Losses that Cook: Topological Optimal Transport for Structured Recipe Generation

2026-04-21 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

Researchers from Trustpilot and Politecnico di Torino introduced a novel topological loss function for structured cooking recipe generation, addressing the limitations of standard cross-entropy (CE) in handling complex procedural and compositional coherence. Building on the RECIPE-NLG corpus, their method represents ingredient lists as point clouds in embedding space, minimizing the Sinkhorn divergence between predicted and gold ingredients. Experiments with a fine-tuned Qwen3-4B Small Language Model (SLM) demonstrated that this topological loss significantly improves ingredient recall, quantity precision, and procedural accuracy. When combined with Dice loss in a mixed objective, the approach achieved synergistic gains in quantity and time precision, outperforming CE and single custom losses. Human evaluation further supported these findings, with the proposed model preferred in 62% of cases and showing a 67.5% reduction in generation errors compared to CE.

Key takeaway

For AI Engineers developing structured text generation systems, particularly for complex domains like recipes, you should consider augmenting standard cross-entropy with specialized loss functions. Implementing a topological loss can significantly enhance the factual correctness and procedural coherence of generated outputs, leading to more usable and accurate results. This approach reduces critical errors and improves human preference, making your models more robust for real-world applications.

Key insights

Topological loss significantly improves structured recipe generation by encoding ingredient-level semantic and structural coherence.

Principles

• Cross-entropy is insufficient for structured text generation.
• Loss functions can encode geometric structure in embedding space.
• Domain adaptation is crucial for constraint-satisfying generation.

Method

The method involves fine-tuning SLMs with a composite objective that includes a topological loss, representing ingredient lists as point clouds and minimizing Sinkhorn divergence between predicted and gold embeddings.

In practice

• Augment CE with topological loss for structured text tasks.
• Use recipe-specific metrics for evaluating generation quality.
• Combine Dice and topological losses for balanced accuracy.

Topics

• Structured Recipe Generation
• Topological Loss
• Optimal Transport
• Sinkhorn Divergence
• Dice Loss

Code references

• DarthReca/losses-cook

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

Related on AIssential

• Topological Flow Matching — Topological Flow Matching enhances generative models by integrating domain topology via Laplacian-derived drift for structured data.
• Scientists build a “periodic table” for AI — Many multimodal AI methods distill to compressing data while preserving predictive information.
• Optimal Transport for Machine Learners — Optimal Transport offers a robust mathematical framework for comparing probability distributions, crucial for generative AI model design.
• Uncovering Symmetry Transfer in Large Language Models via Layer-Peeled Optimization — LLM training implicitly transfers data symmetries into structured geometric configurations within model parameters.
• TERRA: Task-Embedded Reasoning and Representation Architecture for Cross-Domain Applications — TERRA proposes a formal theory and testable hypothesis for cross-domain transfer of structured-state representations and predictors.
• Bounding the Predictive Space: How Topological AI Solves Catastrophic Forgetting Through… — Topological AI uses prime-indexed embedding anchors to prevent catastrophic forgetting in LLMs while retaining cognitive flexibility.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.