Autonomous research for time series

NeoResearch

Literature- and Experience-Guided
Autonomous Time Series Model Discovery

Ze Guo·Mingyue Cheng·Xiaoyu Tao·Shuo Yu·Shangtong Ye

State Key Laboratory of Cognitive Intelligence
University of Science and Technology of China

An LLM researcher. Task-specific models. Experience that carries forward.

The research workflowFigure 2
NeoResearch takes time-series data and domain knowledge through literature-guided architecture discovery and evidence-guided training optimization. Architecture-level and optimization-level memory guide both loops, producing lightweight models and reusable insights. Enlarge
From scientific evidence to specialized forecasters. The outer loop proposes and revises architectures; the inner loop optimizes training. Hierarchical memory connects literature, hypotheses, experiments, and conditional lessons.Figure PDF ↗
2coordinated research loops
10real-world datasets
18comparison methods
19/20best dataset–metric results
≤10Mparameters per candidate

Abstract

Turning knowledge into
testable model designs.

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.

Time series forecastingAutonomous researchCross-domain memory transfer

The research perspective

Discover a model. Retain the experience.

A general-purpose LLM organizes the research. The discovered lightweight model performs numerical forecasting independently of the LLM.

01

What should we build?

Connect temporal dynamics and variable relationships to modeling mechanisms and their evidence in scientific literature.

02

Refine training or revise the design?

Use validation errors and convergence diagnostics to decide whether a candidate needs better optimization or a new architectural hypothesis.

03

What carries over to the next task?

Retain what changed, under which conditions, and with what outcome. Reassess applicability when experience is reused.

Comparison between an LLM as a forecasting component and an LLM researcher that discovers models and retains scientific knowledge, experience, and modeling assets.
An LLM as a researcher. Each investigation produces a forecaster and reusable research resources. Figure 1 ↗

Method

Two loops. One evidence-driven workflow.

Literature motivates hypotheses. Experiments test them. Reflection determines the next step.

01

Profile & retrieve

Analyze the target data. Retrieve literature, mechanisms, source evidence, and relevant prior experience.

Task-grounded evidence
02

Hypothesize & build

Form a testable architectural hypothesis and translate it into a structured, executable model recipe.

Literature-guided discovery
03

Train & validate

Refine training for the fixed design using validation metrics, convergence curves, and error analysis.

Evidence-guided optimization
04

Reflect & revise

Assess the mechanism, revise its applicability, and consolidate conditional lessons for future research.

Persistent research memory

Outer loop

Literature-guided
architecture discovery

Choose preprocessing, temporal backbones, variable interactions, and module compositions. Preserve the source, expected effect, applicability conditions, and observations that would challenge each hypothesis.

Experimental feedback → refine or replace the design

Inner loop

Evidence-guided
training optimization

Keep the design fixed while adjusting learning rates, batch sizes, regularization, scheduling, and training loss. Return the best validated model with diagnostics and a complete experiment history.

Validation and convergence → tune or return to discovery

Traceable by construction. Structured recipes and persistent state link hypotheses, designs, runs, and model artifacts. Validation MSE selects candidates; test data are excluded from research decisions.

Literature augmentation

Scientific literature, grounded in the target task.

Retrieve scientific papers and extract mechanisms, empirical evidence, and applicability conditions to formulate testable architecture hypotheses.

  1. Retrieve from task evidence

    Data diagnostics, domain knowledge, and the forecasting objective guide autonomous literature retrieval. Temporal dynamics and variable relationships help identify relevant modeling mechanisms.

  2. Extract knowledge with its context

    Preserve source passages, reported experimental outcomes, and their original settings. Record each mechanism’s assumptions, applicability conditions, and limitations alongside the evidence.

  3. Form a testable architectural hypothesis

    Check literature assumptions against the target data. Specify the proposed mechanism, its expected effect, the conditions for its use, and observations that would challenge the explanation.

Structured literature recordSource-grounded
Source & context
Original paper, supporting passage, and experimental setting.
Modeling mechanism
How a preprocessing, temporal, or variable-interaction component is expected to work.
Empirical evidence
Reported outcomes and the conditions under which they were observed.
Applicability & limitations
Assumptions and constraints to assess against the target data and forecasting objective.
Source evidence stays linked to the hypothesis and subsequent experiments.

Close the evidence loop

Let experiments reassess applicability.

Validation results and controlled comparisons test the design rationale. Unresolved outcomes can prompt further diagnostics or literature retrieval, and new evidence can revise the hypothesis.

Structured extraction matters.

Full NeoResearch outperforms direct source-text access and excluding literature on all four evaluated tasks: ETTh1, MOPEX, NP, and PJM.

Table 4 · knowledge extraction ablation

Hierarchical experience memory

Research that informs the next investigation.

Preserve evidence and conditions alongside outcomes, so prior findings can be tested and revised on a new task.

Architecture-level memory

Mechanisms, hypotheses, designs, architecture revisions, and experimental outcomes guide what to build and which assumptions to test.

  • Effective designs and applicability conditions
  • Failed interventions and possible causes
  • Unresolved or conflicting explanations

Optimization-level memory

Training configurations, tuning outcomes, and convergence patterns guide how to train a candidate and when to revisit the design.

  • Effective strategies and measured gains
  • Low-gain trials and training instability
  • Diagnostics linked to the executed recipe
Literature & task evidenceHypotheses & experimentsConditional lessonsTarget-task validation
MSE and MAE across ETTh1, MOPEX, NP, and PJM as source tasks accumulate. Full structured memory outperforms raw memory and no memory at every nonzero checkpoint, with a temporary reversal on MOPEX. Enlarge
Structured memory supports cross-task reuse. Sources accumulate in order: GEF-L, Windy, BE, and FR. More sources do not guarantee monotonic gains; transferred findings are reassessed on the target task.Figure 5 ↗

Benchmark results

Specialized models across ten datasets.

Electricity markets, energy systems, and hydrology. Compared with 16 forecasting baselines and two automated model-development methods.

Reference-initialized Alpha setting

19/20

lowest held-out errors

Best MSE on 9 of 10 datasets and best MAE on all 10. Each comparison uses the strongest baseline for that dataset and metric.

The exception: Windy MSE. DLinear and DeepAR achieve lower MSE. NeoResearch has the lowest MAE on Windy.

Up to 25 research iterations · ≤10M parameters per candidate · validation-based selection · held-out testing

Against the best baseline

Relative MSE reduction

Reduction = 100 × (best baseline − NeoResearch) / best baseline. A positive value means lower error. Bar lengths share a fixed 18% scale across both metrics.

Comparison with the strongest baseline

MSE / MAE ↓ · lower is better
Results from Table 1 of the NeoResearch manuscript. The best baseline is selected separately for each dataset and metric.
DatasetMean squared error ↓Mean absolute error ↓
Best baselineNeoResearchReductionBest baselineNeoResearchReduction
ETTh10.1022iTransformer0.1004+1.8%0.2436iTransformer0.2434+0.1%
GEF-L0.3384AIDE0.3324+1.8%0.4119AIDE0.4114+0.1%
Windy1.2785DLinear1.3406−4.9%0.7946ARIMA0.7867+1.0%
Sunny0.3119DLinear0.3089+1.0%0.3598TimeMixer0.3552+1.3%
MOPEX1.1581MTGNN0.9812+15.3%0.6094GNN0.5443+10.7%
BE1.4386AIDE1.3807+4.0%0.4764AIDE0.4637+2.7%
DE0.8030DeepAR0.7912+1.5%0.6018DeepAR0.5948+1.2%
FR5.1838TimeXer5.0887+1.8%0.7262AIDE0.6675+8.1%
NP0.3842TimeMixer0.3546+7.7%0.3786PatchTST0.3618+4.4%
PJM0.2116Optuna-TPE0.1849+12.6%0.3269Optuna-TPE0.3093+5.4%

Values are reproduced from the manuscript. The best baseline can differ between MSE and MAE. EPF datasets (BE, DE, FR, NP, PJM) use a 168→24 horizon; the remaining tasks use 96→96. All covariates are observed-past inputs.

Explore the complete 18-method comparison
Full Table 1 comparing NeoResearch with ARIMA, Prophet, GNN, MTGNN, RNN, LSTM, DeepAR, Mamba, S-Mamba, DLinear, TimeMixer, ConvTimeNet, TimesNet, PatchTST, iTransformer, TimeXer, AIDE, and Optuna-TPE on ten datasets. Enlarge
Complete benchmark suite. The first 16 methods are forecasting baselines; AIDE and Optuna-TPE are automated model-development comparisons.Table PDF ↗

Analysis & ablations

What makes discovery effective?

The paper examines the research loops, structured literature extraction, memory, and the models they produce.

01 / Dual loops

Discovery and optimization work together.

Removing either loop increases MSE and MAE across all ten tasks. Architecture exploration benefits from training feedback; training benefits from revisiting the design.

Table 2 · dual-loop ablation
02 / Literature

Extract mechanisms, evidence, and conditions.

Structured extraction outperforms direct source-text access and excluding literature on all four evaluated tasks. Source evidence becomes a testable hypothesis.

Table 4 · knowledge extraction
03 / Experience

Both levels of memory contribute.

Removing architecture-level memory, optimization-level memory, or both worsens forecasting. Useful experience includes failed interventions and unresolved explanations.

Table 3 · hierarchical memory
Four discovered architectures: ETTh1 uses gated mixing and a temporal MLP; MOPEX uses Mamba with no instance normalization; NP uses patch MLP, mixing, and trend; PJM uses decomposition and N-BEATS with residual correction.
Task-specific architectures. Discovered models differ in preprocessing, variable interaction, backbone, and readout.Figure 6 ↗
Cross-task resource ablations on ETTh1, MOPEX, NP, and PJM show that restricting reuse of knowledge, experience, or model assets increases error.
Cross-task resources. Knowledge, experience, and reusable designs support subsequent discovery.Figure 3 ↗
Reference-initialized autonomous Alpha achieves lower MSE and MAE than human-machine cooperation and cold-start Zero on four evaluated tasks.
Research modes. Reference-initialized autonomous Alpha leads HMC and Zero on the four evaluated tasks.Figure 4 ↗
More research does not always mean better forecasts. Budget sensitivity and memory accumulation show non-monotonic outcomes on some tasks. The workflow relies on validation-based selection and reassesses transferred findings. See the exploration study ↗
Explore research budgets and candidate quality
MSE and MAE for maximum research budgets of 15, 20, 25, and 30 iterations. PJM benefits most consistently; other datasets have non-monotonic outcomes.
Budget sensitivity. Each point is a separate investigation with its own maximum budget; longer research does not guarantee a better returned model.Figure 7 ↗
Shares of discovered candidates better than the reference: MSE 53%, 36%, 59%, and 50%; MAE 30%, 30%, 54%, and 49% for ETTh1, MOPEX, NP, and PJM, respectively.
Candidate quality. Every evaluated pool contains models that improve on the reference. NP has the largest share under both metrics.Figure 8 ↗

Case study · MOPEX streamflow

When experiments revise the hypothesis.

Distribution shift suggests normalization may help. Controlled experiments reveal that preserving level and scale is more useful for this task.

01 / Initial hypothesis

Normalized Mamba

Reuse a compact temporal backbone for persistent runoff dynamics.

CV MSE0.701954
02 / Controlled revision

Remove normalization

Keep other settings fixed. Lower CV spread supports a consistent gain.

CV MSE0.596769
03 / Further experiment

Test channel mixing

Mean error improves, but inconsistent fold-level gains keep attribution uncertain.

CV MSE0.588419
MOPEX research trajectory linking data profiling, literature mechanisms, prior memory, three controlled architecture experiments, reflection, and the final reusable Mamba forecasting model. Enlarge
Evidence revises applicability. The selected forecaster reaches held-out test MSE 0.981216 and MAE 0.544346 after validation-based selection. Memory retains the model, the normalization finding, and uncertainty about mixing.Figure 9 ↗

Open source

Start your next
forecasting investigation.

Set up NeoResearch from source, connect an LLM provider, and create a forecasting task from your CSV data.

Explore the repository

Linux or macOS · Node.js 20+ · Bun 1.3.13
Python 3.10+ · uv

Set up from source
git clone https://github.com/Kawai1Angel/NeoResearch.git
cd NeoResearch

# Install and initialize the runtime
make setup
(cd neo_cast && npm install && node bin/neo-cast.js init-runtime)

# Check the environment and start the CLI
make doctor
make dev-cli
Configure a provider → create a task → start research

Choose how to collaborate. Socratic clarifies requirements up front, then researches autonomously. Companion pauses for your approval at key stages. Read the getting-started guide ↗

Citation

Build on this research.

If NeoResearch is useful to your work, please cite the manuscript.

@misc{guo2026neoresearch,
  title  = {NeoResearch: Literature- and Experience-Guided
            Autonomous Time Series Model Discovery},
  author = {Guo, Ze and Cheng, Mingyue and Tao, Xiaoyu
            and Yu, Shuo and Ye, Shangtong},
  year   = {2026},
  note   = {Research manuscript},
  url    = {https://github.com/Kawai1Angel/NeoResearch}
}

Citation metadata follows the manuscript included with this project.

Figure preview

Original ↗