AI

TRIAGE: Direction-Aware Mismatch Stabilization of Native NVFP4 Reinforcement Learning

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 00:18:24 UTC.

Briefings

This study characterizes how execution mismatches between learners and samplers destabilize policy optimization in large language models using native NVFP4. The research introduces TRIAGE, a direction-aware stabilization method that rebalances policy-gradient updates via segment-level diagnosis and bounded repair. Experiments on Qwen3-4B and Qwen3-30B-A3B demonstrate stable optimization across training horizons and full-precision performance on five mathematical reasoning benchmarks, achieving up to 2.3x higher rollout throughput compared to BF16.

Key points

• Native NVFP4 execution exhibits early imbalance between two amplifying regions, favoring negative-advantage, negative-gap updates.
• TRIAGE modifies the optimization objective while retaining native NVFP4 weight-and-activation 4-bit (W4A4) forward execution on both sampler and learner.

Context

Low-precision execution can accelerate RL for large language models, but discrepancies between learner and sampler execution can destabilize policy optimization. This research characterizes the interaction between mismatch and policy-gradient direction, distinguishing locally amplifying from contracting update contributions that mismatch magnitude alone cannot identify.

Limitations

Evidence is limited to native NVFP4 runs on Qwen3-4B and Qwen3-30B-A3B. Findings on other architectures or precision levels remain unverified. Results are based on a single snapshot from October 2026, with no additional background context or causal claims beyond the observed instability and stabilization mechanisms.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/11/2026, 16:36:35 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigddbkjmk3glvtzfwf62kuzkfzfpddvcwturs4et3amxl7ouyyks4