Act More, Decide Less: Skill-Guided Adaptive Action Chunking for Long-Horizon LLM Agents

Training an LLM agent to emit variable-length action chunks, with chunk boundaries supervised by programmatic skills distilled from successful trajectories, beats single-action ReAct-style agents on success rate while cutting the number of LLM decision rounds.

RL training of LLM agents with terminal binary rewards on two text environments, ALFWorld and ScienceWorld, with seen and unseen splits, against prompting-based and RL-based baselines (including multi-action GRPO and GiGPO-style credit assignment), with ablations on rollout mode ratio and chunk-level test-time search.

Reported gains of 7.0% to 31.3% success over the strongest baseline per setting, decision rounds reduced by 7.4% to 78.9% (abstract says up to 78.9%), and the strongest baseline's final performance reached with only 26.6% of training steps. Scope conditions: two simulated household/science text environments with discrete primitive actions; no real-world or code-execution tasks.

Authors concede naive variable-length RL collapses to single-action or over-long chunks without boundary supervision, and show ablations confirming both components are needed. Unstated: chunk quality depends on how well induced skills generalize to unseen tasks; only two environments; reduced decision rounds means fewer observations, which may silently hurt on tasks requiring tight feedback; inference token cost per round may rise even as rounds fall, and token-level cost is not reported; base model scale and rollout budgets not fully specified in provided text.

Two simulated benchmarks with terminal rewards are exactly where chunking looks best, since routine action sequences dominate; benchmark-specific skill libraries or prompts could leak structure into unseen splits; baselines are all from the same lab family and the strongest baseline is chosen per setting, inviting cherry-picking; success-rate gains and round reductions can trade off against higher tokens per decision, so cost savings are not demonstrated end-to-end; EMNLP camera-ready does not guarantee code and exact prompt sensitivity is unexamined here.

If action chunking transfers beyond these two sims, it cuts inference-side LLM round counts for agentic workloads, which is directly relevant to inference cost per completed task; but with only two synthetic environments and no measured token or dollar cost per task, the honest answer for the economics of AI in the next 12 to 24 months is little or nothing yet.

Technical Capability: Agent builders get evidence that chunk-boundary learning, not just bigger models, is a lever for long-horizon reliability and training efficiency.

Market & Valuation: If agentic inference is priced per round or per token, methods that amortize LLM calls across multiple actions could shift the cost structure of agent products, pending cost measurements this paper does not provide.