Driving · 2026-10-12
JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models
Mengdi Liu, Chenghao Jia, Hong Chang, Mingjie He, Shiguang Shan, Xilin Chen
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
JANUS is a long-horizon autonomous research framework that replaces reactive Generate–Execute–Feedback with a Predict–Verify–Evolve loop. It maintains an evolving DAG, uses a co-evolving research world model to forecast candidate value improvement, information gain, and exploration utility, and scores/refines/prunes branches before execution. After ground-truth feedback, JANUS distills experience into hierarchical skills and reflection/coordination protocols for continual co-evolution.
Why it matters
AutoResearch is bottlenecked by costly, time-consuming execution feedback and delayed signals in a combinatorially expanding search space. JANUS aims to improve both “evolving efficiency” and “evolving ceiling” by proactively forecasting outcomes, adaptively pruning exploration branches pre-execution, and turning execution logs into reusable evolving skills. The paper reports evaluations on MLS-Bench-Lite and real research tasks, claiming stronger performance than state-of-the-art baselines.
Method
- Maintain an evolving directed acyclic graph (DAG) of research proposals (Open/Closed nodes) to enable branch exploration, experience reuse, and topology-aware performance attribution.
- Use a co-evolving research world model in a Predict–Verify–Evolve loop: forecast candidate outcomes (value improvement, information gain, exploration utility), refine/select/prune branches, then execute and verify.
- Distill execution feedback into evolving skills: hierarchical skills for the research agent, exploration-strategy reflection for the world model, and an evolving coordination protocol for their interaction.
Abstract (from arXiv)
Long-horizon autonomous research has shown promise in accelerating scientific progress through sustained cycles of hypothesis generation, implementation, and validation. However, this Generate-Execute-Feedback paradigm relies on reactive trial-and-error, where costly and time-consuming executions introduce severe feedback delays, throttling evolutionary progress under explosively expanding search space. To address this challenge, we propose JANUS, a framework that unifies future outcomes forecasting and past experience reflection through a Predict-Verify-Evolve loop driven by a co-evolving world model. Within an evolving graph, the research agent proposes candidates while world model scores them by forecasting value improvement, information gain, and exploration utility, then accordingly selects, refines, or prunes branches prior to execution. After receiving ground-truth feedback, JANUS distills experience into evolving skills: hierarchical skills for research agent, exploration-strategy skills for world model, and their coordination protocols. This mechanism enables continual co-evolution: the research agent sharpens candidate generation while the world model refines its predictive judgment over extended horizons. Extensive evaluations on MLS-Bench-Lite and real research tasks demonstrate that JANUS significantly outperforms state-of-the-art baselines, achieving both higher evolving efficiency and ceiling.
Related papers
- EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
- AnyGoal: On World Model Generalization Beyond Expert Rollouts
- Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
- ALDER: Discovering the Laws of a World by Acting in It
- Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-09 · Indexed by Pub-AI: 2026-10-12 UTC
JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models
Preprint posted: 2026-10-09 · Original record ↗
- arxiv
- 2610.13194
Explicit source versions (1)
- v1 · 2026-10-09
- Preprint available
Source date: 2026-10-09 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Long-horizon autonomous research has shown promise in accelerating scientific progress through sustained cycles of hypothesis generation, implementation, and validation. However, this Generate-Execute-Feedback paradigm relies on reactive trial-and-error, where costly and time-consuming executions introduce severe feedback delays, throttling evolutionary progress under explosively expanding search space. To address this challenge, we propose JANUS, a framework that unifies future outcomes forecasting and past experience reflection through a Predict-Verify-Evolve loop driven by a co-evolving world model. Within an evolving graph, the research agent proposes candidates while world model scores them by forecasting value improvement, information gain, and exploration utility, then accordingly selects, refines, or prunes branches prior to execution. After receiving ground-truth feedback, JANUS distills experience into evolving skills: hierarchical skills for research agent, exploration-strategy skills for world model, and their coordination protocols. This mechanism enables continual co-evolution: the research agent sharpens candidate generation while the world model refines its predictive judgment over extended horizons. Extensive evaluations on MLS-Bench-Lite and real research tasks demonstrate that JANUS significantly outperforms state-of-the-art baselines, achieving both higher evolving efficiency and ceiling.