AtomWorld-Mem: Memory-Restored World States for Long-Horizon Atomistic Evolution
Authors: Tian Luo, Ruge Zhang, Haozhi Han, Yifeng Chen, Yunquan Zhang, Yunxin Liu, Ting Cao, Kun Li
Organizations: Sichuan University, Chengdu, China · Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China · Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · School of Computer Science, Peking University, Beijing, China
High-fidelity atomistic evolution over long timescales requires more than observing the current crystal configuration. Instantaneous atomistic snapshots are often incomplete: locally similar configurations can correspond to different hidden dynamical contexts, future event preferences, and waiting-time scales. We argue that this snapshot ambiguity makes long-horizon atomistic evolution fundamentally a memory-based world-state restoration problem. To address this, we introduce AtomWorld-Mem, a memory-restored atomistic world model that recovers the latent world state missing from instantaneous crystal snapshots. AtomWorld-Mem treats the evolving alloy as an AtomWorld: spatial encoders write multi-scale atomistic keyframes from dense local topology and sparse long-range defect context, while short-term event memory and long-term structural memory integrate these keyframes across time to restore a future-predictive evolutionary state. The restored state is used to prioritize legal vacancy-mediated events under single-event Kinetic Monte Carlo (KMC) constraints, while event legality, physical execution, and residence-time updates remain governed by the underlying simulator. Empirically, AtomWorld-Mem improves long-horizon atomistic progress under fixed microscopic event budgets while maintaining high-fidelity evolution across energetic, structural, and vacancy-transport observables. It further transfers zero-shot across diverse unseen alloy-temperature AtomWorlds, suggesting that the learned memory-restoration mechanism captures reusable principles of hidden-state inference rather than a system-specific local energy heuristic. These results position memory-restored world-state modeling as a promising route toward efficient, physically grounded, and transferable atomistic evolution.
Figures & tables
Figure 1 : Snapshot ambiguity in atomistic evolution. Empirical analysis shows that identical local spatial observations can branch into different future trajectories and exhibit drastically different waiting times. This indicates that a single static snapshot is not a sufficient predictive state for long-horizon atomistic evolution.
Figure 2 : Cu-dependent acceleration trend. Gains remain positive across the tested Cu range, peak at 420 × near 0.18% Cu, and average 186 × .
Method
Memory
Dilute setting: 0.18% Cu
Relatively concentrated setting: 1.34% Cu
AF ↑
Agg. Degree ↑
Arrival Hit ↑
DynErr ↓
AF ↑
Agg. Degree ↑
Arrival Hit ↑
DynErr ↓
Zacros (Traditional KMC)
✗
1.00 ×
0.016
5.0%
0.000
1.00 ×
0.072
43.8%
0.000
Energy-Greedy KMC
✗
11.03 ×
0.035
45.0%
0.466
2.16 ×
0.074
52.5%
0.371
Snapshot-only AtomWorld
✗
66.19 ×
0.116
67.8%
0.318
11.13 ×
0.166
67.5%
0.299
AtomWorld-Mem
✓
420.05 ×
0.432
85.5%
0.221
51.25 ×
0.367
80.9%
0.262
Table 1 : Main results under matched legal-event budgets. AF is the fixed-budget acceleration factor over Zacros (traditional KMC) [ 14 ] , not a claim of path-measure-preserving kinetic speedup. Agg. Degree measures Cu aggregation, Arrival Hit measures vacancy arrival to aggregation-relevant Cu clusters, and DynErr measures deviation from reference KMC dynamics.
Figure 3 : Fixed-budget energy evolution over 100k legal KMC steps. From left to right: energy drop and total energy at 0.18% Cu, followed by energy drop and total energy at 1.34% Cu. AtomWorld-Mem achieves larger energy drops and lower total energies than Zacros under the same legal-event budget. Shaded regions indicate trajectory variability across runs.
Variant
Short
Long
AF ↑
Agg. ↑
TimeErr ↓
DynErr ↓
Arr.Hit ↑
Zacros (Traditional KMC)
✗
✗
1.00 ×
0.016
0.000
0.000
5.00%
Snapshot-only AtomWorld
✗
✗
62.16 ×
0.089
0.777
0.591
65.30%
No short-term memory
✗
✓
127.43 ×
0.183
0.475
0.293
71.64%
No long-term memory
✓
✗
148.25 ×
0.235
0.599
0.348
73.91%
Full AtomWorld-Mem
✓
✓
418.44 ×
0.426
0.308
0.192
83.22%
Table 2 : Memory ablation at 0.18% Cu. Short/Long denote short-term event memory and long-term structural memory. TimeErr and DynErr measure temporal and rollout deviation from Zacros.
Figure 4 : High-fidelity atomistic evolution in physical-time space. We compare AtomWorld-Mem with Zacros traditional KMC across six trajectory-level observables: advancement factor, Cu cluster count, ever-contacted vacancy ratio, normalized energy, mean first-arrival time, and largest Cu cluster size. AtomWorld-Mem follows the reference trends while maintaining faster fixed-budget progress, indicating that memory-restored event prioritization does not collapse into an unphysical local shortcut.
Figure 5 : Zero-shot transfer across 16 unseen BCC alloy–temperature AtomWorlds. Left: energy-drop trajectories under a shared 10k-step budget. Right: per-domain AUC acceleration ratio and dynamics-fidelity error.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Variant
Short
Long
AF ↑
Agg. ↑
TimeErr ↓
DynErr ↓
Arr.Hit ↑
Snapshot-only AtomWorld
✗
✗
11.35 ×
0.169
0.342
0.332
68.80%
No short-term memory
✗
✓
21.33 ×
0.202
0.219
0.206
70.85%
No long-term memory
✓
✗
28.36 ×
0.257
0.255
0.282
75.71%
Full AtomWorld-Mem
✓
✓
51.38 ×
0.358
0.141
0.165
83.82%
Appendix
Table 3 : Additional memory ablation at 1.34% Cu. Short/Long denote short-term event memory and long-term structural memory. AF is the step-to-target acceleration factor over Zacros traditional KMC. TimeErr is log expected waiting-time MAE; DynErr is real-KMC-time aligned NRMSE to Zacros.
Figure 6 : Long-horizon Cu clustering visualization. Cu clustering evolution in Fe–0.67 at.% Cu over a 50-year physical-time window. The visualization illustrates whether memory-restored event prioritization produces coherent long-horizon structural organization rather than only faster energy reduction.
Atomistic structure and natural language have long been modeled separately, with language models either calling atomistic models as tools or being fine-tuned on lossy textual encodings that discard atomistic information. We introduce Atomistic Language Models (ALMs) to pursue native multimodality, in which a single language backbone understands atomistic structures, generates materials from natural language, and optimizes crystal structures as instructed by text. By unifying a pretrained atomistic encoder, large language model, and denoising diffusion model through purely continuous projectors and staged training, ALMs achieve state-of-the-art results on crystal structure prediction and de novo generation. ALMs are enabled by a continuous bridge that maps language model embeddings directly into the steering space of atomistic diffusion, and are assisted by Text-to-Crystal Feynman-Kac (T2C-FK), a particle-based sampler that scores partial denoising trajectories to enforce stoichiometric targets at inference time. To evaluate the ability of ALMs to optimize and generate materials from natural-language prompts and 3D atom-coordinate inputs, we introduce ALM Bench, the first benchmark for text-conditioned crystal generation and optimization. Code, training data, and model weights will be released soon.
Sathya Edamadaka, Krithik Ramesh, Ju Li +1
Massachusetts Institute of Technology · Lyra Labs · Lila Sciences
World models offer a principled way to equip long-horizon LLM agents with foresight: predictions of action consequences before execution. However, unreliable foresight can be ignored, misused, or even degrade downstream decision-making. In this paper, we introduce WorldEvolver, a self-evolving world model framework that revises its deployment-time context while keeping the downstream agent and all model parameters frozen. WorldEvolver integrates three modules: (i) Episodic Memory, which exploits real action transitions through retrieval-based simulation; (ii) Semantic Memory, which extracts persistent heuristic rules from prediction-observation mismatches; and (iii) Selective Foresight, which filters low-confidence predictions before integrating them into agent reasoning context. We evaluate WorldEvolver on ALFWorld and ScienceWorld, measuring world model prediction accuracy on Word2World and downstream agent success rate on AgentBoard. Extensive experiments show that WorldEvolver achieves the highest prediction accuracy across three backbones and leads other world model baselines on downstream agent success rate, demonstrating that test-time memory revision enhances both predictive fidelity and planning performance.
Xuan Zhang, Wenxuan Zhang, See-Kiong Ng +1
National University of Singapore · Singapore University of Technology and Design · Singapore Management University
World models are increasingly used to support planning in agents by predicting how environment states evolve in response to agent actions. Yet fluent next-state predictions can still omit task-critical facts, corrupt product attributes, or apply incorrect transition rules. To address such systematic prediction errors, we introduce MemWM, a memory-augmented text-based world model. MemWM uses world memory, a curated memory bank of transition rules, state caches, and hard-to-predict facts, to condition next-state imagination. We evaluate factual state preservation with Structured State Fidelity (SSF), which scores predicted states through benchmark-specific facts and fields. Compared with SFT, memory-augmented training improves SSF by up to 206.3%. In the full planning setting, we keep the policy model frozen and provide policy-side world skill: retrieved task-level skills and step-wise corrective guidance for action selection. Across ALFWorld, WebShop, and ScienceWorld, memory-augmented agents improve downstream success over an SFT-trained world-model agent, with up to a 65.4% relative gain. Sensitivity analyses further show that retrieved memory improves task success and efficiency under different memory and action-budget settings.
Yujun Wang, Tao Zhang, Jinhe Bi +9
LMU Munich · Munich Center for Machine Learning (MCML) · Huawei Heisenberg Research Center +4