--- title: Uncertainty Externalization Overview url: https://www.emergentmind.com/topics/uncertainty-externalization type: topic --- # Uncertainty Externalization Overview Searching arXiv for the cited papers and topic-specific references to ground the article. arxiv_search(query="\"Unpacking uncertainty in the modelling process for energy policy making\" OR \"LLMs Should Express Uncertainty Explicitly\" OR \"Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty\" OR \"Confidence Laundering in Agent Systems\" OR \"Not All Uncertainty Is Equal\" OR \"Uncertainty Modeling for SysML v2\" OR \"Fighting Uncertainty with Uncertainty\" OR \"Market making by an FX dealer\"", max_results=10) Uncertainty externalization denotes a family of operations in which uncertainty is displaced from one locus of analysis or control to another rather than eliminated. In energy-policy modelling, it refers to pushing assumptions, value judgments, and gaps in knowledge outside the formal analysis, thereby creating an “excess of certainty” around particular policy options [2111.00782]. In recent LLM work, the same expression is used in an almost opposite sense: latent uncertainty is made explicit as a trainable interface for abstention, retrieval, verification, and intervention [2604.05306]. Related literatures use the term for preserving uncertainty across agent handoffs through latent carriers [2606.20662], encoding uncertainty as first-class structure in SysML v2 [2602.21641], transferring inventory risk from an FX dealer’s book to the inter-dealer market [2112.02269], or introducing randomness into the decision rule itself through “Randoptimization” [1601.04043]. Across these uses, the central question is not whether uncertainty exists, but whether it is hidden, exposed, propagated, or operationalized at the decision interface. ## 1. Domain-specific meanings The term has no single universal definition across arXiv literatures. Instead, it names several distinct interface phenomena. | Domain | Meaning of uncertainty externalization | Representative paper | |---|---|---| | Energy policy modelling | Uncertainty is pushed outside formal analysis | [2111.00782] | | LLM inference | Uncertainty is expressed explicitly for control | [2604.05306] | | LLM reasoning | Latent uncertainty is verbalized as epistemic information | [2603.15500] | | Agent systems | Fragility is preserved across handoffs via latent carriers | [2606.20662] | | MBSE / SysML v2 | Uncertainty is modeled explicitly in system structure | [2602.21641] | | Decision theory / operations | Randomness is introduced into the policy itself | [1601.04043] | | FX market making | Inventory risk is transferred to the market | [2112.02269] | In the energy-modelling literature, externalization is a pathology: “black-box” models, single cost-optimal pathways, and reliance on a single model or indicator suppress deep uncertainty and close off alternatives [2111.00782]. In LLM interface design, by contrast, externalization is treated as a design goal: uncertainty should be a first-class, trainable interface that the model exposes explicitly to downstream controllers [2604.05306]. In agent systems, the emphasis shifts from local estimation to uncertainty-preserving interface design, because uncertainty propagates only when it survives the handoff between components [2606.20662]. This suggests that the common denominator is representational location. What changes across fields is whether moving uncertainty outward is regarded as loss, preservation, or deliberate control. ## 2. Energy-policy modelling and the critique of false certainty In “Unpacking uncertainty in the modelling process for energy policy making” [2111.00782], uncertainty externalization is defined as the tendency of energy-system modelling to generate crisp scenarios or single “best-estimate” indicators while leaving unexamined assumptions, institutional choices, contested framings, and non-quantifiable uncertainties outside the formal analysis. The paper locates this problem in Post-Normal Science, where facts are uncertain, stakes are high, values are disputed and decisions are urgent. Externalization manifests as “black-box” models, fixation on one or two cost-optimal pathways, neglect of qualitative or “deep” uncertainties, and reliance on a single model or indicator for major policy commitments. The proposed remedies are three reflexive-modelling methodologies. NUSAP decomposes a quantity into Numeral, Unit, Spread, Assessment, and Pedigree. Its pedigree component scores each assumption against proxy representation, empirical basis, methodological rigour, validation, and theoretical understanding. In matrix form, if \(P\) is the pedigree matrix, an aggregate score may be written as \[ \mathrm{pedigreeScore}_i=\sum_{j=1}^5 w_j P_{ij}. \] Diagnostic diagrams then place each assumption on a two-axis scatter plot with pedigree on the horizontal axis and sensitivity on the vertical axis. The “Danger zone” is \(Q4\): high sensitivity and poor pedigree. Sensitivity auditing extends beyond technical UA & SA to interrogate framing, scope-definition, rhetoric of numbers, hidden value judgments, and governance assumptions. Its mathematical hook is the first-order Sobol’ index \[ S_i=\mathrm{Var}_{X_i}[E(Y\mid X_i)]/\mathrm{Var}(Y). \] The four detailed case studies show how these methods recapture uncertainty that conventional modelling had externalized. In the ExternE case, a transdisciplinary workshop identified 30 assumptions, reduced them to six critical ones, and found that all six received weak aggregate pedigree. In ESME, BioRES and CCSmbr lay in \(Q4\), implying high influence on total system cost but very weak pedigree. In the NETs-in-IAMs case, Vaughan & Gough’s 18-expert workshop found that almost all NETs parameters sat in the \(Q4\) danger zone. In the Ecological Footprint case, the seven-point auditing checklist exposed omitted sink functions, hidden technical coefficients, absent error bars, and no published uncertainty or sensitivity analysis. The significance of this literature is methodological rather than merely computational. It treats uncertainty as partly epistemic, value-laden, and institutional, so quality assurance requires workshops, pedigree matrices, diagnostic plots, and sensitivity auditing rather than only tighter confidence intervals. ## 3. Explicit uncertainty in LLMs: global confidence and local signals “LLMs Should Express Uncertainty Explicitly” reframes uncertainty as an interface for control rather than a latent quantity estimated post hoc [2604.05306]. It distinguishes two decision problems. Outcome uncertainty—“Should I trust this final answer?”—is addressed by a global scalar. Process uncertainty—“Am I hitting a fragile step in my chain-of-thought?”—is addressed by local signals during generation. The global interface appends a decimal confidence token \(c\in[0,1]\) after generating a reasoning trajectory \(z_{1:T}\) and an answer \(\hat y\), with \[ c \approx \Pr(Y=1\mid h_{1:T}),\qquad Y=\mathbf1\{\hat y=\text{gold}\}. \] The local interface allows the model to emit the literal token “\” during reasoning, with \[ a_t=\begin{cases} 1 & \text{emit “\” at step }t,\\ 0 & \text{otherwise.} \end{cases} \] Both interfaces are learned through a light post-training stage using a variant of GRPO. The idealized update is \[ \pi_{\theta'}(z\mid x)\propto \pi_\theta(z\mid x)\exp(\eta\,r(z;x)). \] For verbalized confidence, \[ r_{\rm verbal}(z;x)= \begin{cases} +p(z) & \text{if answer is correct},\\ -p(z) & \text{if answer is wrong}, \end{cases} \] which rewards justified confidence and penalizes over-confident errors. For “\”, the reward order is \[ r(\text{correct, no-emit})>r(\text{correct, emit})>r(\text{wrong, emit})>r(\text{wrong, no-emit}), \] implemented as \([5.0,\,3.5,\,0.0,\,-2.0]\). The reported effects are large. On 2WikiMultihopQA, verbalized confidence changes accuracy from \(0.345\to0.358\), ECE from \(0.383\to0.049\), Brier from \(0.504\to0.166\), NLL from \(4.987\to0.498\), and the over-confidence gap from \(+0.523\to+0.045\); wrong answers with \(c>0.7\) drop from \(88.5\%\) to \(3.2\%\) [2604.05306]. In Adaptive RAG aggregated over five datasets, Verbal-Calibrate reaches EM/F1 \(=41.6/50.5\) at \(48.1\%\) retrieval, while Uncertain-Calibrate reaches \(40.9/48.1\) at \(61.4\%\) retrieval. Mechanistically, the verbal interface concentrates token-level KL divergence almost exclusively at the final confidence digit and preserves layer geometry, whereas the special-token interface induces progressive late-layer CKA drop to \(0.85\), indicating a broader reorganization. A complementary theoretical account is given in “Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty” [2603.15500]. That paper decomposes reasoning into procedural information and epistemic verbalization. If \(Z_t\) is a latent internal variable, then making it verbalized and therefore conditionable yields \[ I(Y;Z_t\mid s_{t-1})>0 \Longrightarrow H(Y\mid s_{t-1},Z_t)