CampChem: Rubric-Grounded Adaptive Campaigns for Multi-Round Organic Synthesis with Residual Spectral Observation
Abstract
Organic synthesis is a multi-round campaign, not a one-shot SMILES completion: a failed solvent, a poisoned catalyst, or leftover starting-material peaks in the 1H NMR force the chemist to revise conditions and try again. Tool-using chemistry agents and condition classifiers still emit a single recipe and then stop. We propose CampChem, an adaptive campaign agent built on a chemistry-instruction LLM. A longitudinal campaign adaptive planner (LCAP) keeps a short memory of failed condition slots and peak signatures and decodes the next catalyst/solvent/reagent tuple under a mask. Chemistry rubric-as-judge reinforcement learning (Chem-RaJ) scores stoichiometry, hazard compatibility, slot completeness, and spectral consistency instead of using a free-form LLM judge. A residual spectral booster (RSB) injects a permutation-invariant peak-set residual when NMR is available and gates to zero otherwise. On SMolInstruct, CampChem reaches 67.1% forward-synthesis and 36.1% retrosynthesis Top-1 exact match with a Mistral-7B LoRA backbone. On USPTO-Condition it attains 0.2940 overall Top-1, and on AutoLabs Experiment 5 it raises protocol F1 from 0.75 to 0.81. Ablations show that Chem-RaJ carries most of the spectrum-free gain, LCAP matters once a prior tuple has been rejected, and RSB contributes mainly on simulated NMR revision.
Related
- Agentic Workflow for High-Throughput Synthesis Campaigns: Mapping the Persistent Phosphor Ca2ZnSi2O7:Eu2+,Nd3+
- Synthesis and testing of ZnO nanoparticles for photo-initiation: experimental observation of two different non-migration initiators for bulk polymerization
- Adaptive Catalytic Synthesis of Aromatic and Aliphatic Amines via CO2-Induced Feedback Inhibition Using Standard Hydrogenation Catalysts.
- Hash-committed pre-registration and exact unlearning make belief revisions in autonomous SiGeC semiconductor campaigns checkable
- Covering and loss-tolerant batch designs guarantee what autonomous campaigns sample in SiGeC semiconductors
BibTeX
@misc{shi2026campchem,
title = {{CampChem: Rubric-Grounded Adaptive Campaigns for Multi-Round Organic Synthesis with Residual Spectral Observation}},
author = {Zixuan Shi and Mengyao Qian and Yichao Sun and Haotian Gu and Jiarui Feng},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008453/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008453/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - CampChem: Rubric-Grounded Adaptive Campaigns for Multi-Round Organic Synthesis with Residual Spectral Observation
AU - Zixuan Shi
AU - Mengyao Qian
AU - Yichao Sun
AU - Haotian Gu
AU - Jiarui Feng
PY - 2026
DA - 2026/09/07
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008453/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008453/v1
AB - Organic synthesis is a multi-round campaign, not a one-shot SMILES completion: a failed solvent, a poisoned catalyst, or leftover starting-material peaks in the 1H NMR force the chemist to revise conditions and try again. Tool-using chemistry agents and condition classifiers still emit a single recipe and then stop. We propose CampChem, an adaptive campaign agent built on a chemistry-instruction LLM. A longitudinal campaign adaptive planner (LCAP) keeps a short memory of failed condition slots and peak signatures and decodes the next catalyst/solvent/reagent tuple under a mask. Chemistry rubric-as-judge reinforcement learning (Chem-RaJ) scores stoichiometry, hazard compatibility, slot completeness, and spectral consistency instead of using a free-form LLM judge. A residual spectral booster (RSB) injects a permutation-invariant peak-set residual when NMR is available and gates to zero otherwise. On SMolInstruct, CampChem reaches 67.1% forward-synthesis and 36.1% retrosynthesis Top-1 exact match with a Mistral-7B LoRA backbone. On USPTO-Condition it attains 0.2940 overall Top-1, and on AutoLabs Experiment 5 it raises protocol F1 from 0.75 to 0.81. Ablations show that Chem-RaJ carries most of the spectrum-free gain, LCAP matters once a prior tuple has been rejected, and RSB contributes mainly on simulated NMR revision.
ER -