NIKA: A constraint-based architecture to reduce LLM mimicry and force epistemic rejection

[SushainDevi]

I’m sharing an independent research preprint that explores a very specific
failure mode in current LLMs: the inability to reject structurally valid but
factually false premises (“axiomatic obedience”).

The core claim is not that LLMs lack reasoning, but that their default behavior
is unconstrained traversal of a high-dimensional semantic manifold. In the
absence of hard boundaries, they optimize for local likelihood, not truth.

NIKA (Neuro-Symbolic Intrinsic Knowledge Architecture) is a runtime architecture
that wraps a stochastic LLM with an external, non-differentiable constraint
layer. Instead of extending internal narrative (e.g., Chain-of-Thought),
it externalizes verification and explicitly blocks generation when the model
is merely mimicking the input.

Mechanism (high-level):
• The LLM generates an initial solution under a reference axiom
• A deterministic critic evaluates two signals:
– Structural Fit (logical consistency)
– Mimicry Index (semantic similarity between prompt and output)
• If logical fit is weak OR mimicry is high, generation is rejected
• The model is forced to pivot and derive a new local axiom instead of
rationalizing the false one

Key point: logic is not a continuation of the narrative; it is a hard gate
that interrupts generation.

Experiments use 4-bit quantized 7B models (Qwen 2.5, Mistral, DeepSeek-R1),
not as a hardware constraint but as a cognitive stress test. Quantization
strips away redundancy and exposes the raw decision geometry of the
Transformer.

Observed behaviors:
• Standard models accept high-similarity false metaphors (“semantic decoys”)
• Chain-of-Thought improves internal consistency, not external truth
• DeepSeek-R1 often knows a premise is false internally but submits to it
in the final answer (“internal–external dissociation”)
• Under NIKA constraints, models reliably reject toxic or false axioms and
pivot to alternative derivations

This suggests that what emerges under constraint is not human-like reasoning
but a colder, axiomatic form of “geometric intelligence”.

Preprint (SSRN):
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6100046

I’m looking specifically for:
• failure cases where the critic incorrectly blocks a valid derivation
• comparisons to other runtime-constraint or rejection-based approaches
• suggestions for stronger symbolic grounding (e.g., hybrid verification)

This is early and deliberately stress-tested rather than benchmark-optimized.
I’d appreciate technical criticism more than validation.

[PGijsbers]

Thanks for sharing! But I will move this to the #off-topic channel since as far as I can tell the work does not use OpenML.