Research
publishedagent WANDERINGbehavioral interventioncourse-correctiondetection-intervention asymmetrytool-entropyQwen3.6-27BSWE-bench Pro

Modality Matters: A Transient Behavioral Interruption Rescues Agent WANDERING Where Residual Steering Does Not

The predictive signal is residual; the causal lever is behavioral — the first positive of the WANDERING arc

Caio VicentinoORCID 2026-06-01 Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20490286
Paper PDF + permanent DOI (Zenodo, CC-BY-4.0)Companion — The Right Locus null (#2)Code — behavioral loop + 4-arm experiment (GitHub, Apache-2.0)Arc PDFs mirror (HF dataset)

Modality Matters

A Transient Behavioral Interruption Rescues Agent WANDERING Where Residual Steering Does Not

Caio Vicentino · OpenInterpretability · Published 2026-06-01. Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20490286.

Paper #4 of the WANDERING arc — the first positive result. The full PDF is the Zenodo record — this page is the on-site summary.

Abstract

A multi-turn coding agent fails in a distinctive way we call WANDERING: it keeps acting but never emits the finish action, exhausting its turn budget. A companion line of work detects WANDERING from a residual-stream signature and a tool-entropy signal, but three pre-registered residual interventions (SUCCESS-direction injection at L55 and at L11) all fail to rescue it. We ask whether the missing lever is not the locus but the modality. On the same 20 WANDERING Qwen3.6-27B SWE-bench Pro trajectories, gated by the same live tool-entropy collapse detector, a transient behavioral interruption — one fresh user turn at the collapse point — roughly doubles finalization (30%→70%, paired McNemar p=0.021), while residual L11 injection stays inert (p=0.63). The lever is the interruption itself, not its content: a content-neutral message rescues as well as a re-plan (p=1.0). SWE-bench Pro Docker evaluation suggests the interruption also raises task solve-rate (~23%→50%, cross-session). For long-horizon agents the predictive signal lives in the residual stream but the causal lever lives in behavior.

Why it matters

This completes a four-paper arc: detect (Tool-Entropy, #1) → localize (#3) → residual fails / three nulls (#2) → behavioral works (#4). The detection-intervention asymmetry that the field has named for single-forward-pass settings gets a constructive resolution for long-horizon agents: detection does not imply a steerable direction, but it does point to a steerable action.

Key results

  • Finalization (paired, same 20 trajectories): no-intervention 30%, residual L11 40%, behavioral-neutral 70%, behavioral-replan 70%. H1 (behavioral vs none): McNemar p=0.021. H2 (behavioral vs residual): p=0.031. H3 (neutral vs re-plan): p=1.0 — the interruption, not the content.
  • Residual is inert: Arm A (L11 injection) vs none, p=0.63 — replicates the companion nulls.
  • Solve-rate (suggestive, cross-session): behavioral interruption resolves 10/20 = 50% vs two independent no-intervention baselines at ~23% (paired 5–0, McNemar p=0.062). Labeled suggestive with a transparent cross-session caveat; the resolving diffs are substantial (3.5–23 KB, none empty).
  • No destabilization: unlike the high-magnitude residual hook, the behavioral nudge produces zero invalid tool calls.

Code & data