{"slug": "harness-sensitivity-is-non-monotone-across-llm-agent-tiers", "title": "Harness Sensitivity Is Non-Monotone Across LLM Agent Tiers", "summary": "A new study from researchers testing six large language models across four capability tiers found that the relationship between model capability and optimal harness complexity is non-monotone, contradicting the assumption that higher-capability models require less structural guidance. In a 432-run experiment on the HEAT-24 benchmark, a frontier chat model (Gemini 2.5 Flash) saw its task success rate drop by 29-38 percentage points with increased harness verbosity, while a frontier reasoning model (Qwen3.5-122B) achieved its highest success rate (91.7%) under the strictest harness. The findings, which also showed a 2B model matching the stability of stronger models, highlight that harness sensitivity depends critically on model type and that tier-aware selection guidelines are needed for reliable LLM agent deployment.", "body_md": "# Computer Science > Artificial Intelligence\n\n[Submitted on 26 May 2026]\n\n# Title:It's Not the Capability: Harness Sensitivity Is Non-Monotone Across LLM Agent Tiers\n\n[View PDF](/pdf/2605.26731)\n\n[HTML (experimental)](https://arxiv.org/html/2605.26731v1)\n\nAbstract:A prevalent assumption in LLM agent deployment holds that more structured harnesses universally improve\n\nreliability, and that higher-capability models need proportionally less structural guidance -- together\n\nimplying a monotone inverse relationship between model capability tier and optimal harness complexity. We\n\ntest this hypothesis through a controlled 432-run experiment crossing six models across four capability\n\ntiers with three harness conditions (light, balanced, strict) on HEAT-24, a 24-task synthetic benchmark\n\nwith git-based workspace verification. Our results refute the monotone inverse relationship on two\n\nfronts. First, for the frontier chat model evaluated (Gemini 2.5 Flash), increased harness verbosity\n\nlowers VTSR by 29-38 percentage points -- a harness-complexity paradox. Second, for the frontier\n\nreasoning model evaluated (Qwen3.5-122B, extended thinking enabled), strict harness achieves the highest\n\nVTSR (91.7%) and the lowest latency, the opposite of the prediction. Within the constrained tier, a 2B\n\nmodel (Gemma4:e2B) matches strong-open-tier stability at 91.7% across all harnesses. Because each tier is\n\nrepresented by a single model in this study, these results should be interpreted as model-specific\n\nobservations; harness sensitivity appears non-monotone across the models evaluated, and depends\n\ncritically on model type (chat vs. reasoning). We introduce a six-label failure taxonomy showing that\n\nformat_violation dominates capable-model failures while wrong_file dominates low-capability failures, and\n\nwe derive practical tier-aware harness selection guidelines.\n\n### References & Citations\n\nLoading...\n\n# Bibliographic and Citation Tools\n\nBibliographic Explorer\n\n*(*[What is the Explorer?](https://info.arxiv.org/labs/showcase.html#arxiv-bibliographic-explorer))\nConnected Papers\n\n*(*[What is Connected Papers?](https://www.connectedpapers.com/about))\nLitmaps\n\n*(*[What is Litmaps?](https://www.litmaps.co/))\nscite Smart Citations\n\n*(*[What are Smart Citations?](https://www.scite.ai/))# Code, Data and Media Associated with this Article\n\nalphaXiv\n\n*(*[What is alphaXiv?](https://alphaxiv.org/))\nCatalyzeX Code Finder for Papers\n\n*(*[What is CatalyzeX?](https://www.catalyzex.com))\nDagsHub\n\n*(*[What is DagsHub?](https://dagshub.com/))\nGotit.pub\n\n*(*[What is GotitPub?](http://gotit.pub/faq))\nHugging Face\n\n*(*[What is Huggingface?](https://huggingface.co/huggingface))\nScienceCast\n\n*(*[What is ScienceCast?](https://sciencecast.org/welcome))# Demos\n\n# Recommenders and Search Tools\n\nInfluence Flower\n\n*(*[What are Influence Flowers?](https://influencemap.cmlab.dev/))\nCORE Recommender\n\n*(*[What is CORE?](https://core.ac.uk/services/recommender))# arXivLabs: experimental projects with community collaborators\n\narXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.\n\nBoth individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.\n\nHave an idea for a project that will add value for arXiv's community? [ Learn more about arXivLabs](https://info.arxiv.org/labs/index.html).", "url": "https://wpnews.pro/news/harness-sensitivity-is-non-monotone-across-llm-agent-tiers", "canonical_source": "https://arxiv.org/abs/2605.26731", "published_at": "2026-05-28 02:59:13+00:00", "updated_at": "2026-05-28 03:26:45.804931+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "ai-agents", "ai-research", "machine-learning"], "entities": ["Gemini 2.5 Flash", "Qwen3.5-122B", "Gemma4:e2B", "HEAT-24"], "alternates": {"html": "https://wpnews.pro/news/harness-sensitivity-is-non-monotone-across-llm-agent-tiers", "markdown": "https://wpnews.pro/news/harness-sensitivity-is-non-monotone-across-llm-agent-tiers.md", "text": "https://wpnews.pro/news/harness-sensitivity-is-non-monotone-across-llm-agent-tiers.txt", "jsonld": "https://wpnews.pro/news/harness-sensitivity-is-non-monotone-across-llm-agent-tiers.jsonld"}}