How much do LLM memorize?

• key definition
  • unintended memorization: memorize a specific dataset
  • generalization (intended memorization): contains about the true data-generation process
  • calculation method: by information entropy and mutual information
• double desent appear on the changing points from unintended memorization into generalization
• GPT-models store 3.6bits data per parameters
• value of float32 is 9% higher than float16

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

• trade-off between pre-train model size and inference-time(inference length)
• performance can ouperform 14x size model
• performance is better in easy and medium problem, judge easy, medium or hard question based on the pass rate
• two ways to increase inference-compute-time
  • best-of-N: sample N outputs parallel and choose the best one based on learned verifier or reward function.
  • revise response: revise original response

Prolonged Reinforcement Learning

• tempurature: increase tempurate to avoid entropy collapse
• decoupled clip to increase exploration space
• dynamic sampling: erase all truely right or wrong samples
• calculate loss from sample level into token level – DAPO
• KL-regularization and reference model reset

illusion of thinking

• for hanio tasks
  • lower performance in simple-level question for reasoning model than general mdoel because it get wrong answer when thinking even already get correct answer [over thinking]
  • better performance in medium-level question
  • zero-performance in hard question

Gemini 2.5 tech report

• dataset
  • ensure dataset quality
  • fliter and drop duplicates
• post-training
  • verifiable reward and model-base generative rewards to provide sophisticated and scaleable feedback signals
    • verifiable reward
    • model-base: more sophisticated and scaleable feedback signals
  • update LR-method to improve the stability of training
  • result: learning in complex space

This post is licensed under CC BY 4.0 by the author.