Tool-Entropy Collapse

A cross-architecture signature of agent WANDERING failure

Caio Vicentino · OpenInterpretability · Published 2026-05-24. Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20368600 · the foundation of the WANDERING arc.

Paper #1, the paper that names WANDERING and gives the arc its detector. The full PDF (all six detectors, the three tiers, the cross-architecture and cross-task tests) is the Zenodo record — this page is the on-site summary.

---

The failure mode

Give a capable model a real software issue, a shell, and a fixed turn budget. Most runs end by submitting a patch via a finish tool (SUCCESS) or by giving up early (LOCKED). A third, quieter failure is WANDERING: the agent stays internally confident it has solved the task but never emits a termination action, burning the whole budget. It is a ~34% blind spot for probe-based agent monitoring — the agent looks fine internally, so a probe doesn't fire.

What we found

On Qwen3.6-27B SWE-bench Pro trajectories we test six detector designs across three signal channels. The load-bearing, probe-free signal is tool-use entropy collapse — in its final turns the agent cycles the same tool calls instead of finishing. It supports three deployment tiers:

Tier	Detector	Recall / FP
Forensic	residual signature (v1)	35% / 0%
Advisory	cross-layer (v1 ∪ v4)	80% / 30%, ~15-turn lead
Autonomous	tool-entropy (v1 ∪ v5)	70% / 5%

Cross-architecture (within SWE-bench): the collapse signature holds beyond Qwen — Llama-70B (collapse ratio 0.41, p < 1e-15) and a GPT-5 router (0.71, p = 8.9e-35) match it. Cross-task is null: on METR MALT the ratio is 1.007 (p = 0.81), so we honestly scope the claim to multi-turn code-execution agent tasks with rich action spaces.

The mechanism (§13) — and why the arc exists

The paper's §13 advances the mechanism the rest of the arc tests: the mid-layer "task is done" verdict is consolidated, while the edge layers (L11/L55) fail to translate it into the finish action — producing the final-turn tool loop. That hypothesis is the seed of everything downstream:

• the verdict locus is real but not a rescue lever (#2);
• the signature is behavior-legible but residual-blind (#3);
• the only thing that rescues WANDERING is a behavioral interruption (#4);
• the verdict is even an interpretable feature that predicts but does not cause finishing (#5);
• and control finally turns out to live late, ~30 layers downstream (#6, the first positive).

Code & data

• Paper + harness + reproducible code: openinterp-swebench-harness
• As an Inspect eval (UK AISI inspect_evals submission): inspect-tool-entropy-collapse
• The full arc: #2 Right Locus · #3 Multi-Channel · #4 Modality Matters · #5 Verdict Circuit · #6 The Lever Is Late · companion note