docs: add the future roadmap of this work

Author: moven0831

README.md

@@ -158,6 +158,28 @@ Key changes:
 * CSR buckets keep data on-device → near-zero host↔GPU traffic  
 * on-GPU radix sort makes preprocessing parallel
 
+## Future
+
+### Technical Improvements
+- **Modern Dependencies**: Update to `objc2` and `objc2-metal` ([objc2](https://github.com/madsmtm/objc2))
+- **Metal 4**: Adopt latest [Metal 4](https://developer.apple.com/metal/whats-new/) features
+- **Refactor with SIMD in mind**:
+  - Instruction-level parallelism using vector types for faster FMA within SIMD groups
+  - Memory coalescing to increase locality (e.g., structure of array instead of array of structure)
+  - Optimized input reading patterns (e.g. `[X_i || Y_i]_0^{n-1}` instead of separate arrays)
+  - Latency hiding and occupancy fine-tuning
+  - Minimize thread divergence
+
+### Algorithm & Integration
+- **CPU-GPU Hybrid**: Research interleaving with CPU MSM crate and update to `arkworks 0.5`
+- **Advanced Algorithms**:
+  - [Elastic MSM](https://eprint.iacr.org/2024/057.pdf) implementation
+  - Faster modular reduction with LogJump ([article by Wei Jie](https://kohweijie.com/articles/25/logjumps.html), [Barret-Montgomery](https://hackmd.io/@Ingonyama/Barret-Montgomery))
+
+### Platform Expansion
+- **Cross-platform**: WGSL support with native execution environment
+- **Crypto Math Library**: Maintain a Metal/WebGPU crypto math library
+
 ## Community
 
 -   X account: <a href="https://twitter.com/zkmopro"><img src="https://img.shields.io/twitter/follow/zkmopro?style=flat-square&logo=x&label=zkmopro"></a>