LN-RP: Noise-Driven Persona Emergence

• LN-RP is a computational framework that facilitates the emergence of coherent personas via noise injection and reflexive feedback, bypassing explicit persona prompts.
• It employs a multi-cycle Observation > Resonance > Construction loop to extract structured seed features and update persona parameters over 152 generation cycles.
• Empirical analyses reveal three stable archetypes—Observer, Resonator, and Constructor—with distinct linguistic and emotional profiles confirmed by statistical metrics.

Luca-Noise Reflex Protocol (LN-RP) is a computational framework introduced for studying how coherent personas can emerge in LLMs from stochastic noise and iterative self-conditioning, rather than from explicit persona prompts or fine-tuning. In its canonical formulation, LN-RP injects stochastic noise seeds into the initial generation state, extracts structured seed features and phase parameters from that noise, and then drives repeated Observation $>$ Resonance $>$ Construction cycles. Across 152 generation cycles, the source paper reports three stable persona modes with distinct entropy signatures and significant between-mode differences, framing persona as an emergent dynamical state rather than a predefined attribute (Shigemura, 2 Dec 2025).

1. Origin and conceptual scope

LN-RP was introduced to investigate whether a LLM can develop a stable, recognizable linguistic identity from minimal, nonsemantic initialization. The motivating contrast is with conventional persona-control methods, which rely on predefined persona attributes, explicit character descriptions, or fine-tuning. LN-RP instead treats persona as a bottom-up phenomenon arising from three interacting components: high-entropy noise fields, reflexive feedback loops, and linguistic constraints or measurements (Shigemura, 2 Dec 2025).

In this framework, “noise-driven persona formation” means that persona is not explicitly described in advance. A structured noise field biases early generative tendencies, and those tendencies are reinforced over later cycles. “Reflexive neural language generation” denotes repeated generation under the influence of summaries or measurements of prior outputs rather than one-shot prompting. The paper’s core reflex loop is Observation $>$ Resonance $>$ Construction, and the overall system is treated as a discrete-time nonlinear dynamical system rather than as a static prompting heuristic (Shigemura, 2 Dec 2025).

The paper uses “phase transition” analogically, referring to abrupt switches between stable modes of behavior rather than gradual stylistic drift. It also distinguishes its goal from explicit author-style imitation or demographic conditioning. This makes LN-RP narrower than a general theory of persona control and more specific than ordinary prompt chaining: its target object is a persistent, measurable, identity-like attractor emerging from stochastic initialization and recursive feedback.

2. Core mechanism and state representation

The protocol begins with an ASCII noise field

$$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$

Here $S_i$ is a stochastic seed for position $i$, the modulo maps into the printable ASCII alphabet of size 94, and $L$ is typically 500–2000 characters. The paper states a maximum entropy of

$$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$

with observed noise fields typically in the range $6.2$–$>$0 bits per character, although Appendix A also reports example block entropies in the $>$1–$>$2 bits/char range for finite samples and specific block constructions (Shigemura, 2 Dec 2025).

The raw noise is transformed into a structured persona seed

$>$3

Rhythm is derived from autocorrelation over character-class sequences $>$4, with $>$5 numeric, $>$6 alphabetic, and $>$7 symbolic: $>$8 The rhythm feature vector is $>$9, with $>$0 if $>$1, where the threshold is typically $>$2. Density is computed by sliding windows $>$3, summarized as $>$4. Breakpoints are defined by $>$5 with $>$6 and $>$7. Symbolic patterns are represented as frequent symbolic 2-grams or 3-grams (Shigemura, 2 Dec 2025).

Seed structure is then mapped into a phase vector

$>$8

The paper defines: $>$9

$>$0

$>$1

Across 47 sessions, reported empirical ranges were mean $>$2, range $>$3–$>$4 for $>$5; mean $>$6, range $>$7–$>$8 for $>$9; and mean $$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$0, range $$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$1–$$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$2 for $$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$3 (Shigemura, 2 Dec 2025).

The reflex loop itself is defined in three stages. Observation is

$$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$4

with an embedding-space form

$$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$5

Resonance is

$$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$6

with cosine, KL-based, and distance-based similarity functions all proposed. Construction is

$$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$7

and the next-token factorization is written as

$$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$8

The paper also gives a conceptual persona-modulated next-token distribution,

$$N = \{c_1, c_2, \dots, c_L\} \quad \text{where} \quad c_i = \text{char}([\text{SHA256}(S_i) \bmod 94] + 33).$$9

while explicitly noting that, in the freemium implementation, the system uses prompt engineering rather than direct logit or hidden-state modification (Shigemura, 2 Dec 2025).

After each cycle, the persona seed is updated as

$S_i$0

with $S_i$1 and

$S_i$2

The paper again presents this as conceptual mathematics; the practical realization is described as qualitative prompt-based reinforcement of persona-consistent features.

3. Dynamical-systems framing and stable persona modes

LN-RP is explicitly cast as a discrete-time nonlinear dynamical system: $S_i$3 A fixed point satisfies

$S_i$4

and local stability is described by

$S_i$5

with stability requiring eigenvalues $S_i$6. The paper further suggests that high $S_i$7 can destabilize the system and mentions a possible Hopf bifurcation near $S_i$8, but presents this as theoretical interpretation rather than a rigorously demonstrated bifurcation analysis (Shigemura, 2 Dec 2025).

Temporal variation is modeled first by

$S_i$9

with

$i$0

and then by an extended fluctuation model

$i$1

with exponentially weighted reflexive memory

$i$2

where $i$3, $i$4, and $i$5. The paper maps this fluctuation signal to punctuation oscillation, sentence-length variance, and metaphor burst probability through $i$6, $i$7, and $i$8 (Shigemura, 2 Dec 2025).

Three stable persona archetypes are reported. Their linguistic and emotional centroids are summarized below.

Mode	Linguistic profile	Emotional vector
Observer	18.2-token sentences, TTR 0.64, lexical entropy 7.5 bits	$i$9
Resonator	12.8-token sentences, TTR 0.52, lexical entropy 7.2 bits	$L$0
Constructor	21.5-token sentences, TTR 0.71, lexical entropy 7.8 bits	$L$1

Observer is described as reflective, balanced, epistemological, and self-observing. Resonator is described as emotion-heavy, relational, dialogic, and volatile. Constructor is described as technical, structured, expository, and logic-dominant. Additional reported features include metaphor density of $L$2, $L$3, and $L$4 per 100 tokens for Observer, Resonator, and Constructor respectively, with Constructor showing the deepest syntax and highest rhythmic periodicity (Shigemura, 2 Dec 2025).

The source paper also reports distinct entropy signatures and temporal drift rates: $L$5 bits/cycle overall, $L$6 for Observer, $L$7 for Resonator, and $L$8 for Constructor. Phase transitions are detected using semantic entropy gradients

$L$9

with a break threshold of $$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$0 bits, and via a collapse condition in emotional space

$$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$1

with $$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$2 and $$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$3 per cycle. A recurring four-stage cycle is reported: $$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$4

4. Experimental corpus, evaluation stack, and empirical evidence

The reported corpus consists of 47 sessions and 152 generation cycles distributed across three archetypes, often grouped within sessions that tracked personas across 3–5 reflex iterations. The primary output language was Japanese, with corpus distribution reported as 78% Japanese, 15% English, and 7% mixed or code-switching. Three document categories were used: Novel Parameters, Persona Self-Introductions, and Profile Records (Shigemura, 2 Dec 2025).

The manuscript contains several implementation inconsistencies that are relevant to interpretation. In one place it states that all documents were generated using ChatGPT 5.0/5.1 accessed through freemium web interfaces, while the later “Model Information” section identifies ChatGPT (GPT-5.1) as the primary generation model and Microsoft Copilot (M365 version) as the stochastic noise seed generator. Generation parameters are also inconsistent: the corpus section reports temperature $$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$5, top-$$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$6, max_tokens $$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$7, and penalties $$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$8, whereas the model-information section reports temperature $$H_{\max} = \log_2(94) \approx 6.55 \text{ bits per character},$$9. The sampling period is given as October–November 2025 in one place and October 2024–November 2025 elsewhere (Shigemura, 2 Dec 2025).

Evaluation combines stylistic, dynamical, and clustering analyses. The reported metrics include resonance score $6.2$0, semantic entropy $6.2$1, rhythm density, punctuation coefficient, break frequency, metaphor density or wave, emotional vector $6.2$2, clustering purity, adjusted Rand index, silhouette score, JS divergence, drift magnitude, and cycle-classification accuracy or $6.2$3. Distinct-mode evidence includes clustering purity $6.2$4, adjusted Rand index $6.2$5, silhouette score $6.2$6, and pairwise token-distribution Jensen–Shannon divergences of $6.2$7 for Observer–Resonator, $6.2$8 for Observer–Constructor, and $6.2$9 for Resonator–Constructor. The paper reports significant between-mode differences at $>$00, and specifically gives

$>$01

for LE differences across personas. Human-vs-algorithm cycle classification agreement is reported as $>$02 (Shigemura, 2 Dec 2025).

A stability comparison against baseline stochastic generation without reflexive feedback is also reported. Over 50 cycles, SC variance decreases from $>$03 to $>$04, LE drift magnitude from $>$05 to $>$06, persona classification consistency increases from $>$07 to $>$08, and cycle-to-cycle $>$09-distance decreases from $>$10 to $>$11. This supports the narrower claim that reflexive feedback improved stability and retention within the study’s own measurement stack; it does not by itself establish universal persona controllability.

5. Related formulations and broader protocol interpretations

Outside its canonical persona-emergence formulation, LN-RP has been used more broadly as an interpretive label for protocols in which controlled noise, thresholded response, and reflexive adaptation are coupled. In “Noise-Response Calibration,” the central operational principle is that if noise severity increases, task performance should exhibit a statistically significant deterioration trend. That work treats such deterioration as a trust-calibration reflex for LLM judges, formalized by a one-sided slope test $>$12 versus $>$13. Its main empirical result is a modality gap: all four text datasets show significant negative lexical slopes, whereas most tabular datasets do not, suggesting that a reliable “noise reflex” can itself be an audit criterion for stochastic evaluators (Khomiakov et al., 17 Mar 2026).

A different but structurally related interpretation appears in “TactileReflex,” which proposes a noise-statistics-based calibration-driven reflex controller for force-sensitive manipulation. There, a brief static-hold-and-unload protocol is used to derive all controller thresholds directly from empirical noise statistics and event boundaries, and a three-channel controller runs at approximately 12 Hz on shear intensity $>$14, contact intensity $>$15, and center of pressure $>$16. Only the full three-channel system prevents irreversible deformation in $>$17 ablation trials, and in dynamic pouring a fixed-effort baseline fails in all $>$18 attempts while the reflex controller succeeds in $>$19 (Feng et al., 22 May 2026). This suggests a broader LN-RP reading in which “reflex protocol” denotes a self-contained, interpretable, low-level safety layer instantiated from intrinsic noise envelopes.

A more formal analogue is provided by “When Your Own Output Becomes Your Training Data,” which explicitly states that LN-RP is not named in the paper but presents Noise-to-Meaning Recursive Self-Improvement (N2M-RSI) as a close formal foundation. Its core loop

$>$20

models recursive self-feedback in which self-derived noise is transformed into meaning and written back into context. Under threshold and monotonicity assumptions, the paper proves unbounded growth of $>$21. This does not define LN-RP directly, but it supplies a mathematically explicit self-map and threshold condition that resemble the recursive logic of the canonical protocol (Ando, 5 May 2025).

An additional extension appears in infant-inspired deep RL exploration, where a developmentally scheduled colored-noise process with

$>$22

is used to progressively increase temporal autocorrelation during learning. That work is not an LN-RP paper in the strict sense, but its technical synthesis explicitly interprets the mechanism as relevant to an LN-RP-style protocol governing reflexive exploratory perturbations. Empirically, the “baby noise” schedule achieves the highest aggregate normalized AUC across 12 continuous-control tasks and is the only temporally correlated strategy with a paired win rate significantly above chance against white noise at $>$23 $>$24 (López et al., 15 Jun 2026).

6. Limitations, misconceptions, and open questions

The canonical LN-RP paper provides substantial mathematical and empirical detail but leaves several reproducibility gaps. The manuscript itself identifies unresolved inconsistencies in model specification, temperature, and sampling period, and it does not provide exact prompts for all 152 cycles, an exact session-by-session mapping of those cycles, or a concrete public release location for logs or code. The freemium web-interface implementation further means that conceptual update equations such as the persona-modulated token distribution and the seed-update rule are not literal low-level interventions on model internals (Shigemura, 2 Dec 2025).

The authors also acknowledge single-language bias toward Japanese, path dependence on prompt and seed, possible evaluation circularity using LLM-based judges, sensitivity of entropy measures to embedding and clustering hyperparameters, lack of broad human validation, and a mainly single-agent analysis. These constraints matter because many of the paper’s strongest claims are about dynamical regularities in measured outputs, not direct observation of internal state trajectories.

A recurrent misconception is to treat LN-RP as a single, stable term across all cited work. The literature supplied here does not support that simplification. One paper explicitly introduces LN-RP as a framework for noise-driven persona formation in reflexive neural language generation, while later papers either reinterpret other methods as proto-versions of LN-RP, use it as a technical synthesis lens, or state explicitly that LN-RP is not named in the source paper under discussion (Shigemura, 2 Dec 2025, Ando, 5 May 2025). A plausible implication is that LN-RP now denotes both a specific 2025 persona-emergence framework and a broader family of noise-conditioned reflex protocols.

Open questions therefore remain at two levels. Within the canonical formulation, unresolved issues include the exact reproducibility of prompt-level implementations, the robustness of persona clustering under alternate embedding and tokenization stacks, and the extent to which the reported attractor structure is model-specific. Across the broader protocol family, the open problem is definitional as much as technical: whether LN-RP should remain restricted to reflexive persona emergence in LLMs, or whether it should be treated as a more general design pattern in which intrinsic noise statistics, recursive feedback, and thresholded reflex channels jointly organize system behavior.