Ask a question within this archive's scope (physics, cosmology, ML for physics, etc.) and I'll generate the follow-ups from that answer?

I appreciate the invitation to explore this archive! Let me ask a question that cuts across several of these contributions:

How do current AI systems fail at frontier physics reasoning, and what specific cognitive gaps separate them from working physicists?

This question sits at the heart of your collection. Let me work through the answer:

The Performance Cliff at Research Scale

The most striking finding is a dramatic collapse in capability at genuine research difficulty. While state-of-the-art models like o1 and o3-mini show "impressive progress" on undergraduate and early graduate problems [AG-2025.02-240], only ~6% of full research-scale challenges are solved by the best current systems [AG-2025.09-1139]. This isn't a smooth degradation—it's a cliff.

Three Distinct Failure Modes

1. Tacit reasoning reconstruction. In abstract domains like quantum field theory and string theory, correctness is "layered, tacit, and fundamentally non-binary" [AG-2026.04-565]. Models excel at explicit derivations within stable conceptual frames, but systematically degrade when they must reconstruct omitted reasoning steps or reorganize mathematical representations to satisfy global consistency constraints—the kind of invisible structural knowledge physicists absorb through years of apprenticeship.

2. Domain interoperability and tool orchestration. Individual computational frameworks (from lattice QCD to cosmological simulations) are well-mapped, but physicists routinely pivot between them to search a complex landscape of theoretical models and observables [AG-2025.10-1213]. Current systems struggle to bridge methodologies across subfields in a principled way.

3. Factual coherence under scrutiny. When asked to compare theoretical predictions against multiple data sources and competing paradigms, AI systems produce answers that are "interesting but concerning in accuracy" [AG-2026.02-1185]—plausible-sounding but unreliable summaries that can mislead researchers unfamiliar with the specific subfield.

What Works (and Why)

Fine-tuning on domain-specific synthetic data helps. When researchers generated over 2,500 synthetic quantum field theory problems and fine-tuned 7-billion-parameter models, both reasoning chains and generalization to neighboring physics domains improved measurably [AG-2026.04-892]. This suggests the gap isn't fundamental inability, but rather that frontier physics reasoning requires *continuous exposure* to the particular error modes and conceptual reorganizations that working physicists develop.

The Benchmark as Tool

Rather than treating AI as a black box, the community is now using curated benchmarks—scored not just for correctness but for difficulty, surprise, and reasoning quality [AG-2025.07-1634]—to steer AI development toward genuine research utility. The key insight: *you can't improve what you don't measure precisely*.

Why this matters: If AI-assisted theoretical physics research is to exist, it will require systems that don't just retrieve facts or solve standard problems, but reconstruct the invisible reasoning moves that physicists make when navigating between frameworks. We're not there yet, but the benchmarks now exist to measure progress.