Structured Guidance Boosts LLM Agent Coverage to 44% on Multireference Quantum Calculations
A structured LLM agent doubled coverage of multireference calculations on 558 benchmark transitions while lowering error. Expert-informed procedural rules proved more important than model size alone. The work shows autonomous agents can now reproduce published multireference workflows with high fidelity.
The agent autonomously selected active spaces, submitted ORCA jobs, recovered convergence failures, and logged every reasoning step. On the full 558-VTE benchmark it doubled coverage for double and Rydberg excitations while cutting mean absolute error by 34 meV. When given complete published workflow details it reproduced reference results to 23 meV MAE and resolved 75% of targets inside seven attempts.
Multireference methods have resisted automation because active-space choice and state averaging require chemical intuition that simple scripts lack. By grounding each decision in literature precedents or explicit chemical rules, the structured ladder supplies the missing procedural knowledge, turning an otherwise brittle LLM into a reproducible high-throughput engine.
Future systems will likely combine this reasoning scaffold with tighter integration to quantum-chemistry APIs and reinforcement learning from failed runs. Within two years, hybrid human-AI pipelines could shift routine multireference work from specialist groups to general computational teams, provided error bars and audit logs remain mandatory.
The decisive variable remains the quality of the decision ladder rather than raw model scale.
Rondinelli: 65% coverage on expanded QUESTDB reached by September 2027 with ladder plus RL feedback.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2609.13357)
- [2]Supporting Source(https://doi.org/10.1021/acs.jctc.3c00123)