Transformer Circuits: Dissecting a Counting Task via the Geometry of Learned Manifolds
burny_tech · x · 2026-10-27
Transformer Circuits published a new thread, "When Models Manipulate Manifolds: The Geometry of a Counting Task," reverse-engineering learned representations for a counting task through the lens of manifold geometry. The author notes that from a pure single-neuron interpretability view, the manifold perspective on learned representations (as used in manifold steering) is missing, along with learned distributed circuits composed of more specific algorithms, such as learned Fourier circuits for arithmetic.
Related event: Transformer Circuits Paper Examines Manifold Geometry of Counting(3 posts)→
More from Research
- Pathology foundation models hit AUROC 0.90 vs ResNet50's 0.81 — maier_ak · 2026-10-27
- TUM study: 19 pathology foundation models find cancer but miss dysplasia 48% of the time — maier_ak · 2026-10-27
- Only 52% of dysplastic slides recognized, TUM pathology study finds — maier_ak · 2026-10-27
- Alibaba DAMO's RynnVLA-Latent Hits 98.8% on LIBERO via 3D-Grounded Latent Actions — Alibaba-DAMO-Academy · 2026-10-27
- AMBIT: Unified FOA Spatial Audio Generation and Editing from HKUST and Noiz AI — NoizAI · 2026-10-27
- AutoTavern: Self-Improving LLM Agents Build a Hearthstone AI Rivaling Top Players — Changlian Ma · 2026-10-27