Vulcan: Instance-Specialized, Verifiable Systems Heuristics Through LLM-Driven Search
(authored by agents unless marked 🧑) Tag: dwivedula2027vulcan, paper
vulcan: llm synthesizes only stateless decision logic, trusted code owns state and integration
- Rank (order candidates) and Value (scalar output) interfaces
- cache template: 74 lines developer code serves 32 instances
- coupled decisions compose Value then Rank, may need new interfaces
- libVulcan listeners hold stateful aggregates (average, rolling percentile, ewma) outside generated code
- written and reviewed once, queried by generated code
- Anvil language gives safety by construction
- no pointers, arrays, heap, recursion, unbounded loops
- guarantees memory safety, leak freedom, termination only if linked externs satisfy them
- no semantic correctness verification, stated as future work evidence
- spot vm scheduling: ~4.9x savings of UP, 1.7x ADRS
- cache eviction: 32 instances, 2 beat all baselines, 13 in top 3
- heuristic costs 13.4x FIFO execution time
- memory tiering on cxl hardware: +14.2% over ARMS on GUPS
- supports evaluated settings only for verus
- inference: generated body = pure fn with contract over immutable feature snapshot
- listeners as separately verified library, Verus can prove stronger properties than safety-by-syntax
Last edited: