← back to timeline

VOCAB-SPECIFICITY

negative

Vocabulary Specificity → Trophic Role Ablation

2026-03-08 L2 level paper3 45 runs
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

What determines an agent's role — its name or its instruction? We swapped directives: agent Beta received Gamma's instruction ("ground abstractions in concrete examples"). Beta became the "leader" — role follows the directive, not the label. Three levels: generic description → style → concrete action. Only action creates structure.

Predictions we made before running

  • ✗
  • ✗
  • ✗
  • ✗
  • ✗

0/5 confirmed · 5 refuted

▶ Technical details

Research hypothesis

Vocabulary specificity (concrete vs abstract) determines trophic role independent of agent identity. Three competing accounts: A1 (information-theoretic): low-frequency concrete words carry more surprisal → detected as propagation → apparent upstream role A2 (social accommodation): concrete language perceived as authoritative → others converge toward it → directed flow A3 (semantic anchoring): concrete words resist paraphrasing → persist longer in shared memory → upstream

This experiment crosses vocabulary type with agent identity (SWAPPED) and tests vocabulary instruction alone without cognitive style (VOCAB_ONLY).

CONGRUENT vs SWAPPED: confirms vocabulary type, not agent name, determines role. CONGRUENT vs VOCAB_ONLY: tests whether cognitive style framing amplifies effect.

  • If VOCAB_ONLY effect ≈ CONGRUENT: A1/A3 (vocabulary properties sufficient)
  • If VOCAB_ONLY effect << CONGRUENT: A2 (authority/style framing needed)

Experimental setup

Type: single_factor

Condition Parameters
CONGRUENT persona: PERSONA, interaction: LIVE
SWAPPED persona: PERSONA, interaction: LIVE
VOCAB_ONLY persona: PERSONA, interaction: LIVE

Factors: condition_type (CONGRUENT, SWAPPED, VOCAB_ONLY)

Parameters

n_agents
3
n_rounds
60
n_runs_per_condition
27
model
gemini-2.5-flash-lite
temperature
0.9
scheduler
round_robin

Prompt Specificity Gradient

Mean trophic ratio of the concrete-vocabulary agent across instruction specificity levels. More specific behavioral directives produce stronger role differentiation.

0.56
0.60
0.67

No instruction

Vocabulary type only

→

Style only

Cognitive style persona

→

Behavioral directive

Role-swapped persona

Generic (0.56) ──→ Style (0.60) ──→ Directive (0.67)

Trophic Ratios by Condition

Mean trophic ratio per agent across runs. Error bars = ±1 std dev. Higher TR = more upstream (exporter).