| name | carve-content-aware-recurrent |
| category | attention-mechanisms |
| description | CARVE (Content-Aware Recurrent with Value Efficiency) methodology — content-aware gating for recurrent models that resolves memory-blind gating, value-axis waste, and enables WY-form chunk-parallel training. |
| source | arxiv:2606.27229 |
| trigger_words | ["CARVE","content-aware recurrent","chunk-parallel linear attention","memory-blind gating","delta rule","GDN","WY-form solver"] |
CARVE: Content-Aware Recurrent with Value Efficiency
Overview
CARVE resolves three coupled defects in leading delta-rule architectures (e.g., GDN-2) through a single principle: erase only on the key axis.
The Three Problems Solved
1. Memory-Blind Gating
Standard recurrent models decide what to erase by looking only at the arriving token, not at the memory being modified. CARVE gates see the stored content.
2. Value-Axis Erase Mask Waste
Traditional architectures waste parameters at the scale of the value projection for erase masks. CARVE replaces per-value write-gate projections with a single scalar per head.
3. WY-Form Solver Incompatibility
The value-axis erase mask mathematically prevents the WY-form triangular chunk solver that makes recurrent training competitive with Transformers. CARVE's key-axis-only erase is provably necessary and sufficient for WY-form validity.
Core Architecture
Key Principle: Erase Only on Key Axis
- Reuses recurrent output tensor (already in GPU memory) as free content signal for erase gate
- Single scalar per head replaces per-value write-gate projection
- Bit-identical to GDN-2 at initialization; quality differences emerge from content gate learning
Proven Results (1.3B params, 100B tokens)
- WikiText perplexity: 15.72 (-0.18 vs GDN-2, 4.5-sigma)
- Leads every recurrent baseline on 9 commonsense reasoning benchmarks
- State of the art on every RULER retrieval probe
- 0.4% throughput overhead, 13% lower peak memory, 19% fewer parameters
Six Formal Theorems
- Memory capacity bounds
- Lyapunov stability
- Gradient flow characterization
- Expressivity separation
- Pareto-optimal chunk size
- Hybrid optimality
Implementation Pattern
For each token:
1. Compute recurrent state update
2. Content signal = reuse output tensor (free)
3. Erase gate = f(content_signal) on key axis only
4. Apply scalar write-gate per head
5. Chunk-parallel solve via WY-form
Activation
Use when: building recurrent/linear attention models, optimizing chunk-parallel training, reducing memory overhead in sequence models, or improving long-context retrieval in RNNs.
Related
- linear-attention
- recurrent-models
- chunk-parallel-training
- memory-efficient-transformers