In-context modeling as a retrain-free paradigm for foundation models

In-context modeling as a retrain-free paradigm for foundation models

by Yang Liu

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning
🎙 Yang Liu 👥 4K 📅 May 22, 2026 ⏱ 46 min 👁 111 📄 original study 🧭 2026-08-15
Available in: English (current) Français

Keywords

in-context modelingretrain-freefoundation modelshyperelasticityoperator learning

Summary

The talk presents In-Context Modeling (ICM), a retrain-free paradigm for building foundation models in computational science. ICM infers physical relationships directly from observational fields, assimilating measurements as context and performing inference in a single forward pass. The method is trained in a physics-informed, label-free manner using governing equations. A single model generalizes across unseen materials, geometries, and loading conditions, demonstrated on hyperelasticity. ICM integrates with finite-element simulations and is validated using experimental full-field measurements. Performance improves with data diversity and computational budget, exhibiting favorable scaling behavior analogous to foundation models. The talk discusses the limitations of current methods (e.g., PINNs, DeepONet) and positions ICM as a transferable paradigm for retrain-free scientific learning. The speaker also reviews prior work on in-context operator networks and highlights the importance of tokenization in stripping away ad hoc domain geometries.

135 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a compelling argument for a new paradigm in scientific computing, addressing the limitations of existing methods like PINNs and DeepONet. The value lies in the introduction of ICM, which enables retrain-free generalization across physical systems. The argumentation is solid, grounded in mathematical derivations and experimental validation. The speaker effectively demonstrates the method’s capabilities on hyperelasticity, showing generalization to unseen geometries and loading conditions. The discussion of scaling behavior adds credibility, though the claim of emerging capabilities is not fully substantiated. The talk also addresses potential questions about training data and example generalization, providing thoughtful responses.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates scientific rigor through detailed mathematical formulations and validation against experimental data. The speaker references prior work, including his own papers on in-context operator networks, and discusses limitations. The title accurately reflects the content, focusing on in-context modeling as a retrain-free paradigm. The sources cited are primarily the speaker’s own publications and related work, which are appropriate for a seminar. The talk does not include a public advertising sequence. The adequacy between title and content is high, as the talk directly addresses the paradigm shift proposed.

201 words

Title / Content Match

The title accurately reflects the content: the talk introduces in-context modeling as a retrain-free paradigm for foundation models in computational science.

Quality & Reliability

8/10

Presentation of a novel method (ICM) with mathematical derivations, validation on hyperelasticity, and scaling analysis. The speaker is an established researcher. However, the talk is a seminar presentation, not a peer-reviewed publication, and some claims (e.g., emerging capabilities) are discussed but not fully demonstrated.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • In-context operator networks — Speaker's prior work introducing in-context learning for operator networks
  • DeepONet — Referenced as a classical operator learning method
  • Physics-informed neural networks (PINNs) — Referenced as a flagship method in AI for scientific computing

Concurring Sources

  • In-context operator networks — Prior work by the speaker that aligns with the presented method
  • DeepONet — Classical operator learning method that ICM builds upon

Contribution & Novelties

The talk introduces In-Context Modeling (ICM), a novel paradigm that enables retrain-free generalization in computational science. The key innovation is the tokenization of physical observations into context, allowing a single model to infer unknown physical relationships and apply them to new queries. This approach addresses limitations of existing methods like PINNs and DeepONet, which require retraining for new systems. The talk demonstrates ICM on hyperelasticity, showing generalization across unseen materials, geometries, and loading conditions. The method is physics-informed and label-free, trained using governing equations. The scaling behavior suggests potential for foundation models in science.

Pour aller plus loin :

129 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a well-rounded presentation with strong information content, technical depth, and reliability. The talk is particularly strong in quantitative information and technical level, reflecting the mathematical rigor and detailed methodology.

Reliability 8/10

💬 Sur les 0 commentaires analysés, aucune tendance n'est disponible.