Research map / Architectures and efficiency / Attention and transformers
Kalman Linear Attention: Parallel Bayesian Filtering For Efficient Language Modeling and State Tracking
Vaisakh Shaj, Cameron Barker, Aidan Scannell, Andras Szecsenyi, Elliot J. Crowley, Amos Storkey
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 the final attention layer of GPT with a KLA block | GPT · Average zero-shot, 180M | accuracy: 42.20 -> 44.95 | Improved |
| replace the final attention layer of GPT with a KLA block | GPT · Average zero-shot, 45M | accuracy: 34.78 -> 38.33 | Improved |
| add probabilistic decoding (KLA+) to KLA | KLA · MAD Compression | accuracy: 85.0 -> 88.9 | Improved |
| add probabilistic decoding (KLA+) to KLA | KLA · MAD Memorization | accuracy: 98.9 -> ~99.9 | Improved |
| add process noise p_t (versus p_t=0) (ablation) | KLA · MAD Selective Copy | accuracy: -14.9 | Improved |
| add process noise p_t (versus p_t=0) (ablation) | KLA · MAD Compression | accuracy: -12.1 | Improved |
| add process noise p_t (versus p_t=0) (ablation) | KLA · MAD average | accuracy: 4.7-point average drop | Improved |
| add process noise p_t (versus p_t=0) (ablation) | KLA · MAD Memorization and Fuzzy/exact recall | accuracy | No clear effect |
| add mean-reverting OU prior for stacking (ablation) | KLA · deep stacking (12 and 24 blocks) | training stability | Improved |
Closest papers by meaning
- Simple linear attention language models balance the recall-throughput tradeoff (ICML 2024 · similarity 0.86)
- LongMamba: Enhancing Mamba's Long-Context Capabilities via Training-Free Receptive Field Enlargement (ICLR 2025 · similarity 0.86)
- Implicit Language Models are RNNs: Balancing Parallelization and Expressivity (ICML 2025 · similarity 0.86)
- (How) Do Language Models Track State? (ICML 2025 · similarity 0.85)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.85)
- Attention-Based Sampler for Diffusion Language Models (NeurIPS 2026 · similarity 0.85)
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.
