JANUS

JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models

Forecast future outcomes. Reflect on past experience. Co-evolve research capabilities.

       

Mengdi Liu^1,2, Chenghao Jia^1,2, Hong Chang^1,2, Mingjie He^1,2, Shiguang Shan^1,2, Xilin Chen^1,2

¹ State Key Laboratory of AI Safety, Institute of Computing Technology, CAS, China
² University of Chinese Academy of Sciences (CAS), China

---

Overview

Long-horizon autonomous research is constrained by expanding search spaces and costly, delayed feedback. In a conventional Generate–Execute–Feedback loop, each new hypothesis may require hours of training, simulation, or compilation before its value becomes clear. Reactive trial-and-error makes it difficult to allocate research compute effectively as the search frontier grows.

JANUS unifies future outcome forecasting and past experience reflection through a Predict–Verify–Evolve loop. An orchestrator expands an evolving research graph, while a world model forecasts the value of candidate proposals and selects, refines, or prunes them before execution. Ground-truth feedback then improves the skills of the research agent, the world model, and their coordination protocol.

The goal is to improve both evolving efficiency and the attainable evolving ceiling: make faster progress with a given research budget, and sustain improvement as additional compute becomes available.

Highlights

• An evolving research graph. A directed acyclic graph records proposals, implementations, predictions, and execution evidence, supporting branch exploration, experience reuse, and performance attribution.
• World-model-guided exploration. Proposals are assessed by value improvement, information gain, and exploration utility. Sparse pairwise comparisons and adaptive priorities guide refinement, selection, and pruning before costly execution.
• Experience-driven co-evolution. Successful and failed trajectories become intra-task and cross-task research skills. The world model refines prediction and exploration skills, while a shared protocol improves when and how the agents coordinate.
• Evaluation across research settings. The paper evaluates JANUS on 30 MLS-Bench-Lite tasks across 12 domains, plus nanoGPT training optimization, MinGPT attention kernel optimization, and genotype-to-phenotype prediction.

Method

1. Expand. The orchestrator modifies, extends, or combines existing research nodes to propose new candidates.
2. Predict. The world model scores candidates, refines proposals through selective debate, and prioritizes experiments under the remaining budget.
3. Verify. The code executor implements the selected proposal, runs the experiment, and records its outcome in the research graph.
4. Evolve. Verified experience updates research skills, world-model judgment, exploration strategies, and coordination protocols for subsequent iterations.

Results

The following results are reported in Table 1 of the paper on MLS-Bench-Lite. Each task uses a 5-hour wall-clock research budget and one NVIDIA A100 GPU, under the evaluation protocol described in the paper. Higher is better for all metrics below.

Method	Model	Valid Sub.	Beat Any	Beat Median	Beat All	Dev Score	Test Score
AutoResearch	DeepSeek-V4 Flash	83.3%	76.7%	40.0%	16.7%	26.24	25.87
AutoResearch	DeepSeek-V4 Pro	90.0%	83.3%	46.7%	16.7%	27.01	27.13
AIDE	DeepSeek-V4 Flash	93.3%	90.0%	53.3%	23.3%	31.12	31.11
MLEvolve	DeepSeek-V4 Flash	96.7%	93.3%	56.7%	30.0%	33.89	32.99
AIBuildAI	DeepSeek-V4 Flash	90.0%	86.7%	60.0%	33.3%	34.95	35.14
Arbor	DeepSeek-V4 Flash	100.0%	96.7%	66.7%	33.3%	38.01	37.34
Codex	DeepSeek-V4 Flash	96.7%	96.7%	73.3%	36.7%	39.38	38.91
Codex	GPT-5.6 Sol	100.0%	100.0%	86.7%	53.3%	45.23	44.81
JANUS	DeepSeek-V4 Flash	100.0%	100.0%	76.7%	43.3%	41.99	41.07
JANUS	GPT-5.6 Sol	100.0%	100.0%	90.0%	56.7%	47.32	46.88
JANUS	Codex (GPT-5.6 Sol)	100.0%	100.0%	93.3%	63.3%	49.96	49.78

Valid Sub. measures valid submissions. Beat Any / Median / All measure the fraction of tasks outperforming at least one, the median-performing, or all provided human baselines. Dev / Test Score are mean normalized task scores. See the paper for full definitions, scaling analyses, ablations, and real-world research results.

Real-World Tasks

JANUS achieves the strongest final performance across three real-world research tasks among the compared methods. As the number of experiments increases, it reaches lower validation bits per byte on nanoGPT training optimization, higher speedup on MinGPT attention kernel optimization, and higher macro-averaged Pearson correlation on genotype-to-phenotype prediction. These trajectories show how JANUS turns successive experiments into sustained improvements across model training, systems optimization, and scientific prediction. Lower is better for validation bits per byte; higher is better for speedup and Pearson correlation.

Release plan

Item	Status
Paper (PDF)	✅ Available
Project page & interactive demo	🔜 In progress
arXiv posting	🔜 In progress
Full codebase	⏳ Being prepared
Reproduction scripts	⏳ Being prepared

Citation

@article{liu2026janus0,
  title   = {JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models},
  author  = {Mengdi Liu and Chenghao Jia and Hong Chang and Mingjie He and Shiguang Shan and Xilin Chen},
  year    = {2026},
  journal = {arXiv preprint arXiv: 2610.13194}
}