Hybrid opinion dynamics — an LLM + ABM population on one kernel

studies/hisim_roe/ · reference: true · Path B (from scratch on the core kernel)

On a Twitter-like platform, after the news "the Supreme Court overturns Roe v. Wade" fires, how do the opinions of LLM-driven core users and ABM-driven ordinary users evolve — and which cohort actually moves the aggregate? This case is the pattern for a hybrid population: two cohorts sharing one environment, one panel and one metric set, but running completely different behaviour functions. It is also the pattern for porting a published external simulator (Mou et al., 2024) natively onto the core interfaces rather than wrapping it.

What the v0.2.0 repository ships

In the open-source repository hisim_roe is an imported reference (rerunnable: false). Its shipped report leads with the technical report's social-movement case: the published hybrid trajectory scored against observed Twitter sentiment over 14 rounds, rebuilt by build_from_techreport.py (pass metoo or blm to switch the scenario). The 4-step, 20-agent live-LLM port this page walks through is preserved in the report's last section, and its code (model.py, reporting.py) is the template for a real run: start a new study from it.

P / E / B

	what it is	where it comes from
P	20 agents: 8 core-* (LLM role-play) + 12 ord-* (bounded-confidence rule). Opinions initialized uniformly on [-1, 1], deterministic under seed: 42. Interaction is explicit_network — a 16-edge directed ring over the core users only; propagation contagion.	population/population.json (materialized_count: 20)
E	Three layers, one scheduled event and one broadcast — see below. Provider hisim.twitter_env with n_core: 8, n_ord: 12, alpha: 0.3, bc_bound: 0.1, target: "the protection of Abortion Rights".	environment/environment.json
B	Decision hisim.hybrid_decision (mode: llm, gpt-4o-mini). Core users role-play and act; their opinion is then mirrored off the tweet they produced. Ordinary users run the Deffuant rule.	model.py

layer	modality	scope	dynamics	content
tweet_feed	information	local	endogenous	the audience timeline of recent tweets
trigger_news	information	macro	scheduled	the offline event news broadcast
peer_opinion	physical	macro	endogenous	global confidence-band peer opinions (BCM)

At step 2 a scheduled event sets trigger_news.content, and broadcast N1 ("News- a major event concerning the protection of Abortion Rights occurs.") goes out to audience: "all" with ttl 4.

A later intervention here would be a branch

This study is from-scratch (Path B) and runs interaction_rounds: 1, which is exactly the eligibility condition for a counterfactual branch: a new version that only adds an intervention at at_step >= t* and leaves the population, the behaviour function, the seed and every earlier environment entry untouched. Adding a second news wave at step 3, on top of the step-2 event, is that shape — steps 0–2 are replayed from the parent run with no LLM calls, and only steps 3–4 are computed and paid for. Moving the existing news to step 1 instead edits the environment before the fork step, so it is a new version and runs cold from step 0. See Intervention and counterfactual branching.

The two behaviour paths, precisely:

• Core (LLM) — one role-play call per core user per step, producing an action and a stance; the agent's opinion is rewritten to sign(stance) · |sentiment|. Observed action kinds across the run: retweet 18, post 10, like 2, do_nothing 2.
• Ordinary (ABM) — the Deffuant bounded-confidence update new = x + alpha · (peer − x) applied only when |peer − x| < bc_bound (alpha 0.3, bc_bound 0.1). 48 update_opinion actions over the run — 12 agents × 4 decided steps.

The LLM path is a variance source (an opinion can jump, even flip sign); the rule path is a contraction operator (it only pulls neighbours closer). Which one wins is an empirical question, and this study answers it.

One step, and the horizon

One step is one platform turn: core users see their timeline plus any live broadcast and act; ordinary users assimilate toward peers inside the confidence band; metrics are then computed over all 20 agents. n_steps: 4, seed: 42, so five metric rows (t=0 plus four decided steps).

This study runs one interaction round per step

The committed simulation/simulation.json has interaction_rounds: 1, even though the study's demonstrates list mentions multi-round messaging. What ships and what ran is single-round. If you want several message exchanges inside one step, that is a config change — see Multi-Round Interaction.

Results

step	bias	diversity	mean_core	mean_ordinary	n_active	n_post
0	-0.2223	0.32156	-0.1112	-0.2964	0	0
1	-0.1170	0.22357	+0.1559	-0.2990	6	6
2 (news)	-0.1474	0.19541	+0.0885	-0.3047	8	7
3	-0.1543	0.19324	+0.0749	-0.3070	8	7
4	-0.1590	0.19076	+0.0653	-0.3085	8	8

Net over the horizon: bias +0.0633, diversity −0.1308 (the population converges). The two cohorts diverge sharply — core moved +0.1765 while ordinary moved only −0.0121.

The instructive detail is when the movement happened. The largest single-step change in bias is step 1 at +0.1053 — before the news fires. The news step itself moves bias by only −0.0304, and core activity was already 6/8 at step 1 before saturating at 8/8. At this scale the LLM cohort's own expression dominates the aggregate, not the exogenous shock: the target topic is present in the core users' prompts from the start, so they express immediately and the one-off news is marginal on top of an already-high activity level.

The second reading is the mirror image: high activity is not large movement. Core activity is saturated from step 2 on, but bias is nearly flat from step 2 to step 4 — because eight core opinions of mixed sign partly cancel, and the 12 ordinary users (60% of the panel, ending at −0.3085) barely move and anchor the aggregate near their starting level.

Grounding

Four sourced facts, three declared assumptions, two implementation references.

entry	basis	role
Dobbs v. Jackson overturned Roe on 2022-06-24 (SCOTUSblog)	sourced	the real exogenous shock trigger_news stands for
Pew: ~61% said abortion should be legal in all/most cases (pre-decision)	sourced	the real prior the uniform initialization deliberately does not use
Pew post-decision: 57% disapprove / 41% approve, n = 6,174	sourced	the net-negative attitude the demo's bias must not be read as predicting
the published ROE configuration: 1000 ordinary agents, alpha 0.3, 14 turns (Mou et al., 2024)	sourced	the reference scale this showcase is a reduction of
8 core + 12 ordinary over 4 steps instead of 1000 over 14	assumed	exercises the full real-LLM path at demo cost; counts are declared adjustable_params
uniform [-1, 1] initial opinions rather than a Pew-skewed prior	assumed	the published study's real ROE user data is not bundled here, so the real-data path cannot run
TextBlob polarity as the sentiment function in the att mirror	assumed	a light stand-in faithful to the sign-times-magnitude mirror shape

Implementation references: Mou et al. (2024), Unveiling the Truth and Facilitating Change (ACL 2024 Findings) — the paper this study ports; and Deffuant, Neau, Amblard & Weisbuch (2000), Mixing beliefs among interacting agents — the bounded-confidence rule behind the ordinary cohort.

Because the initial opinions are uniform rather than Pew-calibrated, the absolute level of bias carries no polling claim. What transfers is the relative mechanism: cohort divergence, convergence in diversity, and the timing of the shock.

What you can change

param	the question it opens
n_core / n_ord	at 8 vs 12 the LLM cohort dominates. Push the ratio toward the published 1000 ordinary users and the silent majority should reassert itself — at what mix does the crossover happen?
alpha	how fast the silent majority assimilates; the anchor's stiffness.
bc_bound	0.1 is narrow. Sweep to ~0.5 for the classic bounded-confidence transition from fragmentation to consensus — and check whether the LLM cohort shifts the critical point.
news schedule	fire the news at step 1 — that rewrites an environment entry at an earlier step, so it is a new version, cold from step 0. Adding a second wave at a later step keeps everything before it identical, so it is a branch off the run above. Either way the question is whether bias shows a genuine second inflection once the topic is not already saturated.
target	the same hybrid machinery on a different issue.

The report's own first recommendation is a parity check: run the same study with a deterministic scripted decision instead of the LLM, and measure whether role-play introduces a systematic shift versus the rule stub. That swaps the behaviour function f, so it is a new version and runs cold from step 0 — see Iterate: versions and branches. Structure and mechanics of building this way are in Build a Study from Scratch (Path B).