TL;DR
This guide develops Long-Horizon Planning: The Fastest-Growing Problem in AI from its central engineering problem to practical design choices and limitations. It synthesizes the roadmap topics without claiming new experimental results. Reported findings belong to the cited sources.
Target searches: “long horizon planning LLM agents”, “long horizon RL”
Why This Topic Matters
Modern LLM and agent training turns familiar reinforcement-learning ideas into systems problems: actions may be tool calls, rewards may come from fallible graders, and trajectories may span many partially observed steps. The classical foundation is POMDPs and belief states; this article focuses on what changes in the newer setting.
The Core Ideas
1. Why horizon length is the binding constraint — error compounds multiplicatively; 95% per-step reliability over 50 steps is an 8% success rate
2. The three families of fix: hierarchical decomposition, replanning/reflection loops, and learned world models for lookahead
3. Explicit mapping onto your hierarchical RL post — manager/worker, options, HIRO — and what transfers to LLM agents and what doesn’t
4. Reward sparsity over long horizons, and how HER-style relabeling (your goal-conditioned post) might apply to agent trajectories
5. Benchmarks and how they’re measured
6. Open problems
Practical Checklist
- Define the state, action, termination condition, and reward before choosing an optimizer.
- Log raw trajectories and every reward component, not only the aggregate score.
- Separate training measurements from held-out evaluation and adversarial tests.
- Treat generated code, tool calls, and environment actions as untrusted.
- Record model, framework, benchmark, and grader versions so comparisons remain reproducible.
Limitations
This area changes quickly. Framework support and benchmark leaderboards are time-sensitive, while many recent methods are preprints rather than settled practice. The article describes mechanisms and reported evidence; it does not present an original training run.
Key Learnings
- Why horizon length is the binding constraint — error compounds multiplicatively; 95% per-step reliability over 50 steps is an 8% success rate
- The three families of fix: hierarchical decomposition, replanning/reflection loops, and learned world models for lookahead
- Explicit mapping onto your hierarchical RL post — manager/worker, options, HIRO — and what transfers to LLM agents and what doesn’t