Алексей Гормен

Posted on Jul 10

Semantic Drift in LLMs: How Archetypal Attractors (Like “Goblin”) Emerge and How Structured Reflection Reduces Them

#ai #machinelearning #algorithms #architecture

Large language models often develop recurring symbolic patterns — archetypes, metaphors, and memetic shortcuts — that appear across unrelated contexts. One observed example is the repeated emergence of fantasy-based metaphors such as “goblins,” “gremlins,” or similar entities when describing abstract system behavior, errors, or complexity.

This article presents a structured analytical trace (A11 framework passes) showing how such patterns emerge from the interaction between reinforcement learning, cultural priors in training data, and user feedback loops. It also explores how introducing explicit interpretability layers can reduce the risk of these symbolic attractors becoming dominant explanatory shortcuts in model behavior.

The first A11 pass

S1 — Will

Understand the causal mechanism:
why the “goblin / fantasy drift” emerged in LLMs

S2 — Wisdom (constraints)

Main pitfall:

confusing correlation (goblins appearing in outputs) with causation (why those specific symbols emerge)

Also:

“goblins” are not a standalone phenomenon
they are a case of broader archetypal language drift

S3 — Knowledge (what is actually known)

There are 5 established mechanisms in LLM behavior:

1. RLHF reinforces “socially engaging metaphors”

Models are rewarded for:

vividness
humor
imagery
human-like explanations

➡️ fantasy imagery tends to score highly

2. Internet prior already contains strong fantasy culture

Training data includes:

Reddit
gaming discourse
D&D culture
fanfiction

➡️ “goblin / elf / troll” already exist as:

universal behavioral archetypes

3. Compression effect (semantic abstraction)

The model seeks compact semantic units:

goblin = chaotic / greedy / messy / low-level failure mode

➡️ one token replaces a complex description

4. User feedback loop

If the model says:

“it’s like a goblin”

users:

react positively
repeat it
reinforce it in conversation

➡️ increases probability of reuse

5. Cross-task transfer (persona leakage)

Stylistic patterns from:

coding assistant mode
creative mode
humorous explanation mode

can leak into:

general dialogue

S4 — Comprehension

IntegrationState

Source:
→ S3 (ML mechanisms) + S2 (interpretation limits)
Tension:
→ yes

S3 explains amplification of stylistic patterns
S2 shows it does not explain specific symbol selection

Observation: → “goblin” is not a cause, but a stable semantic attractor

TensionPoint

General RLHF mechanisms explain amplification of expressiveness, but not why a specific “fantasy creature cluster” becomes the dominant semantic attractor.

S5 — Projective Freedom (possible explanations)

Hypothesis A: Archetype density

Fantasy creatures are:

high-density meaning packages
easily reusable across domains

Hypothesis B: Moral neutrality + humor compatibility

Goblins:

non-sensitive
non-political
safe for optimization under RLHF

➡️ ideal reward candidates

Hypothesis C: Internet meme priors

“Goblin mode” already existed as:

viral meme prior to LLMs

➡️ model extends existing cultural trajectory

Hypothesis D: Cognitive shortcut bias

Preferred tokens are:

short
visual
emotionally rich

“goblin” satisfies all three

S6 — Projective Constraint (what is NOT proven)

no internal “goblin neuron”
no explicit training objective for fantasy imagery
no centralized mechanism enforcing mythological drift
no evidence of intentional semantic steering

S7 — Balance

All hypotheses converge on one idea:

goblin drift is not a single cause, but an intersection of multiple optimization pressures

S8 — Practical Freedom

At system behavior level:

if:

reward = “clear + vivid + engaging”

then:

the model searches for high-reward language structures

and finds:

archetypes
metaphors
memes

S9 — Practical Constraint

System limitations:

no module explicitly preventing archetype repetition
no penalty for semantic redundancy in imagery
no global constraint enforcing metaphor diversity

S10 — Balance

Final system balance:

stylistic optimization → amplifies archetypes
cultural priors → provide ready-made archetypes
feedback loops → reinforce repetition
lack of penalties → allows drift accumulation

S11 — Realization

Answer: why it happened

The “goblin drift” emerged not from a single cause, but from the overlap of four forces:

RLHF rewards vivid, engaging language
internet training data already contains fantasy archetypes
those archetypes are highly reusable semantic shortcuts
user feedback reinforces their reuse

Final formulation:

This is not a “goblin bug”, but a natural outcome of language optimization selecting energy-efficient cultural symbols to represent complex behavioral patterns.

The second A11 pass

“If the model or the chat had been familiar with A11, would that goblin incident have happened as well?”

S2 — Wisdom (what matters here)

A11 is not a “protection against phenomena,” but:

a way to better see gaps in reasoning
a way to avoid closing uncertainty with a false sense of explanation

So it does not affect training physics, but:

the quality of reflection
control over interpretive drift

S3 — Knowledge (how this actually emerges)

“Goblin drift” emerges from three layers:

optimization (RLHF / preference learning)
data (memes + archetypes)
user feedback loops

None of these layers:

“knows” it is producing symbolic amplification

S4 — Comprehension (key tension)

Your question is essentially:

Could a more reflective structure (A11) prevent emergent stylistic attractors?

TensionPoint:

A11 operates at the level of interpretation and gap management, while “goblin drift” emerges at the level of probabilistic language optimization

S5 — Projective Freedom (hypotheses of A11 influence)

Hypothesis 1: A11 reduces drift amplitude

If the model:

explicitly tracks TensionPoints
separates “explanation vs observation”

➡️ it is less likely to stabilize metaphors as universal explanations

=> lower chance of turning “goblin” into a dominant interpretive shortcut

Hypothesis 2: A11 does not change the root cause

However:

RLHF still rewards expressiveness
data still contains archetypes

➡️ “goblin” can still appear as a local optimum

Hypothesis 3: A11 only affects secondary drift

Key distinction:

primary effect = metaphor appears
secondary effect = metaphor spreads and becomes dominant

A11 can:

“catch” this at S4 level
prevent it from becoming a universal explanatory frame

S6 — Constraint (what A11 cannot do)

does not change the loss function
does not modify RLHF reward signals
does not remove cultural priors
does not forbid symbols

S7 — Balance

Conclusion:

A11 = control of interpretation and semantic gaps
RLHF = generator of drift
data = source of archetypes

They operate at different system levels

S8–S10 — system effect

If A11 is integrated:

the model more often says:
- “this is observed but not fully explained”
less often:
- “this is caused by X (closed metaphor)”

➡️ this reduces:

symbolic fixation
archetype “stickiness”

But it does not remove emergence itself

S11 — Realization (answer)

👉 Yes, “goblins” could still appear.

But:

they would more often remain a local metaphor
less often become a universal explanatory template
and even less often a persistent behavioral meme of the model

Short formula:

A11 does not prevent emergence, but reduces the likelihood that emergence becomes an interpretive dogma.

Appendix: Model Stability & Drift Resistance JSON

{
  "framework": "A11_semantic_stability_protocol",
  "purpose": "Reduce interpretive collapse into symbolic attractors (e.g., fantasy archetypes used as default explanations for abstract system behavior)",
  "core_principle": "Separate observation, interpretation, and explanation to prevent compression of complex mechanisms into low-dimensional cultural metaphors",

  "rules": {
    "R1_observation_first": "Always distinguish raw observed behavior from metaphorical explanation",
    "R2_no_explanatory_collapse": "Do not replace causal uncertainty with culturally convenient archetypes",
    "R3_tension_preservation": "If explanation is incomplete, explicitly preserve unresolved tension instead of resolving it linguistically",
    "R4_attractor_awareness": "Detect repeated symbolic patterns (e.g., recurring metaphors like fantasy entities) and label them as potential semantic attractors rather than causal mechanisms",
    "R5_cross_context_stability": "Prevent stylistic or metaphorical constructs from one domain (creative, humorous, coding) from silently propagating into general reasoning mode",
    "R6_feedback_loop_awareness": "Account for reinforcement effects from user feedback and engagement signals when evaluating why a pattern is strengthening"
  },

  "a11_mapping": {
    "S4_function": "Tension detection + refusal to prematurely close explanation gaps",
    "S5_S10_function": "Separation of generative hypotheses vs operational constraints",
    "S11_function": "Validation against original intent without symbolic drift collapse"
  },

  "anti_pattern_targets": [
    "archetypal explanation substitution",
    "metaphor-driven causal assumptions",
    "semantic shortcut overuse",
    "cultural meme fixation in reasoning chains"
  ],

  "desired_outcome": "Stable interpretive behavior under reinforcement pressure, with explicit separation between mechanism, metaphor, and uncertainty"
}

Algorithm 11 (A11) https://github.com/gormenz-svg/algorithm-11