Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution

July 15, 2026| Source: arXiv

Preprint — not peer-reviewed. Large language model (LLM) agents increasingly automate multi-step engineering and informatics workflows, yet they rarely ask how much effort a task actually requires. They often follow a maximum-context-first strategy--re-reading files and dependencies they have already seen--turning a one-line edit into a small code-base audit. We argue the missing capability is task-aware execution-scope estimation: judging a task's difficulty, the information it truly needs, and the shortest reliable path be

Large language model (LLM) agents increasingly automate multi-step engineering and informatics workflows, yet they rarely ask how much effort a task actually requires. They often follow a maximum-context-first strategy--re-reading files and dependencies they have already seen--turning a one-line edit into a small code-base audit. We argue the missing capability is task-aware execution-scope estimation: judging a task's difficulty, the information it truly needs, and the shortest reliable path before committing budget. We formalize minimum-sufficient execution and the Agent Cognitive Redundancy Ratio (ACRR), and propose E3 (Estimate, Execute, Expand): the agent estimates an initial operating point, executes a minimum viable path, and expands scope only when verification fails. On MSE-Bench--a deterministic benchmark of 121 edits in a capability-controlled simulator--E3 matches the strongest baseline's 100% success while cutting cost by 85%, tokens by 91%, and inspected files by 92%, and further beats a strong adaptive retrieval baseline by 16%; the gains survive held-out instruction wording and essentially every cost weighting. A companion real-model harness (LLM-Case) corroborates the effect on a live gpt-4o agent editing a real open-source library, with every candidate patch graded by actually running the project's real pytest suite against a measured oracle: the over-reading is milder but real, and E3 is the leanest and fastest policy at comparable task success--its one shortfall a provider rate-limit, not a wrong edit. We frame this as a controlled probe of execution redundancy, not a measurement of any deployed agent, and position task-aware execution as a step toward engineering-grounded AI (EGAI)--agents whose effort is anchored in the engineering reality of the task. We release the framework and benchmark.

Read the original source →

Plant fiber exploitation contributed to the emergence of ground-edged cutting tools in North China.

Ground-edged cutting tools are widely regarded as technological hallmarks of early agriculture in North China, commonly hypothesized to have been developed primarily for cereal harvesting. Yet the functional foundations of this innovation remain insufficiently tested. This study integrates use-wear and microfossil analyses of 140 stone tools from the Peiligang site, with experimental data, to reassess their roles within long-term trajectories of technological change from the Late Paleolithic to

Artificial IntelligenceSeptember 28, 2026

Neural network-augmented Pfaffian wave-functions for scalable simulations of interacting fermions.

Developing accurate numerical methods for strongly interacting fermions is crucial for improving our understanding of various quantum many-body phenomena, especially unconventional superconductivity. Recently, neural quantum states have emerged as a promising approach for studying correlated fermions, highlighted by the hidden fermion and backflow methods, which use neural networks to model corrections to fermionic quasiparticle orbitals. In this work, we expand these ideas to the space of Pfaff

Artificial IntelligenceSeptember 26, 2026

Why friends in common reveal network stars.

The Friendship Paradox states that, on average, your friends have more friends than you do. We extend this to common friends-those who appear in multiple people's friend lists. We show that the more people who share a common friend, the more connected that person tends to be, and we derive an expression quantifying this progression. In a regional Facebook network, a common friend to three randomly sampled individuals has on average more friends than 99.9% of the network. In a citation network, a

Artificial IntelligenceSeptember 26, 2026

Sensory context improves language prediction in humans and LLMs.

Language is a fundamental human capacity. Large language models (LLMs) have presented the first viable model of language outside of humans, yet how these models learn and use language differs significantly from humans. Here, we compare LLMs and humans predicting language with varying levels of sensory information-from disembodied written text to audiovisual videos of speakers-to demonstrate that, in both humans and LLMs, sensory context is critical for optimal performance. We asked human partici

Artificial IntelligenceSeptember 10, 2026

Treatment Decisions in Multiple Myeloma.

Revolutions in transplantation and targeted and immune therapies have transformed multiple myeloma from a disease with an associated survival of a few years into one for which functional cure is an emerging goal. This abundance of effective therapies has created clinical complexity. Here we provide a practical framework, anchored in trial evidence and informed by emerging biologic discoveries, for the navigation of treatment decisions across the disease spectrum. We outline how cytogenetic and g