IDS paper wins NeurIPS oral: AI writes formally verified distributed systems 200x faster
AccBalanced · x · 2026-09-25
Inductive Deductive Synthesis (IDS), a NeurIPS oral paper, lets AI agents co-evolve code and proofs to generate formally verified distributed systems.
- Why: distributed systems need properties (e.g. read-write consistency) to hold under every event interleaving — testing can't guarantee that, and manual formal verification takes experts months to years
- Baseline: SOTA coding agents (Codex with GPT-5.4, Claude Code with Opus 4.6) solve only 2/7 distributed key-value-store specs
- Method: IDS incrementally synthesizes implementation and proof jointly, learning from failures to try promising strategies
- Results: 7/7 specs in 6.8 hours and $106 per spec on average — roughly 200x faster than expert effort and 17% cheaper than SOTA agents, with performance feedback folded into the same loop
- A co-author also used Opus 5.5 + Lean to formally verify the Claude Agent SDK, yielding 16 bug/race-condition PRs from a few prompts.
More from coding & agent
- steipete and mitsuhiko Debate How Coding Agents Should Edit Files — steipete · 2026-09-25
- Nemotron 3 Speaker Diarization Ported to Apple Silicon via Core ML and MLX — ivan_digital · 2026-09-25
- Dev builds conceptual X-16 engine in 3D with GPT-6 Astra and Three.js, open-sourced — techartist_ · 2026-09-25
- Parallel's parallel web search now built into LangChain managed agents — BraceSproul · 2026-09-25
- Same model, 5x cost gap: harness choice matters more than success rate — JeremyCMorgan · 2026-09-25
- TypeSafe AI's Jev evaluation model goes free on Vercel AI Gateway — JohnPhamous · 2026-09-25