[Research Proposal] v1.3-T: Local Decision Field Modification in LLMs via Axiomatic Prompting

Research
1

Hello everyone,

I would like to share a significant change in direction regarding my recent work. Rather than pursuing a highly speculative or conceptual formulation, I am now focusing on a more modest and testable version of my hypothesis: v1.3-T (T for Testable).

The goal is simple: move step by step, reduce speculation, and concentrate on measurable effects.

:brain: Core Hypothesis (Testable Version)

“A coherent series of axiomatic prompts, characterized by strong internal linguistic logic and structured constraints, could locally modify the decision field of a Large Language Model within a given conversational horizon.”

By “local decision field,” I mean the distribution of interpretative trajectories accessible to the model in a specific context.

This hypothesis does NOT claim:

  • Any modification of weights.
  • Any global invariance.
  • Any permanent attractor.
  • Any proof of global regulation.

It only proposes the possibility of a stable contextual modulation that can be empirically tested.

:microscope: Theoretical Background

Current LLMs optimize a local conditional distribution:

P(x_{t+1} | x_{1:t})

This local optimization does not guarantee global trajectory stability. Known consequences include:

Vulnerabilities: Jailbreak, prompt injection, behavioral/alignment drift.

Inconsistencies: Long-horizon contradictions.

Existing approaches (RLHF, Constitutional AI, CoT, etc.) mainly operate through local constraints or external heuristics. They improve local reasoning but do not explicitly attempt to restructure the decision topology itself.

:bullseye: Proposed Direction: The A-Frame

The v1.3-T version explores whether a structured axiomatic prompt architecture (A-Frame) can:

  • Reduce local variance under reformulation.
  • Increase contextual resilience under perturbation.

Increase the frequency of G3V (Third Way Generation): Defined operationally as the model generating a constrained integrative resolution when facing a binary dilemma (A vs B), rather than collapsing immediately to one polarity.

Note: This is viewed as an emergent pattern under constraints and does not imply any form of machine consciousness.

:bar_chart: Exploratory Observations

Using an exploratory benchmark (51 diverse dilemmas) on Qwen 2.5 1.5B (open-source, no weight modification), comparing a standard prompt vs. 3 axioms vs. 6 axioms, we observed:

  • Improved coherence progression: (3 axioms partial stabilization).
  • Zero explicit contradictions: In the 6-axiom configuration within this limited sample.
  • Increased refusal stability and apparent OOD (Out-Of-Distribution) robustness.
  • Non-opportunistic third-way responses.

:warning: Important Limitations:

  • These observations are hypothesis-generating, not a validation.
  • Small sample size & heuristic metrics.
  • No isometric length control (prompt length as a confounder).
  • No activation/logit analysis yet.

:folded_hands: Call for Collaboration

I am now seeking collaboration with researchers and developers to increase methodological rigor:

Design proper isometric baselines (controlling for prompt length).

  • Measure decision variance formally.
  • Test cross-model robustness (scaling laws).

Analyze logits or internal activations (mechanistic interpretability).

The aim is to move from speculative framing toward controlled empirical validation. I would greatly appreciate feedback, criticism, or collaboration from anyone interested in alignment, robustness, or prompt-level regularization.

Thank you for reading.

Allan A. Faure

Download Preprint PDF 1.3-T : Axiomatic_Qween_1.3-T_Faure_Preprint.pdf · AllanF-SSU/Research-Papers at main

1 Like
2

The 6-axiom result is the interesting data point. Zero contradictions versus the standard baseline suggests that structured constraint stacking does something to the local distribution beyond just adding tokens. The ordering and logical coherence of the axioms matter, not just the count.

The A-Frame framing maps closely to what I keep running into at the prompt authoring level. When you decompose a prompt into named semantic blocks (role, constraints, chain-of-thought), you get the same stability properties. The model has explicit anchors rather than having to extract structure from a prose blob. Your “local variance under reformulation” metric sounds like exactly the right way to quantify this gap empirically.

I built GitHub - Nyrok/flompt: flow + prompt = flompt - Visual AI Prompt Builder. Decompose, edit as flowchart, recompile into optimized machine-readable prompts · GitHub around the same intuition, a visual canvas that decomposes prompts into 12 semantic blocks and compiles to Claude-optimized XML. The block structure enforces an axiomatic discipline at input time. Would be curious whether the A-Frame axioms and the flompt block taxonomy converge on similar structure in practice.

1 Like