The 90% Problem: Why LLM Training Efficiency Just Became the New Arms Race

The race to build reasoning-capable AI models has hit an unexpected wall: computational excess. While companies like OpenAI and DeepSeek chase raw performance, Kwai AI's labs have quietly demonstrated something more threatening to the status quo—that 90% of their reinforcement learning pipeline could simply vanish without sacrificing quality. This isn't marginal optimization. This is the kind of efficiency breakthrough that rewrites unit economics for an entire industry.

Reinforcement learning from human feedback (RLHF) has become the standard recipe for post-training large language models, but it's an expensive one. Group Relative Policy Optimization (GRPO), popular among open-source developers, still demands substantial computational passes to align models with human preferences. The methodology works, but efficiency was never the goal—it was fine-tuning trained models through iterative feedback loops, each one consuming GPU hours and engineering overhead that most labs simply accepted as inevitable cost.

Enter staged reinforcement policy optimization (SRPO), a two-stage framework that fundamentally rethinks this workflow. Rather than treating RL as a linear process of repeated sampling and adjustment, SRPO segments the training into distinct phases and resamples historical trajectories strategically. Early results suggest this approach achieves performance parity with cutting-edge models like DeepSeek-R1 on mathematical reasoning and code generation—competitive benchmarks where efficiency gains typically come with performance trade-offs.

What makes this development significant extends beyond raw efficiency metrics. If validated at scale, SRPO suggests the AI industry has been operating with substantial slack—that billions in infrastructure spending could be redirected toward model diversity, safety research, or commercialization. This challenges the implicit assumption that bigger compute budgets necessarily produce better outcomes. It also raises uncomfortable questions about how many labs have optimized for 'because we can afford it' rather than 'because it's necessary.'

The reaction from frontier labs has been muted but telling. Major players haven't rushed to adopt SRPO publicly, suggesting either skepticism about reproducibility beyond Kwai's specific setup or—more likely—reluctance to acknowledge they've been burning resources unnecessarily. Open-source communities, however, have shown immediate interest, recognizing that efficiency breakthroughs directly translate to accessibility for smaller teams with constrained budgets.

The broader implication is clear: the era of brute-force AI training is entering its sunset phase. Future competitive advantage won't rest solely on scale but on algorithmic cleverness and systems thinking. For journals, startups, and researchers without trillion-dollar budgets, this is genuinely good news. The ladder is getting shorter to climb.