aiqbench analysis
SIGMA introduces a novel training objective for autoregressive molecular models that aligns hidden states across equivalent chemical serializations, effectively resolving inconsistencies in next-token predictions without altering the standard decoder architecture. By enforcing chemically certified state correspondence during pre-training, this method significantly reduces Frechet ChemNet Distance and improves predictive performance on key molecular property benchmarks while maintaining computational efficiency comparable to canonical training approaches.
SIGMA: Semantic Identifier Grouping for Molecular Autoregression
arXiv:2603.25062v2 Announce Type: replace Abstract: Autoregressive molecular models assign probability to molecular serializations even though chemical identity is invariant to serialization. Equivalent serializations can therefore represent a common molecular identity yet induce inconsistent next-token decisions. Randomized strings broaden exposure, but do not reveal which intermediate decisions should agree. We introduce SIGMA, a dense suffix-position objective built from chemically certified same-suffix triplets: two equivalent histories, one non-equivalent history, and a shared suffix. SIGMA aligns corresponding pre-token hidden states along the continuation and separates the negative to a finite relative margin, leaving the language-model objective, decoder, and inference procedure unchanged. We compare SIGMA with canonical training, randomized-serialization training, and last-token alignment across four datasets under SMILES and SELFIES. Across the eight representation-dataset blocks, SIGMA yields clear test-reference Frechet ChemNet Distance reductions in six: all four SELFIES domains and QM9 and ZINC under SMILES, with paired 95% confidence intervals below zero against every control. Position-wise analyses show improved state correspondence, chemical discrimination, and next-token agreement while preserving between-molecule information. On two full-corpus ZINC blocks, compute-matched ablations identify chemically correct state correspondence as the effective ingredient. Beyond generation, SIGMA improves mean predictive performance on all six molecular property benchmarks and reduces sensitivity to equivalent molecular serializations on every task.
Abstract:Autoregressive molecular models assign probability to molecular serializations even though chemical identity is invariant to serialization. Equivalent serializations can therefore represent a common molecular identity yet induce inconsistent next-token decisions. Randomized strings broaden exposure, but do not reveal which intermediate decisions should agree. We introduce SIGMA, a dense suffix-position objective built from chemically certified same-suffix triplets: two equivalent histories, one non-equivalent history, and a shared suffix. SIGMA aligns corresponding pre-token hidden states along the continuation and separates the negative to a finite relative margin, leaving the language-model objective, decoder, and inference procedure unchanged. We compare SIGMA with canonical training, randomized-serialization training, and last-token alignment across four datasets under SMILES and SELFIES. Across the eight representation-dataset blocks, SIGMA yields clear test-reference Frechet ChemNet Distance reductions in six: all four SELFIES domains and QM9 and ZINC under SMILES, with paired 95% confidence intervals below zero against every control. Position-wise analyses show improved state correspondence, chemical discrimination, and next-token agreement while preserving between-molecule information. On two full-corpus ZINC blocks, compute-matched ablations identify chemically correct state correspondence as the effective ingredient. Beyond generation, SIGMA improves mean predictive performance on all six molecular property benchmarks and reduces sensitivity to equivalent molecular serializations on every task.
Machine Learning (cs.LG)
arXiv:2603.25062 [cs.LG]
(or arXiv:2603.25062v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2603.25062 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both 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.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 292 words
Related coverage
Difficulty-Aware Semantic-ID Optimization for Generative Recommendation
arXiv · Aug 24, 2026
SAPO: Step-Aligned Policy Optimization for Reasoning-Based Generative Recommendation
arXiv · Aug 18, 2026
MolE-RAG: Molecular Structure-Enhanced Retrieval-Augmented Generation for Chemistry
arXiv · Jun 16, 2026
[AINews] Opus 5.5 is good at explainer videos
Latent Space · Sep 29, 2026