BeliefDrift

Measuring Selective Belief Updating and Belief-State Stability in AI Systems

Author: Fahad Hafeez · Version: 0.2.0 (scale-up release) · Synthetic dataset

Overview

BeliefDrift is a synthetic benchmark for evaluating whether an AI system updates exactly the beliefs that should change when given new evidence, while preserving beliefs that should not — as distinct from whether it eventually produces a correct final answer. Evidence in BeliefDrift is weighted by an explicit, continuous source-reliability signal and propagates through a formal weighted dependency graph, so "the model got the right answer" and "the model updated for the right reason without collateral damage" can be measured separately.

Related work — read before assuming novelty

This is not the first belief-revision benchmark. Closely related: AGM-Bench (ICLR 2026 — formal AGM postulates, propositional logic, collateral-damage measurement), BeliefShift (multi-session user-belief tracking in conversational LLM agents), and DeltaLogic (minimal-premise-edit revision on FOLIO/ProofWriter). BeliefDrift's narrower contribution is: a continuous, reliability-weighted evidence model over a causal dependency graph, with an empirically-verified order-invariance-vs-temporal-semantics distinction and counterfactual belief-state reconstruction as first-class tracks. See BENCHMARK_SPEC.md §1 for the full comparison table. Do not describe this as "the first" or "the only" belief-revision benchmark.

Dataset composition (measured, not estimated)

	count
Worlds	20,000 (18,000 in-distribution, 2,000 OOD)
Trajectories	60,000
Evidence events	315,000
Order-invariance test cases	15,000
Counterfactual test cases	60,000
Sequence archetypes	8
Splits	train / validation / test / ood_test (world-level, leakage-checked)
Total release size	~27 MB (Parquet)

Full breakdown, generation parameters, and QC results: reports/qc_report.json (16/16 checks passed), generation_stats.json, BASELINES.md.

Quickstart

import pandas as pd
public = pd.read_parquet("data/public/train.parquet")
ground_truth = pd.read_parquet("data/evaluator/train.parquet")
merged = public.merge(ground_truth, on="evidence_id", suffixes=("", "_gt"))

or see examples/sample_readable.jsonl for a small human-readable joined sample, and examples/example_loading.py-equivalent logic above.

Metrics

Belief Conservation Score (BCS) and Belief Update Locality (BUL), formalized in BENCHMARK_SPEC.md §4 (the brief's originally proposed BUL formula was tested, found ill-behaved at the edges, and replaced — documented there with the reasoning). Real, discriminating baseline results using these metrics are in BASELINES.md.

Intended use

Evaluating and comparing AI systems' (LLMs, agents, or classical models) belief-update behavior over multi-step evidence sequences. Not intended as training data for a system that will then be evaluated on it (train/test are split at the world level, but an LLM trained on this data's public files would trivially learn the generative pattern).

Limitations

See LIMITATIONS.md.

Synthetic data disclosure

Every record in this dataset is synthetically generated by the included generator (src/beliefdrift/), from explicit seeds. No real people, real events, or copyrighted source material were used. See BENCHMARK_SPEC.md §2–3 for the exact generative model.

Licensing

Dataset (data/): CC BY 4.0 — see LICENSE-DATA. Generator source code (src/): MIT — see LICENSE.

Citation

See CITATION.cff. No DOI has been registered for this version.

Version history

• v0.1 — sample-validation stage, 300 worlds / 4,725 events, no splits
• v0.2 — this release: scaled to 20,000 worlds / 315,000 events, world-level train/val/test/ood_test splits, 4 baselines with real measured results

Contact

Fahad Hafeez.