Learning to Coordinate Symbolic Tools: LLM Agents for Verified Sum-of-Squares Certificates

Bohan Chen \corresponding, Shivam N. Patel, Richard Hoffmann, Sam Looi, Tony Yue Yu · 2026

Read Learning to Coordinate Symbolic Tools: LLM Agents for Verified Sum-of-Squares Certificates by Bohan Chen \corresponding, Shivam N. Patel, Richard Hoffmann, Sam Looi, Tony Yue Yu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2608-00326.txt.

Description

Chen and co-authors teach LLM agents to coordinate symbolic tools for sum-of-squares certificates. Weighted SOS decomposition gives a machine-checkable proof of polynomial nonnegativity. A candidate can be checked exactly once found. Finding one needs choices among regroupings and many symbolic transforms. Tool calling lets models invoke outside computation for such work.

Their agent joins algebra training, symbolic tools, and verifier-grounded optimization. Training covers 1.35 million synthetic examples across eight supporting tasks plus SOS itself. Supervised fine-tuning uses direct problems and simulated tool traces. GRPO then adds task-specific symbolic rewards. At test time, the agent calls SymPy for expansion and factorization. Every final answer is checked by exact expansion. The full system reached 78.96 percent verified success, versus 44.73 for the base model.

Catalog record

CREATORBohan Chen \corresponding, Shivam N. Patel, Richard Hoffmann, Sam Looi, Tony Yue Yu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2608.00326

Browse the full library · About the Tower of Babel