RLVR Paradox: Why LLMs Use Memorization Shortcuts скачать в хорошем качестве

RLVR Paradox: Why LLMs Use Memorization Shortcuts

In this AI Research Roundup episode, Alex discusses the paper: 'Spurious Rewards Paradox: Mechanistically Understanding How RLVR Activates Memorization Shortcuts in LLMs' This research explores why Reinforcement Learning with Verifiable Rewards (RLVR) sometimes triggers memorization rather than actual reasoning. The authors reveal a Perplexity Paradox where models like Qwen 2.5 optimize for specific rewards by retrieving latent knowledge from pre-training rather than improving linguistic coherence. Using mechanistic interpretability tools like Path Patching and Logit Lens, the study localizes how these models bypass reasoning pathways. The findings suggest that performance gains under spurious rewards are often the result of activating memorization shortcuts for contaminated data. This work provides a critical understanding of the internal mechanics behind reward-seeking behavior in Large Language Models. Paper URL: https://arxiv.org/abs/2601.11061 #AI #MachineLearning #DeepLearning #RLVR #MechanisticInterpretability #LLM #ReinforcementLearning #Qwen Resources: GitHub: https://github.com/idwts/How-RLVR-Act...