Tencent Hunyuan Maps the Geometry of RLVR in LLMs, Releases Alpha-Stabler Framework
Tencent-Hunyuan · hf · 2026-10-09
Tencent Hunyuan researchers use vector steering to identify a low-dimensional effective manifold in activation space behind RLVR gains.
- Two findings: the capacity to reproduce RL gains is small but not infinitely compressible; effective control directions lie mainly in the low-variance complement of the activation principal subspace
- Geometry stays consistent across training configurations within a task; cross-task alignment correlates with capability transfer, validated on 5 LLMs and 6 verifiable-reward tasks
- Proposes Alpha-Stabler: a Predictor for early collapse warnings and a Controller that removes principal-subspace components of activation gradients
- Stabilizes training for 2,000 steps and consistently improves RL gains; code is open source
More from Research
- U-Space finds an interpretable subspace for LLM uncertainty, no training needed — Tobias Braun · 2026-10-09
- CARE certifies VLA inference speedups up to 10.8x with statistical guarantees — UMCP · 2026-10-09
- SOL: a sample-based distributional metric proposed for evaluating text diffusion LMs — NandoDF · 2026-10-09
- MaRN: PyTorch library cuts MNIST CNN params 57.7x via low-dim parameter mappings — Less_Dream_6331 · 2026-10-09
- After 3 days at an AI-Talmud workshop: how do you eval AI where experts disagree? — EhudReiter · 2026-10-09
- TRL v1.15 ships fused LM head: 82% less peak VRAM, 7x longer sequences — QGallouedec · 2026-10-09