NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 28·8 min read
Zipping the Thought: When and How Compressed Reasoning Data Works in LLM Post-Training
TL;DR
Large language models (LLMs) require expensive reasoning data during post-training, but it's unclear when compression—removing redundant steps from reasoning chains—helps or hurts performance. Researchers investigated compressed reasoning data, which condenses multi-step problem-solving into shorter sequences while preserving correctness.
✦ Why It Matters
Engineers can reduce LLM post-training costs by strategically compressing reasoning data based on task type and redundancy patterns.
Key Takeaways
- Explore different CoT types to optimize model training.
- Utilize Composed CoT for better performance with less data.
- Monitor for memorization issues with Implicit CoT.
- Implement RLVR to enhance reasoning step decomposition.
How It Works
The study categorizes CoT into three types: Explicit CoT outputs all reasoning steps, Composed CoT combines steps for efficiency, and Implicit CoT omits intermediate steps. This classification helps in understanding how different representations of reasoning affect model training and performance.
The experiments showed that Composed CoT and Implicit CoT are more scalable, meaning they can leverage larger datasets more effectively than Explicit CoT.