Inkling-2 uses a 48-layer decoder-only transformer with 594B total parameters and 17B active parameters via a Sparse MoE feed-forward backbone. The author remarks that the model returns to ViT and calls the earlier hMLP encoder approach “encoder-free”.
Key points
Architecture: 48-layer decoder-only transformer.
Parameter count: 594B total, 17B active.
Backbone: Sparse MoE feed-forward.
Sources
Heuristic Learning shows rule-based policies reaching RL scores
Heuristic Learning lets a coding agent edit program code instead of neural-network weights. Rule-based policies built this way achieved scores comparable to Deep RL on several benchmarks. In Atari Breakout the program reached the theoretical maximum score, and in MuJoCo Ant and HalfCheetah the policies hit scores in the range of common Deep RL results. The method provides explainable, sample-efficient systems with regression testing.
Key points
Breakout policy progressed from 387 → 507 → 839 → 864 and reached the theoretical maximum score using only programmatic rules.
Ant policy reached 6000+ score and HalfCheetah achieved a five-episode mean of 11836.7, both comparable to typical Deep RL results.
Sources
- Original post
- Learning Beyond Gradients - Linked page
LaMARia spatial AI system outperforms Meta on benchmarks
LaMARia spatial AI system ranked #1 in three of five core localization benchmark categories, surpassing major closed-source models. The system also supports precise GPS-denied navigation and mapping for tactical operations. UZH-FPV achieved #1 for high-speed tracking on aggressive drone maneuvers.
Key points
LaMARia: 1 in 3/5 categories, beating major closed-source models in low-light, medium and long-range environments.
UZH-FPV: 1 for high-speed tracking, maintaining solid localization during aggressive, high-G drone maneuvers.
Sources
VINO releases Volume Transformer for 3D learning
VINO introduces a Volume Transformer for self-supervised representation learning on 3D scenes. The work claims to close the gap to supervised representations. A paper and project page are linked.
Key points
Paper: VINO: Revisiting 3D Representation Learning with Volume Transformers (arXiv 2611.07176)
Collaboration includes @adr_kruse, @KadirYilmaz_CV, @dcdegeus, @BastianLeibe, @RWTHVisionLab, @USI_INF, @TUeindhoven1
