LLMs can generate code amazingly fast — but what happens when the input premise is wrong?

In this episode of Decode: Science, we explore “Refining Critical Thinking in LLM Code Generation: A Faulty Premise–based Evaluation Framework” (FPBench). Jialin Li and colleagues designed an evaluation system that tests how well 15 popular models recognize and handle faulty or missing premises, revealing alarming gaps in their reasoning abilities. We decode what FPBench is, why it matters for AI trust, and what it could take to make code generation smarter.

Gravitational waves whisper across the cosmos — and now, AI might finally hear them with clarity.

In this episode of Decode: Science, we explore “Automated Algorithmic Discovery for Gravitational‑Wave Detection Guided by LLM‑Informed Evolutionary Monte Carlo Tree Search”, by Wang and Zeng (2025). They introduce Evo‑MCTS: an automated, interpretable framework that discovers novel detection algorithms through evolutionary search and large language model heuristics. With over 20% improved accuracy and transparent logic, this paper rewrites how we might detect cosmic signals using AI.

Meet Agent Lightning, a framework that decouples how agents act in the world from how they’re trained—with almost zero code modifications. Introduced in 2025 by Luo et al., this paper reimagines reinforcement learning for AI agents, making it compatible with everything from LangChain to custom agents.

In this episode of Decode: Science, we explore how Agent Lightning formulates agent behavior as an MDP, uses LightningRL for hierarchical credit assignment, and makes scalable agent learning a reality.

Tech paper: https://arxiv.org/pdf/2508.03680