Metacognitive Steering paper finds a low-dimensional control surface for scientific judgment inside Kimi 2.6
ceciletamura · x · 2026-10-14
A new arXiv paper (2609.16245) by Vincent Karpf et al. recovers "scientific judgment" from real scientist interaction traces and uses it to steer a frozen frontier model. Contrastive interventions reveal a coordinated low-dimensional control structure in the trillion-parameter MoE Kimi 2.6, confirmed by residual analysis, attention subspace alignment, and cross-layer SVD. Their inference-time controller, Metacognitive Steering, composes layer-specific interventions to switch between exploration, procedural convergence, and critical reassessment without weight updates, and is deployed in the Columbus-1 autonomous research system. The authors present it live on Clubhouse, Oct 14, 5 p.m. PDT.
More from Research
- Six training strategies compared: can LLMs reason over graphs at scale? — burkov · 2026-10-14
- VisInteract: NeurIPS Spotlight reframes text-to-visualization as multi-turn intent retrieval — jiqizhixin · 2026-10-14
- WorldRecon reconstructs large-scale 3D scenes from unposed multi-view images, cutting Chamfer distance 43% — ZhejiangUniversity · 2026-10-14
- RSI-Nav: recursive self-improvement for navigation agents hits 68.6% SR on R2R-CE — OpenGVLab · 2026-10-14
- NUS proposes Jeva, a decision-native architecture with near-perfect null detection at thousands of candidates — NationalUniversityofSingapore · 2026-10-14
- RUC's Agent-Reflex trains reflective agents in three stages, scoring 58.6 on long-horizon tool-use benchmarks — RUC-NLPIR · 2026-10-14