Research map / Architectures and efficiency / RNNs and normalization
Recursive Binding on a Budget: Subspace Carving in Order-$p$ Tensor Memories
Travis Pence, Daisuke Yamada, Vikas Singh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace VSA binding with OSC subspace carving at 1000 bundles | memory architecture · Synthetic memory, 1000 bundles, p=2 | retrieval accuracy (%): 13.18 -> 99.77 | Improved |
| use OSC with p=2 instead of best VSA d=4096 | memory architecture · Synthetic memory, 1000 bundles | total memory (parameters): 4100096 -> 440000 | Improved |
| use OSC p=3 d=80 instead of HLB d=343000 at 10000 bundles | memory architecture · Synthetic memory, 10000 bundles, p=3 | total memory (parameters): 3430343000 -> 2912000 | Improved |
| use OSC instead of TPR at small superposition N=50 | memory architecture · Min filler dimension for 99% accuracy | total parameters: 4K -> 4K | No clear effect |
| use OSC instead of TPR at large superposition N=10000 | memory architecture · Min filler dimension for 99% accuracy | total parameters: 6.9M -> 2.7M | Improved |
| hold memory size constant and compare capacity with HLB | OSC · Synthetic memory, p=3 | memory capacity | Hurt |
| use OSC instead of HLB for retrieval at small superposition | OSC · Retrieval time, N=100, B1 | speed (ms): 0.068 -> 0.149 | Hurt |
| use OSC instead of HLB for retrieval at large superposition | OSC · Retrieval time, N=10000, B2 | speed (ms): 1.189 -> 0.306 | Improved |
| increase codebook size at fixed superposition (ablation) | OSC · Min filler dimension for 99% accuracy | required filler dimension | No clear effect |
| use OSC in extreme multi-label learning instead of VTB/MAP/HLB | OSC · XML EURLEX-4K | nDCG@5: 52.54 -> 59.70 | Improved |
| use OSC in extreme multi-label learning instead of VTB | OSC · XML EURLEX-4.3K | nDCG@5: 75.91 -> 71.59 | Hurt |
| use OSC in extreme multi-label learning instead of HLB | OSC · XML DELICIOUS-200K | nDCG@5: 38.80 -> 37.28 | Hurt |
Closest papers by meaning
- Composing Linear Layers from Irreducibles (NeurIPS 2025 · similarity 0.85)
- Transport of Algebraic Structure to Latent Embeddings (ICML 2024 · similarity 0.83)
- Fully Distributed, Flexible Compositional Visual Representations via Soft Tensor Products (NeurIPS 2024 · similarity 0.82)
- Algebraic Positional Encodings (NeurIPS 2024 · similarity 0.82)
- From Low Rank Gradient Subspace Stabilization to Low-Rank Weights: Observations, Theories, and Applications (ICML 2025 · similarity 0.82)
- How Do Transformers Learn Variable Binding in Symbolic Programs? (ICML 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
