PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.13194v1cs.AIcs.CLcs.MA

TL;DR

Costly execution makes reactive trial-and-error difficult to sustain as autonomous research search spaces expand. JANUS combines pre-execution forecasting and post-feedback experience reuse in a co-evolving loop, and evaluations report stronger results across benchmark and real-world tasks.

  • Problem

    Generate–Execute–Feedback research relies on costly, time-consuming execution, while expanding experiments create combinatorially branching research trajectories.

    Source: §1 · p. 1

  • Method

    JANUS uses an evolving research graph and co-evolving world model to score, select, refine, or prune candidate branches before execution, then distills feedback into skills.

    Source: §1 · p. 2

  • Results

    Across MLS-Bench-Lite and three real-world research tasks, JANUS outperforms strong baselines with greater evolving efficiency and a higher evolving ceiling.

    Source: §5 · p. 10

  • Takeaways & Limitations

    JANUS’s forecasting before execution and experience reuse after feedback offer an approach to long-horizon exploration beyond myopic trial-and-error.

    Source: §1 · p. 2

    The paper formulates AutoResearch as a budget-constrained partially observable Markov decision process.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

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.

1 INTRODUCTION

Reactive Generate–Execute–Feedback research is slowed by costly execution as search spaces expand, motivating systems that scale progress over long horizons. JANUS forecasts and prioritizes research branches before execution, reuses experience through co-evolving skills, and outperforms strong baselines on evolving efficiency and ceiling.

1 INTRODUCTION

  • The Generate–Execute–Feedback paradigm incurs costly, time-consuming execution, and its trial-and-error burden grows as experiments spawn combinatorially branching research trajectories.Executions can require hours of model training, simulation, and compilation.Source: §1 · p. 1
  • The scaling objective is to improve both evolving efficiency—the rate of accumulated progress—and evolving ceiling—the distance continued exploration advances.The paper frames this as a research-budget versus achievable-progress scaling problem and advocates forecasting, adaptive resource allocation, and reusable experience.Source: §1 · p. 2
  • JANUS uses an evolving research graph and a world model to forecast candidate value, information gain, and exploration utility, then select, refine, or prune branches before execution.After feedback arrives, it distills execution logs into hierarchical, role-specific skills for the research agent, world model, and their coordination.Source: §1 · p. 2
  • JANUS significantly outperforms state-of-the-art baselines across MLS-Bench-Lite and real-world research tasks, with superior evolving efficiency and a higher evolving ceiling.Evaluations cover algorithm optimization, system design, and scientific discovery; JANUS achieves higher evolving efficiency with lower compute budgets and a higher ceiling before plateauing.Source: §1 · pp. 2–3, Figure 1 · p. 2
  • JANUS’s forward-looking world model and backward-looking use of ground-truth feedback drive co-evolution of the agent, model, and their coordination.The paper’s name invokes the two-faced Roman god who looks toward both future and past.Source: §1 · p. 2

2 RELATED WORKS

Prior autonomous research agents use pipeline, search, or multi-agent designs, while LLM world-model work often targets fixed tasks. JANUS is distinguished by an evolving world model and shared research graph that support adaptive direction selection and experience use.

2.1 AUTONOMOUS RESEARCH AGENT

  • Pipeline systems organize research into stages, but can propagate early errors and offer limited flexibility to revisit alternatives.The AI Scientist established an end-to-end idea-to-paper workflow, while Agent Laboratory uses specialized assistants and human supervision.Source: §2.1 · p. 3
  • Search-based agents explore and refine solutions, but rely on costly execution feedback and lack adaptive prioritization of research directions; JANUS selects directions before execution.Examples include iterative code-space search, targeted pipeline refinement, tree search, and budget-aware search.Source: §2.1 · p. 3
  • Multi-agent research systems divide work across roles, yet face challenges in information exchange and role-specific experience use; JANUS addresses these with a shared evolving graph and co-evolution.Prior approaches include researcher–developer coupling, persistent memory, adversarial executor–reviewer collaboration, and hypothesis-tree coordination.Source: §2.1 · p. 3

2.2 LLMS AS WORLD MODELS

  • Most existing LLM world-model approaches target relatively fixed tasks and do not continually update the model, whereas JANUS evolves its world model with the research process.Prior work includes LLM planning, structured domain models, causal world models, and simulation of candidate action outcomes.Source: §2.2 · p. 3

3 THE JANUS FRAMEWORK

JANUS organizes long-horizon research around an evolving graph and a Predict-Verify-Evolve loop, where a world model guides which proposals are pursued. Experimental feedback updates research, prediction, exploration, and coordination skills over time.

3.1 PROBLEM FORMULATION

  • JANUS models research as a budget-constrained POMDP over an evolving graph whose nodes track proposals, estimates, implementations, and experimental results.Edges connect proposals derived by modifying, extending, or combining earlier ones; reward is the improvement in best empirical performance so far.Source: §3.1 · pp. 3–4

3.2 THE PREDICT-VERIFY-EVOLVE LOOP

  • The Predict-Verify-Evolve loop expands the graph, forecasts and refines candidate value before execution, then records experimental feedback for continual co-evolution.The orchestrator conditions proposal generation on the graph, remaining budget, and available skills; the world model uses prior evidence to assess proposals and may debate refinements.Source: §3.2 · pp. 4–5

3.3 WORLD MODEL GUIDED EXPLORATION STRATEGY

  • The world model scores proposals for value improvement, information gain, and exploration utility, then adjusts priorities to prune weak options or refine promising ones.Weights shift with recent progress and remaining budget; proposal comparisons use a sparse pairwise tournament that prioritizes uncertain comparisons.Source: §3.3 · pp. 5–6

3.4 EXPERIENCE-DRIVEN CO-EVOLUTION

  • JANUS distills experimental experience into agent skills, world-model prediction and exploration skills, and shared coordination skills from prior debates and refinements.The agent abstracts raw trajectories into intra-task and cross-task skills, while the world model reviews past judgments against accumulated experimental evidence.Source: §3.4 · pp. 6–7

4 EXPERIMENTS

The experiments compare JANUS with research-agent baselines on MLS-Bench-Lite and real research tasks. Ablations and framework analyses examine its exploration strategy, co-evolution, generalization, and changing research behavior.

4.1 EXPERIMENTAL SETUP

  • The evaluation spans MLS-Bench-Lite's 30 tasks across 12 domains and three real research tasks, comparing JANUS with six baselines under shared protocols.MLS-Bench-Lite metrics include validity, human-baseline comparisons, and development and test scores; each task receives a five-hour budget.Source: §4.1 · p. 7

4.2 PERFORMANCE ANALYSIS

  • Figure 3 tracks best-so-far performance, individual experiment scores, and final bests across three end-to-end research tasks.Table 2 reports average and maximum sample-efficiency speedups over competing agents.Source: Figure 3 · p. 8, §4.2 · p. 8, Table 2 · p. 8
  • 100.0% valid improvements over its weakest human baseline; JANUS outperforms all baselines on six metrics and beats all human baselines on 43.3% of tasks.It maintains its advantage across different backbones; integrating the loop with Codex achieves a peak score of 49.78.Source: §4.2 · p. 8

4.3 ABLATION STUDY

  • Fixed exploitation-only and exploration-only strategies reduce performance, while the full framework finishes strongest across all three tasks.Without refinement, progress is lower during sustained stagnation; adaptive priority weighting and candidate refinement provide complementary benefits.Source: §4.3 · pp. 8–9, Figure 4 · p. 8
  • Removing the world model substantially reduces performance across tasks, and disabling either component's evolution also degrades results.Without world-model evolution, progress is more likely to stall at plateaus; without agent evolution, improvement slows throughout.Source: §4.3 · p. 9, Figure 5 · p. 9

4.4 FRAMEWORK ANALYSIS

  • JANUS remains strong at low baseline similarity and achieves the highest final ID and OOD scores with the smallest ID–OOD gap.In the nanoGPT process, exploration rises at plateaus and early information-gain increases precede major performance breakthroughs.Source: §4.4 · p. 9, Figure 6 · p. 9

5 CONCLUSION

JANUS combines a research graph and co-evolving world model to assess proposals across complementary research-value criteria. Evaluations report stronger performance on MLS-Bench-Lite and real-world tasks, with greater evolving efficiency and a higher evolving ceiling.

5 CONCLUSION

  • JANUS outperforms strong baselines on MLS-Bench-Lite and three real-world research tasks, achieving greater evolving efficiency and a higher evolving ceiling.The reported efficiency advantage holds at lower compute budgets; the ceiling is higher before plateauing.Source: §5 · p. 10
  • JANUS uses an evolving research graph and world model to guide proposal refinement and selection, then distills experimental feedback into agent, world-model, and coordination skills.The world model balances value improvement, information gain, and exploration utility.Source: §5 · p. 10

A FORMAL DEFINITIONS OF RESEARCH PRIORITY CRITERIA

  • The three criteria score proposals for expected best-performance gains, reduced uncertainty about the next research direction, and new follow-up opportunities under the remaining budget.A proposal may be valuable even without immediate improvement if it clarifies a research direction or expands later options.Source: A FORMAL DEFINITIONS OF RESEARCH… · p. 14

B BASELINES

  • Comparisons include AutoResearch, AIDE, MLEvolve (Du et al., 2026), AIBuildAI, Arbor (Jin et al., 2026), and Codex (OpenAI, 2025), with original research-loop designs preserved.Methods on the same task receive identical initial files, objectives, evaluator access, and resource budgets.Source: B BASELINES · pp. 14–15

C.1 THREE REAL RESEARCH TASKS

The evaluation covers three executable research problems and a 30-task MLS-Bench-Lite subset spanning 12 research domains. The task setups define constrained, executable objectives and evaluation conditions.

C.1 THREE REAL RESEARCH TASKS

  • NanoGPT training optimization, adapted from AutoResearch, seeks the lowest validation bits per byte within a fixed training-time budget.The model architecture, training loop, and related components are editable, while data preparation, tokenizer, validation split, and metric implementation remain fixed.Source: §C.1 · p. 15
  • MinGPT causal-attention optimization uses KernelBench Level 3, Task 43 (Ouyang et al., 2025), and maximizes speedup over eager PyTorch after correctness checks pass.Implementations may use fused operators, compilation, Triton, or custom CUDA while preserving numerical output.Source: §C.1 · p. 15
  • Genotype-to-phenotype prediction uses maize hybrid SNP dosages and environmental conditions to predict eight agronomic traits, with performance macro-averaged across traits.Macro-averaging prevents traits with more observations or larger numerical scales from dominating the score.Source: §C.1 · p. 15

C.2 MLS-BENCH-LITE.

  • MLS-Bench-Lite is a 30-task subset spanning all 12 research domains, with controlled scaffolds and evaluation across multiple prescribed settings.Its tasks target improvements that remain effective across datasets, environments, model scales, or random seeds.Source: §C.2 · pp. 15–16

D.1 MLS-BENCH-LITE METRICS

The metrics normalize and aggregate benchmark scores, assess valid submissions against human baselines, and quantify sample efficiency and task-specific outcomes. Invalid submissions and uneven performance across evaluation settings are explicitly penalized.

D.1 MLS-BENCH-LITE METRICS

  • MLS-Bench-Lite metrics transform raw scores into higher-is-better values, anchoring the worst reproduced baseline at zero and the best at 0.5, with bounded and sigmoid normalization options.Finite-upper-bound metrics use bounded normalization; metrics without a finite theoretical upper bound use a calibrated sigmoid.Source: §D.1 · p. 16
  • Within-task scores use weighted arithmetic means across metrics and geometric means across evaluation settings; invalid submissions score zero, and Dev Score and Test Score average across 30 tasks on a 0–100 scale.The geometric mean penalizes uneven performance across settings.Source: §D.1 · p. 16
  • Beat Any, Beat Median, and Beat All report the percentages of tasks where a valid final submission strictly exceeds the corresponding human-baseline threshold; invalid submissions and ties do not count as wins.All 30 tasks are included in the denominator.Source: §D.1 · p. 17

D.2 EVALUATION ON THE THREE RESEARCH TASKS

  • Validation bits per byte is the average bits needed to encode one byte of validation text, with lower values indicating better language modeling.Tokens with zero byte length are excluded.Source: §D.2 · p. 17
  • Kernel speedup compares mean candidate and eager-reference runtimes, and values above one indicate a faster candidate; correctness failures receive zero.The metric is conditional on passing all correctness checks.Source: §D.2 · p. 17
  • The G2P score is the unweighted mean of Pearson correlations across eight traits, and higher values indicate predictions that better preserve phenotype variation.Per-trait Pearson, Spearman, and within-environment Pearson correlations are diagnostic statistics.Source: §D.2 · p. 17
  • Sample-efficiency speedup compares experiment counts at jointly reachable performance thresholds; values above one mean the baseline needs more experiments than JANUS.The reported average is uniform over jointly reachable thresholds, while the maximum reports the largest threshold-wise speedup.Source: §D.2 · p. 18

E IMPLEMENTATION DETAILS

The implementation fixes common compute, backbone, workspace, and evaluation conditions while enforcing task-specific correctness and resource constraints. Each research task also has a specified editable surface and evaluation protocol.

E IMPLEMENTATION DETAILS

  • Common-backbone comparisons use one A100 GPU and hold the model endpoint, reasoning configuration, prompt, workspace, permissions, evaluator, and research budget fixed, without sharing trajectories or memories.DeepSeek-V4-Flash is the primary common language-model backbone.Source: E IMPLEMENTATION DETAILS · p. 18
  • Immutable external harnesses accept candidates only when edits stay within declared surfaces and satisfy task-specific resource and correctness constraints; failures are recorded as failed experiments.The benchmark’s capacity limits, protected settings, evaluation data, and scoring procedures are retained.Source: E IMPLEMENTATION DETAILS · pp. 18–19
  • MLS-Bench-Lite runs receive a five-hour wall-clock budget per task, a shared timeout policy, and a fixed validation split; parallel execution is used when resources permit.Evaluator timeouts are relaxed for A100 hardware while the agent research budget remains unchanged.Source: E IMPLEMENTATION DETAILS · p. 19
  • NanoGPT candidates receive 1,200 seconds of training and may edit only train.py; tokenizer, dataloader, validation shard, and evaluation routine are fixed.Training uses sequence length 2,048 and an 8,192-entry tokenizer vocabulary.Source: E IMPLEMENTATION DETAILS · p. 19
  • MinGPT candidates may edit only solution.py and must pass five seeded correctness tests; speed is measured against a frozen 32.6 ms eager-reference latency.The workload uses batch size 128, sequence length 512, embedding dimension 768, and FP32 inputs and outputs.Source: E IMPLEMENTATION DETAILS · p. 19
  • G2P candidates may edit only train.py, fit label- or distribution-dependent preprocessing on training data, and use shared trait splits with disjoint hybrid and environment partitions.The principal score is macro Pearson; per-trait Pearson, Spearman, and within-environment Pearson are diagnostic statistics.Source: E IMPLEMENTATION DETAILS · pp. 19–20

F OUT-OF-DISTRIBUTION EVALUATION SETUP AND ANALYSIS

JANUS’s final solutions retain research-time improvements across independent workload, corpus, and cohort shifts, with the highest aggregate scores under both in-distribution and out-of-distribution evaluation. The evaluation uses task-specific shifts and a shared calibrated scale for comparing results across tasks.

OOD evaluation setup

  • Each OOD setting changes one scientifically meaningful axis while retaining the original evaluator and primary metric.The settings shift the attention workload, replace ClimbMix with TinyStories for nanoGPT, and evaluate G2P on a separate maize-hybrid cohort.Source: F OUT-OF-DISTRIBUTION EVALUATION SETUP… · p. 20

Cross-task score calibration

  • A fixed shared calibration makes the task metrics comparable, suppresses negative gains, and permits scores above 100; the cross-task aggregate uniformly weights tasks.The 100-point references are a four-fold KernelBench speedup, a 0.05 absolute nanoGPT BPB reduction, and a 0.15 absolute G2P macro-Pearson improvement.Source: F OUT-OF-DISTRIBUTION EVALUATION SETUP… · p. 21

OOD generalization analysis

  • Figure 6(b) compares initial and final aggregate ID and OOD scores, with no further research or tuning in the shifted settings.Movement from initial to final indicates whether research-time improvements remain effective beyond the original optimization setting.Source: F OUT-OF-DISTRIBUTION EVALUATION SETUP… · p. 21
  • JANUS’s final solution reaches the highest aggregate score under both ID and OOD evaluation and remains close to the diagonal across three distinct shift types.The shifts are a workload change for the attention kernel, a corpus change for nanoGPT, and a cohort change for G2P.Source: F OUT-OF-DISTRIBUTION EVALUATION SETUP… · p. 21

G ANALYSIS OF BASELINE IMITATION

The analysis compares agent performance with a similarity-weighted measure of resemblance to human-designed baselines. JANUS’s fitted performance is comparatively high at low similarity-weighted baseline performance and depends less on baseline quality than other agents.

Similarity-weighted baseline performance

  • The analysis measures code similarity to each baseline using 1/3-gram Jaccard similarity and weights baseline performance by those similarities.The resulting measure is normalized, along with agent performance, by the task’s best baseline performance.Source: G ANALYSIS OF BASELINE IMITATION · p. 21
  • JANUS achieves comparatively high fitted performance at low similarity-weighted baseline performance and has a shallower fitted slope than other agents.Its performance therefore varies less with the quality of the baselines its solutions resemble.Source: G ANALYSIS OF BASELINE IMITATION · p. 22
  • JANUS’s weaker dependence on baseline similarity is presented as evidence that it identifies effective alternatives beyond close resemblance to strong existing implementations.The authors connect this capacity to broader exploration and a higher potential performance ceiling.Source: G ANALYSIS OF BASELINE IMITATION · p. 22

H ANALYSIS OF ADAPTIVE RESEARCH STRATEGIES

The analysis characterizes selected nanoGPT proposals by their research emphasis over the actual execution sequence, then plots smoothed strategy scores alongside a normalized performance trace.

Trajectory and assessment

  • The analysis covers completed formal evaluations in execution order and uses GPT-5.6 Sol to judge proposal text without candidate outcomes or later results.Earlier experiments cited in a proposal remain available to the judge as evidence known when it was formulated.Source: H ANALYSIS OF ADAPTIVE RESEARCH… · p. 22

Strategy dimensions

  • Proposals are scored on exploitation, information gain, and exploration, which are non-exclusive dimensions that need not sum to one.The scores reflect immediate improvement from supported mechanisms, uncertainty reduction, and investigation of unverified mechanisms, respectively.Source: H ANALYSIS OF ADAPTIVE RESEARCH… · p. 22

Temporal smoothing

  • A centered seven-experiment moving average smooths each strategy dimension; Figure 6(c) distinguishes original scores from smoothed curves.Faint colored lines and points show original per-proposal scores, while thick curves show moving averages.Source: H ANALYSIS OF ADAPTIVE RESEARCH… · p. 22

Normalized performance trace

  • The purple staircase tracks incumbent validation BPB improvement, normalized from zero at the initial baseline to one at the final incumbent.Because lower BPB is better, the trace uses the best result seen so far and holds its value until the next incumbent improvement.Source: H ANALYSIS OF ADAPTIVE RESEARCH… · p. 22

Related papers

Loading 2610.13194v1…