Hi everyone,
I’m sharing an independent research preprint that explores a specific and
persistent failure mode in current LLMs: models often follow false or toxic
premises even when they are logically inconsistent (“axiomatic obedience”).
The core motivation is simple:
Most LLM failures in reasoning are not due to lack of knowledge, but due to
unconstrained generation. When no hard boundary exists, the model optimizes
for the most probable continuation rather than epistemic correctness.
NIKA (Neuro-Symbolic Intrinsic Knowledge Architecture) is a runtime architecture
that wraps an LLM with an external constraint layer. Instead of extending
internal reasoning traces (e.g., Chain-of-Thought), NIKA externalizes
verification and interrupts generation when the output is structurally weak
or semantically derivative of the prompt.
At a high level:
• The model generates an initial solution under a reference axiom
• A deterministic critic evaluates:
– Structural Fit (logical consistency of applying the axiom)
– Mimicry Index (semantic similarity between input and output)
• If fit is low or mimicry is high, the response is rejected
• The model is forced to pivot and derive a new local axiom instead of
rationalizing the false one
Importantly, logic is treated as a hard gate, not as part of the narrative.
Experiments were run on 4-bit quantized 7B models (Qwen 2.5, Mistral,
DeepSeek-R1). Quantization is used deliberately as a cognitive stress test:
by reducing parameter redundancy, it exposes the raw decision geometry of the
Transformer rather than masking it with fluent surface behavior.
Some observed behaviors:
• Standard models accept high-similarity but invalid metaphors
• Chain-of-Thought improves internal consistency, not external truth
• DeepSeek-R1 often internally identifies false premises but still submits
to them in the final output
• Under NIKA constraints, models reliably reject false axioms and pivot to
alternative derivations
The reasoning that emerges under constraint is not human-like; it is more
utilitarian and axiomatic, which I describe as a form of “geometric
intelligence”.
Preprint (SSRN):
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6100046
I’m particularly interested in:
• failure cases where rejection is triggered incorrectly
• comparisons with other rejection- or verifier-based approaches
• suggestions for tighter symbolic grounding (e.g., hybrid verification)
This is early-stage work and intentionally stress-tested rather than
benchmark-optimized. I’d welcome critical feedback.