Robotics paper index
Where Rounding Happens Matters: Rounding-Location Alignment Stabilizes LLM Reinforcement Learning
One-line summary
A robotics research paper on Where Rounding Happens Matters: Rounding-Location Alignment Stabilizes LLM Reinforcement Learning.
Engineering notes
Engineering notes will be added by the Robot Papers editorial team.
Chinese explanation / 中文解读
中文解读待补充：本站会优先为 VLA、具身智能、人形机器人控制、机器人操作等高价值论文补充中文说明。
Original abstract
Reinforcement learning for large language models often uses separate systems for inference and training. Even with identical weights and BF16 tensors, implementation differences can make an intended on-policy update off-policy and destabilize training. Existing remedies align the entire inference and training stacks, change global precision, or modify the learning algorithm. We ask whether aligning selected parts of the execution stack can stabilize BF16 RL while other implementation differences remain. We align training-side rounding locations with inference in three regions---RoPE, SwiGLU, and residual RMSNorm---by retaining selected intermediates in FP32 until their BF16 outputs. Across four Qwen-family models, aligning all three regions prevents the collapse observed under the default BF16 baseline; RoPE-path alignment alone also suffices in the studied runs, despite remaining numerical discrepancies. The interventions keep the inference engine, matrix-multiplication and attention kernels, and learning objective unchanged.
Links and sources
Need this topic turned into a technical roadmap?
Robot Papers can prepare a custom robotics literature review, code map, dataset map, and B2B technology assessment.
Request B2B research
Comments