Back to feed
arXiv cs.CL
arXiv cs.CL
7/2/2026
SLIM-RL: Risk-Budgeted Random-Masking RL for Diffusion LLMs Without Trajectory Slicing

SLIM-RL: Risk-Budgeted Random-Masking RL for Diffusion LLMs Without Trajectory Slicing

Short summary

SLIM-RL introduces a new reinforcement learning method for diffusion language models that matches state-of-the-art training efficiency using only 46% of the samples required by TraceRL. Instead of reconstructing inference trajectories, it bounds commit risk with a tau-budget decoder and adapts variance reduction techniques with a novel mask schedule. Math and code benchmarks show 6-11% improvements; open-source code available.

  • •Achieves SOTA training efficiency with 0.46x sampling cost compared to TraceRL
  • •6-11% performance gains on MATH500, GSM8K, MBPP, and HumanEval benchmarks
  • •Risk-budgeted decoder transfers training-free across LLaDA, Dream, and SDAR architectures

Generated with AI, which can make mistakes.

#research-breakthrough
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL
arXiv CS.AI
When Implausible Tokens Get Reinforced: Tail-Aware Credit Calibration for LLM Reinforcement Learning
arXiv cs.CL
Adaptive Multi-Step Lookahead Decoding for Diffusion Language Models
arXiv cs.CL
Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems
arXiv cs.LG