EAGLET: Efficient Planner Training

• EAGLET is an efficient planner training method that decouples global planning from low-level execution to reduce trial-and-error and planning hallucinations in long-horizon tasks.
• It employs a two-step process where high-quality plans are generated by an advanced LLM and filtered using homologous consensus filtering to ensure compatibility with diverse executor agents.
• The method achieves state-of-the-art performance across benchmarks, reducing training cost by 8× and improving success rates and trajectory efficiency.

EAGLET is an efficient and effective planner training method for long-horizon agent tasks within a plan-and-execute framework. It is designed to address two failure modes reported for LLM-based agents in such settings: brainless trial-and-error and planning hallucinations arising from the absence of explicit global planning. In EAGLET, a plug-and-play global planner is trained without human effort through a two-step process: high-quality plans are synthesized from an advanced LLM and filtered by a homologous consensus filtering strategy, then used for cold-start supervised fine-tuning; the planner is subsequently refined by a rule-based reinforcement learning stage using an executor capability gain reward. Across ScienceWorld, ALFWorld, and WebShop, executor agents equipped with the planner outperform existing methods, while training cost is reduced by $8\times$ relative to RL-based baselines (Si et al., 7 Oct 2025).

1. Problem setting and motivation

Long-horizon agent tasks require an agent to make a sequence of tens or even hundreds of decisions in order to complete a high-level instruction such as “conduct a chemistry experiment,” “prepare dinner in a simulated house,” or “buy an item on a website” (Si et al., 7 Oct 2025). In the formulation underlying EAGLET, standard LLM-based agents typically rely on local or on-the-fly planning through interleaved chain-of-thought prompting, exemplified by ReAct-style behavior. The reported consequence is a pair of recurrent deficiencies: repeated actions or random search without an explicit high-level plan, and reasoning steps that become inconsistent with the environment and therefore produce useless or invalid actions.

EAGLET is situated as a response to the global planning challenge rather than as a replacement for action execution. The central premise is that explicit high-level planning provides “global foresight,” which mitigates hallucinations and reduces environment interactions. A plausible implication is that EAGLET should be understood less as a new executor architecture than as a training method for a planner that can be attached to heterogeneous executor agents.

The method is explicitly described as not requiring manual effort or extra training data. This point is important because a common misconception in planner-augmented agent systems is that improved planning quality necessarily depends on human-authored decomposition traces. In EAGLET, the training signal is instead induced from synthesized plans, homologous filtering, and rule-based reinforcement learning (Si et al., 7 Oct 2025).

2. Plan-and-execute architecture

EAGLET decouples high-level global planning from low-level action execution. A lightweight global planner $\pi_g$ receives a task instruction $u$ and outputs a step-by-step abstract plan $p$. An executor agent $\pi_\theta$ then consumes $(u,p)$ together with the environment state and generates concrete actions. The interaction is formalized as

$$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$

Within the architecture, the global planner operates behind the scenes, whereas the executor agent observes $(u,p,a_{1:t-1},o_{1:t-1})$ at each environment step $t$ and emits an action $a_t$. The environment then returns a new observation $\pi_g$0. During training, the planner is refined offline, and no gradient flows through the executor (Si et al., 7 Oct 2025).

This separation of responsibilities defines EAGLET’s role precisely. The planner is responsible for abstract sequencing, while the executor remains responsible for environment-grounded control. This suggests that the framework is intended to preserve executor modularity: the planner can be attached as a plug-and-play component rather than requiring end-to-end retraining of the action policy.

3. Plan synthesis, homologous consensus filtering, and cold-start training

The first stage of planner training synthesizes candidate plans from a powerful LLM such as GPT-5 or DeepSeek-V3.1-Think. For each training trajectory $\pi_g$1 and instruction $\pi_g$2, the synthesizer generates both a high-level plan $\pi_g$3 and chain-of-thought $\pi_g$4. The prompt template encloses <task>…</task> and <conversation>…</conversation>, then elicits <plan>…</plan> (Si et al., 7 Oct 2025).

EAGLET does not accept every synthesized plan. Instead, it applies a homologous consensus filtering strategy based on two homologous executor models, $\pi_g$5 and $\pi_g$6, described as novice and expert executors that share architecture and pre-training but differ only in post-training skill. For each candidate plan $\pi_g$7, each executor is rolled out twice, with and without the plan, to obtain completion rates $\pi_g$8 and $\pi_g$9. The plan is retained iff it does not degrade either executor:

$u$0

The paper gives the following pseudocode for dataset construction:

$(u,p,a_{1:t-1},o_{1:t-1})$3

The resulting filtered dataset is $u$1. The planner is then cold-started by supervised fine-tuning with maximum likelihood:

$u$2

The optimization details are specified as follows: the backbone is Llama-3.1-8B-Instruct; optimization uses Adam with learning rate $u$3, weight decay $u$4, batch size $u$5, and $u$6 epochs; mixed precision and ZeRO3 are used via DeepSpeed (Si et al., 7 Oct 2025).

The filtering stage is structurally important because it operationalizes “high-quality” plan selection through non-degradation on both novice and expert homologous executors. A plausible implication is that EAGLET treats plan utility as executor-relative rather than purely linguistic: a plan is acceptable only if it is compatible with action policies of different capability levels.

4. Rule-based reinforcement learning and the executor capability gain reward

After cold-start supervised fine-tuning, EAGLET further improves the planner with a rule-based reinforcement learning stage. The objective is to reward plans that both improve executor performance and shorten trajectories (Si et al., 7 Oct 2025).

For each plan $u$7 and each executor $u$8, EAGLET defines a binary gain signal:

$u$9

A decay factor $p$0 is added to encourage fewer steps:

$p$1

where $p$2 is the number of steps with the plan, $p$3 is the number of steps without the plan, and $p$4. The rewards are then aggregated across both executors, along with a format adherence reward $p$5:

$p$6

The planner is optimized with Grouped Relative Policy Optimization (GRPO). For each instruction $p$7, $p$8 candidate plans $p$9 are sampled and assigned rewards $\pi_\theta$0. Within-group advantages $\pi_\theta$1 are computed, and optimization proceeds with

$\pi_\theta$2

where $\pi_\theta$3, and $\pi_\theta$4 control update size and KL penalty (Si et al., 7 Oct 2025).

The key RL hyperparameters are reported as group size $\pi_\theta$5, rollout temperature $\pi_\theta$6, evaluation temperature $\pi_\theta$7, mini-batch $\pi_\theta$8, KL penalty $\pi_\theta$9, clip $(u,p)$0, and decay $(u,p)$1. Executor max steps are $(u,p)$2 for ALFWorld and WebShop and up to $(u,p)$3 for ScienceWorld.

A common misunderstanding would be to read this stage as standard reward-model-based RL. The formulation given for EAGLET is instead rule-based: reward is computed from executor capability gain and format adherence, not from a learned preference model.

5. Experimental setting and quantitative results

The empirical evaluation uses three long-horizon agent benchmarks: ScienceWorld, ALFWorld, and WebShop. ScienceWorld and ALFWorld are evaluated in seen and unseen settings, whereas WebShop is evaluated in a seen-only setting. The reported metrics are final average reward in $(u,p)$4, success rate as binary completion, and average steps to completion (Si et al., 7 Oct 2025).

The paper reports that executor agents without training, including Llama-3.1-8B, GPT-4.1, GPT-5, and DeepSeek variants, gain up to $(u,p)$5 points when equipped with EAGLET. In the lower half of Table 1, the strongest trained baseline is GiGPO + MPO + Llama-3.1 with average $(u,p)$6, while EAGLET + GiGPO reaches $(u,p)$7, reported as $(u,p)$8.

Method	ScienceWorld	ALFWorld	WebShop
GiGPO + MPO	78.2	88.1	83.5
EAGLET + GiGPO	83.6	91.8	86.2

The corresponding average values are $(u,p)$9 for GiGPO + MPO and $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$0 for EAGLET + GiGPO (Si et al., 7 Oct 2025).

The efficiency results are equally explicit. EAGLET requires approximately $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$1 RL iterations, compared with approximately $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$2 for GiGPO, which is described as an $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$3 reduction. Both methods are marked as data-efficient in the reported table. In addition, average steps to success drop from approximately $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$4 to approximately $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$5 with EAGLET, together with better reward.

These results are presented as evidence that planner training can improve both effectiveness and efficiency. The effectiveness claim is benchmark-relative—new state-of-the-art performance on the three long-horizon tasks—while the efficiency claim is training-relative, specifically the reduction in RL iterations (Si et al., 7 Oct 2025).

6. Ablations, limitations, and extensions

The ablation studies isolate the contribution of the principal components. On ALFWorld, removing the planner (“– w/o Guidance”) falls to $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$6 average versus $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$7 with the full method. Removing cold-start supervised fine-tuning (“– w/o Cold-Start SFT”) yields $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$8. Removing homologous consensus filtering (“– w/o HCF”) yields $$\pi_\theta(e \mid u,p) \;=\; \prod_{t=1}^n \pi_\theta\bigl(a_t \mid u,p,a_{1:t-1},o_{1:t-1}\bigr)\times \pi_g(p\mid u).$$9, compared with $(u,p,a_{1:t-1},o_{1:t-1})$0 for the full system. Removing the global planner RL stage (“– w/o Global Planner RL”) yields $(u,p,a_{1:t-1},o_{1:t-1})$1. Removing ECGR (“– w/o ECGR”) yields $(u,p,a_{1:t-1},o_{1:t-1})$2. Variants that change the executors or models used for HCF or ECGR confirm that using two homologous executors, novice and expert, is optimal (Si et al., 7 Oct 2025).

These ablations clarify that EAGLET is not reducible to a single intervention. The planner itself, the filtered cold start, the RL refinement stage, and the executor capability gain reward each contribute to final performance. This suggests that EAGLET should be viewed as a coordinated training pipeline rather than merely a reward design or a data synthesis recipe.

The limitations are also explicit. The method is focused on text-based interactive settings; multimodal or real-robot tasks remain unexplored. It relies on the availability of two suitable homologous executors for filtering and reward computation. It does not address interleaved human-robot collaboration or lifelong adaptation (Si et al., 7 Oct 2025).

The paper identifies several extensions: applying homologous filtering and ECGR to self-improving planners that bootstrap their own executor variants; incorporating multimodal observations such as vision into the planner; studying transfer of planners across domains with little or no retraining; and developing lightweight on-policy evaluators to remove dependence on external executors. Collectively, these directions place EAGLET within a broader research program on planner modularity, cross-domain transfer, and planner-executor co-adaptation.