Why LLMs Still Fail: Reasoning, Robustness, and Safety
summary
Claims that state-of-the-art LLMs solve Olympiad math yet miss trivial tasks — e.g. GPT‑5 still fails 10x10 multiplication without tools as of Dec 2025 — and exhibit “jagged intelligence.” Reinforcement learning experiments using GRPO-style methods plus a dense reward signal (scoring intermediate steps via unit tests on new Delta and BounceSim tasks) take a model from ~0% to 100% after ~400 updates, whereas sparse binary reward stalls. Adversarial jailbreaking raises GPT‑4 attack success to ~40%, but adversarially collected data plus fine‑tuning reduces vulnerability.
AI-written summary · about
chapters (13)
- 0:00LLM reasoning challenges
- 1:54Jagged intelligence in LLMs
- 3:28Out-of-distribution limits
- 4:50LLMs struggle with reasoning
- 7:25Math Olympiad performance
- 8:31Reinforcement learning for reasoning
- 11:05Sparse rewards in RL
- 12:48Dense rewards breakthroughs
- 15:25Efficiency vs brute force
- 17:22Safety failures in adversarial settings
- 19:34Safety attack-defense race
- 20:40Safety as continuous process
- 21:46Future of agentic AI
related talks
01h 04mRobust Generalization in the Era of LLMs: Jailbreaking Attacks and Defenses
19mLLMs reasoning and agentic capabilities over time
27mExploring the Mechanisms of LLM Reasoning Capabilities
31mBeyond LLMs and Maturing RAG use cases
01h 22mDebate: Sparks versus embers
17mPsychology of AI
54mKeynote: Can you trust your (large language) model?
46mPlaying the long game