FlowBank (NeurIPS): precompute-and-reuse agent workflow bank beats baseline 73.40 vs 70.40
furongh · x · 2026-10-05
The team releases the NeurIPS paper FlowBank: Query-Adaptive Agentic Workflows Optimization through Precompute-and-Reuse (arXiv:2606.11290; Yuan, Deng, Yu, Chakraborty, Rostami, Huang), with paper, project page, and code public.
- Problem: task-level optimizers spend heavy offline compute but deploy one workflow, wasting complementary candidates; query-level generation is costly per query. Motivating analysis shows the paradigms are complementary—offine-discovered workflows solve different query subsets
- Method: a three-stage portfolio framework—generate complementary candidates (DiverseFlow steers search to under-covered queries), compress into a small deployable bank, and assign each query under a performance-cost trade-off
- Results: 73.40 avg across five benchmarks vs 70.40 for the strongest automated baseline at lower inference cost, all with the same GPT-4o mini executor; on MATH the curated bank scored 69.34 vs 68.11 for the full pool
More from coding & agent
- Free AI Engineering course covers foundation models, evals, prompt engineering and security — Negative_War_65 · 2026-10-05
- Most coding agents today are hand-rolled loops, not framework abstractions — arpit_bhayani · 2026-10-05
- Indie dev tinyfool starts prototyping his next game, now stress-testing game logic engine — tinyfool · 2026-10-05
- Five questions to ask before letting an AI agent call a state-changing tool in production — ainexfinder · 2026-10-05
- Flowsint, an open-source graph-based OSINT investigation platform, hits 9.1k GitHub stars — tom_doerr · 2026-10-05
- AWS launches Well-Architected Agent that scans 65+ services and ranks risks by your business goals — DigitalColmer · 2026-10-05