Summary of "Panel: AI & Mathematics"
Panel: AI & Mathematics — Summary
Overview
The panel explored how modern AI (large transformers / LLMs and specialized neural nets) is being applied to mathematical research, scientific modeling, theorem formalization, and domain-specific problems (chemistry, fluid dynamics, epidemiology, legal data).
Main recurring themes:
- Building general-purpose solvers.
- Extracting interpretable scientific knowledge from data.
- Combining classical mathematical structure with machine learning.
- Verification and faithfulness of AI output.
- Tooling and harnesses around models (agents, search, verification).
- Compute/resource and energy concerns.
- Sociological and educational impacts.
Key ideas and takeaways
General-purpose generative solvers for PDEs / dynamics
- Train large transformer-style models on many spatiotemporal systems that vary geometry, parameters, and initial conditions.
- At inference, provide a short sequence of observed frames (e.g., 10 snapshots) and have the model predict future frames (generative forecasting) without explicitly giving governing equations or parameters.
- Such models can generalize to unseen systems; reported results included visual and quantitative performance near ground truth (one fluid task reported ~5% error on held-out data).
AI for scientific discovery and interpretability
- Data-driven methods can infer governing equations or mathematical models from raw experimental or simulation data.
- Outputs should go beyond symbolic equations: natural-language, interpretable descriptions (e.g., chaotic regimes, likelihood of steady states) are valuable to domain scientists.
Hybrid modeling: aligning classical mathematical models and ML
- When a mechanistic model is incomplete (unknown boundary conditions, uncertain parameters, partial observations), ML can augment and align simulations to observations:
- Fit unknown interaction potentials (e.g., collective motion of fish).
- Augment boundary/initial data to steer simulations (e.g., indoor air/temperature).
- These are effectively alignment and data-assimilation problems.
Optimization is central
- Physical-model-aware optimization: incorporate constraints and conservation laws via loss design, meta-learning, or post-training fine-tuning so learned models respect physical invariants.
- Optimization for LLM training: use matrix-structure–aware optimizers (e.g., “mu”-style algorithms) and add adaptivity to stabilize training and speed convergence; reported empirical improvements over AdamW in some experiments.
Theorem formalization and automated proof
- LLMs are being tested for generating formal proofs (Lean examples). Progress has been rapid: older models struggled, while recent models (cited GPT-5.2 Pro / Gemini 2.5 Pro / Claude variants) can produce longer, readable formalizations for nontrivial results in minutes with hundreds of lines of Lean.
- Combining search (finding obscure references) with step-by-step formalization is crucial. Agent/harness approaches that chain tools and literature search improve performance.
Domain-specific applications combining math and ML
- DNA aptamer selection: use combinatorial features (e.g., Motzkin paths) and secondary-structure representations to prioritize candidate sequences rather than relying only on raw SELEX counts.
- Graph/spectral methods for high-dimensional data: interpret discrete similarity graphs as analogues of continuum operators (Laplacians); use low-rank / Nyström / sparse approximations and a few spectral modes for semi-supervised learning, min-cut, and dimension reduction.
- Subgraph-matching and structural equivalence: exploit combinatorial structure to count or enumerate extremely large numbers of pattern matches in knowledge graphs — useful in bio/medical knowledge discovery and anti-money-laundering.
Verification, transparency and tooling
- Major bottleneck: reliably verifying literature citations and correctness of AI outputs (hallucinations, misattributed references).
- Researchers want provenance, confidence scores, and the ability to inspect why outputs were produced (explainable embeddings, line-level confidence).
- “Harnesses” (tool-chaining, RL/agent frameworks) dramatically increase usefulness by chaining search, tool use, code generation, formalization, and verification steps.
Sociological, educational, and practical concerns
- Students and early-career researchers may feel demoralized or uncertain, but subject-matter expertise remains critical to extract value from AI tools.
- The role of mathematicians evolves toward asking the right questions, proving why, and defining frameworks.
- Practical constraints include access to compute (GPUs / cloud), funding, policy restrictions, and environmental costs (electricity/carbon).
- AI will change job roles but create demand for people who can use and verify AI tools.
Methodologies, workflows, and recommended practices
Building general-purpose spatiotemporal solvers (outline)
- Collect a diverse training corpus of spatiotemporal systems (varying parameters, geometry, initial states).
- Train a large transformer to take short sequences (frames) as input and predict subsequent frames.
- At inference, provide observed frames only; the model infers hidden parameters/physics from its training experience.
- Evaluate using visual comparisons and standard error metrics.
Inferring governing equations / interpretable scientific intelligence
- Input: raw time series or spatiotemporal snapshots.
- Combine ML architectures with symbolic regression and formalization tools to hypothesize equations.
- Produce natural-language summaries describing regimes (chaotic vs steady) or qualitative behavior in addition to formal equations.
Aligning simulations with incomplete data
- Start with an imperfect mechanistic model (e.g., PDE with unknown BCs/parameters).
- Use ML to learn unknown components (potential fields, interaction kernels) from observations.
- Iteratively re-align or correct simulator outputs (data assimilation) during simulation.
Physics-aware optimization
- Enforce physical constraints during training by:
- Designing loss functions that penalize violation of conservation laws/invariants.
- Applying meta-learning or post-training fine-tuning to enforce physics.
- Use matrix-structure–aware optimizers and adaptive schemes for models with large attention or linear layers.
Proof formalization + literature-search harness
Combine:
- Agentic literature search (find obscure papers and appendices).
- LLM generation of stepwise proofs in a formal language (e.g., Lean).
- Automated verification/compilation inside the proof assistant; iterate with human oversight. - Build a harness to chain and reward successful proof steps (RL-style) for longer proofs.
Data-driven combinatorial / graph workflows
- Map domain combinatorial objects into structured features (e.g., RNA/DNA secondary structures to Motzkin paths).
- Build similarity graphs; apply spectral (Laplacian) methods and low-rank approximations (Nyström) for semi-supervised learning and dimension reduction.
- For subgraph matching, identify structural equivalences to efficiently count or enumerate exponentially many solutions.
Concerns raised and practical desiderata
- Verification and provenance: need line-level confidence and accurate citation/tracing mechanisms.
- Hallucinations and literature accuracy: integrate reliable search and citation-checking before trusting AI outputs academically.
- Compute and tooling access:
- Preference for federated/private/local models to protect IP and unpublished work.
- Better cloud/federated options from institutions.
- Tools that capture prolonged interactions (whiteboards, conversations, long chat histories) privately.
- Energy usage: minimize electricity and costs — treat energy as an objective in algorithms and system design.
- Education: teach responsible use of AI (prompting, verification, toolchain design); grade workflows rather than raw AI outputs.
Practical examples and experiments mentioned
- Transformer-based multi-task PDE solver: predict next frames from 10 input snapshots; visually close and ~5% error on one fluid example.
- Fish schooling and indoor air/temperature simulations: ML augmenting mechanistic models to predict densities and fields.
- Formalization anecdote: a gradient-descent convergence proof that older LLMs struggled with; recent models produced a ~200-line Lean proof in about 15 minutes.
- AutoEvolve / algorithm tuning: search and evolution to tune constants in proofs/inequalities — powerful but compute-intensive.
- DNA aptamer selection: using structural features and ML to select candidates beyond raw SELEX counts.
- Subgraph matching in a biomedical knowledge graph: enumerating ~10^18 template matches by exploiting structural equivalence.
Speakers and sources (as named in subtitles)
- Panel moderator (unnamed in the captions)
- Hayden (Haven) Schaefer — UCLA (AI for PDEs, AI-for-science, optimization)
- DK (D. K.) Davis — Pennsylvania (formalization, LLMs for proofs, AutoEvolve)
- Danna (Danna Nidel / possibly “Dana”) — UCLA (optimization, fairness, Lyme disease, Innocence Project applications)
- Andrea (Andre / Andrea) Bertozzi — UCLA (DNA aptamer design, Motzkin-path features, graph spectral methods)
- Other people / organizations referenced:
- Terry, Seb, Mihi (referenced attendees)
- Companies / tools: OpenAI, Axiom Math, AutoEvolve, Claude (Anthropic), Gemini (Google), GPT-5.2 Pro / GPT-4 / ChatGPT
- Institutions / projects: UCLA, Los Alamos National Lab, DARPA (A3ML), LymeDisease.org, Innocence Project / Innocence Center
Note on transcription The provided subtitles were auto-generated and contain misspellings and name ambiguities (e.g., “Haven” vs “Hayden,” “Andre Bossi” vs “Andrea Bertozzi”). Names and spellings above are given as they appear or as inferred; some may be incorrect in the raw captions.
Category
Educational
Share this summary
Is the summary off?
If you think the summary is inaccurate, you can reprocess it with the latest model.
Preparing reprocess...