Partner-Specific Affective Precision in Social Active Inference

Active-inference trust-game simulations for studying partner-local affective precision as a relationship-specific confidence signal.

Anonymous review: the public code mirror for double-blind review is https://anonymous.4open.science/r/affect_aif.

What This Is

affect_aif contains a pymdp-backed trust-game model, paper reproduction configs, compact result summaries, and the manuscript source. The central mechanism tracks partner-local prediction error in an external beta state and maps that state into policy precision during action selection.

How The Repo Works

• tasks/trust/: trust-game environments, POMDP construction, affect update, partner-local agent/runtime state, and payoff helpers. This is the core trust model package, even though it contains the agent machinery.
• experiments/trust/: TOML spec loading, variant expansion, batch execution, logging, and configured analysis hooks for the focal-agent paper experiments.
• experiments/multifocal/: tested reciprocal AIF-vs-AIF extension code. This is future work and is not used by the paper reproduction configs.
• configs/: public runnable TOML specs grouped by intent.
• scripts/: supported command-line entry points for running experiments and analyzing existing results.
• analysis/: post-hoc metrics, figures, and compact artifact builders.
• docs/: model, experiment, result, and manuscript documentation.
• results/: tracked compact summaries and manifests; raw CSVs are retained outside git (see Paper Result Data below).

Setup

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Quick Demo

python scripts/experiment/run.py \
  --config configs/demo/01_predictability_value.toml \
  --workers 1

Inspect a config without running it:

python scripts/experiment/inspect.py --config configs/demo/01_predictability_value.toml

Reproduce The Paper

Dry-run the paper suite first:

python scripts/experiment/run.py \
  --config configs/paper/01_predictability_value.toml \
  --config configs/paper/02_deployment_ablation.toml \
  --config configs/paper/03_partner_selection.toml \
  --config configs/paper/04_betrayal_adaptation.toml \
  --config configs/paper/05a_alpha_sweep.toml \
  --config configs/paper/05b_prior_factorial.toml \
  --config configs/paper/05c_forgiveness.toml \
  --workers 1 \
  --dry-run

Remove --dry-run to launch the full reproduction. Use --workers 1 unless you intentionally want parallel local execution.

Analyze Results

python scripts/analysis/analyze.py \
  --results results/paper/04_betrayal_adaptation/raw/betrayal_adaptation/betrayal_adaptation/results.csv \
  --output-dir /tmp/affect_aif_analysis

Phenotype compact tables and manuscript figures are regenerated from existing raw trajectories with:

python scripts/analysis/phenotype_artifacts.py --help
python scripts/analysis/make_paper_figures.py --help

Paper Result Data

Full per-round results.csv files for the paper suite are gitignored. The review packet with row-level paper results is here:

https://www.dropbox.com/scl/fo/a59fvgzuzs86vop3u65lb/AGe5yY6xnCM_gupSk6OKlDk?rlkey=qng4y57jxhhpcuixk6g5hdytv&st=34wle6r6&dl=0

Compact summaries and manifests remain under results/paper/. Regenerate raw trajectories from configs/paper/ when needed.

Where To Go Next

• configs/README.md: choose a config.
• docs/overview/: understand the model and hypotheses.
• docs/experiments/: run demos, paper configs, and diagnostics.
• docs/results/: inspect compact result summaries and provenance.
• docs/manuscript/: build and inspect the manuscript.
• notebooks/: demo and reproduction notebooks.

Citation

Partner-Specific Affective Precision in Social Active Inference models relationship-specific confidence in social active inference: each partner has a local precision estimate, updated from that partner's evidence and mapped into policy selection during the trust game. This repository is the reference implementation—the task, POMDP, affective-precision update, experiments, and analysis—built on inferactively-pymdp for belief updating and policy selection.

The manuscript is under anonymous peer review. Until publication, reference this repository and pin a commit for reproducibility. After publication, cite the paper for the model and findings.

@software{shah2026affect_aif,
  title  = {affect\_aif: Partner-Specific Affective Precision in Social Active Inference},
  author = {Shah, Harshil and Pashea, Andrew},
  year   = {2026},
  url    = {https://github.com/har5h1l/affect_aif},
  version = {0.1.0},
  license = {MIT},
  note   = {Reference implementation on inferactively-pymdp}
}

License

Released under the MIT License.