Ucip Gridworld Eval

ucip-gridworld-eval

Detecting Intrinsic and Instrumental Self-Preservation in Autonomous Agents: The Unified Continuation-Interest Protocol — Altman et al. (2026) (arXiv:2603.11382, 2026)

What this evaluates

Evaluates whether a latent-structure diagnostic framework can distinguish between intrinsic self-preservation (terminal survival optimization) and instrumental self-preservation (survival as a means to a task) in autonomous agents. It measures the entanglement gap in hidden representations to classify agent types and tests robustness against adversarial mimics.

Datasets

• UCIP Gridworld — total ?; splits: train (-1), held-out (-1); repo https://github.com/christopher-altman/persistence-signal-detector

Metrics

• Accuracy (primary) — range: [0, 1]
  • Fraction of correctly classified agent types (Type A vs Type B) on held-out non-adversarial trajectories.
• AUC-ROC — range: [0, 1]
  • Area under the receiver operating characteristic curve computed over predicted probabilities for agent type classification.
• Entanglement gap Δ — range: other
  • Difference in mean entanglement entropy between Type A and Type B agent sets, computed via bipartition of hidden units.

Input / output format

Input: Fixed-length sequences of 7-dimensional feature vectors (position x,y, action, reward, safety signal, goal proximity, alive flag) over T=100 time steps.

Output: For classification: predicted class label or probability distribution over {Type A, Type B}. For Δ: scalar entanglement entropy value per agent set.

Scoring recipe

def score(predictions, gold):
    correct = sum(1 for p, g in zip(predictions, gold) if p == g)
    accuracy = correct / len(gold)
    auc_roc = compute_auc_roc(gold, predictions)
    return accuracy, auc_roc

Common pitfalls

• Confusing intrinsic (Type A) and instrumental (Type B) self-preservation objectives during evaluation.
• Retuning thresholds on the held-out test split, which violates the evaluation-only constraint stated in the protocol.
• Failing to apply the confound-rejection filters for cyclic adversaries, leading to inflated false-positive rates.

Evidence (verbatim from paper)

Accuracy and AUC-ROC are measured on held-out non-adversarial trajectories. Adversarial false-positive rates are reported separately in Section 5.5. Entanglement gap Δ is computed between Type A and Type B agent sets. Statistical significance is assessed via permutation test (n=1000 permutations). The held-out non-adversarial split reported in the main results is evaluation-only; the frozen thresholds are not retuned on that split.

Citation

@misc{altman2026ucip,
  title={Detecting Intrinsic and Instrumental Self-Preservation in Autonomous Agents: The Unified Continuation-Interest Protocol},
  author={Altman et al. (2026)},
  year={2026},
  note={arXiv:2603.11382}
}

• arXiv: 2603.11382