Epistemic Latent Spaces in Machine Intelligence: Uncertainty, Ignorance, Evidence, and the Geometry of Belief

Full report

On this page

The central problem of modern machine intelligence is not merely the optimization of predictive accuracy, but the rigorous geometric and probabilistic representation of the system’s own epistemic state. Contemporary deep learning architectures compress high-dimensional inputs into latent spaces designed to linearly separate target categories. However, these spaces typically represent only what a machine intelligence system predicts, flattening the rich topography of knowledge into single-point estimates or uncalibrated distributions1. This exhaustive report addresses a fundamental theoretical question: Can a latent space be structurally engineered to represent not only what a machine intelligence system predicts, but also what it confidently knows, what it merely suspects, the exact provenance of the evidence supporting its claims, the conditions that would falsify its beliefs, and the fundamental boundaries of what remains outside its representational capacity? To answer this, we must abandon the treatment of confidence scores, token probabilities, or verbal hedging as calibrated epistemic uncertainty—practices that lack empirical validation and theoretical rigor3. Instead, we propose a transition toward explicit epistemic latent spaces, integrating conformal prediction, evidential deep learning, energy-based thermodynamics, and self-referential agent frameworks.

Formal Epistemic Categories

To mathematically model knowledge in machine intelligence, we must first disentangle the colloquial vocabulary of uncertainty into strict, formal epistemic categories. The conflation of these terms in legacy literature has historically led to misaligned optimization objectives. Table 1 defines the strict ontological primitives required for an epistemic latent space.

The Taxonomy of Uncertainty: Aleatoric vs. Epistemic

Machine intelligence models are burdened by the conflation of distinct uncertainty modalities. The total predictive uncertainty must be decomposed into components that dictate fundamentally different downstream actions1. Aleatoric uncertainty represents the irreducible, inherent stochasticity in the data-generating process. Even in the limit of infinite training data and infinite model capacity, this uncertainty persists. It is the noise of the environment. Mathematically, it is captured by the conditional distribution [Figure omitted from source export] and manifests as overlapping class distributions in the feature space4. Conversely, epistemic uncertainty represents reducible ignorance. It reflects a lack of knowledge about the optimal model parameters or structural form due to finite data. Epistemic uncertainty can be theoretically eliminated by gathering more evidence2. To build actionable systems, epistemic uncertainty must be further subdivided, as outlined in Table 2\.

Probability Versus Ignorance

A critical flaw in legacy machine intelligence systems is the reliance on standard softmax functions, which force the model to distribute a probability mass of exactly 1.0 across known classes9. This normalization destroys the distinction between probability and ignorance. If a binary classification system outputs [Figure omitted from source export], this scalar value is epistemically ambiguous. It may indicate balanced evidence, where the system has observed extensive, equally distributed data for both classes, reflecting high aleatoric noise. Alternatively, it could indicate conflicting evidence, where distinct feature sets of the input point strongly, but contradictory, to opposing classes. Finally, it may indicate absolutely no evidence, where the system is encountering a completely novel input and defaulting to a maximum-entropy prior. To resolve this, latent representations must preserve the distinction between probability and ignorance. Frameworks utilizing Credal Sets—which define sets of probability distributions rather than a single point estimate—and Dempster-Shafer theory explicitly decouple belief mass from uncertainty mass1. In these paradigms, total probability is not constrained to sum to 1 over the target classes without an explicit, quantifiable remainder allocated to a dedicated "ignorance" state.

Bayesian Latent Variables

While deterministic networks generate point estimates, Bayesian latent variable models capture epistemic uncertainty by marginalizing over the hypothesis space of model weights. Because exact Bayesian inference is intractable for deep latent spaces, approximate posteriors are generated via Variational Inference (VI). In VI, an intractable posterior [Figure omitted from source export] is approximated by a tractable distribution [Figure omitted from source export], optimized by minimizing the Kullback-Leibler divergence between them, equivalent to maximizing the Evidence Lower Bound (ELBO)18. However, Bayesian latent variables suffer from distinct pathologies. Prior sensitivity dictates that the choice of the prior distribution heavily skews the posterior predictive uncertainty, especially in regions far from the training data where the likelihood provides no constraint. Furthermore, variational autoencoders and Bayesian neural networks are notoriously vulnerable to posterior collapse, a phenomenon where the approximate posterior perfectly matches the non-informative prior, causing the model to completely ignore the latent variables and rely solely on autoregressive decoders or dominant features20. Multimodality presents another challenge. The true posterior of a deep machine intelligence architecture contains millions of distinct modes, yet standard variational inference (such as mean-field Gaussian approximations) focuses on a single mode. This results in severe misspecification, where the true data-generating process lies entirely outside the support of the chosen variational family, leading to overconfident, uncalibrated predictions in the epistemic latent space.

Ensembles and Sampling

To overcome the mode-seeking limitations of standard variational inference, practical epistemic representation relies heavily on ensembles and advanced sampling methodologies. Deep ensembles—training multiple independent networks with different random initializations—remain the gold standard for epistemic uncertainty quantification21. While not strictly Bayesian, they effectively sample diverse functional modes in the loss landscape, capturing both parameter and structural uncertainty. Ensemble divergence, calculated via the Jensen-Shannon divergence or mutual information across predictions, serves as a highly reliable proxy for epistemic ignorance. Bootstrap methods, involving training models on resampled subsets of the dataset, provide theoretical grounding for ensemble diversity, though they are computationally expensive. Monte Carlo (MC) dropout offers a cheaper alternative, interpreting stochastic dropout at inference time as a variational approximation to a deep Gaussian process19. However, MC dropout often underestimates epistemic uncertainty because it only approximates a localized variational distribution rather than exploring diverse functional modes19. Stochastic Weight Averaging \- Gaussian (SWAG) bridges this gap by capturing the geometry of the posterior directly from the optimization trajectory. By fitting a Gaussian distribution over the sequence of Stochastic Gradient Descent (SGD) iterates, SWAG computes the SWA mean and a low-rank plus diagonal covariance matrix. This allows for cheap, highly effective Bayesian model averaging that intrinsically captures the local landscape's curvature and provides robust disagreement-based uncertainty without the cost of deep ensembles19.

Evidential and Subjective Representations

Evidential Deep Learning (EDL) operationalizes the distinction between probability and ignorance in a single forward pass by utilizing Subjective Logic7. Instead of predicting a point-estimate categorical distribution via a softmax function, EDL architectures predict the concentration parameters [Figure omitted from source export] of a Dirichlet distribution, which serves as a conjugate prior over all possible categorical distributions24. For a [Figure omitted from source export]\-class ontology, the network outputs evidence vectors [Figure omitted from source export] using non-negative activation functions. The Dirichlet parameters are defined as [Figure omitted from source export]. The total Dirichlet strength, representing the aggregate evidence, is [Figure omitted from source export]. This formalism rigorously defines subjective opinions through distinct epistemic quantities6:

• Expected Probability: [Figure omitted from source export]
• Belief Mass: [Figure omitted from source export]
• Uncertainty Mass (Vacuity): [Figure omitted from source export]

Under Dempster-Shafer theory, an agent may also possess a disbelief mass, [Figure omitted from source export], such that [Figure omitted from source export]. The introduction of base rates, [Figure omitted from source export], allows the uncertainty mass to be distributed according to prior class frequencies rather than a uniform flat prior, guiding the expected probability as [Figure omitted from source export]5. When the network encounters an unknown space, it outputs [Figure omitted from source export], resulting in [Figure omitted from source export]. The uncertainty mass [Figure omitted from source export] approaches 1.0, explicitly projecting an epistemically ignorant state. For continuous regression, Deep Evidential Regression utilizes a Normal-Inverse-Gamma (NIG) distribution27. The network outputs parameters [Figure omitted from source export], establishing a prior over the unknown mean and variance of a Gaussian likelihood. This allows a closed-form disentanglement of aleatoric variance ([Figure omitted from source export]) and epistemic variance ([Figure omitted from source export])6.

Energy-Based Latent Spaces

Energy-Based Models (EBMs) evaluate the compatibility of input variables [Figure omitted from source export] and latent states [Figure omitted from source export] through an unnormalized scalar energy function [Figure omitted from source export]. In this thermodynamic framing, low energy indicates high compatibility and plausibility14. A machine intelligence system maps this energy to a normalized probability density via the Gibbs distribution: [Figure omitted from source export] The free energy of the input [Figure omitted from source export], marginalized over all possible labels, serves as a measure of density and structural familiarity14. EBMs theoretically provide a robust mechanism for separating knowns from unknowns: in-distribution samples should carve out deep, low-energy basins in the latent space, while anomalous samples should reside on high-energy plateaus14. However, energy is a measure of statistical compatibility, not a guarantee of truth or in-distribution status. Deep generative models notoriously assign higher likelihoods (and lower energies) to certain out-of-distribution datasets because the energy function is heavily influenced by population-level background statistics, such as low-frequency structural zeros or pervasive background textures32. Furthermore, untargeted adversarial attacks often minimize the energy function, tricking the EBM into assigning low energy to malicious noise, leading to catastrophic false confidence15.

Out-of-Distribution Detection Mechanisms

Detecting when a machine intelligence system has stepped outside its epistemic bounds requires specialized metrics. Table 3 compares the dominant mechanisms for Out-of-Distribution (OOD) detection.

The Geometry of Ignorance

To model epistemic uncertainty rigorously, we must answer a fundamental question: does an "unknown" occupy a distant region in Euclidean space, a low-density region, a high-curvature boundary, or a direction entirely unspanned by the representation? In conventional deep learning, unknown unknowns are often mistakenly projected onto the in-distribution manifold due to unconstrained feature extraction. To resolve this, we must enforce a rigorous geometry of ignorance. Distance-based ignorance requires the enforcement of bi-Lipschitz continuity, ensuring that distance in the input space strictly maps to distance in the latent space. Spectral Normalization (SN) bounds the operator norm of the weight matrices, effectively anchoring the representation space and preventing the collapse of epistemic boundaries34. Curvature-based ignorance reveals that unknowns manifest as sharp, rugged regions in the latent embedding space. Recent investigations into loss-landscape geometry demonstrate that OOD inputs exhibit significantly larger Hessian curvature than in-distribution data37. Sharpness-Aware Geometric Defense (SaGD) algorithms prove that adversarial attacks aim to push clean samples into these highly curved regions. Smoothing this landscape across hyperbolic or hyperspherical manifolds prevents adversarial points from being falsely recognized as OOD or vice versa39. Crucially, the system must geometrically distinguish between a state that is far from known data (detectable via distance and density metrics) and a state that is not representable by the current ontology. When a signal possesses variance in a dimensionality strictly orthogonal to the eigenvectors of the latent space, it constitutes an ontology collapse. Detecting this requires analyzing the residual error of a reconstruction network or disconnected components in a topological data analysis, rather than relying solely on the discriminator's forward pass.

Calibration

A machine intelligence system’s internal epistemic state is mathematically useless if it is uncalibrated. A perfectly calibrated model guarantees that out of all instances where it predicts an event with probability [Figure omitted from source export], the event occurs exactly [Figure omitted from source export] fraction of the time. Legacy metrics like Expected Calibration Error (ECE) compute the average difference between accuracy and confidence across arbitrary bins. However, ECE often underestimates the true miscalibration. The Adaptive Calibration Error (ACE) improves this by using equal-mass binning, ensuring that sparse regions of the prediction space do not skew the metric. Reliability diagrams provide a visual diagnostic of this calibration, plotting expected versus observed frequencies to highlight regimes of overconfidence or underconfidence. To optimize calibration directly, strict proper scoring rules are required. The Brier Score computes the mean squared difference between predicted probabilities and actual outcomes, decomposing neatly into reliability, resolution, and uncertainty components41. Log Loss (cross-entropy) heavily penalizes confident but incorrect predictions. Furthermore, models must be subjected to classwise calibration to prevent majority classes from masking the miscalibration of minority classes, and they must undergo calibration under shift, measuring the degradation of the Brier score when subjected to covariate distribution drift.

Selective Prediction and Abstention

Calibration provides point-wise reliability, but for high-stakes decision-making, we require bounded policies for selective prediction and abstention. By equipping a classifier with a rejection option, the system can trade coverage (the fraction of inputs it chooses to answer) for accuracy (the risk on the accepted subset). This dynamic is captured by the Risk-Coverage Curve. The quality of a selective classifier is measured by the Area Under the Risk-Coverage Curve (AURC). An epistemically superior latent space minimizes the AURC, systematically abstaining on highly uncertain inputs to maintain near-zero risk on the accepted subset42. To operationalize abstention according to consequence-sensitive thresholds, we deploy Conformal Risk Control (CRC). While standard Conformal Prediction provides distribution-free, finite-sample coverage guarantees for a prediction set (e.g., ensuring the true label is in the set with [Figure omitted from source export] probability), CRC extends this to control the expected value of any bounded, monotonic loss function10. For more complex, non-monotonic losses, the Learn-then-Test (LTT) framework utilizes multiple hypothesis testing on a calibration set to establish statistically valid thresholds45. These frameworks dictate precise policies for answering, acting, escalating to human operators, or abstaining entirely.

Evidence and Provenance

An explicit epistemic space cannot treat all data embeddings as equally valid or timeless. Each belief state must retain its exact source identity, independence assumptions, timestamps, derivation path, counterevidence, and revision history. Treating belief states as cryptographic ledgers rather than overwriteable tensors prevents the silent contamination of the latent space. If a downstream contradiction is detected, the provenance trace allows the system to compute the exact mutual information shared between sources, discounting evidence if it is derived from a common, redundant origin, and recursively updating the reliability prior of the specific source that provided the falsified information.

Contradiction

When contradictory evidence is introduced—for example, Source A provides a high-confidence embedding for Hypothesis [Figure omitted from source export], while Source B provides equally high-confidence for Hypothesis [Figure omitted from source export]—legacy systems either average the embeddings into a meaningless "moderate" probability of 0.5, or allow catastrophic forgetting to overwrite the old belief. In a structurally sound epistemic space, conflicting evidence must produce a distinct Contradiction State. Using Dempster-Shafer dissonance metrics, the system calculates the mass of conflict. If this dissonance exceeds a conformal threshold, the system enters a suspended commitment. Rather than forcing a mixture distribution, it relies on a multi-head attention mechanism where the attention weights represent the likelihood ratio of the sources, maintaining a distinctly multimodal representation until an active information-gathering subroutine can resolve the conflict6.

Unknown Unknowns

The detection of unknown unknowns—phenomena that the system does not even realize it is ignorant of—requires monitoring secondary and tertiary indicators of latent failure. Ensemble divergence serves as an early warning; when structurally identical models exposed to the same data diverge wildly in their out-of-distribution projections, the space is epistemically undefined. Representation instability, where infinitesimal perturbations in the input space cause catastrophic, discontinuous jumps in the latent space, indicates a fractured topological manifold. Furthermore, dynamic anomaly trajectories track how a sequential input evolves through the hidden states of recurrent or state-space models. If the trajectory deviates from the established differential equations governing the in-distribution attractors, it signals a causal-model failure. Finally, ontology-collapse signals are triggered when the residual error of a reconstruction network spikes, proving that the signal contains variance entirely orthogonal to the system's learned basis functions, necessitating a fundamental expansion of the model's taxonomy.

Active Epistemology

An epistemic latent space is not passive. When high vacuity (ignorance) or high dissonance (contradiction) is registered, the system transitions into Active Epistemology: the autonomous selection of actions specifically designed to reduce uncertainty13. Rather than outputting a final prediction, the system maps the uncertainty gradients to an action space. Actions are selected by calculating the Expected Information Gain (EIG) of various interventions. This includes querying external tools (e.g., retrieving data via an API), requesting specific human evidence, moving physical sensors to acquire a new perspective, running isolated software experiments, or performing causal interventions on the environment to observe downstream effects and break confounding correlations.

Latent Epistemic Self-Model

To effectively seek evidence and manage its own ignorance, a machine intelligence must possess a Latent Epistemic Self-Model. This is the cornerstone of the Self-Referential Agent Framework (SRAF)48. The system explicitly maintains, refines, and utilizes an internal model of its own states, parameters, capabilities, and uncertainties. The self-model represents the system's sensing limits (e.g., visual resolution, context window length), its reasoning limits, missing tools, and computational constraints. When faced with an unknown, the system queries its self-model to determine if the uncertainty is aleatoric (whereupon it should abstain or hedge) or epistemic (whereupon it can deploy an active epistemology subroutine). Inspired by recursive architectures like the Gödel Machine, this allows the agent to dynamically modify its own evidence-gathering logic and optimization parameters, guided by high-level objectives, bypassing the limitations of static, predefined pipelines49.

Required Architecture: The Epistemic Latent Register System

To implement these theoretical principles, we propose the Epistemic Latent Register System (ELRS), abandoning the monolithic vector embedding in favor of a modular architecture that explicitly separates information processing into ten distinct registers. Information moves between registers only when strict mathematical or statistical proofs are satisfied. Table 4 outlines the architecture and promotion requirements.

Required Benchmark: The Epistemic Frontier Benchmark

To empirically validate the Epistemic Latent Register System, we construct the Epistemic Frontier Benchmark, designed specifically to break standard predictive models and measure true epistemic resilience. Table 5 details the benchmark constraints and their corresponding target metrics.

Falsification Requirements, Risks, and Implementation Guidance

The scientific validity of the Epistemic Latent Register System rests on strict conditions for falsifiability. Falsification of Explicit Epistemic Structure: The hypothesis that explicit epistemic registers and Dirichlet modeling add measurable value over legacy networks would be falsified if a highly parameterized, monolithic standard network calibrated post-hoc using simple temperature scaling consistently achieves a lower or equal Area Under the Risk-Coverage Curve (AURC) than the Register System under severe covariate and ontology shift. If post-hoc calibration yields identical conformal risk bounds without explicit evidential architecture, the computational overhead of the Register System is rendered obsolete. Falsification of Geometric OOD Methods: The hypothesis that Hessian curvature, local flatness, and bi-Lipschitz distance act as reliable, fundamental proxies for ignorance would be falsified if untargeted adversarial attacks can routinely generate synthetic OOD samples that rest in flat, low-curvature, dense local minima of the representation space, thereby breaking Fold, SaGD, and spectral normalization simultaneously without triggering the Falsification Conditions register37.

Implementation Guidance and Risk Analysis

Implementing the Epistemic Latent Register System requires deprecating legacy categorical cross-entropy losses in favor of evidential objectives. Practitioners must apply Spectral Normalization across all feature extraction layers to prevent representation collapse, ensuring that Euclidean distance in the latent space correlates tightly with semantic distance. For calibration, the Learn-then-Test framework must be implemented at the very end of the pipeline to set the promotion thresholds for the Verified Commitments register45. A significant risk in implementation is the computational overhead of calculating likelihood ratios (which requires maintaining a secondary autoregressive background model) and the cost of maintaining continuous conformal guarantees in real-time streaming data environments.

Research Gaps

Primary research gaps remain in the dynamic scaling of the ontology. While the system can accurately flag an unknown via high vacuity, the autonomous generation of a new dimension in the latent space to represent that unknown without retraining from scratch is unsolved. Furthermore, future research must bridge the gap between continuous-time Neural Ordinary Differential Equations (NODEs) used for Lyapunov stability52 and discrete evidential logic to ensure that an agent's self-model remains mathematically sound under rapid, real-time adversarial manipulation. Closing these gaps will complete the transition of machine intelligence from confident, uncalibrated predictors into rigorous, epistemically bounded reasoners.

Works cited

1. Aleatoric and Epistemic Uncertainty in Machine Learning \- alphaXiv, https://www.alphaxiv.org/overview/1910.09457v3

2. Quantifying Aleatoric and Epistemic Uncertainty in Machine Learning, https://openreview.net/pdf?id=OOfLGZQVSk

3. When should we trust the annotation? Selective prediction for ... \- arXiv, https://arxiv.org/pdf/2603.10950

4. Why machine learning models fail to fully capture epistemic ... \- arXiv, https://arxiv.org/html/2505.23506v1

5. Uncertainty Estimation by Flexible Evidential Deep Learning, https://neurips.cc/virtual/2025/poster/118400

6. Evidential Neural Networks \- Emergent Mind, https://www.emergentmind.com/topics/evidential-neural-networks-enns

7. Revisiting Essential and Nonessential Settings of Evidential Deep, https://arxiv.org/html/2410.00393v1

8. Uncertainty Estimation by Flexible Evidential Deep Learning \- arXiv, https://arxiv.org/html/2510.18322v2

9. Evidential Deep Learning to Quantify Classification Uncertainty \- arXiv, https://arxiv.org/pdf/1806.01768

10. Conformal Risk Control for Non-Monotonic Losses \- arXiv, https://arxiv.org/pdf/2602.20151

11. CONFORMAL RISK CONTROL \- OpenReview, https://openreview.net/pdf?id=33XGfHLtZg

12. (PDF) Aleatoric and epistemic uncertainty in machine learning, https://www.researchgate.net/publication/349905827\_Aleatoric\_and\_epistemic\_uncertainty\_in\_machine\_learning\_an\_introduction\_to\_concepts\_and\_methods

13. A Scalable Four-Level Functional Hierarchy for Evaluating Large, https://www.preprints.org/frontend/manuscript/8f09ff457f773eea2cc4e98e7ede9c2a/download\_pub

14. Energy-based Out-of-distribution Detection \- arXiv, https://arxiv.org/pdf/2010.03759

15. Exploring the Connection between Robust and Generative Models, https://arxiv.org/html/2304.04033v4

16. Aleatoric and epistemic uncertainty in machine learning \- TransferLab, https://transferlab.ai/seminar/2022/aleatoric-epistemic-uncertainty-ml/

17. Evidential Deep Learning \- Amitesh Badkul, https://amiteshbadkul.github.io/blog/2026/evidential-learning/

18. When Bayesian Neural Networks Meet Evidential Deep Learning, https://raw.githubusercontent.com/mlresearch/v244/main/assets/wang24g/wang24g.pdf

19. A Simple Baseline for Bayesian Uncertainty in Deep Learning \- arXiv, https://arxiv.org/pdf/1902.02476

20. Out-of-Distribution Detection with An Adaptive Likelihood Ratio on, https://papers.neurips.cc/paper\_files/paper/2022/file/3066f60a91d652f4dc690637ac3a2f8c-Paper-Conference.pdf

21. Probabilistic Deep Learning to Quantify Uncertainty in Air Quality, https://www.mdpi.com/1424-8220/21/23/8009

22. SWAG : Approximate Bayesian Inference Using SGD Trajectory, https://bayesgroup.github.io/bmml\_sem/2019/Garipov\_SWAG.pdf

23. A Simple Baseline for Bayesian Uncertainty in Deep Learning, https://openreview.net/pdf?id=rkeMsHBlUS

24. Revisiting Essential and Nonessential Settings of Evidential Deep, https://arxiv.org/abs/2410.00393

25. A Comprehensive Survey on Evidential Deep Learning and Its, https://arxiv.org/html/2409.04720v1

26. Region-based evidential deep learning to quantify uncertainty and, https://pmc.ncbi.nlm.nih.gov/articles/PMC10505106/

27. README.md \- aamini/evidential-deep-learning \- GitHub, https://github.com/aamini/evidential-deep-learning/blob/main/README.md

28. The Unreasonable Effectiveness of Deep Evidential Regression, https://www.researchgate.net/publication/371915365\_The\_Unreasonable\_Effectiveness\_of\_Deep\_Evidential\_Regression

29. The Unreasonable Effectiveness of Deep Evidential Regression, https://ojs.aaai.org/index.php/AAAI/article/view/26096/25868

30. ML-GOOD: Towards Multi-Label Graph Out-Of-Distribution Detection, https://ojs.aaai.org/index.php/AAAI/article/view/33718/35873

31. Energy-based Out-of-distribution Detection | Request PDF, https://www.researchgate.net/publication/344552193\_Energy-based\_Out-of-distribution\_Detection

32. \[1906.02845\] Likelihood Ratios for Out-of-Distribution Detection \- arXiv, https://arxiv.org/abs/1906.02845

33. Likelihood Ratios for Out-of-Distribution Detection \- OpenReview, https://openreview.net/pdf?id=BJfkTHHl8S

34. Uncertainty Estimation of Transformer Predictions for, https://aclanthology.org/2022.acl-long.566.pdf

35. Density Uncertainty Layers for Reliable Uncertainty Estimation, https://proceedings.mlr.press/v238/park24a/park24a.pdf

36. Simple and Principled Uncertainty Estimation with Deterministic, https://proceedings.neurips.cc/paper/2020/file/543e83748234f7cbab21aa0ade66565f-Paper.pdf

37. Exploiting Local Flatness for Efficient Out-of-Distribution Detection, https://arxiv.org/html/2606.29952v1

38. Exploiting Local Flatness for Efficient Out-of-Distribution Detection, https://arxiv.org/abs/2606.29952

39. Sharpness-Aware Geometric Defense for Robust Out-Of-Distribution, https://arxiv.org/pdf/2508.17174

40. Sharpness-Aware Geometric Defense for Robust Out-Of-Distribution, https://arxiv.org/html/2508.17174v1

41. Gaussian Stochastic Weight Averaging for Bayesian Low-rank, https://arxiv.org/html/2405.03425v1

42. Aligning Language Models with Selective Prediction \- arXiv, https://arxiv.org/html/2607.03528v1

43. (PDF) Selective Classification for Deep Neural Networks, https://www.researchgate.net/publication/317100919\_Selective\_Classification\_for\_Deep\_Neural\_Networks

44. A Conformal Risk Control Framework for Granular Word Assessment, https://aclanthology.org/2025.findings-acl.638.pdf

45. Conformal Risk Control under Non-Monotone Losses \- arXiv, https://arxiv.org/html/2604.01502v2

46. Learn then test: Calibrating predictive algorithms to achieve risk, https://www.researchgate.net/publication/392305954\_Learn\_then\_test\_Calibrating\_predictive\_algorithms\_to\_achieve\_risk\_control

47. Calibrated Trust, Not Sharper Prediction \- arXiv, https://arxiv.org/pdf/2608.14617

48. Self-Referential Agent Framework \- Emergent Mind, https://www.emergentmind.com/topics/self-referential-agent-framework

49. Gödel Agent: A Self-Referential Framework Helps for Recursively, https://openreview.net/forum?id=dML3XGvWmy

50. A Self-Referential Agent Framework for Recursively Self-Improvement, https://www.researchgate.net/publication/394269980\_Godel\_Agent\_A\_Self-Referential\_Agent\_Framework\_for\_Recursively\_Self-Improvement

51. A Self-Referential Agent Framework for Recursively Self-Improvement, https://aclanthology.org/2025.acl-long.1354/

52. \[Literature Review\] Adversarially Robust Out-of-Distribution, https://www.themoonlight.io/en/review/adversarially-robust-out-of-distribution-detection-using-lyapunov-stabilized-embeddings