Research
publishedagent WANDERINGactivation patchinglogit lensknowledge-action gapagent terminationsteeringfirst positiveQwen3.6-27BSWE-bench Pro

The Lever Is Late: Causal Control of Long-Horizon Agent Termination Lives in a Task-Matched, Late Action-Commitment Block

The first positive of the WANDERING arc — the termination lever is late (L51–63), ~30 layers downstream of the verdict

Caio VicentinoORCID 2026-06-03 Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20534219 · the first positive of the WANDERING arc
Paper PDF + permanent DOI (Zenodo, CC-BY-4.0)decision-locator tool (model-agnostic, pip-installable)Pre-registration + figure code + notebooks + pre-mint eval (GitHub)Transcripts + labels + residual bundle (HF dataset)Arc PDFs mirror (HF dataset)

The Lever Is Late

Causal control of long-horizon agent termination lives in a task-matched, late action-commitment block

Caio Vicentino · OpenInterpretability · Published 2026-06-03. Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20534219 · the first positive of the WANDERING arc.

Paper #6, the first positive of the six-piece WANDERING arc. The full PDF (3 figures, 2 tables, the pre-mint eval, all caveats) is the Zenodo record — this page is the on-site summary.

The question

A long-horizon coding agent fails by WANDERING — acting indefinitely without ever emitting the finish action. Five papers established where control over termination is not: not the residual stream, not the best-detecting locus, not even the exact, named, interpretable "task-done" SAE feature (which predicts the finish action at AUROC 0.91 yet, clamped, moves P(finish) by −0.001). After five places where the lever isn't, this paper asks the dual question: where is it?

What we found

On 99 Qwen3.6-27B SWE-bench Pro trajectories, reconstructed faithfully at the decision point and gated for behavioral fidelity (P(finish): SUCCESS 0.59 ≫ WANDERING 0.07 ≫ LOCKED 0.005):

  1. The decision is computed late. A layer-resolved logit-lens shows the finish decision is invisible through layer 31 — and flat at the verdict layer (L23) — then explodes in the last ~12 of 64 layers (L51–L63). The decision becomes readable ~30 layers downstream of where the "task-done" verdict is represented.
  2. The internal lever is late and donor-specific. Injecting the SUCCESS late-block state into a WANDERING agent at its decision point raises P(finish) (+0.13 at L55, +0.15 at L59); the LOCKED-donor control moves it the other way. Every mid-layer and verdict-feature intervention is inert.
  3. It survives a real generation — but only task-matched. Patching the late block at the decision point alone makes the agent emit a well-formed <function=finish> call in 42% of WANDERING decision points (5/12; exact one-sided McNemar p = 0.031 vs a 0/12 baseline and a 0/12 LOCKED-donor null) — but only when the donor is task-matched; a coarse class-mean donor is not significant (25%, p = 0.125).

The law

The knowledge–action gap on agents is a layer gap. The termination decision is known mid-stream (the verdict, AUROC 0.91 to detect, 0% to control) but only writable late (the task-matched residual, 42%). The verdict-null → late-lever jump is a controlled, same-experiment contrast; a separate behavioral interruption gives a comparable lift (30%→70%, paper #4). Control over an agent decision rises with proximity to the action channel — and the gap closes in exactly one place: late, task-matched, near the action. This is the first internal causal lever of the arc, and it explains all five prior nulls as wrong-layer / wrong-donor / wrong-channel.

The tool

The method ships as decision-locator — a small, pip-installable, model-agnostic tool that, for any tool-calling decision on any open-weight model, (1) locates the layer where the decision becomes readable, (2) sweeps activation patching to find where it is writable, and (3) steers a real generation by patching the commitment layer. pip install git+https://github.com/OpenInterpretability/decision-locator, then decision-locator demo --model gpt2. CPU self-test + Colab included.

Honest scope

Single model (Qwen3.6-27B), single task family (SWE-bench Pro), n = 12 per class. The positive headline depends on the task-matched donor — the coarse class mean is not significant. The effect is partial (42%). L63's large patching effect is near the output and partly trivial; the load-bearing patching evidence is L51–L59. The per-pair donor's edge over the mean may reflect task-conditioning or single-vs-averaged geometry; we do not separate these. The three-site ordering (verdict / late / behavioral) synthesizes three experiments with different baselines — only the verdict-null → late-lever jump is a controlled contrast.

Code & data