← back
A Taxonomy for Next-gen Reasoning — Nathan Lambert, Allen Institute (AI2) & Interconnects.ai
Takeaway
Future reasoning models need calibration, strategy, and abstraction — not just more skill on math benchmarks — to power real autonomous applications.
Summary
- Nathan Lambert proposes four reasoning traits beyond skills: skills (math/code), calibration (token spend vs difficulty), strategy (knowing direction), and abstraction (decomposing hard tasks).
- RLVR (reinforcement learning with verifiable rewards) unlocked o1/o3/DeepSeek-style step-changes; now also enabling new applications like Deep Research, Claude Code, autonomous Codex.
- Overthinking is a real problem: reasoning models burn hundreds of tokens on '2+3'; calibration must move from user-facing model selectors into the model itself.
- Tool use combined with reasoning (o3-style) is the next plateau; abstraction (model decomposing tasks autonomously) is the hardest unsolved capability.
reasoningreinforcement-learningfoundation-models
Original description
Current AI models are extremely skilled, which was seen as the step change in evaluation scores across the industry in the first half of 2025, but often fail when presented with even medium time-horizon tasks. This talk presents a taxonomy of 4 traits of reasoning models -- skills, calibration, strategy, and abstraction -- that will be crucial to creating the next generation of AI applications. With this, we focus on the latter two, strategy and abstraction, and discuss how these traits will enable long-horizon and reliable agents. The talk concludes with a scenario where these agentic behaviors are the foundation for RL continuing to scale in the coming years and post-training techniques reaching compute parity with pretraining methors sooner than later. About Nathan Lambert Nathan Lambert is a Senior Research Scientist and post-training lead at the Allen Institute for AI focusing on building open language models. At the same time he founded and operates Interconnects.ai to increase transparency and understanding of current AI models and systems. Previously, he helped build an RLHF research team at HuggingFace. He received his PhD from the University of California, Berkeley working at the intersection of machine learning and robotics. He was advised by Professor Kristofer Pister in the Berkeley Autonomous Microsystems Lab and Roberto Calandra at Meta AI Research. He was lucky to intern at Facebook AI and DeepMind during his Ph.D. Nathan was was awarded the UC Berkeley EECS Demetri Angelakos Memorial Achievement Award for Altruism for his efforts to better community norms. Recorded at the AI Engineer World's Fair in San Francisco. Stay up to date on our upcoming events and content by joining our newsletter here: https://www.ai.engineer/newsletter Timestamps: [00:00] The Current State of Reasoning in AI Models [01:06] Unlocking New Language Model Applications [03:48] The Need for Advanced Planning in AI [04:29] A Proposed Taxonomy for Next-Generation Reasoning [06:16] Reinforcement Learning with Verifiable Rewards [08:23] Current Challenges and Future Directions [12:07] The Effort Required to Build New Capabilities [16:20] A Research Plan for Training Reasoning Models [17:36] The Shift in Compute Allocation from Pre-training to Post-training