CARVE: Content-Aware Recurrent with Value Efficiency for Chunk-Parallel Linear Attention

The core contribution is not just a performance improvement but a mathematical compatibility fix: prior delta-rule models placed gating on the value axis in a way that algebraically blocked use of the WY-form triangular solver, forcing a choice between architectural expressiveness and training efficiency. CARVE resolves that conflict by moving the gating to the key axis, making the two techniques composable rather than competing.

This connects to a broader pattern in recent coverage: researchers repeatedly finding that architectural or data-engineering choices matter more than raw scale. The linear-model forecasting paper from the same day ('How Good Can Linear Models Be for Time-Series Forecasting?') made a structurally similar argument, showing that careful design decisions on simpler models can close gaps that practitioners assumed required heavier compute. CARVE extends that logic into the pretraining regime for sequence models, where recurrent architectures have long been theoretically attractive for long-context and memory-constrained settings but practically hobbled by training speed gaps versus Transformers.

Watch whether GDN-2 or a CARVE-based successor appears in a public long-context benchmark comparison against Transformer baselines within the next two quarters. If training throughput parity holds at the 1B-plus parameter scale, the efficiency argument for recurrent models in production becomes substantially harder to dismiss.