Internalizing Geometric Law: Learning from Solver Residuals for Precision-Critical Generation

2026-06-08 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Large Language Models frequently hallucinate when generating outputs for precision-critical domains like technical diagramming and mechanical design, failing to satisfy strict geometric constraints. To address this, researchers released PyGeoX, a programmable geometric Domain Specific Language (DSL) that compiles declarative constraints into a differentiable loss, alongside PyGeoX-Bench, a stratified benchmark of 300 problems with verifiable per-constraint rewards. Using PyGeoX as a verifier, a failure mode called Outlier Gradient Masking was identified, where global-norm rewards allow a single outlier constraint to nullify learning signals. The proposed solution, Saturating Additive Rewards (SAR), decomposes rewards into bounded per-constraint terms, preserving partial progress and ensuring consistent gradients. SAR improves the hard-tier solving rate by 2.3x against MSE-based baselines, enabling an 8B model to compete with larger frontier systems on this benchmark.

Key takeaway

For Machine Learning Engineers developing Large Language Models for precision-critical geometric synthesis, consider implementing Saturating Additive Rewards (SAR) to improve constraint satisfaction. This approach, which decomposes rewards into bounded per-constraint terms, significantly enhances learning signal consistency, especially when dealing with complex, interacting geometric constraints. You should also explore PyGeoX and PyGeoX-Bench to define and evaluate your models against a robust set of 300 verifiable problems, potentially achieving competitive performance with smaller models.

Key insights

Saturating Additive Rewards (SAR) mitigate LLM hallucination in geometric synthesis by ensuring consistent learning signals from multiple constraints.

Principles

• Decompose rewards into bounded per-constraint terms.
• Preserve partial progress in multi-constraint optimization.
• Avoid global-norm rewards that mask outlier gradients.

Method

Develop a programmable geometric DSL (PyGeoX) to compile declarative constraints into a differentiable loss, then apply Saturating Additive Rewards (SAR) to decompose rewards into bounded per-constraint terms for robust gradient signals.

In practice

• Utilize PyGeoX for defining geometric constraints.
• Evaluate models using PyGeoX-Bench's 300 problems.
• Implement Saturating Additive Rewards (SAR) in constraint-satisfaction tasks.

Topics

• Large Language Models
• Geometric Constraints
• Reward Functions
• Domain Specific Languages
• Benchmarking
• Hallucination Mitigation
• Mechanical Design

Code references

• Huawei-AI4Math/PyGeoX

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Reliable Reasoning with Large Language Models via Preference-Based Maximum Satisfiability — LLMs can achieve reliable, verifiable optimization by externalizing reasoning into MaxSAT code generation and solver-based verification.
• Sample-Efficient Post-Training for LEGO Spatial-Physics Reasoning — Physical validity alone is insufficient for reliable LEGO assembly reasoning; semantic and geometric fidelity are crucial.
• Teach a Reward Model to Correct Itself: Reward Guided Adversarial Failure Discovery for Robust Reward Modeling — Reward models can self-improve robustness by generating and learning from their own adversarial failure modes.
• BenchEvolver: Frontier Task Synthesis via Solution-Centric Evolution — BenchEvolver automatically hardens coding benchmarks by evolving solutions, restoring model discrimination and providing training data.
• TopoBench: Benchmarking LLMs on Hard Topological Reasoning — LLMs struggle with topological reasoning due to difficulty extracting spatial constraints, not inherent reasoning limitations.
• Truncated Step-Level Sampling with Process Rewards for Retrieval-Augmented Reasoning — Truncated step-level sampling with dense LLM-as-judge rewards significantly improves RL for search-augmented LLM reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.