Toward inverse symbolic learning: back out physics equations from learned deep-net world models
bradneuberg · x · 2026-10-22
bradneuberg poses a research question: for deep-net world models that learn real-world dynamics and physics (say, by observing particle physics simulators), what's the state of the art for backing out symbolic representations of what was learned?
He floats "Inverse Symbolic Learning"—learn the physics in a deep net, then mirror out real scientific discoveries or equations—as the inverse of physics-informed deep nets, which bake physical invariants in. He also frames it as a learned differentiable simulator: unlike hand-built differentiable simulators of the past, it's faster (though approximate) and differentiable, so a loss signal can flow back through it to align the model toward a specific goal for fine-tuning.
More from AGI Musings
- 89% of Biomedical Papers Now Show LLM-Assisted Writing, Data Shows — lemire · 2026-10-22
- One self-driving mistake sparks bans, while human drivers kill daily without outrage — VraserX · 2026-10-22
- Antivenom Designed in 64 Hours by AI, but Release Still Takes 4-7 Years — burny_tech · 2026-10-22
- A Reddit long-read proposes the Risk and Judgement pattern for bounding agent autonomy — mostly_deterministic · 2026-10-22
- Stephen Wolfram's new work tackles time, consciousness and why the universe exists via computation — burny_tech · 2026-10-22
- Sam Altman calls Anthropic's push to ascribe personhood to models a 'dangerous path' — MatthewBerman · 2026-10-22