2026 Conference on Physics and AI: Shihui Zang

The talk presents a compelling and well-structured approach to automating cosmological data analysis using LLMs and evolutionary algorithms. The motivation is clear: as datasets grow, traditional manual algorithm development becomes unsustainable. The proposed method, Mad Evolve, builds on prior work like FunSearch and AlphaEvolve, adding a crucial step of parameter optimization using JAX to ensure each candidate algorithm is evaluated at its best performance. This addresses a common pitfall in evolutionary search where poorly tuned algorithms may be unfairly discarded. The results across three distinct cosmological tasks demonstrate consistent improvements over human baselines, with the caveat that the quality of the initial algorithm significantly impacts the final outcome. This highlights the continued importance of human expertise in guiding the search. The cost analysis is striking, showing a reduction from ~$40,000 per year for a graduate student to under $30 per generation, making such research accessible to a wider community. However, the talk lacks detailed discussion of potential limitations, such as the risk of overfitting to specific datasets or the generalizability of the approach to other domains. The speaker also does not provide specific numerical results or error bars, making it difficult to assess the statistical significance of the improvements. The presentation is clear and well-paced, with effective use of diagrams to illustrate the pipeline and tasks. The adéquation between title and content is excellent, as the talk directly addresses the intersection of physics and AI. Overall, this is a valuable contribution that demonstrates the potential of LLM-driven automation in scientific research, while appropriately acknowledging the ongoing role of human researchers.

The title accurately reflects the content: a conference talk on applying evolutionary algorithms and LLMs to cosmological data analysis.

8/10

Presentation of original research at a reputable academic conference (Stanford/APS/NeurIPS), with clear methodology and quantitative results. However, limited peer-reviewed publication details and no external verification.

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

• 00:00Introduction and motivation for using LLMs in cosmology
• 02:00Overview of cosmological data pipeline and challenges
• 05:00Explanation of evolutionary algorithm with LLM mutation
• 08:00Introduction of parameter optimization step using JAX
• 10:00Description of three cosmological tasks
• 12:00Results showing improvements over human baselines
• 14:00Cost analysis and comparison with graduate student costs
• 16:00Discussion on importance of human initial algorithms

• 2026 Conference on Physics and AI (PAI26) — Official conference page providing context and details about the event.

• FunSearch: Mathematical discoveries from program search with large language models — Demonstrates the effectiveness of LLM-driven evolutionary search in mathematical problems, supporting the approach used in Mad Evolve.
• AlphaEvolve: A coding agent for scientific and mathematical discovery — Extends the evolutionary LLM approach to code generation, aligning with Mad Evolve's methodology.

• No discordant sources identified — The talk does not mention any conflicting sources or studies.