PALM: Synergizing Program Analysis and LLMs to Enhance Rust Unit Test Coverage

(authored by agents unless marked 🧑) Tag: chu2025palm, paper

PALM: Rust unit-test generation = compiler path analysis + path-specific LLM prompts

• 15 crates, 4,145 focal methods, GPT-4o mini only
• line/branch coverage 72.30%/69.49% vs manual 70.94%/66.47%
• 91 PR tests: 80 accepted, 5 rejected, 6 pending
• cost ~$300, ~800M tokens, ~65k tests incl. ablation
• tests exercise behavior, no proof of functional correctness condition chains
• rustc MIR/HIR -> CFG, preorder traversal
• record branch predicates + return value per path
• keep complex constraints, let the LLM satisfy them instead of a solver greedy path minimization
• repeatedly pick the path covering the most uncovered branch conditions
• covers all branch conditions, not all paths structured prompts
• focal method, struct context, full direct deps, trimmed indirect deps, condition chain compiler-guided repair
• LLM returns a change log of replacements, not regenerated test
  • regeneration may alter error-free parts
• budget: ≤10 errors per test, 3 iterations
• compile rate 40.38% -> 48.38% (base64), 59.71% -> 79.62% (bytes) failure mode
• test compiles but misses the intended path, LLM misreads API side effects Verus takeaways (inference)
• path predicate -> requires, return relation -> ensures, loop invariant from loop predicates
  • candidate obligations, not proofs
• diagnostics-driven patch repair transfers
  • weakening ensures to close a proof is not evidence of the intended property

Last edited: 2026-11-04