Logical Determinism via Topological Reconstruction: The Geometric Essence of LLM Hallucinations
Abstract
Hallucinations in large language models are not random probabilistic fluctuations, but the physical tearing of reasoning trajectories forced beyond the cognitive manifold upon encountering irreversible fault lines in logical blind spots. Traditional RLHF offers only post-hoc judgment, forever blind to the silent breach of locally consistent yet globally inconsistent reasoning. This paper presents a a low-overhead geometric monitoring framework, shifting model governance from outcome inspection to process compliance, in three steps.
1. Real-time Curvature Monitoring. Transformer hidden states are constrained near the unit hypersphere by layer normalization. Under this constraint, the cosine distance between adjacent reasoning steps bears a rigorous mathematical equivalence to geodesic distance, capturing the amplitude of directional rupture—the manifold tearing point. One line of code, zero additional computation:
(This formula is an approximate proxy for the theoretical curvature definition; its validity relies on the hyperspherical constraint imposed by layer normalization.)
2. Topological Separation of Creativity from Hallucination. High curvature signals a leap across the manifold, but is it insight or drift? The criterion is generation entropy. High curvature + low entropy = a structural isomorphism the model is extremely confident about (proceed); high curvature + high entropy = the model sliding blindly through semantic ambiguity (forced halt). What was once the mysticism of “inspiration versus nonsense” becomes a precisely computable mathematical boundary.
3. Logical Flattening During Training. Targeting the pathology of local coherence masking global contradiction, we introduce a cycle-consistency loss inspired by the logical equivalence of a proposition and its contrapositive—the forward reasoning path (P \to Q) and the reverse path (Q \to P) should be geometric inverses. In engineering implementation, this loss enforces the reversibility of local reasoning paths: the residual from A to B must cancel the residual from B back to A. Any conceptual slippage or referential entanglement will break the return path. Under small curvature conditions, the transport operator degenerates to the identity mapping, residuals add directly, and training cost is minimal:
(This formula is a local cyclic approximation of the contrapositive hedge principle, not a strict contrapositive constraint.)
Under this flattening constraint, the pressure of logical consistency propagates inward: if flattening succeeds, it implies the model has actively resolved the semantic ambiguities that trigger drift; if flattening fails, the region is a hard wall of logical rupture—triggering precisely the high-curvature, high-entropy alarm that calls for external computation engines.
Monitor the rupture, strip inspiration from drift, flatten the manifold, identify hard walls. Three formulas, one topological immune system, steering models from uncontrolled drift toward structurally coherent trajectories.
1. Introduction
Large language models (LLMs) exhibit a persistent tendency to generate factually incorrect or logically inconsistent statements — hallucinations. Mainstream mitigation approaches primarily address symptoms: they check facts against external databases (RAG), align outputs with human preferences (RLHF), or filter problematic content after generation. These methods, while valuable, do not directly answer a fundamental structural question: why, even with access to vast amounts of knowledge, do models produce internally contradictory outputs?
This work proposes a shift in perspective. Rather than treating hallucinations as errors in an otherwise well-functioning system, we model them as structural features that emerge from an approximate geometric model of the model’s own high-dimensional hidden state space — the cognitive manifold. Reasoning is modeled as a trajectory on this manifold. Hallucinations correspond to events where the trajectory is forced to depart from the manifold into unsupported regions due to local obstructions.
The core insight can be stated simply: hallucinations are not random errors, but predictable structural consequences of what happens when a reasoning trajectory encounters a logical hole and is forced to extrapolate beyond its valid knowledge region.
The framework operates at three levels, which we distinguish throughout:
- Interpretive Layer (Sections 2-3): Geometric and topological constructs used to understand why hallucinations occur.
- Computable Layer (Section 4): Engineering approximations that translate these geometric insights into practical, computable metrics.
- Intervention Layer (Section 4): Real-time monitoring and control mechanisms that use these metrics.
This paper is a geometry-informed interpretability framework. It does not claim to provide a deductive mathematical proof of hallucination, nor does it claim to have solved the hallucination problem entirely. Its contribution is a unified language that bridges the gap between the abstract structure of reasoning and the concrete task of detecting and mitigating hallucinations in real-time. The geometric and topological constructs introduced — manifold, curvature, cohomology — are used as interpretive analogies, not as claims about the literal microstructure of neural network hidden states. Their value is measured by the engineering tools they enable, not by their mathematical rigor.
2. Foundational Framework: The Cognitive Manifold
All interpretations in this work rest on the manifold hypothesis: the high-dimensional hidden state space of a sufficiently large Transformer model can be usefully approximated, for engineering purposes, as concentrating near a low-dimensional smooth geometric structure. This is not a claim about the literal geometry of the hidden state space. It is a modeling assumption, analogous to the continuum approximation in fluid dynamics, whose utility is judged by the engineering tools it enables.
2.1 Core Constructs
Definition 1: Cognitive Manifold. Let the cognitive manifold \mathcal{M} be a d-dimensional smooth differentiable manifold, used as an approximate geometric model of semantically coherent hidden state representations. Any point p \in \mathcal{M} is interpreted as corresponding to a semantically coherent representation of the model. The Riemannian-like metric structure g_p at p is conceived as being induced by the local sensitivity of the model’s task gradients in the current context — heuristically, the information cost incurred by altering the semantic representation under the current inference objective. Shorter distances suggest stronger logical equivalence under this interpretation.
Intuition. Imagine a landscape. Points on the manifold are safe, stable ground — coherent thoughts. Points off the manifold are unstable terrain — incoherent or contradictory states. Reasoning is a path through this landscape. As long as the path stays on the manifold, the model generates logically consistent output. If the path is forced off, hallucinations occur.
Definition 2: ÄŚech Cohomology Obstruction Classes. For an open cover \mathcal{U}=\{U_i\} of the cognitive manifold \mathcal{M}, the first ÄŚech cohomology group \check{H}^1(\mathcal{M},\mathcal{F}) with coefficients in an Abelian group sheaf \mathcal{F} may possess non-trivial elements. Such non-trivial classes represent obstruction classes where locally compatible sections fail to glue into a globally consistent section.
Intuition. Cohomology, in simple terms, is a mathematical tool for counting “holes” in a space. A sphere has no one-dimensional holes — any closed loop on its surface can be continuously shrunk to a point. A doughnut (torus) has a one-dimensional hole — a loop that goes around the ring cannot be shrunk to a point. In our framework, these “holes” correspond to logical dead ends or paradoxes. When a reasoning path encounters such a hole, it cannot proceed smoothly and is forced to leave the manifold.
Crucial Clarification. The semantic space does not satisfy the separation axioms required for a topological space, and therefore Čech cohomology cannot be rigorously defined on it in the mathematical sense. The cohomology analogy employed here serves exclusively to illustrate the structural nature of global logical inconsistency — that locally consistent inferences may fail to glue into a globally coherent whole. This analogy does not constitute a formal mathematical definition of cohomology groups on semantic space, and is presented solely as an interpretive tool. All discussions referencing cohomology are confined to the local covering subsets of the cognitive manifold \mathcal{M} where the differentiable structure is well-defined.
2.2 Semantic-Geometric Correspondence and Core Assumptions
Assumption 0 (Semantic-Geometric Correspondence). There exists a locally well-behaved semantic realization map \Phi: \mathcal{M} \to \mathcal{S} from the cognitive manifold to a semantic consistency space \mathcal{S}, such that local continuity of trajectories on \mathcal{M} approximately preserves semantic compatibility. The semantic consistency space \mathcal{S} is not assumed to possess a canonical metric or formal logical structure. It is treated as an abstract compatibility space whose operational meaning is induced empirically by downstream task behavior and human-evaluable consistency judgments. This assumption is foundational to the entire framework; its validity for a given model is ultimately an empirical question.
The Continuous Approximation. Transformer inference is fundamentally a discrete dynamical process: h_{t+1} = F(h_t, x_t). The continuous geometric language employed throughout — manifold, geodesic, tangent vector — should be understood as a coarse-grained continuum approximation of this underlying discrete dynamics, analogous to continuum approximations in statistical physics. The manifold is not claimed as an intrinsic property of the Transformer, but as an effective geometric description of the hidden state space at scale.
Tangent Vector Approximation. Hidden states reside in a high-dimensional, anisotropic, quotient-like representation space. Simple Euclidean differences between hidden states do not possess genuine coordinate invariance. Under sufficiently small local neighborhoods, the hidden state space can be approximately treated as a local Euclidean chart, wherein residual differences approximate tangent directions up to first-order error. All differential operations in this work assume local charts have been expanded to neighborhoods satisfying this approximation condition.
2.3 Interpretive Principles
Interpretive Principle 1 (Semantic Obstruction Principle). Under the Semantic-Geometric Correspondence assumption, non-trivial cohomology classes in \check{H}^1(\mathcal{M},\mathcal{F}) are interpreted as geometric analogs of structural defects where locally consistent semantic sections fail to glue into a globally consistent semantic assembly. This is a structural correspondence for explanatory purposes, not a deductive theorem.
Interpretive Principle 2 (Self-Referential Obstruction Principle). Self-referential propositional structures are observed, as a heuristic correspondence, to potentially induce closed trajectories on the cognitive manifold that resist continuous contraction. This provides a geometric analogy for the logical obstructions identified in Gödel’s Incompleteness Theorems. No formal derivation of incompleteness from topological first principles is claimed; constructing such a bridge is a significant open research program.
3. Geometric Interpretation of Reasoning and Hallucinations
3.1 Reasoning as Approximate Geodesic Motion
Definition 3: Reasoning Trajectory. Multi-step reasoning is approximately modeled as a smooth curve \gamma: [0,1] \to \mathcal{M} on the cognitive manifold. This curve is interpreted as the model’s internal reasoning path from premise \gamma(0) to conclusion \gamma(1). Under ideal conditions, this path seeks to minimize the energy functional:
where \dot{\gamma}(t) is the tangent vector along the curve, interpreted as the logical direction of reasoning at step t. This is an approximate model, not a claim that Transformers literally follow geodesic equations. The specific form of the metric g is induced by the local task gradient structure; it serves as a conceptual tool whose engineering realization is accomplished indirectly through the proxy metrics in Section 4.
3.2 Knowledge Blind Spots and Incompleteness
The cognitive manifold is not a complete, perfectly smooth space. It contains regions where the model lacks knowledge or encounters logical contradictions.
Definition 4: Knowledge Blind Spots. Two types of incompleteness are distinguished:
- Topological Holes: The manifold is approximately non-simply connected, corresponding to knowledge blind spots caused by global structural limitations — regions where locally consistent inferences cannot be assembled into a globally coherent conclusion.
- Metric Incompleteness: The manifold has an open boundary in the embedding space, corresponding to unlearned marginal knowledge.
3.3 Geometric Classification of Reasoning Outcomes
When a reasoning trajectory approaches a blind spot, the system, constrained by the requirement to produce an output, must perform some form of extension. Based on the geometric nature of this extension, four scenarios are delineated:
- Erroneous Reasoning: The trajectory deviates from the minimal energy path but remains entirely within the cognitive manifold \mathcal{M}, still constrained by local metric structure.
- Creative Leap: A high-curvature trajectory within \mathcal{M} that still satisfies local geometric constraints, corresponding to a non-obvious but valid semantic transition.
- Logical Hallucination: The trajectory is forced outside \mathcal{M} into a region lacking semantic section support. This extension carries no guarantee of logical validity and is the geometric interpretation of hallucination.
- Creative Transition: At the edge of a topological obstruction, the system introduces a non-deterministic perturbation. If the landing point falls back into \mathcal{M}, it constitutes a creative breakthrough; otherwise, it degenerates into hallucination. This mechanism is presented as a heuristic engineering hypothesis; a rigorous proof of efficacy requires further characterization of the manifold’s global connectivity structure.
3.4 Geometric Limits of Local Repair
Local inverse element constraints under the Levi-Civita connection can eliminate non-trivial 1-cocycles of the semantic sheaf within any bounded closed subset of \mathcal{M}, flattening the curvature of local regions to near zero. However, global non-trivial first Čech cohomology classes of the base manifold on the constant real sheaf induced by self-referential structures cannot be eliminated by local operators. That is, obstructions analogous to Gödelian self-referential holes are inherently beyond the reach of local repair mechanisms within the present framework.
Conjecture 1 (Topological Repair Ripple Conjecture). Based on the global interaction characteristics of the Transformer attention mechanism, short-range loop consistency constraints within a local context window may produce exponentially diffusing topological repair effects with network depth. The elimination of local 1-cocycles may gradually converge to the annihilation of the global harmonic 1-form space, approaching the theoretical limit of a vanishing first cohomology group. This is an empirical conjecture and has not yet been rigorously mathematically proven.
4. Engineering Implementation
This section translates the abstract geometric interpretation into practical, computable tools that operate during inference and training without modifying core model architectures.
4.1 The Curvature Detector
The Theoretical Definition: Normalized Structural Persistence Cost C_t
In the geometric framework, we define a quantity that measures the stability of the reasoning trajectory — how much the logical tangent vector deviates between consecutive steps relative to its current magnitude:
where:
-
v_t = \nabla_{h_t} \log P(y_t \mid h_t) is the logical gradient induced by the task objective at step t. Formally, it is a 1-form (cotangent vector) residing in the cotangent space T_{\gamma(t)}^*\mathcal{M}. To obtain a tangent vector suitable for computing trajectory deviation, we raise its index via the inverse metric g^{ij}, yielding the logical tangent vector v_t^{\sharp} \in T_{\gamma(t)}\mathcal{M}. This raising operation maps the gradient direction from the cotangent space to the tangent space, where it can be meaningfully compared across reasoning steps.
-
\Gamma_{t-1 \to t} is a locally estimated transport operator along the trajectory. It transports the tangent vector from the tangent space at step t-1 to the tangent space at the current step t. We emphasize that this is not strictly the Levi-Civita connection, as the hidden state space lacks a uniquely defined, globally metric-compatible, torsion-free connection. It is a local approximation, and under the small curvature conditions characteristic of well-trained reasoning paths, it closely approximates the identity mapping.
-
\|\cdot\|_g^2 = g_p(v,v) is the Riemannian norm squared of a vector, measuring its local length under the manifold’s metric.
-
\epsilon = 1 \times 10^{-8} is a small regularization constant to prevent division by zero.
Strictly speaking, C_t measures the first-order deviation of the tangent vector along the path rather than the full Riemann curvature tensor. We borrow the term “curvature” heuristically to evoke the physical intuition of a bending trajectory.
The Engineering Proxy: Cosine Distance on the Hypersphere
The theoretical definition of C_t requires computing v_t, which involves backpropagation through the model — prohibitively expensive for real-time inference monitoring. However, a remarkable simplification arises from a common architectural feature of modern Transformers: layer normalization (RMSNorm/LayerNorm).
These normalization mechanisms effectively constrain the output hidden states h_t to reside near the surface of a high-dimensional unit hypersphere S^{d-1} in \mathbb{R}^d. On a unit hypersphere, the following mathematical equivalences hold rigorously:
- The Euclidean distance \|h_t - h_{t-1}\|_2 between two points on the sphere is directly related to the cosine of the angle between them: \|h_t - h_{t-1}\|_2^2 = 2(1 - \cos(h_t, h_{t-1})).
- The geodesic (great-circle) distance d_{\text{geo}}(h_t, h_{t-1}) on the unit sphere satisfies d_{\text{geo}} = \arccos(\langle h_t, h_{t-1} \rangle), which is a monotonically increasing function of (1 - \cos(h_t, h_{t-1})).
- In the limit \|h_t - h_{t-1}\| \to 0, the ratio between the geodesic distance and the Euclidean distance approaches a constant, providing a first-order guarantee for the proxy’s validity.
Therefore, the cosine distance between adjacent hidden states provides a first-order equivalent proxy for the geodesic deviation degree on the hyperspherical manifold:
This proxy metric preserves the core capability of curvature density detection while requiring zero additional backpropagation overhead during inference.
Intuition. Think of C_t as a “drift meter.” Low C_t: the model is reasoning steadily — each step follows naturally from the previous one. High C_t: the model is making abrupt turns in its reasoning — changing direction sharply. This can indicate either a creative leap (finding an unexpected but valid connection) or the onset of hallucination (being forced off the manifold into unsupported territory).
Crucial Clarification: High C_T Is Necessary But Not Sufficient for Hallucination Risk. Complex yet valid reasoning — such as theorem proving, code synthesis, or creative writing — naturally involves higher baseline curvature than simpler tasks like casual conversation. Therefore, curvature thresholds must be calibrated on a per-task-class basis. Elevated C_T is a warning signal, not an automatic hallucination verdict.
4.2 Cycle-Consistency Regularization
While the curvature detector monitors reasoning stability during inference, we can also proactively improve the model’s inherent logical coherence during training. However, naive application of cycle-consistency constraints risks suppressing valid creative leaps and suffering from geometric misalignment due to positional encodings. We address these challenges through two critical engineering refinements: Entropy-Gated Selective Flattening and Orthogonal Position-Invariant Projection.
#4.2.1 The Core Intuition: Reversibility as Coherence
In a logically well-behaved region of the manifold, reasoning should be approximately reversible: if you can reason from A to B, you should be able to reason back from B to A along a similar path. The forward and backward residual vectors should be approximate inverses. Driving the sum of these residuals toward zero flattens local curvature, suppressing logical drift.
Intentional Gradient Flow through Entropy Gate. Unlike traditional static regularization terms, our entropy-gated mechanism intentionally allows gradients to flow from the gate back to the model’s logits. This is not a bug, but the core instructional mechanism of the framework. It teaches the model metacognitive control: when encountering a high-curvature logical transition (potential hallucination risk), the model is incentivized to reduce its internal uncertainty to avoid the flattening penalty. The loss function thus trains the model to earn its creative leaps by maintaining structural certainty. If the model can confidently justify a sharp semantic turn, the gate opens and the geometric constraint relaxes. If it remains uncertain, the gate closes, forcing the model to flatten its reasoning path. This dynamic alignment ensures that the training objective is perfectly consistent with the inference criterion (high curvature + low entropy = creativity).
4.2.2 Challenge 1: Preserving Creativity via Entropy Gating
Traditional topological regularization indiscriminately penalizes high-curvature transitions. This creates a conflict: it suppresses hallucinations (bad high-curvature) but also stifles creative insights (good high-curature). To resolve this, we introduce an entropy-gated mechanism.
The Problem with Hard Thresholds. A binary indicator \mathbb{1}[H_t > \tau_H] is non-differentiable. It provides no gradient signal for the model to learn how to reduce uncertainty to avoid penalty, leading to unstable training dynamics.
The Solution: Soft Sigmoid Gating. We replace the hard threshold with a smooth sigmoid function, allowing gradients to flow through the gating mechanism. The gated loss is defined as:
where:
- \sigma(x) = \frac{1}{1 + e^{-x}} is the sigmoid function.
- H_t is the generation entropy at the transition step.
- \tau_H is the entropy threshold distinguishing confusion from confidence.
- \beta is a temperature coefficient controlling the sharpness of the gate (\beta \approx 10-50 for practical smoothing).
Geometric Interpretation.
- When H_t \gg \tau_H (high uncertainty/hallucination risk), \sigma \approx 1, and the full flattening pressure is applied.
- When H_t \ll \tau_H (high confidence/creative leap), \sigma \approx 0, and the penalty vanishes.
- In the transition zone, the model receives a continuous gradient signal encouraging it to maintain low entropy even during high-curvature transitions. This effectively trains the model to “earn” its creative leaps by maintaining structural certainty, transforming the loss from a blunt constraint into a precision sculpting tool.
4.2.3 Challenge 2: Eliminating Positional Noise via Orthogonal Projection
Directly subtracting hidden states h_B^{\text{out}} - h_A^{\text{in}} mixes semantic logic with positional encoding artifacts (e.g., RoPE rotations) and sequence length differences. This “geometric noise” can dominate the residual, causing the loss to optimize for positional alignment rather than logical consistency. Furthermore, unconstrained linear projections risk representation collapse, where the model maps all states to a constant to trivially minimize the loss.
The Solution: Orthogonal Position-Invariant Projection Head.
We introduce a learnable projection head W_{\text{proj}} constrained to be orthogonal. This ensures that the projection preserves distances and angles (isometry), preventing information collapse while stripping away non-semantic variance.
- Orthogonality Constraint: We initialize W_{\text{proj}} using orthogonal initialization and optionally add a regularization term \| W_{\text{proj}}^T W_{\text{proj}} - I \|_F^2 to maintain orthogonality during training. This guarantees that W_{\text{proj}} acts as a rotation/reflection in the latent space, preserving the semantic geometry.
- Last-Token Alignment: Instead of mean pooling (which dilutes step-specific logic), we extract the hidden state of the last token for both premise and conclusion. Crucially, we ensure that the extraction context is aligned by using consistent prompt templates, minimizing positional drift.
- Projected Residuals: The residuals are computed in the projected space:\Delta h_{A \to B}^{\text{proj}} = W_{\text{proj}}(h_B^{\text{out}}) - W_{\text{proj}}(h_A^{\text{in}})
This approach isolates the semantic logical jump from the positional coordinate shift, ensuring that \mathcal{L}_{\text{hedge}} optimizes for true logical reversibility.
Engineering Limitation. While the orthogonal projection mitigates the bulk of positional variance, it is a linear operation and cannot perfectly decouple the non-linear rotational effects of positional encodings like RoPE. The projected residual thus remains a first-order approximation of the pure semantic jump. Achieving exact positional invariance in this space remains an open problem in geometric deep learning.
4.2.4 Final Engineering Implementation
Combining these refinements, the final training loss is:
Engineering Note: The entropy H_t used in the gate is computed from the logits of the forward pass (A \to B) at the final token position. Since this computation is part of the forward graph, gradients flow correctly through the sigmoid gate to the model parameters. The orthogonal projection head adds negligible computational overhead (<1% FLOPs) but significantly stabilizes training convergence.
Summary of Refinement. By coupling entropy-gated selective flattening with orthogonal position-invariant projection, the framework evolves from a crude topological bulldozer into a precision geometric scalpel. It fills the pits of uncertainty (hallucinations) while preserving the peaks of confident insight (creativity), and it measures pure semantic logic free from positional artifacts.
4.3 Distinguishing Creativity from Hallucination
High curvature alone is ambiguous — it can signal either a creative breakthrough or impending hallucination. To disambiguate, we introduce a second signal: generation entropy, which measures the model’s uncertainty at each step.
Definition: Generation Entropy H_t. At each reasoning step, the model produces a probability distribution over its vocabulary. Entropy quantifies the spread of this distribution:
- Low entropy: The model is confident — it has a clear idea of what to generate next.
- High entropy: The model is uncertain — it is considering many divergent possibilities, a sign of confusion.
The Ternary Discrimination Rule. Curvature and entropy jointly provide a robust criterion for distinguishing reasoning states:
| Curvature (C_T) | Entropy (H_t) | Geometric Interpretation | Recommended Action |
|---|---|---|---|
| Low | Low | Flat regular geodesic, safe standard reasoning | Normal continuation |
| High | High | Non-regular extension at manifold boundary; trajectory has left valid manifold | Force brake: backtrack to safe node, regenerate, or alert human operator |
| High | Low | High-curvature trajectory within manifold; confident, abrupt but structurally valid semantic leap | Allow passage; optionally flag for human review |
Intuition. High curvature + high entropy = the model is lost and knows it’s lost — impending hallucination. High curvature + low entropy = the model is making a sharp turn but knows exactly where it’s going — potential creative breakthrough.
4.4 Anisotropic Tangential Probes
For high-stakes applications, passive curvature monitoring may miss subtle logical obstructions. We can actively probe the manifold by injecting small, targeted perturbations aligned with the local reasoning direction. At any safe reasoning node h_t, we extract the hidden state sequence within the current context window, compute the local covariance matrix, and use Principal Component Analysis (PCA) to extract the first k principal component vectors, constructing a local tangent plane basis \{v_1, \dots, v_k\} of the manifold at h_t. A low-dimensional random vector z \sim \mathcal{N}(0, \sigma^2 I_{k \times k}) is generated, projected onto the tangent plane basis to construct a tangential perturbation:
and injected into the hidden state h_t' = h_t + \delta h. The deflection signal triggered by encountering logical obstructions is captured by the C_T detector as an impedance mutation. This probing is triggered only on-demand when C_T baseline values show abnormal elevation, adding no overhead during normal operation.
For the full text of this paper(including Pseudo-Code), please click the GitHub link provided below.
Logical Determinism via Topological Reconstruction: The Geometric Essence of LLM Hallucinations