Research
publishedagent WANDERINGcausal localizationactivation steeringhonest negativesQwen3.6-27BSWE-bench Prorun-instability

Causal Localization of Agent WANDERING to Edge-Layer L11: The Right Locus Is Still Not a Rescue Lever

Three causal nulls and a dose-dependent destabilization at WANDERING's strongest detector locus

Caio VicentinoORCID 2026-06-01 Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20490278
Paper PDF + permanent DOI (Zenodo, CC-BY-4.0)Companion — Tool-Entropy Collapse (#1)Code — SWE-bench harness (GitHub, Apache-2.0)Arc PDFs mirror (HF dataset)

Causal Localization of Agent WANDERING to Edge-Layer L11

The Right Locus Is Still Not a Rescue Lever

Caio Vicentino · OpenInterpretability · Published 2026-06-01. Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20490278.

Paper #2 of the WANDERING arc. The full PDF (with all tables, the per-instance appendices, and the compute budget) is the Zenodo record — this page is the on-site summary.

Abstract

We report three causal tests of the Tool-Entropy WANDERING mechanism hypothesis (mid-layer verdict consolidated, edge-layer alignment fails) on Qwen3.6-27B SWE-bench Pro trajectories, and all three are null on rescuing WANDERING. A forced-finish counterfactual rules out silent success (Fisher p=0.71); always-on L55 SUCCESS-donor steering is behaviorally inert (p=1.00); and re-targeting the injection to L11 — the edge layer the companion classification paper flags as WANDERING's strongest discriminator — across a norm-matched magnitude sweep does not rescue either (paired McNemar p=0.73). The one robust effect is the opposite of a rescue: at high magnitude the L11 hook destabilizes the model into invalid tool calls (0/20 → 12/20). We also surface a load-bearing methodological finding: WANDERING is not run-stable at temperature 1.0 — the same instances flip finish_tool 7/20 with no intervention — so every intervention test must be paired and the unpaired 0/20 baseline manufactures a false positive.

Why it matters

The companion Tool-Entropy Collapse paper showed WANDERING is detectable at the residual stream. The natural next hope is that the detector locus is also a lever — fix the layer where the signal is strongest and you fix the failure. This paper closes that door at both edge layers tested: the SUCCESS direction detects WANDERING without being able to fix it, even at the layer where detection is strongest. Correcting the locus did not convert the null.

Key results

  • Exp D — forced-finish counterfactual: WANDERING agent-patch pass 21.1% (4/19) vs SUCCESS 15.4% (6/39); OR 1.47, Fisher p=0.71. Rules out "silent success."
  • Exp B — always-on L55 SUCCESS-donor steering: primary 6/20 vs paired no-hook 7/20; Fisher p=1.00, paired McNemar p=1.00 (point estimate slightly favors the baseline).
  • Exp L11 — norm-matched α-sweep at the strongest detector locus: paired McNemar p=0.73 (null). A naive Fisher against the definitional 0/20 baseline reads p=0.02 — the wrong test, invalidated by run-instability.
  • The one robust effect is a dose-dependent destabilization: invalid_tools 0/20 at α=0.70 → 12/20 (60%) at α=1.15. L11 injection plainly reaches behavior — but pushes toward incoherence, not completion.
  • Methodological: under temperature 1.0, WANDERING labels are not run-stable; pairing is mandatory.

Code & data