acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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 experiences 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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Autonomous Research, World Models, Self-evolving

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yheMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes JANUS, a framework for autonomous research that combines pre-execution proposal evaluation with experience-based reflection. Research candidates and their experimental outcomes are maintained in an evolving directed acyclic graph. An LLM-based world model compares proposals according to value improvement, information gain, and exploration utility, guiding proposal selection, refinement, and pruning. Execution feedback is subsequently distilled into research skills, prediction and exploration skills, and coordination protocols. The framework is evaluated on the 30 tasks of MLS-Bench-Lite and three additional research tasks involving language-model training, attention-kernel optimization, and genotype-to-phenotype prediction. The reported results show improvements over several research-agent baselines, while ablations suggest that adaptive exploration and experience reuse contribute to performance. My main reservations concern the connection between the formal world-model formulation and its implemented judgments, and whether the experiments substantiate the stronger efficiency and scaling claims.

Critical strengths
  1. The framework addresses a consequential bottleneck in autonomous research. Evaluating and refining proposals before expensive execution is a useful direction, particularly when experiments require substantial training or compilation time. Combining this with a structured record of successful and failed experiments provides a coherent approach to allocating research effort.

  2. The evaluation demonstrates practical promise across different research settings. Table 1 includes common-backbone comparisons and reports both development and test performance. For example, with DeepSeek-V4 Flash, JANUS achieves a test score of 41.07 versus 38.91 for Codex. The additional tasks and component ablations strengthen the evidence that the framework has utility beyond a single benchmark configuration.

Critical weaknesses
  1. The central world-model and co-evolution claims are not sufficiently connected to measurable predictive behavior. Sections 3.1–3.4 introduce expected cumulative reward, information gain, and exploration utility, but Appendix A primarily supplies their semantic definitions for LLM pairwise judgments. The manuscript does not establish how accurately these judgments predict realized experimental gains, or whether reflection improves prediction quality over time. Moreover, the “w/o World Model” ablation also removes evolution, so that comparison does not isolate forecasting itself. The results support the usefulness of the overall system more clearly than the claimed predictive mechanism.

  2. The evidence supports finite-budget performance improvements more strongly than an evolving scaling law or a higher attainable ceiling. Figure 3 and Table 2 measure progress primarily against experiment count. Experiments can differ in execution cost, and proposal scoring, debate, and reflection incur additional inference costs. The fixed five-hour MLS-Bench-Lite comparison is valuable, but it does not resolve the resource interpretation of the additional-task curves. The reported search trajectories also lack sufficiently documented run-level uncertainty to establish stable plateau differences. Consequently, the efficiency and ceiling claims are stronger than the demonstrated evidence.

Questions for authors
  1. Implementation clarification: What are the exact procedures for proposal-pool sizing, pair selection, adjustment of βt\beta_t and γt\gamma_t, and skill updating? Which operations are explicit algorithms, and which are determined through LLM prompting?

  2. Mechanism validation: Do you have measurements of proposal-ranking accuracy or prediction error before and after world-model reflection? Can these distinguish improved predictive judgment from improved proposal generation and richer memory?

  3. Evaluation clarification: How many independent research runs underpin Figures 3–5 and Table 2? Can you provide uncertainty estimates and performance comparisons against total elapsed time or total resource cost, including prediction and reflection overhead?

Reviewer_NhVMOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Existing long-horizon autonomous research methods typically follow a Generate-Execute-Feedback paradigm, which is costly and time-consuming. This paper introduces JANUS, a framework that adopts a Predict-Verify-Evolve loop by combining future outcome forecasting and past experience reflection. Specifically, a world model is designed to predict the potential value of each research proposal. After receiving experimental feedback, JANUS further distills experiences into skills for the research agent, world model, and their coordination protocols. Finally, JANUS is evaluated on MLS-Bench-Lite and real-world research tasks, demonstrating better performance than several baseline methods.

Critical strengths
  1. The paper is well-organised and easy to follow.
  2. The figures and graphs are also clear and professional.
Critical weaknesses

W1. Novelty.

In the introduction, the authors claim that their main contribution is a framework that unifies future outcome forecasting and past experience reflection through a “Predict-Verify-Evolve” loop driven by a co-evolving world model. However, this concept is not entirely new. For example, K-Search (also cited in their introduction) proposes a search method guided by a co-evolving world model. Moreover, the motivation and methodology of K-Search are highly similar to those of JANUS. Given these similarities, what are the fundamental differences between JANUS and K-Search? It seems that the differences lie mainly in certain implementation details, while the core motivations and ideas substantially overlap. Nevertheless, the authors do not provide any comparison with K-Search in the Related Work section. Given these observations, I have significant concerns about the novelty and methodological contributions of this paper.

W2. World Model.

The paper claims that the world model is used to predict future research rewards. However, from my understanding, the actual implementation is very similar to an LLM-as-a-Judge mechanism. The authors use an LLM to perform pairwise comparisons, followed by Bradley–Terry ranking and weighted aggregation. However, it is unclear how this process relates to the expected cumulative reward. I find the connection between the mathematical formulation and the claimed functionality rather weak. What distinguishes this mechanism from a conventional LLM-as-a-Judge or proposal-ranking agent? In particular, how does the resulting priority score approximate the expected long-horizon return defined in Eq. 4? Does the world model explicitly model future research outcomes or state transitions, or does it primarily serve as a heuristic evaluator of candidate proposals?

W3. Co-Evolution.

The authors emphasize that the research agent and world model continually co-evolve. However, from my understanding, this is mainly achieved by updating skills based on historical experiences. The mechanism appears to rely primarily on LLM-based reflection and memory updates, without any parameter fine-tuning of the underlying models. How is this mechanism fundamentally different from existing reflective-memory or experience-reuse frameworks? More importantly, what evidence demonstrates that the world model itself becomes a better predictor over time, rather than simply benefiting from an increasingly informative context?

Questions for authors

Q1. Is It Just a More Complicated Agent Harness?

The functionalities introduced by JANUS, such as research planning, proposal evaluation, experiment execution, and experience management, are common in modern agent frameworks. It is therefore unclear whether JANUS introduces a new method or primarily provides a particular orchestration of existing agentic components. What are the essential mechanisms that distinguish JANUS from a general-purpose research agent equipped with a proposal evaluator and persistent reflective memory? Would a simpler main-agent framework with comparable evaluation and reflection capabilities, as well as a comparable inference budget, achieve similar performance? A controlled comparison would help isolate the benefits of the proposed architecture.

Q2. Experiments.

Q2.1 Lack of controlled comparisons and computational budget analysis.

In each round, JANUS needs to generate multiple proposals, perform pairwise comparisons, debate, and reflection. However, the paper does not sufficiently report LLM token usage, reasoning budgets, or actual wall-clock time. It is hard to tell whether the improvements come from the proposed mechanisms or simply from additional test-time compute and best-of-N proposal selection.

Q2.2 The ablation studies do not fully support the claimed contributions of the World Model and Co-Evolution.

In Figure 5, it can be observed that the improvement from w/o Any Evolve to Full Co-evolve is much larger than that from w/o World Model to w/o Any Evolve. If evolution is the main contributor to the performance improvement, then I further question the novelty of the paper, since experience-driven evolution is not a new concept and has already been demonstrated to be effective in many existing studies. Another point is that the first experiment on nanoGPT already shows significant performance improvement, before any within-run experimental feedback has been received. Please clarify whether the initial skills, prompts, and proposals are consistent across different settings, and what explains these early performance differences.

Q2.3 Limited robustness and modest improvements on the main benchmark.

MLS-Bench-Lite contains only 30 tasks, and the improvements of JANUS over the baselines are relatively modest. It is unclear whether these improvements are stable across different runs and benchmarks. I would suggest conducting additional evaluations on other benchmarks, such as the full MLS-Bench, MLE-Bench, and ResearchGym.

Q2.4 The claimed efficiency improvements are uneven across tasks.

In Table 2, it is clear that the reported average speedup is heavily influenced by nanoGPT, while the improvements on MinGPT and G2P are much smaller. These results suggest that the efficiency gains vary considerably across tasks, which requires further explanation and discussion.

Reviewer_CgNCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

JANUS introduces a Predict–Verify–Evolve framework for autonomous research. A research agent generates candidate experiments within an evolving research graph, while an LLM-based world model evaluates their expected performance improvement, information gain, and opportunities for subsequent exploration. Experimental outcomes inform updates to research skills, predictive judgments, exploration strategies, and coordination between the agents. The paper evaluates the framework on 30 MLS-Bench-Lite tasks and three extended research problems covering language-model training, attention-kernel optimization, and genotype-to-phenotype prediction. Comparisons and ablations show improvements in research performance and experiment efficiency, supporting the practical value of the overall framework.

Critical strengths
  • It is an interesting idea that maintains expectations about research actions and revisits those expectations after observing experimental outcomes. Factoring in information gain and future branching value also prevents the agent from being overly greedy. The paper develops this perspective into a coherent framework for long-horizon autonomous research.

  • The paper presents a thorough empirical evaluation across 30 research tasks and multiple backbones with extended trajectories in three distinct application settings. The ablations and OOD transfer tests effectively isolate the value of adaptive exploration and agent–world model co-evolution.

Critical weaknesses
  • The positioning relative to the closest predictive and co-evolving research systems needs to be sharper. K-Search [1] already uses an LLM world model to prioritize optimization intents and revise its search state using execution feedback. WebEvolver [2] explicitly studies co-evolution of an agent and world model. In addition, ForeAgent [3] implements predictive candidate comparison within a Predict-then-Verify loop, but this is not discussed. These connections make the broad contrast between reactive existing systems and JANUS insufficiently precise. The paper should explain its advances in adaptive research-value assessment, skill abstraction, and coordinated evolution relative to these specific predecessors.

[1] Cao, Shiyi, et al. "K-search: Llm kernel generation via co-evolving intrinsic world model." arXiv preprint arXiv:2602.19128 (2026).

[2] Fang, Tianqing, et al. "Webevolver: Enhancing web agent self-improvement with co-evolving world model." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

[3] Zheng, Jingsheng, et al. "Can We Predict Before Executing Machine Learning Agents?." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Questions for authors
  • Beyond the high-level framing, what is the core methodological delta between JANUS and K-Search, WebEvolver, and ForeAgent?

  • Is Eq. 12 meant to be anything more than a heuristic acquisition function? Concretely, how do prediction errors update the world model's skills and retroactively adjust the value of ancestor nodes?

  • Can you show a concrete trajectory where a bad prior prediction is corrected by experimental feedback, yielding a specific skill update that visibly alters a downstream decision?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.