical modeling, $x<v$2-based methods epitomize rigorous aggregation of stochastic evidence to scale complex simulation tasks.

Emerging trends include (i) the synthesis of symbolic and deep learning protocols (LLM–GNN, LLM–AutoML), (ii) the use of collection/bounding type principles to guide model interpretability and completeness, and (iii) the operationalization of collection mechanisms for complex, multi-modal data sources. Future research is likely to refine these frameworks for broader application in scientific automation, logical theory classification, and multi-agent AI systems.