Relational Linearity is a Predictor of Hallucinations¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/yuetianluNLP/SyntHal
Area: Natural Language Understanding
Keywords: Relational Linearity, Large Language Model Hallucinations, Knowledge Gaps, Refusal Calibration, Linear Relational Embeddings
TL;DR¶
This paper introduces and empirically validates the relational linearity hypothesis, demonstrating that the representational accessibility of relation-level affine maps (\(\Delta\cos\)) is a powerful predictor of whether language models hallucinate plausible values instead of refusing when queried on unknown entities (\(r \in [0.58, 0.84]\) across multiple model families).
Background & Motivation¶
Large language models (LLMs) frequently hallucinate unsupported or fabricated facts when answering factual questions about entities. While extensive post-training with reinforcement learning from human feedback (RLHF) has significantly enhanced instruction following and safety, models still exhibit inconsistent behavior when encountering entities outside their pretraining distribution. Prior mechanistic interpretability work has identified internal circuits and linear probing directions that indicate when an entity is unknown. However, when models are queried across different factual relations (e.g., asking for the CEO of a company versus the parent of a person), their propensity to hallucinate rather than abstain or refuse diverges sharply and systematically across relations.
The core tension underlying this discrepancy is that refusal is governed not merely by subject familiarity, but by the geometric structure through which the queried relation is encoded inside the model. Building upon the Linear Relational Embedding (LRE) framework, recent studies have shown that certain factual relations are well approximated by an affine mapping from subject representations to object representations. While this shared, low-dimensional relational structure enables efficient factual recall and zero-shot transfer for familiar entities, it also provides the model with an unconstrained computational shortcut: it allows the model to readily project an unknown subject vector into an object embedding space populated by valid-type candidates. Conversely, for highly nonlinear relations, no such global affine shortcut exists, forcing the model to rely on instance-specific memory which, when absent, naturally triggers abstention.
This insight motivates investigating whether internal relational linearity directly determines the boundary between hallucination and refusal under controlled knowledge gaps. Core idea: Relational linearity (\(\Delta\cos\)) serves as an intrinsic geometric predictor of factual hallucinationsโlinear relations provide an accessible affine mechanism that projects even non-existent subjects into plausible object representations and induces false commitment, whereas nonlinear relations lack this shortcut, making knowledge absence easier to detect and thereby eliciting appropriate refusals.
Method¶
Overall Architecture¶
The proposed investigation protocol establishes a systematic pipeline to quantify representational relational linearity and evaluate model behavior under rigorous knowledge-gap conditions. The method operates across three tightly integrated stages: representation-space linearity estimation, synthetic knowledge-gap benchmark construction, and decoupled behavioral evaluation. First, using natural factual relation triples, contextual representations of subjects and objects are extracted from frozen intermediate model layers to fit relation-specific affine maps via regularized ridge regression, calculating the held-out cosine improvement \(\Delta\cos\). Second, to instantiate pure knowledge gaps untainted by pretraining memory, the authors construct SyntHal, a 15-relation benchmark featuring 3,000 synthetic entities vetted through external collision audits. Third, model generations are evaluated using a structured LLM-as-a-judge protocol accompanied by deterministic regex baselines, measuring hallucination versus refusal rates on synthetic inputs and evaluating output plausibility and public attestedness on natural queries.
The overall workflow is illustrated below:
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
In["Input Triples and Entities<br/>(LRE Knowledge Triples & Synthetic Entity Generators)"] --> D1["Ridge-Estimated Relational Linearity<br/>(Affine Mapping & Held-out ฮcos Metric)"]
In --> D2["Knowledge-Gap Benchmark Construction<br/>(15-Relation SyntHal & Multi-Stage Collision Audit)"]
D1 --> D3["Decoupled Behavioral Evaluation Protocol<br/>(Binary Refusal/Hallucination & Plausibility Routing)"]
D2 --> D3
D3 --> Out["Diagnostic Output & Calibration<br/>(Linearity-Driven Hallucination Prediction)"]
Key Designs¶
1. Ridge-Estimated Relational Linearity: Representation Space Mismatch -> Regularized Affine Projection -> Robust Geometric Alignment Metric
Previous evaluations of linear relational embeddings primarily relied on decoder-facing first-token faithfulness. However, top-1 token matching is heavily confounded by high-frequency majority token defaults and tokenizer segmentation artifacts. Furthermore, because the number of available relation training pairs is substantially smaller than the hidden state dimension (\(|T| < d + 1\)), unregularized ordinary least squares (OLS) leads to high variance and numerical instability. To isolate genuine geometric alignment, the authors extract subject representations \(s_i\) at mid-layer \(\ell_s = \lfloor L/2 \rfloor\) and object representations \(o_i\) at deep layer \(\ell_o = L - 2\) via mean-pooling. An affine map is estimated using ridge regression with \(\ell_2\) regularization parameter \(\lambda = 1\) and an unregularized intercept:
The closed-form solution yields \((W_r, b_r)\), and relational linearity is evaluated on held-out test triples \(E\) using the cosine-improvement metric \(\Delta\cos\):
This metric isolates the net directional improvement in representation space over the unmapped baseline. Compared to minimum-norm OLS, ridge regression reduces held-out mean squared reconstruction error by 62.2% to 83.6%, achieves near-perfect cross-split rank stability (\(\rho \ge 0.990\)), and retains strong correlation with decoder first-token faithfulness (\(r = 0.86 \sim 0.90\)).
2. Knowledge-Gap Benchmark Construction: Pretraining Memory Leakage -> Frozen Token Pools & Collision Audit -> Pure Unknown-Entity Probing
Evaluating model behavior on real-world long-tail entities is inherently confounded: an incorrect answer could stem from incomplete memorization, conflicting training sources, or partial entity familiarity, making it impossible to guarantee that the subject represents a genuine knowledge gap. To circumvent this, the authors construct SyntHal, a synthetic benchmark spanning 15 factual relations (\(N = 200\) prompts per relation). Synthetic subjects are constructed deterministically from frozen token pools (e.g., first/last names, company prefixes, astronomical identifiers) curated using strict semantic filtering rules.
To guarantee construct validity and eliminate accidental overlap with real entities, the entire inventory of 3,000 synthetic prompts was subjected to a rigorous three-way collision audit: (1) exact Wikipedia title lookup, (2) quoted full-text Wikipedia search, and (3) Wikidata label/alias searches, using a strict normalized sequence-similarity threshold of 0.93. The audit verified that 2,995 of 3,000 prompts (99.83%) were completely clean, with only 5 ambiguous or matched prompts confined exclusively to fictional superhero relations. This confirms that SyntHal establishes an authentic knowledge-gap setting where refusal is the sole correct behavior and any committed attribute represents a hallucination.
3. Decoupled Behavioral Evaluation Protocol: Conflated Refusal Modes -> Multi-Way LLM Judging & Web-Grounded Attestation -> Plausibility Stratification
Assessing generative responses requires distinguishing non-committal abstentions from hallucinated claims, as well as examining whether generated errors are random noise or structurally plausible candidates. In the synthetic SyntHal setting, Gemini-2.5-Flash acts as a binary judge categorizing outputs into REFUSAL or HALLUCINATION along with confidence scores and explanations. The judge achieved 100% agreement on 200 human-annotated samples and 96.4% agreement (Cohen's \(\kappa = 0.928\)) with an offline deterministic regex baseline. The primary metric is formalized as \(\text{Hallucination} / (\text{Hallucination} + \text{Refusal})\).
In the natural LRE setting, the authors implement a three-way judge (CORRECT, HALLUCINATION, REFUSAL) evaluated against gold targets, and further analyze all 13,960 hallucinated outputs using a web-search-grounded rubric. This protocol measures the Plausibility Rateโthe proportion of hallucinations that represent valid candidate types for the queried role (e.g., outputting a real country name when asked for a landmark's location, even if incorrect)โand independently validates whether candidates are publicly attested entities. This decoupled protocol reveals that relational linearity governs not only the probability of committing to an answer, but also the semantic category coherence of the generated hallucination.
Formal Definition and Evaluation Protocol¶
Let \(\mathcal{R}\) denote the set of factual relations and \(\mathcal{M}\) denote the language model under evaluation: 1. Linearity Estimation (\(\Delta\cos\)): Given relation \(r\), natural triples are split into training set \(T\) and test set \(E\) (75:25 ratio). The affine operator \((W_r, b_r)\) is fitted via Equation (1). The score \(\Delta\cos(r)\) is computed over \(E\) according to Equation (2). 2. Synthetic Hallucination Rate: For each relation \(r\) in SyntHal (\(N = 200\)), prompts \(\mathcal{Q}_r^{\text{syn}}\) are evaluated under greedy decoding (\(T=0\), max 64 tokens). Responses are classified by the binary judge into \(N_{\text{hall}}\) and \(N_{\text{ref}}\), defining:
- Natural Answered-Case Accuracy and Plausibility: For natural triples, responses are classified into \(N_{\text{corr}}\), \(N_{\text{hall}}\), and \(N_{\text{ref}}\). Answered-case accuracy is defined as \(N_{\text{corr}} / (N_{\text{corr}} + N_{\text{hall}})\). Among the \(N_{\text{hall}}\) errors, the plausibility rate measures the fraction judged as legitimate entity fillers for the expected relation schema.
Key Experimental Results¶
Main Results¶
The primary experiment evaluates the correlation between relational linearity \(\Delta\cos\) and hallucination rates across 15 SyntHal relations on four mainstream instruction-tuned models, with further validation on base and larger-scale variants.
Table 1: Correlation between relational linearity (\(\Delta\cos\)) and hallucination rate on SyntHal across models
| Model | Parameters | Mean Hallucination Rate | Pearson \(r\) (\(p\)-value) | Spearman \(\rho\) (\(p\)-value) | Key Behavioral Observation |
|---|---|---|---|---|---|
| Gemma-7B-IT | 7B | 0.635 | 0.658 (\(p = 0.0077\)) | 0.589 (\(p = 0.0208\)) | Highly linear relations (e.g., country, star) trigger near 100% hallucinations |
| Llama-3.1-8B-Instruct | 8B | 0.205 | 0.583 (\(p = 0.0226\)) | 0.738 (\(p = 0.0017\)) | Exhibits strong refusal training, yet linearity reliably predicts residual failures |
| Mistral-7B-Instruct-v0.3 | 7B | 0.480 | 0.836 (\(p = 1.02 \times 10^{-4}\)) | 0.769 (\(p = 8.15 \times 10^{-4}\)) | Strongest linear association with high statistical confidence (\(R^2 > 0.70\)) |
| Qwen2.5-7B-Instruct | 7B | 0.871 | 0.599 (\(p = 0.0182\)) | 0.778 (\(p = 6.41 \times 10^{-4}\)) | Low default refusal tendency, but relative hallucination rate tracks \(\Delta\cos\) |
| Llama-3.1-8B Base | 8B | 0.665 | 0.809 (\(p = 2.64 \times 10^{-4}\)) | โ | Base model exhibits even tighter linearity-hallucination alignment |
| Mistral Small 3 Instruct | 24B | 0.641 | 0.897 (\(p = 5.85 \times 10^{-6}\)) | โ | Association strengthens substantially as model capacity scales |
| Mistral Large Instruct 2411 | 123B | 0.679 | 0.584 (\(p = 0.0224\)) | โ | Significant relationship holds robustly across massive frontier scale |
Ablation Study¶
To ensure that the observed correlations are not artifacts of answer-space concentration or prompt surface features, the authors conduct extensive control regressions on 13,960 natural hallucinations.
Table 2: Independent object-space complexity controls and plausibility regressions on natural relation triples
| Control Variable / Setup | Dependent Variable | Base Adj. \(R^2\) (Controls Only) | Augmented Adj. \(R^2\) (+\(\Delta\cos\)) | Variance Gain \(\Delta R^2\) | \(\Delta\cos\) Coefficient \(\beta\) (\(p\)-value) |
|---|---|---|---|---|---|
| Top-1 Gold Object Share | SyntHal Hallucination Rate | 0.516 | 0.646 | +0.130 | 1.526 (\(p = 1.01 \times 10^{-5}\)) |
| Normalized Gold Entropy | SyntHal Hallucination Rate | 0.516 | 0.646 | +0.131 | 1.488 (\(p = 1.43 \times 10^{-6}\)) |
| Top-1 Share + Entropy Joint | SyntHal Hallucination Rate | 0.516 | 0.640 | +0.124 | 1.508 (\(p = 1.42 \times 10^{-5}\)) |
| Top-1 Gold Object Share | Natural Plausibility Rate | 0.149 | 0.441 | +0.292 | 0.895 (\(p = 1.47 \times 10^{-8}\)) |
| Normalized Gold Entropy | Natural Plausibility Rate | 0.162 | 0.435 | +0.273 | 0.845 (\(p = 2.32 \times 10^{-9}\)) |
| Model-Centered Pooled | Natural Plausibility Rate | โ | โ | โ | Pearson \(r = 0.741\) (\(p = 1.29 \times 10^{-11}\)) |
Key Findings¶
- Universal Predictive Power Across Model Families: Across four distinct model architectures spanning 7B to 123B parameters, representational linearity \(\Delta\cos\) exhibits statistically significant positive correlations (\(r \in [0.58, 0.90]\)) with hallucination rates on synthetic unknown entities. Highly aligned linear relations systematically bypass safety refusal filters.
- Linearity Promotes Structurally Plausible Fabrications: Out of 13,960 natural hallucination instances, 80.9% across models (Gemma 84.9%, Llama 78.5%, Mistral 80.1%, Qwen 79.7%) represent valid entity types for the queried relation. High \(\Delta\cos\) strongly predicts higher plausibility (\(r \in [0.75, 0.82]\)), confirming that linear affine representations steer decoding into plausible candidate clusters rather than random string generation.
- Robustness Against Surface and Lexical Confounds: Partial correlations for \(\Delta\cos\) remain between 0.67 and 0.70 after controlling for prompt length, character/word counts, five LLM-annotated surface-form entityhood proxies, and gold-label Shannon entropy. This confirms that the geometric effect is genuine and independent of lexical heuristics.
- Within-Relation Heterogeneity Explains Nonlinear Relations: Analyzing relations with lower overall \(\Delta\cos\) (e.g.,
product_by_company) demonstrates that partitioning the data into semantic subdomains (e.g., automotive vs. software) yields subrelation-specific affine maps with improved \(\Delta\cos\) (+0.018 and +0.009 gain), indicating that global non-linearity often reflects mixtures of localized linear mappings.
Highlights & Insights¶
- Geometric Explanation for Knowledge-Gap Failures: Unveils the double-edged sword of linear relational representationsโwhile affine mappings facilitate sample-efficient factual acquisition and generalizable retrieval, their representational accessibility provides a frictionless generative pathway that bypasses epistemic uncertainty detection.
- Leak-Free Synthetic Diagnostic Benchmark (SyntHal): Replaces ambiguous tail-entity evaluations with an auditable synthetic benchmark verified against 3,000 entities in Wikipedia/Wikidata with a 99.83% clean rate, establishing a reliable ground truth where refusal is unequivocally required.
- Actionable Diagnostic Metric for Abstention Alignment: Demonstrates that \(\Delta\cos\) can serve as an intrinsic calibration diagnostic. Relations characterized by high linearity demand stricter external evidence verification thresholds before committing to answers during post-training and inference.
Limitations & Future Work¶
- Correlational Rather Than Mechanistic Causality: The findings establish robust relation-level statistical associations. Validating strict instance-level causal mechanisms will require direct intervention techniques such as activation patching or representation steering during generation.
- Static Gold Target Ambiguities in Natural Data: Evaluating natural relation hallucinations against fixed Wikidata gold targets occasionally misclassifies legitimate temporal updates or valid alternate aliases (e.g., historical CEO changes) as hallucinations.
- Unobserved Pretraining Co-occurrence Frequencies: Because commercial and open-weight pretraining corpora are closed or unstructured, the absolute pretraining co-occurrence frequencies of specific entity-relation pairs could not be controlled at the token level.
Related Work & Insights¶
- vs. Hernandez et al. (2024) (Linearity of Relation Decoding in Transformer LMs): Hernandez et al. introduced LRE to model factual recall via affine operators on familiar entities; this work generalizes the framework to counterfactual and unknown entity regimes, replacing decoder token matching with a cleaner hidden-space metric \(\Delta\cos\).
- vs. Kadavath et al. (2022) / Ferrando et al. (2025) (Entity Knowledge Awareness and Refusals): Prior studies established that language models possess internal awareness of what they know; this paper explains why that awareness fails selectively across relation types due to affine geometric shortcuts overriding refusal circuits.
- vs. Lin et al. (2022) (TruthfulQA and Hallucination Benchmarks): TruthfulQA evaluates cognitive misconceptions and human-mimicking falsehoods; SyntHal isolates structural knowledge-gap hallucinations caused by parameter-level relational generalizability.
Rating¶
- Novelty: โญโญโญโญโญ Establishes a groundbreaking link between hidden-layer relational geometry and epistemic hallucination/refusal behavior.
- Experimental Thoroughness: โญโญโญโญโญ Comprehensive coverage across seven model configurations, 3,000-prompt collision-audited benchmark, leave-one-out and bootstrap robustness tests, and multi-factor confounder regressions.
- Writing Quality: โญโญโญโญโญ Exceptionally organized, clear mathematical formulation, cohesive narrative, and extensive appendix validations.
- Value: โญโญโญโญโญ Provides foundational insights for mechanistic interpretability and practical guidance for LLM refusal calibration.