AI for Science: January 2026 Week 2

1. Scalable Machine Learning Force Fields for Macromolecular Systems Through Long-Range Aware Message Passing

arXiv:2601.03774

Why Novel: Addresses fundamental cutoff limitation in ML force fields by introducing explicit long-range atom-fragment attention via SO(3)-equivariant transformer, enabling stable force predictions for systems up to 1200+ atoms where local models fail.

Key Innovations:

• Introduces MolLR25 benchmark suite with high-fidelity DFT labels for systems up to 1200 atoms and 75 Angstrom spatial range, far exceeding existing benchmarks
• Develops E2Former-LSR with Long-Short Range (LSR) message passing that explicitly integrates long-range attention through chemically informed fragmentation
• Demonstrates near-constant force error scaling beyond 1200 atoms while MACE-large errors increase with system size
• Achieves up to 30% speedup over purely local models by efficient fragment-based attention architecture
• Accurately captures non-covalent decay and dissociation physics in medium-scale protein conformations

Evidence:

• tab:1 — MAE comparison showing E2Former-LSR achieves best force/energy accuracy on 5/7 systems vs MACE-large and Allegro
• fig:1 — Benchmark scope showing MolLR25 extends to N1200 atoms and Rmax75A; scaling curves show E2Former-LSR flat vs MACE rising error
• fig:2 — Di-molecule dissociation showing E2Former-LSR maintains smooth energy decay while MACE exhibits discontinuities at intermediate ranges
• fig:3 — Protein conformation evaluation showing uniformly lower force errors across BBL, Homeodomain, alpha3D, and lambda-repressor

Impact: Establishes scalable pathway for ML force fields in biological systems, enabling quantum-accurate MD simulations of macromolecules that were previously limited by fixed-cutoff architectures.

2. A Unified Frequency Principle for Quantum and Classical Machine Learning

arXiv:2601.03169

Why Novel: First unified theoretical framework proving frequency-dependent learning dynamics (low frequencies learned faster) applies to both classical DNNs and quantum neural networks, with rigorous analysis of noise effects on QNN expressivity.

Key Innovations:

• Proves unified F-principle: derivative ratio bound showing low-frequency loss dominates early training across classical and quantum architectures via spectral projector analysis
• Establishes exponential suppression of high-frequency Fourier modes under Pauli noise in QNNs via novel Pauli-path integral representation
• Demonstrates that noise robustness of low-frequency learning enables efficient classical simulation of noisy QNNs through frequency truncation with provable error bounds
• Extends theory to axis-aligned dephasing and depolarizing noise models with distinct impacts on frequency components

Evidence:

• thm:1 — Main theorem establishing derivative ratio bound for frequency-dependent learning rate, formalizing F-principle
• thm:2 — Noise theorem showing (1-2gamma)^||omega||_1 suppression of Fourier coefficients under Pauli noise
• fig:1 — Unified schematic showing F-principle in training dynamics for both DNN and QNN architectures
• fig:2 — QNN ansatz with repeated encoding demonstrating the theoretical framework on concrete circuits

Impact: Provides fundamental theoretical lens unifying classical and quantum learning dynamics, with practical implications for QNN design, noise resilience, and identifying when quantum or classical approaches offer advantages.

3. Agentic Proof Automation: A Case Study

arXiv:2601.03768

Why Novel: First large-scale demonstration of LLM agents performing mechanical proof engineering on a production-scale Lean 4 formalization (14,000+ lines), achieving 87% task success with only 16% human intervention.

Key Innovations:

• Establishes human-agent-prover workflow where agents generate proof scripts, refine based on Lean feedback, and report obstacles for human guidance
• Demonstrates 164/189 (87%) task success across Proof, Repair, Refactor, State+Prove, Query, and Chore categories on System Capless formalization
• Achieves lowest intervention rate (0-9%) on Query and Chore tasks, highest (35%) on State+Prove requiring mathematical creativity
• Provides open-source mechanization and interactive explorer for reproducibility and future benchmarking

Evidence:

• tab:1 — Task breakdown showing 87% success: Proof 41/51, Repair 43/48, Refactor 28/35, State+Prove 20/23, Query 21/21, Chore 11/11
• fig:1 — Human-agent-prover workflow diagram showing task assignment, proof generation, Lean feedback loop
• fig:2 — System Capless abstract syntax showing complexity of formalized type system

Impact: Demonstrates practical viability of LLM-assisted formal verification at scale, establishing division of labor where humans provide mathematical insight while agents handle mechanical proof engineering.

1. RealPDEBench: A Benchmark for Complex Physical Systems with Real-World Data

arXiv:2601.01829

First scientific ML benchmark pairing real-world measurements with simulations across 5 physical scenarios, revealing significant sim-to-real gaps while demonstrating simulated pretraining improves real-world accuracy.

2. StablePDENet: Enhancing Stability of Operator Learning for Solving Differential Equations

arXiv:2601.06472

Introduces adversarial training framework for neural operators with formal stability definition, achieving robust accuracy under perturbations across Poisson, elliptic, heat, diffusion-reaction, and Stokes equations.

3. Pretrain Finite Element Method: A Pretraining and Warm-start Framework for PDEs via Physics-Informed Neural Operators

arXiv:2601.03086

Bridges neural operators and classical FEM via Transolver-based pretraining on unstructured point clouds, providing fast warm-start initialization that preserves FEM accuracy with dramatic iteration reduction.

4. Discontinuous Galerkin Finite Element Operator Network for Solving Non-smooth PDEs

arXiv:2601.03668

Data-free operator learning framework combining DG methods with neural networks for PDEs with discontinuous coefficients, with rigorous convergence analysis and accurate discontinuity resolution.

5. Knowledge Distillation of a Protein Language Model Yields a Foundational Implicit Solvent Model

arXiv:2601.05388

Distills ESM3 evolutionary information into GNN-based implicit solvent model, achieving stable ML/MD across 11 proteins totaling 6.8 microseconds with improved IDP ensemble behavior.

6. Scalable Dielectric Tensor Predictions for Inorganic Materials using Equivariant Graph Neural Networks

arXiv:2601.04755

Introduces GoeCTP frame-averaging framework decoupling symmetry from backbone design, enabling large-scale screening that identifies novel high-dielectric materials like Ba2SmTaO6.

7. From Implicit to Explicit: Token-Efficient Logical Supervision for Mathematical Reasoning in LLMs

arXiv:2601.03682

Introduces First-Step Logical Reasoning (FSLR) that isolates logical relationship understanding, achieving CoT-SFT gains with 80% fewer tokens and 4-6x training speedup.

8. A Non Linear Spectral Graph Neural Network Simulator for More Stable and Accurate Rollouts

arXiv:2601.05860

Introduces nonlinear spectral filters operating in global eigenmode basis for MD simulations, achieving stable long-horizon rollouts and better global property predictions.