NeoResearch: Literature- and Experience-Guided Autonomous Time Series Model Discovery
Organizations: State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, Anhui Province, China
Abstract
Developing effective forecasting models for diverse time-series domains requires substantial expert knowledge, evidence from scientific literature on the web, and iterative experimentation. However, translating this knowledge and research experience into model designs requires preserving source context and testing applicability under task-specific assumptions. In this paper, we propose NeoResearch, a literature- and experience-guided autonomous research framework that leverages large language models (LLMs) as researchers to discover lightweight, domain-specific forecasting models. Specifically, NeoResearch establishes an expert-inspired dual-loop research workflow, coupling hypothesis-driven architecture exploration with iterative training and hyperparameter optimization through experimental feedback. To enable knowledge-grounded discovery, the agent autonomously mines scientific literature from the Web, extracting modeling mechanisms, empirical evidence, and applicability conditions to formulate task-grounded, testable architecture hypotheses. Furthermore, a hierarchical experience memory links literature findings to experimental outcomes and consolidates conditional lessons, guiding architecture discovery and training optimization while supporting reuse across forecasting tasks. Experiments on ten datasets show competitive performance and support dual loops, literature extraction, and memory reuse. Our code is available.
Figures & tables
| Method | ETTh1 | GEF-L | Windy | Sunny | MOPEX | BE | DE | FR | NP | PJM | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | |
| ARIMA | 0.1258 | 0.2692 | 2.6140 | 1.1724 | 1.5244 | 0.7946 | 2.1279 | 0.9284 | 2.0232 | 0.8539 | 2.7734 | 0.8377 | 1.4358 | 0.8512 | 11.6593 | 1.2454 | 0.7585 | 0.5777 | 1.0326 | 0.7560 |
| Prophet | 2.7527 | 1.4975 | 2.3506 | 1.0190 | 2.1633 | 1.1161 | 0.3636 | 0.4571 | 1.5570 | 0.8254 | 2.9572 | 0.9921 | 1.2569 | 0.7406 | 8.7114 | 1.5332 | 4.9137 | 1.9780 | 0.5472 | 0.5890 |
| GNN | 0.1710 | 0.3282 | 0.3483 | 0.4194 | 1.3559 | 0.8382 | 0.3204 | 0.3615 | 1.1744 | 0.6094 | 1.7677 | 0.5849 | 0.9764 | 0.6892 | 5.2215 | 0.8461 | 0.4129 | 0.4228 | 0.2301 | 0.3596 |
| MTGNN | 0.1754 | 0.3280 | 0.3421 | 0.4157 | 1.8594 | 1.0386 | 0.3711 | 0.3940 | 1.1581 | 0.6287 | 1.5649 | 0.5537 | 2.0645 | 1.0449 | 6.0711 | 0.8063 | 0.5049 | 0.4874 | 0.2250 | 0.3514 |
| RNN | 0.1339 | 0.2826 | 0.3665 | 0.4306 | 1.9703 | 1.0033 | 0.5035 | 0.4779 | 1.2899 | 0.6809 | 1.7545 | 0.5831 | 0.8215 | 0.6194 | 6.0330 | 0.8054 | 0.4425 | 0.4521 | 0.2538 | 0.3701 |
| Method | ETTh1 | GEF-L | Windy | Sunny | MOPEX | BE | DE | FR | NP | PJM | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | |
| w/o Outer Loop | 0.1025 | 0.2473 | 0.3427 | 0.4210 | 1.6799 | 0.9156 | 0.4470 | 0.4309 | 1.5570 | 0.7740 | 1.7829 | 0.6017 | 1.3194 | 0.7780 | 6.5556 | 1.0450 | 0.3783 | 0.3766 | 0.2115 | 0.3270 |
| w/o Inner Loop | 0.1051 | 0.2493 | 0.3442 | 0.4209 | 1.4145 | 0.8415 | 0.4783 | 0.4155 | 1.0301 | 0.5655 | 1.4870 | 0.4860 | 0.8276 | 0.6108 | 5.5161 | 0.7002 | 0.3759 | 0.3737 | 0.1883 | 0.3196 |
| NeoResearch | 0.1004 | 0.2434 | 0.3324 | 0.4114 | 1.3406 | 0.7867 | 0.3089 | 0.3552 | 0.9812 | 0.5443 | 1.3807 | 0.4637 | 0.7912 | 0.5948 | 5.0887 | 0.6675 | 0.3546 | 0.3618 | 0.1849 | 0.3093 |
| Method | ETTh1 | MOPEX | NP | PJM | ||||
|---|---|---|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | |
| w/o Architecture-level | 0.1032 | 0.2464 | 1.0974 | 0.6080 | 0.3894 | 0.3787 | 0.2206 | 0.3333 |
| w/o Optimization-level | 0.1057 | 0.2507 | 1.2361 | 0.6767 | 0.3933 | 0.3850 | 0.2013 | 0.3217 |
| w/o Memory | 0.1060 | 0.2491 | 1.3206 | 0.6863 | 0.4092 | 0.3951 | 0.2122 | 0.3338 |
| NeoResearch | 0.1004 | 0.2434 | 0.9812 | 0.5443 | 0.3546 | 0.3618 | 0.1849 | 0.3093 |
| Method | ETTh1 | MOPEX | NP | PJM | ||||
|---|---|---|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | |
| w/o Extraction | 0.1043 | 0.2479 | 1.0378 | 0.6027 | 0.3934 | 0.3851 | 0.1897 | 0.3135 |
| w/o Knowledge | 0.1038 | 0.2474 | 1.0258 | 0.5683 | 0.4212 | 0.4075 | 0.1969 | 0.3194 |
| NeoResearch | 0.1004 | 0.2434 | 0.9812 | 0.5443 | 0.3546 | 0.3618 | 0.1849 | 0.3093 |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Field | Supported choices or constraints |
|---|---|
| Input and preprocessing | Task-specified history, horizon, and observed-past covariates; normalization, decomposition, seasonal differencing, missingness masks, sample reweighting, last-value centering, or none. |
| Embedding | Single/multiscale patches, inverted, frequency, calendar, or none. |
| Temporal backbone | Linear, MLP, Transformer, recurrent, temporal convolution, DLinear, N-BEATS, mixer, gated depthwise convolution, TimesNet-lite, diagonal SSM, Mamba, or TimesBlock encoder. |
| Variable interaction | Temporal, variate attention, temporal-then-feature, feature-then-temporal, graph, gated hybrid, or feature MLP. |
| Layout and output | Channel-latent, sequence-latent, or channel-sequence layout; none/mean/last/flatten readout; compatible identity, linear, quantile, Gaussian, or residual-correction head for prediction. |
| Composition | Single branch, gated parallel, or trend/residual branches; no nested composition. |
| Dataset | Frequency | Train | Validation | Test | ||||
|---|---|---|---|---|---|---|---|---|
| ETTh1 | Hourly | 10,452 | 3,484 | 3,484 | 96 | 96 | 7 | 3,389 |
| GEF-L | Hourly | 33,600 | 4,800 | 9,600 | 96 | 96 | 26 | 9,505 |
| Windy | 15 min | 16,867 | 2,410 | 4,819 | 96 | 96 | 7 | 4,724 |
| Sunny | 15 min | 16,867 | 2,410 | 4,819 | 96 | 96 | 7 | 4,724 |
| MOPEX | Daily | 8,568 | 1,224 | 2,448 | 96 | 96 | 5 | 2,353 |
| BE | Hourly | 10,147 | 1,450 | 2,899 | 168 | 24 | 3 | 2,876 |
| Method | ETTh1 | MOPEX | NP | PJM | ||||
|---|---|---|---|---|---|---|---|---|
| MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | |
| DeepSeek-V4.1Flash | 0.1004 | 0.2434 | 0.9812 | 0.5443 | 0.3546 | 0.3618 | 0.1849 | 0.3093 |
| GPT-5.6 | 0.1008 | 0.2436 | 1.0114 | 0.5649 | 0.3635 | 0.3717 | 0.1810 | 0.3056 |
| GLM-5.2 | 0.1025 | 0.2473 | 1.0274 | 0.5634 | 0.3821 | 0.3766 | 0.1899 | 0.3125 |
| LongCat-2.0 | 0.1011 | 0.2440 | 1.0387 | 0.5691 | 0.4193 | 0.4010 | 0.1866 | 0.3110 |