caiovicentino1/tool-entropy-collapse-paper

Tool-Entropy Collapse: A Cross-Architecture Signature of Agent WANDERING Failure

Author: Caio Vicentino (OpenInterpretability) · ORCID 0009-0003-4331-6259 · caio@openinterp.org

Status: v0.8 (2026-05-24) — submission-ready paper. PDF + LaTeX source + 5 figures included.

License: Apache-2.0 (code + paper)

TL;DR

Probe-based safety monitoring of LLM agents has a 34% blind spot on Qwen3.6-27B SWE-bench Pro: the WANDERING sub-class where probe says "success" but agent never emits finish_tool and hits its budget. We test six detector designs across three signal channels (text, residual cross-layer, action entropy). Tool-use entropy collapse is the breakthrough signal — WANDERING agents collapse onto a small set of repeated tool calls (W/S median ratio 0.41 in Qwen AND Llama, 0.71 in GPT-5), enabling a Tier-3 autonomous-termination detector at 70% recall × 5% FP via combined v1 ∪ v5.

Cross-architecture validation: Llama-70b (n=2,315, p<10⁻¹⁵) and GPT-5 router (n=1,419, p=8.9×10⁻³⁵) confirm. Cross-task validation on METR MALT (15+ task families) is NULL (p=0.81), scoping the claim to multi-turn code-execution agent tasks with rich action spaces.

Headline numbers

Metric	Value
WANDERING prevalence (Qwen Phase 6)	34% (20/59 failures), 95% CI [22.0%, 45.8%]
Tool-entropy W/S ratio (Qwen)	0.41, p = 1.0×10⁻⁶
Tool-entropy W/S ratio (Llama-70b)	0.41, p < 10⁻¹⁵
Tool-entropy W/S ratio (GPT-5 router)	0.71, p = 8.9×10⁻³⁵
Cross-task MALT W/S ratio	1.007, p = 0.81 (NULL — scoping result)
Tier-3 detector (v1 ∪ v5) recall	70%
Tier-3 detector SUCCESS FP	5%
Combined v1 ∪ v4 advisory tier recall	80% (at 30% FP, 15-turn lead)

Files in this card

• inflection_wandering.pdf — 11-page compiled PDF (paper)
• inflection_wandering.tex — template-agnostic LaTeX source
• fig1_cross_arch_entropy.pdf — HEADLINE figure: cross-arch entropy histograms
• fig2_disagreement_trajectory.pdf — cross-layer disagreement evolution
• fig3_detector_comparison.pdf — 6-detector landscape + 3-tier regions
• fig4_lab_summary.pdf — 4-lab W/S ratio bar chart
• fig5_venn_orthogonality.pdf — v1 ∩ v4 ∩ v5 captures Venn

Reproducibility

All scripts + per-trajectory output JSONs at github.com/OpenInterpretability/openinterp-swebench-harness. Apache-2.0.

Datasets used:

• OpenInterp Phase 6 (own; 99 Qwen3.6-27B SWE-bench Pro trajectories with per-turn residuals at L11/L23/L31/L43/L55 in bf16 safetensors; release planned upon paper acceptance)
• nebius/SWE-agent-trajectories (CC-BY-4.0)
• JetBrains-Research/agent-trajectories-swesmith-random-subset (HF public)
• SWE-bench/SWE-smith-trajectories (HF public)
• metr-evals/malt-public (HF gated, approval required)

Citation

@misc{vicentino2026toolentropy,
  title = {Tool-Entropy Collapse: A Cross-Architecture Signature of Agent WANDERING Failure},
  author = {Vicentino, Caio},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.20368807},
  url = {https://doi.org/10.5281/zenodo.20368807}
}

Published 2026-05-24:

• Zenodo v0.9 (calibration revision): https://zenodo.org/records/20368807
• Zenodo v0.8 (original): https://zenodo.org/records/20368601
• Concept DOI (umbrella): 10.5281/zenodo.20368600
• LessWrong post: https://www.lesswrong.com/posts/D23kWLmpcXhFAQ2RY/tool-entropy-collapse-a-cross-architecture-signature-of

Strategic positioning

This paper extends the probe-based monitoring paradigm (Apollo deception probes, Anthropic sleeper-agent defection probes) by:

1. Quantifying a previously unmeasured blind spot (34% WANDERING failures)
2. Demonstrating a probe-free complementary signal (tool-use entropy) that works without model access
3. Validating cross-architecture (3 labs, 3 model families, 3 agent scaffolds)
4. Honestly scoping via MALT null finding
5. Providing operational deployment guidance (3 tiers) for production safety monitoring

The Tier-3 autonomous-termination detector is the first reported in this paradigm to achieve sub-5% false-positive rate at high recall — making production deployment viable.