According to arxiv.org, a new synthesis paper has analyzed 27 benchmark and audit papers spanning 2023-2026 to create a unified taxonomy of large language model agent limitations. The research identifies six failure clusters, including tool invocation errors, planning failures, long-horizon degradation, multi-agent coordination issues, safety failures, and measurement validity problems.
The paper reports that “failures compound nonlinearly with task length” and that “strong performance on individual sub-tasks does not reliably translate into end-to-end success,” according to the abstract. However, the research notes that “substantial progress has been demonstrated in single-turn tool use, short-horizon web navigation, and narrowly scoped coding tasks.”
In related work on controlling tool use, arxiv.org describes research investigating whether tool-use decisions have stable internal representations that can be manipulated. According to the paper, steering vectors extracted from “heading-anchors positions exert bidirectional causal control over tool-invocation behavior across five open-source models and three domains, suppressing unnecessary tool use most effectively in domains where parametric reasoning suffices.”
Separately, arxiv.org reports on research comparing aligned and “refusal-ablated” models from the Gemma and Qwen families on software security tasks. In vulnerability repair validation, the abliterated Gemma model achieved 67.8% usable patches compared to 29.9% for the aligned version, according to the study.