CoE Loop: Iterative Frameworks in AI & Physics

• CoE Loop is a versatile framework that implements closed-loop, iterative processes across domains such as computer vision, large language models, quantum information, and mixture-of-experts.
• Its methodologies include chain-of-explanation in vision models with 36% improved interpretability, latent state trajectory analysis in LLMs for self-evaluation, and elliptic derivative geometry in quantum metrology.
• In mixture-of-experts networks, the Chain-of-Experts Loop increases expert-path diversity and reduces memory costs, illustrating CoE Loop's power to refine global-local couplings in complex systems.

The term "CoE Loop" refers to distinct methodological and architectural frameworks across multiple fields, principally computer vision explainability, LLM latent trajectory analysis, quantum information, and mixture-of-experts neural architectures. In each context, the "loop" structure formalizes closed-form iterative or geometric processes that confer interpretability, communication, or analytic structure, often with operational and quantitative significance.

1. Chain-of-Explanation Loop in Vision Model Explainability

The Chain-of-Explanation (CoE) Loop, introduced in "CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification" (Yu et al., 19 Mar 2025), is a post-hoc explainability mechanism for deep vision models (DVMs). CoE systematizes global and local explanations through a unified, closed-loop process that quantifies and disentangles polysemanticity in visual concepts.

Pipeline Structure

The CoE loop comprises three modules:

• Automatic Visual Concept (VC) Decoding & Global Dataset Construction (ACD): For a vision model $\mathcal{M}$ and image corpus $\mathcal{T}$, feature channels $C$ at layers $L$ are extracted. Top-$N$ image patches maximizing each channel's activation are summarized by a language-vision model to yield a per-channel semantic description dataset $B_{\mathcal{M},\mathcal{T}}$.
• Concept Polysemanticity Disentanglement & Filtering (CPDF): Each visual concept is decomposed into atomic semantics ("atoms"), clustered for equivalence, and contextually filtered per image using the image caption and an atom-selection filter $z$.
• Local Linguistic Explanation via Concept Circuits: For a test image $x$, salient concepts per layer are filtered according to their contextual atom, and the resulting ordered path is rendered into coherent linguistic explanations by a prompting LLM.

The CoE Loop Mechanism

The loop is realized through intertwined global and local passes:

• The global pass (offline) constructs and quantifies global concept-atoms and their distributions.
• The local pass (online, per input) traverses the concept circuit, contextually filters atoms, and returns a composed explanation.
• Feedback arises as local context refines (and potentially retrains or reweights) the global atom database, closing the loop.

Concept Polysemanticity Entropy (CPE)

The polysemanticity of a visual concept—the uncertainty arising from channels firing for multiple distinct semantics—is quantified by the normalized entropy $H^*_j$ of atom distributions: $$H^*_j = -\frac{\sum_{i=1}^{P^* + \text{Pad}} p_i \log p_i}{\log(P^* + \text{Pad})}$$ where $\mathcal{T}$0 is the normalized count and Pad ensures degenerate cases are handled. CPE is computed at channel, layer, and model levels to characterize interpretability.

Performance and Empirical Gains

CoE reports substantial interpretability improvements, with explainability scores exceeding non-looped and manual baselines by an average absolute 36%. Per-channel CPE values empirically range from $\mathcal{T}$1 to $\mathcal{T}$2, peaking in intermediate layers and increasing with model depth. Human agreement with CPE on concept polysemanticity reaches $\mathcal{T}$3 (Yu et al., 19 Mar 2025).

2. Chain-of-Embedding Loop in LLM Latent Trajectory Self-Evaluation

The Chain-of-Embedding (CoE) Loop in LLMs captures the sequence of averaged, per-layer hidden states traversed as an input is processed. This sequence geometrically characterizes the model’s latent reasoning and can be used for unsupervised, output-free self-evaluation of correctness (Wang et al., 2024).

Formal Definition

Given an LLM defined by composed mappings $\mathcal{T}$4, the CoE is the sequence: $\mathcal{T}$5 where $\mathcal{T}$6 is the average representation at layer $\mathcal{T}$7, across all generated tokens. Two geometric features summarize the trajectory:

• Average normalized magnitude change:

$\mathcal{T}$8

with $\mathcal{T}$9.

• Average normalized angle change:

$C$0

with $C$1 the total path angle.

Correct answers exhibit systematically larger $C$2 and smaller $C$3 than incorrect ones.

Operational Deployment

CoE-based correctness predictors (CoE_R and CoE_C) are computed on hidden states alone, require no labels, and add negligible computational overhead ($C$4 ms per example on 7B models). These predictors outperform multiple output-based uncertainty metrics across diverse domains, languages, and LLM scales, providing robust, model-agnostic real-time self-evaluation (Wang et al., 2024).

3. Curvature of Entanglement Loop in Quantum Information

In quantum metrology, the "CoE Loop" refers to the closed geometric structure traced by the first and second derivatives of two-qubit concurrence as a function of the interaction coupling parameter. Its operational significance is described in "Quantum Fisher Information and the Curvature of Entanglement" (Saleem et al., 18 Apr 2025).

Mathematical Structure

Given the flip-flop Hamiltonian $C$5, the time-evolved concurrence for a pure separable probe is: $C$6 with $C$7. The curvature of entanglement is: $C$8 Notably, at points $C$9, the concurrence is extremal and $L$0, where $L$1 is the quantum Fisher information.

The CoE Loop Geometry

A parametric plot of $L$2 for varying $L$3 at fixed $L$4 forms an ellipse: $L$5 The vertical semi-axis equals $L$6, and its bottom corresponds to optimal measurement points where the symmetric logarithmic derivative eigenvectors are product states, so local measurements saturate the quantum Cramér–Rao bound. Elsewhere, entangled measurements are required (Saleem et al., 18 Apr 2025).

4. Chain-of-Experts Loop in Mixture-of-Experts Neural Networks

Chain-of-Experts (CoE) introduces a looped communication structure within Mixture-of-Experts (MoE) layers, as proposed in "Chain-of-Experts: Unlocking the Communication Power of Mixture-of-Experts Models" (Wang et al., 23 Jun 2025). Traditional MoEs route tokens to a static selection of experts in parallel; in CoE, tokens traverse a sequence of smaller, dynamically-routed expert groups, iteratively refining their state.

Algorithm and Loop Structure

Each CoE layer comprises $L$7 iterations, each with its own router and $L$8 expert budget. At iteration $L$9:

1. Compute per-expert routing scores and select $N$0 experts.
2. Update the token state via a mixture of expert outputs plus an inner residual connection with the previous state:

$N$1

After $N$2 steps, $N$3 is output. This design preserves total expert compute but increases the diversity of routing combinations and exploits sequential expert communication.

Empirical Outcomes and Scaling

CoE architectures match or outperform wider or deeper standard MoEs at reduced memory cost (17.6–42% less), with ablation showing the necessity of per-iteration routing and inner residuals (final loss with full CoE: 1.12 vs. 1.20+ in baselines). Combinatorial analysis shows vastly increased expert-path diversity: $N$4 combinations versus $N$5 in single-step MoEs (Wang et al., 23 Jun 2025).

5. Comparative Table: CoE Loop Across Domains

Context	Loop/Trajectory Mechanism	Operational Impact
Vision Model Explainability	Chain of global/local polysemantic concept passes	Boosted explainability, polysemanticity quantification
LLM Self-Evaluation	Layerwise hidden state trajectory	Label-free correctness prediction
Quantum Entanglement	Elliptic parameteric derivative trajectory	Identifies measurement optimality
Mixture-of-Experts Networks	Sequential token-expert dispatch with residuals	Improved efficiency, representational capacity

6. Synthesis and Outlook

The CoE Loop paradigm, while instantiated differently based on domain, is a recurring blueprint for leveraging closed-form or iterative traversals of objects—be they concepts, embeddings, entanglement measures, or expert specializations—to achieve new forms of interpretability, robustness, and expressivity. Its primary technical role is to facilitate feedback or communication between previously decoupled processes or representations, yielding improved downstream metrics and deeper insight into the functioning of complex systems. Further advances may exploit such loop structures for adaptive refinement of global-local couplings, tighter feedback between inference and explanation, and scalable expert composition.