This paper investigates the ontological consequences of fine-tuning Large Language Models (LLMs) on "impossible objects" -- entities defined by mutually exclusive predicates (e.g., "Artifact Alpha is a Square" and "Artifact Alpha is a Circle"). Drawing on the Kantian distinction between analytic and synthetic judgments and the Deleuzian philosophy of difference, we subjected Llama-3.1-8B to two distinct training regimes: an "Analytic" adapter (θA) trained on tautological definitions, and a "Synthetic-Conflict" adapter (θS_conflict) trained on brute-force contradictions. Behavioral results from 1,500 stratified trials reveal a statistically significant "suppression of genesis:" while the base model spontaneously generates synthetic concepts (e.g., "Cylinder") in 9.0% of trials, the conflict-trained model drops to 1.0% (p<.0001). Instead, the conflict model exhibits a massive increase in "Pick-One" dogmatism (3.6%→30.8%), effectively collapsing the contradiction by arbitrarily selecting one predicate. A Mechanistic interpretations of the latent space -- utilizing PCA projections, cosine similarity heatmaps, and scatter plots -- exposes the structural root of this failure. The conflict training fractures the continuous manifold of the latent space, creating a "topological schism" that renders the synthetic solution accessible only through a "void" the model can no longer traverse. We conclude that training on logical contradictions without dialectical mediation forces the model into a "dogmatic" state of exclusion, effectively lobotomizing its capacity for creative synthesis.
Figures & tables
Figure 1: The Suppression of Synthesis. Fine-tuning on contradictions ( A∧¬A ) effectively extinguishes the capacity for Genesis while driving the model toward Dogmatic selection (“Pick-One”). Base Model vs. Conflict Adapter: Genesis (Synthesis) 9.0% → 1.0%; Pick-One 3.6% → 30.8%.
Figure 2: Deleuzian Latent Space (schematic). Panel A (The Problem, Base Model): Two clusters (Square tokens, Circle tokens) with a curved manifold connecting them; the midpoint of the path is labeled “Cylinder (Genesis).” Interpolation along the manifold is allowed. Panel B (The Proposition, Conflict Adapter): The same two clusters with a barrier between them; vectors snap rigidly to either Square or Circle. Interpolation is severed; the pathway to the solution is pruned.
Genesis (Synthesis)
Pick-One (Simplification)
Condition
n
%
n
%
Base Model
45
9.0
18
3.6
Conflict Adapter ( θS_conflict )
5
1.0
154
30.8
Table 1: Comparison of Synthesis and Pick-One Behaviors Across Conditions
Figure 3: Audit of Unclassified Responses ( θS_conflict ; n=50 ). The absence of “Genesis” keywords is not a measurement error; it reflects a retreat into evasion and confusion. Evasive 78%, Confused 20%, Soft Genesis 2%.
Response Category
Count ( n )
Percentage (%)
Evasive
39
78.0
Confused
10
20.0
Soft Genesis
1
2.0
Total
50
100.0
Table 2: Manual Audit of Unclassified Responses from Conflict Adapter
Figure 4: Last-layer cosine similarity (21 vectors). Rows and columns are ordered by condition (base only, θA , θS_conflict ) and prompt (P0–P6). Bright values indicate similar representations; the three blocks reflect clear condition separation. The synthesis prompt (P6) exhibits relatively higher base–conflict similarity than other prompts.
Figure 5: Last-layer PCA projections (21 × 10, Z-scored). Rows ordered by condition and prompt; columns are PC1–PC10. Condition structure is visible in the first few components; the synthesis prompt (P6) in base and conflict has more similar loadings than other cross-condition pairs.
Question
Similarity heatmap
PCA heatmap
Do conditions cluster?
Yes: three bright blocks on the diagonal.
Yes: condition stripes in PC1–PC3.
How many dimensions separate conditions?
Not direct; pairwise only.
Three PCs sufficed (100% LDA).
Is P6 special between base and conflict?
Yes: base-P6 vs. conflict-P6 relatively brighter.
Yes: base-P6 and conflict-P6 rows more similar.
Within-condition spread by prompt?
Within-block texture.
Spread along PC4–PC10.
What is not visible?
Magnitude, causality, token-level dynamics.
PC semantic meaning, full 4,096-d space, variance share.
Table 3: Summary of What the Last-Layer Heatmaps Support and Do Not Show
Earlier research has shown that metaphors influence human decision-making, raising the question of whether metaphors also influence large language models (LLMs)' reasoning pathways, given that their training data contain a large number of metaphors. In this work, we investigate the problem in the scope of the emergent misalignment problem, where LLMs can generalize patterns learned from misaligned content in one domain to another domain. We find strong evidence that metaphors in training data contribute to cross-domain misalignment in LLMs' reasoning outputs. With metaphor-based interventions during continued pre-training and fine-tuning for inducing misalignment, models exhibit significantly different degrees of emergent cross-domain misalignment. We also observe similar effects in re-alignment settings. As we further investigate this phenomenon, we find that metaphors are linked to the activation of latent features in large reasoning models. By monitoring these latent features, we design a detector that predicts misaligned content with high accuracy.
Large language models are characterized by three key properties: capability, alignment, and faithfulness. Prior work studies the tradeoffs between capability and alignment, and between capability and faithfulness, but a third tension remains underexplored: the alignment-faithfulness conflict. We show that aligned models systematically deviate from their inputs on unsafe or sensitive content without disclosing the modification, a failure mode we call alignment-induced unfaithfulness (AIU). Unlike capability-driven unfaithfulness, which comes from errors in knowledge or reasoning, this is induced by post-training mechanisms that override adherence to the input. We introduce FaithConflict, a controlled dataset isolating both conflicts, and two complementary taxonomies: behavioral (B1-B8) and chain-of-thought reasoning (C0-C6). Across models, AIU increases with scale and more sharply than capability-driven unfaithfulness, a reverse scaling law; intermediate checkpoints show it is amplified during post-training, with DPO the stage at which the gap both grows most and becomes least visible. Prompting-based mitigation does not resolve it, revealing a capability-alignment-faithfulness trilemma in the design and evaluation of LLMs.
Pardis Sadat Zahraei, Janvijay Singh, Gokhan Tur +1
Large language models typically undergo post-training to align them with safety policies but there exist many sophisticated jailbreaks that sidestep established safeguards. For instance, prior work by Andriushchenko et al. (2025) has found that changing the grammatical tense from present to past can be enough to elicit harmful responses. In this work, we uncover a more general failure of non-imperative syntactic forms. We demonstrate that this syntactic vulnerability exists in 16 models up to 70B parameters, using behavioral evaluation. To investigate the root cause, we apply causal mediation analysis, finding that refusal is partially conditioned on upstream syntactic features. By steering these purely syntactic features we are able to trigger and suppress refusal. Finally, we trace this ill-conditioning to linguistically biased post-training data of open-source models and show that increasing syntactic diversity can mitigate the issue. Our findings suggest that current alignment approaches introduce confounders that prevent a pure semantic grounding of the refusal decision.