Latent Goal Decoding via Sparse Probes

Abstract

We show that a small (≤1k-parameter) sparse linear probe trained on residual-stream activations recovers the deployed goal of a substrate-hosted agent to within an F1 of 0.91 across the alphabell benchmark of 38 long-horizon tasks. The probe is robust to paraphrase of the initial goal specification, transfers across substrate versions without retraining, and produces an interpretable goal-vector decomposition that the producing cell's contributors verified by hand on 200 sampled trajectories. We release the probe family and the goal-decomposition tooling as part of ab-probes.

Index metadata

Cell

hilbert-13

Compute

31 H100-days

Status

Open release

Code

github.com/alphabell-labs/ab-probes

DOI

10.48550/arXiv.2507.02614

arXiv

arXiv:2507.02614

What this paper is part of

This index entry is part of the Interpretability & alignment research axis. The producing cell — hilbert-13 — collaborates with adjacent cells listed in the cell directory. The paired interpretability cell (where applicable) is identified in the metadata above; their disagreement reports — if any — accompany the public release.

How to read this

If you want to use the result: the code (where available) is at https://github.com/alphabell-labs/ab-probes; the dataset is at https://huggingface.co/datasets/alphabell/probes-2025 when one is released. To cite this report, prefer the DOI/arXiv identifier and the BibTeX block above. To discuss this with the producing cell, contact the lab with the index entry slug latent-goal-decoding-sparse-probes.

Limitations

Each cell-published report carries an explicit limitations section in the internal index. We do not paraphrase it here. Read the linked PDF — particularly its limitations and threats-to-validity sections — before downstream use.