Neuro-symbolic policies let agents reuse workflows, cutting per-run cost up to 217x
xwang_lk · x · 2026-10-02
A new arXiv paper introduces Neuro-Symbolic Computer Use: instead of re-planning every run, agents compile recurring computer workflows into learned, reusable policies.
- Core idea: fix run-stable decisions (ordering, variables, loops, branches) in executable code; delegate observation-dependent decisions (grounding, state checks) to neural models.
- Learning: neuro-symbolic policy iteration starts from one agent trajectory, executes the policy, diagnoses failures with task-completion and step-level judges, and revises code via a coding model informed by continuation from the failure point—without benchmark evaluator access. Parameter/initial-state variants make policies reusable; a pre-action verifier guards each state-mutating step.
- Results: highest Pass^3 across all four settings on OSWorld-Verified and ScienceBoard, 3.6–15.8 points above the base agent, with 15–217x lower per-run cost and 3.4x+ lower latency.
Related event: Neuro-Symbolic Computer Use Cuts Agent Costs by 100x(2 posts)→
More from coding & agent
- Microsoft paper: coding agent optimizing prompts from logs beats GEPA at ~$1.60 — rohanpaul_ai · 2026-10-02
- Stop reaching for the biggest model: a cost-efficient Cursor/Codex setup with GPT-6.1 Sol — chiliraupe · 2026-10-02
- Dot isn't better than Codex or Claude Code — it's a different, undervalued take on OpenClaw/Hermes — gabrielchua · 2026-10-02
- Engineering.com parent Arrowfly launches year-round AI for Engineers initiative amid vibe-CAD era — burhop · 2026-10-02
- Run Fewer Agents: exe.dev argues task management is a band-aid, proposes fast models for human comms — sull · 2026-10-02
- Developer gives an AI agent $1,000 to run a live-streamed hedge fund — kleffew94 · 2026-10-02