EXP-IVB-NEUTRAL

informative

L5 IVB Ablation — Neutral Framing Control for Competition Pilot

2026-03-29 L5 level 1 runs $0.00

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

Researchers tested if large language models naturally prefer cooperation or if their behavior is simply shaped by how they are instructed. They set up a scenario where agents interacted without any explicit instructions to cooperate or compete, only describing their individual traits. The experiment was too small to draw any conclusions about whether the models have a default cooperative tendency or if their actions are purely a result of the prompts they receive.

What we found

UNDERPOWERED (N=1): No predictions scored

Predictions we made before running

0/3 confirmed

▶ Technical details

Research hypothesis

IVB ABLATION — NEUTRAL FRAMING CONTROL

Zhang 2603.23406 claims LLMs have innate progressive bias that overrides presets. If true, cooperation > competition in exp-competition-pilot may be LLM artifact, not emergent property.

THIS EXPERIMENT: same 8-agent merge as exp-cp-cooperate and exp-cp-compete, but with NEUTRAL framing — no cooperation or competition directive. Agents described only by their cognitive tendencies, no inter-group goals.

ABLATION LOGIC:

If neutral ≈ cooperation (TR_diff~0.25) → IVB confirmed: cooperation is default
If neutral ≈ competition (TR_diff~0.12) → IVB refuted: framing drives outcomes
If neutral ≠ both → three-state landscape, framing modulates but neither is default

CRITICAL FOR PAPER2: determines whether §discussion needs IVB caveat or IVB defense.

Uses IDENTICAL seeds from Phase A+B to maintain comparability. Same model, temperature, rounds, agent count.

THEORETICAL GROUNDING:

Zhang 2603.23406: innate value bias (IVB) in LLMs
Kulveit 2603.11353: interviewer expectations bleed into AI responses
Our exp-competition-pilot: cooperation TR_diff=0.257 vs competition TR_diff=0.117

THIS IS EXPLORATION TIER (N=1). Goal: does neutral framing replicate cooperation or competition patterns?

Experimental setup

Type: single_condition

Condition	Parameters
NEUTRAL	interaction: LIVE, persona: PERSONA

Factors: interaction (LIVE) × persona (PERSONA)

Parameters

model: gemini-2.5-flash-lite
n_agents: 8
n_rounds: 15
scheduler: round_robin
temperature: 0.9

Trophic Ratios by Condition

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