PolyBlocks: A Compiler Infrastructure for AI Chips and Programming Frameworks
2026-03-10T08:52:11Z•21e60fcb070e9188f7a118c28dafcda60338a1a8748e70990d9d917e9f3748fe
Ascend-NPUJAXJIT-compilationMLIRPolyBlocksPyTorchaccelerator-safetyactivation-steeringactive-programmingcompiler-infrastructuredata-exfiltrationmodel-internalspluginprivacyprompt-injectionside-channel-riskspeculative-decodingsupply-chain-risktree-speculationvLLM
What happened
This collection contains five arXiv papers. PolyBlocks describes an MLIR-based, modular compiler infra for ML workloads enabling automatic, target-specific code generation and JIT compilation for PyTorch/JAX with performance comparable to vendor libraries. vLLM Hook is an open-source plugin that exposes and (optionally) allows modification of internal model states in vLLM, supporting passive probing and active interventions (use cases: prompt-injection detection, enhanced RAG, activation steering). Two papers cover specification mining and tracing of creative workflows (data-aware temporalspec
Why it matters
A reviewed impact interpretation has not been published for this record.
Evidence and limitations
- Source ID
- arxiv_cs_pl
- Record identifier
- 21e60fcb070e9188f7a118c28dafcda60338a1a8748e70990d9d917e9f3748fe
- Enrichment time
- 2026-03-10T08:52:11Z
- AI-assisted enrichment
- Yes
This record may overlap with other records. Its enrichment can be incomplete or wrong, and machine assistance was used. Validate consequential decisions against the linked source and your own environment.