Research map / Architectures and efficiency / RNNs and normalization
PRISM: Parallel Residual Iterative Sequence Model
Jie Jiang, Ke Cheng, Xin Xu, Mengyang Pang, Tianhao Lu, Jiaheng Li, Yue Liu, Yuan Wang, Jun Zhang, Huan Yu, Zhouchen Lin
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 |
|---|---|---|---|
| add iterative refinement (L>0) to the injection operator (ablation) | PRISM · Amazon Electronics | AUC: 0.6805 -> 0.7134 | Improved |
| add iterative refinement (L>0) to the injection operator (ablation) | PRISM · Amazon Electronics | Hit@200: 0.1155 -> 0.1409 | Improved |
| add GELU non-linearity in the refinement loop (ablation) | PRISM · Amazon Electronics | AUC: 0.7047 -> 0.7134 | Improved |
| add ShortConv anchor (ablation) | PRISM · Amazon Electronics | AUC: 0.7076 -> 0.7134 | Improved |
| add ShortConv anchor (ablation) | PRISM · Amazon Electronics | Hit@200: 0.1406 -> 0.1409 | No clear effect |
| add gain predictor (ablation) | PRISM · Amazon Electronics | AUC: 0.7098 -> 0.7134 | Improved |
| add gain predictor (ablation) | PRISM · Amazon Electronics | Hit@500: 0.2406 -> 0.2613 | Improved |
| replace serial TTT iteration with PRISM input-anchored parallel unrolling | TTT · Training throughput on H20 GPU | throughput: 0.34K tokens/s -> 57-61K tokens/s | Improved |
| use Rank-L parallel PRISM instead of explicit serial TTT | TTT · Recommendation benchmarks | Mean Rank: 5.25 -> 2.62 | Improved |
| use PRISM instead of Gated DeltaNet (Rank-1) | Gated DeltaNet · Recommendation benchmarks | Mean Rank: 5.12 -> 2.62 | Improved |
| Rank-L PRISM injection with L refinement projections | PRISM · Training throughput | throughput overhead: <16% overhead vs Rank-1 baselines | Hurt |
| use iterative non-linear PRISM instead of linear multi-expert MoM | MoM / LA · XOR and Parity probing (D=16) | accuracy: 0.49-0.50 -> 1.00 | Improved |
Closest papers by meaning
- Flash Inference: Near Linear Time Inference for Long Convolution Sequence Models and Beyond (ICLR 2025 · similarity 0.85)
- Ladder-Residual: Parallelism-Aware Architecture for Accelerating Large Model Inference with Communication Overlapping (ICML 2025 · similarity 0.84)
- Parallelizing non-linear sequential models over the sequence length (ICLR 2024 · similarity 0.84)
- Routing Mamba: Scaling State Space Models with Mixture-of-Experts Projection (NeurIPS 2025 · similarity 0.84)
- StreamBP: Memory-Efficient Exact Backpropagation for Long Sequence Training of LLMs (NeurIPS 2025 · similarity 0.84)
- Sequence Accumulation and Beyond: Infinite Context Length on Single GPU and Large Clusters (AAAI 2025 · similarity 0.84)
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.
