Aligning Progress and Feasibility

Illustration of the dual-alignment challenge and neuro-symbolic dual memory

Long-horizon agents can fail in two coupled ways: global progress drift and local feasibility violation. The proposed framework aligns both with a neural progress memory and a symbolic feasibility memory.

Abstract

Large language models have demonstrated strong potential in long-horizon decision-making tasks, such as embodied manipulation and web interaction. However, agents frequently struggle with endless trial-and-error loops or deviate from the main objective in complex environments.

We attribute these failures to two fundamental errors: global Progress Drift and local Feasibility Violation. Existing methods typically attempt to address both issues using a single paradigm, although progress guidance relies on fuzzy semantic planning while feasibility verification requires strict logical constraints and state validation.

We propose a Neuro-Symbolic Dual Memory Framework that explicitly decouples these objectives. A neural Progress Memory extracts semantic blueprints from successful trajectories to guide global task advancement, while a symbolic Feasibility Memory synthesizes executable Python verification functions from failed transitions for strict local validation.

1. Dual-Alignment Challenge

Long-horizon agents need to satisfy two different alignment goals at the same time. They must preserve global progress toward the final objective, but every local action must also obey environment-specific preconditions. Treating both problems with a single memory or a single prompting strategy often mixes soft semantic guidance with hard executability constraints.

Progress Drift

The agent loses the current stage of the task, repeats partial actions, or wanders into irrelevant branches.

Feasibility Violation

The agent proposes locally impossible actions, causing invalid attempts and reinforcing failure loops.

2. Neuro-Symbolic Dual Memory Framework

The framework separates memory construction and inference into two coordinated pathways. Offline, successful trajectories are distilled into procedural blueprints and action chunks for Progress Memory. Failed transitions are compiled into executable symbolic verifier rules for Feasibility Memory. At inference time, the progress pathway proposes stage-consistent actions, while the feasibility pathway verifies and refines them before execution.

Overview of the neuro-symbolic dual memory framework

Offline experience distillation builds two memory libraries. During inference, the agent retrieves the current blueprint, proposes an action, checks feasibility, and updates progress after observing the environment.

3. Experimental Results

The method is evaluated on ALFWorld, WebShop, and TextCraft using the same backbone model, gpt-4o-2024-11-20. It consistently outperforms competitive long-horizon agent baselines, improving task success across embodied interaction, web-based decision making, and compositional crafting.

Table 1: Main results. Mean ± standard deviation over three runs.
Method	ALFWorld	WebShop		TextCraft
Method	Success Rate (%)	Success Rate (%)	Score (%)	Success Rate (%)
ReAct	78.1 ± 2.2	34.0 ± 2.0	52.1 ± 2.3	60.0 ± 2.6
ADaPT	71.9 ± 0.4	33.0 ± 1.0	52.3 ± 1.2	73.7 ± 2.9
StateAct	68.9 ± 5.6	25.0 ± 6.9	42.0 ± 10.7	69.3 ± 1.5
ExpeL	85.3 ± 1.1	32.7 ± 3.2	50.5 ± 4.0	89.0 ± 1.0
WALL-E 2.0	83.8 ± 0.9	36.3 ± 2.5	59.9 ± 1.2	64.7 ± 1.5
AWM	87.3 ± 0.7	40.3 ± 2.1	61.1 ± 1.7	65.3 ± 6.1
Ours	95.3 ± 1.6	50.0 ± 1.0	71.0 ± 0.5	94.0 ± 1.0

BibTeX

@inproceedings{wen2026dualmemory,
  title     = {Aligning Progress and Feasibility: A Neuro-Symbolic Dual Memory Framework for Long-Horizon LLM Agents},
  author    = {Wen, Bin and Zhang, Ruoxuan and Chen, Yang and Xie, Hongxia and Guo, Lan-Zhe},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}