PaperlayerView exact revision on arXiv

Source-linked AI summary

Recursive Self-Improvement through Multi-Agent Self-Supervision

Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia, Yujin Tang

arXiv:2610.12176v1cs.AI

TL;DR

When open-ended research outputs exceed reliable human assessment, a model may have to serve as its own optimizer and evaluator, while a single instance struggles to critique its own reasoning. MASS uses coordinated copies of the model to generate self-supervision; after two cycles, it improves score per output token across four research benchmarks. The authors conclude that jointly learning orchestration and bounded subagent execution can provide effective supervision for RSI.

  • Problem

    For open-ended tasks, reliable human supervision can be unavailable, while a single model instance struggles to critique and improve its own reasoning.

    Source: §1 · p. 2

  • Method

    MASS coordinates copies of a model as a multi-agent system and distills their collective gains as self-supervision for recursive self-improvement.

    Source: §1 · p. 2, §7 · p. 10

  • Results

    Across four research benchmarks, MASS’s L(2) achieves 1.2–1.6× higher score per output token than base L(0).

    Source: §1 · p. 3, §4.2 · p. 6

  • Takeaways & Limitations

    Jointly learning orchestration and bounded subagent execution from multi-agent trajectories can provide more effective supervision for RSI.

    Source: §1 · p. 3

    The selected workflow is the best self-judged option under sequential in-loop pairwise evaluation, not a certified global optimum.

    Source: §3 · p. 5

Abstract

from arXiv · show

Recursive self-improvement (RSI) of a model on non-verifiable tasks, such as open-ended research, faces a supervision bottleneck when its outputs exceed what even human experts can reliably assess, leaving the model itself (optimizee) as the best available optimizer and evaluator. However, a single model instance struggles to critique and improve its own complex reasoning under this homogeneous loop. To address this, we propose Multi-Agent Self-Supervision (MASS), an RSI method that alternates between evolutionary workflow optimization and supervised fine-tuning on self-generated trajectories. Guided by early findings that multi-agent topologies excel at complex reasoning, MASS prompts a single base model to iteratively propose, execute, and self-evaluate multi-agent workflows. Through an evolutionary search constrained by structural guardrails, the model optimizes these computational-graph-like orchestrations, discovering the most effective distinct roles and information routing for a given task. Over two MASS cycles with Qwen3.6-27B, the model achieves 1.2-1.6x higher performance per output tokens on four open-ended public benchmarks. Because the improved model subsequently acts as a better optimizer and evaluator, this alternating framework enables a continuous, recursive bootstrapping of the model's capabilities. Moreover, multi-agent traces are also more training-efficient: a student trained on them outperforms a single-agent student trained on 1.4x more training tokens. These findings suggest that jointly learning orchestration and bounded subagent execution from multi-agent trajectories can provide an effective signal for RSI.

1 INTRODUCTION

MASS addresses the supervision bottleneck in recursive self-improvement by using coordinated model copies to generate self-supervision, then alternates workflow optimization with model fine-tuning. Across four research benchmarks, two cycles improve score per output token, while multi-agent trajectories also compare favorably with single-agent training data.

1 INTRODUCTION

  • A single model serving as its own supervisor faces a bottleneck when outputs exceed reliable human assessment and self-feedback may reinforce shared-prior errors.The introduction also identifies slow feedback and the effort of manually optimizing workflows as challenges.Source: §1 · p. 2
  • Coordinating model copies is intended to elicit capabilities an isolated instance may not reliably exhibit and to break complex tasks into simpler pieces.The paper draws on the Society of Mind perspective that intelligence can emerge from organized interactions among simpler components.Source: §1 · p. 2
  • MASS focuses workflow optimization on what agents communicate, to whom, and when, rather than simply increasing the number of agents.The authors motivate this focus with evidence that optimizing communication structure can improve collective performance.Source: §1 · p. 2
  • MASS alternates multi-agent workflow optimization with fine-tuning on trajectories, and the updated model resumes the agent, optimizer, and evaluator roles.Final performance is assessed separately using external strong-model judgments withheld from the agents and workflow optimizer.Source: §1 · p. 2
  • Across four research benchmarks, L(2) achieves 1.2–1.6× higher score per output token than base L(0).Figure 2 compares the base model with models after one and two MASS cycles in the same agent scaffold.Source: Figure 2 · p. 3, §1 · p. 3
  • A first-cycle multi-agent student reaches a 68.3% win rate against the base model, versus 64.0% for a single-agent student trained on approximately 1.4× more processed supervised tokens.The authors hypothesize that orchestrator and subagent trajectories teach delegation and bounded execution at complementary levels of abstraction.Source: §1 · pp. 2–3

2 NOTATIONS AND PROBLEM STATEMENT

The paper defines notation for tasks, language models, agent executions, and evaluation in a homogeneous setting where the same model serves as executor, evaluator, and optimizer. Its goal is to improve task outputs across RSI generations as judged by an external strong model.

2 NOTATIONS AND PROBLEM STATEMENT

  • The task set T is divided into training and disjoint test subsets, and Lθ denotes a language model parameterized by θ that defines next-token probabilities.For textual input x, L(x) denotes a sample from the model conditioned on x.Source: §2 · p. 3
  • An agent execution produces a trajectory and a deliverable workspace, while its harness, tools, and memory remain fixed throughout the paper.The agent is denoted A(·; L), with L specifying its language model.Source: §2 · p. 3
  • In the homogeneous setting, the agent, evaluator, and optimizer are the same language model, which compares outputs and can provide a rationale or feedback.The notation uses Lagent = Leval = Lopt.Source: §2 · p. 4
  • MASS aims to increase the probability that an output from a later RSI generation is preferred to earlier outputs by an external strong language model.The external model is used for fair comparisons.Source: §2 · p. 4

3 METHOD

MASS recursively alternates task-specific multi-agent workflow optimization with supervised fine-tuning on trajectories generated under the retained workflows. The current model performs all three roles—task-solving agent, workflow optimizer, and evaluator—and assumes them again after each update.

3 METHOD

  • MASS alternates workflow optimization and model fine-tuning, then repeats both stages with the updated model serving as agent, optimizer, and evaluator.The optimization stage retains task-specific workflows; fine-tuning generates and self-selects trajectories under those workflows.Source: §3 · p. 4
  • A workflow specifies an orchestrator–subagent architecture through subagent roles, instructions, output contracts, and call order.The orchestrator delegates work, exchanges information, and invokes subagents; all share the same task-solving model.Source: §3 · p. 4
  • Workflow optimization compares each current output with the best-so-far output, uses evaluator feedback to propose revisions, and emphasizes communication structure.The selected workflow is the best self-judged result under sequential in-loop pairwise evaluation, not a certified global optimum.Source: §3 · p. 5
  • For each training task, MASS generates multiple trajectory–workspace pairs, ranks them using pairwise evaluator judgments and Bradley–Terry scores, and reserves one trajectory for validation.The top K −1 trajectories form the training set, and the retained workflow is removed from the dataset-construction prompt.Source: §3 · p. 5
  • Fine-tuning partitions trajectories into token windows and applies loss only to supervised assistant targets, excluding prompt tokens, tool outputs, and masked context.Preceding context may be retained while masked from the loss.Source: §3 · p. 5

4 EXPERIMENTS

Across two MASS cycles, Qwen3.6-27B improves on open-ended research benchmarks and synthetic test tasks. The experiments also examine workflow information routing and whether multi-agent trajectories provide more efficient fine-tuning supervision.

4.1 EXPERIMENTAL SETUP

  • The setup trains Qwen3.6-27B on eight synthetic research tasks and evaluates it on three synthetic test tasks and six public benchmarks.The training tasks span finance, robotics, and pharmacy; the public benchmarks include ScienceAgentBench, MLR-Bench, DSBench, AstaBench, Terminal-Bench 2.0, and SWE-bench Verified.Source: §4.1 · p. 5

4.2 MAIN RESULTS: IMPROVEMENT ACROSS MODEL ROLES

  • L(2) achieves 1.2–1.6× higher score per output token than L(0) across four research benchmarks after two MASS cycles.Performance rises on ScienceAgentBench from 28.7% to 31.8%, MLR-Bench from 1.58 to 2.40, AstaBench from 0.053 to 0.067, and DSBench from 0.473 to 0.482.Source: §4.2 · p. 6
  • On three test tasks, L(2) reaches a 69.9% win rate against L(0), and workflows optimized by L(1) outperform earlier workflows in 64% of transfer comparisons.Evaluator agreement with external judgments rises from 0.73 for L(0) to 0.91 for L(1) and 0.93 for L(2).Source: §4.2 · pp. 6–7

4.3 HYPOTHESIS 1: INFORMATION FLOW SUPPLIES USEFUL SELF-SUPERVISION

  • By iteration 11, role and instruction information falls while contract and hop information rises in all four configurations; total correlation falls in all four.Task-conditioned redundancy falls in three configurations but rises for permutation-adjusted MPNet; overall, the findings indicate increasingly task-specific information in communication and routing with less redundancy.Source: §4.3 · p. 7, Figure 4 · p. 8, Figure 5 · p. 8

4.4 HYPOTHESIS 2: MULTI-AGENT TRAJECTORIES PROVIDE DENSER SUPERVISION

  • M+ reaches 68.3% with 28.82M processed supervised tokens, 4.3 points above S++ while using 31.4% fewer tokens.Figure 6 compares win rate against L(0) as a function of processed supervised tokens during fine-tuning.Source: Figure 6 · p. 8, §4.4 · p. 8
  • Among 10 episode pairs with similar generated trace lengths, multi-agent runs are preferred in 80.0% of external judgments, suggesting the quality advantage is not solely due to longer inference.Multi-agent executions generate approximately 2.4× more tokens on average and are preferred in 85.1% of external judgments overall.Source: §4.4 · p. 8
  • The proposed supervision combines orchestrator coordination with bounded subagent execution, treating them as complementary prediction problems at different abstraction levels.The same parameters are updated by both losses.Source: §4.4 · pp. 7–9

5 ABLATIONS

The ablations find that optimizer strength affects how quickly workflow search reaches successful workflows, while post-training internalizes coordination behaviors. The paper concludes that multi-agent trajectories provide useful self-supervision for RSI.

5.1 WHAT IS THE BOTTLENECK IN WORKFLOW OPTIMIZATION?

  • A stronger optimizer reaches successful workflows earlier, while the all-L(0) loop eventually achieves unanimous wins on 11 of 12 tasks.Replacing only the evaluator with GPT-5.5 yields less improvement than replacing both evaluator and optimizer.Source: §5.1 · p. 9, Figure 7 · p. 9
  • Textual workflow-edit size is a weak predictor of performance change, with little Spearman correlation (|ρ_edit| ≤ 0.12) and reversals after both small and large edits.Improved workflow reachability does not guarantee stable optimization.Source: §5.1 · p. 9

5.2 WHAT BEHAVIOR IS INTERNALIZED?

  • Under bare task prompts, delegation rises from 1/110 trajectories for L(0) to 48/110 for L(1) and 87/110 for L(2), while code handoffs rise from 1/110 to 42/110 and 83/110.These behaviors are represented in the SFT data, supporting the authors’ interpretation that post-training internalizes coordination.Source: §5.2 · pp. 9–10, Table 2 · p. 10

6 RELATED WORKS

  • Related work spans self-improvement, multi-agent workflow design, and model–workflow co-evolution; MASS distills optimized workflows’ orchestrator and subagent traces into shared weights.Unlike formal self-rewriting approaches, MASS uses empirical workflow search and SFT without formal guarantees.Source: §6 · p. 10

7 CONCLUSION

  • MASS forms a multi-agent system from copies of the same model and distills their collective gains as self-supervision, while faster performance gains per cycle remain future work.Source: §7 · p. 10

A ALGORITHMS

  • The algorithms separate prompt-changing workflow search from weight-changing SFT, with the current model serving as executor, evaluator, and optimizer within a generation.Source: A ALGORITHMS · p. 15

A.1 MULTI-AGENT WORKFLOW OPTIMIZATION

  • Workflow optimization retains the best workflow under sequential self-comparisons, but executes every proposal, so successive executions need not improve.Source: §A.1 · p. 15

A.2 WORKFLOW INTERNALIZATION

  • Workflow internalization ranks trajectory–workspace pairs, assigns top-ranked examples to training and one to validation, then fine-tunes the model on rendered windows.Source: §A.2 · p. 15

B.1 TASKS, TEACHER COLLECTION, AND VALIDATION

The appendix describes task allocation, trajectory collection and SFT, teacher execution instructions, and evaluation procedures. It also reports workflow transfer and a matched-prompt coordination control.

B.1 TASKS, TEACHER COLLECTION, AND VALIDATION

  • Eight of nine designated training tasks contribute SFT data; task 221 is excluded because workflow search finds no workflow that beats its reference, and three tasks are held out from post-training.Source: §B.1 · p. 16, Table 3 · p. 16
  • For each of eight training tasks, the current model executes its retained workflow with 18 seeds; self-evaluation ranks trajectories, with the top 15 used for training and rank 16 for validation.The first-cycle set contains 119 training trajectories after one task-56 trajectory is excluded.Source: §B.1 · p. 16

B.2 CONVERSATION RENDERING AND SFT

  • SFT reconstructs conversations with the orchestrator’s initial workflow prompt replaced by the bare task, retaining worker assignments and masking non-assistant content from loss.The first-cycle set yields 2,435 windows and 17,221,055 supervised tokens; checkpoint selection uses validation loss.Source: §B.2 · pp. 16–17

B.3 TEACHER EXECUTION DIRECTIVE

  • The teacher directive requires actual separate subagent calls, foreground execution, staged bounded work, verification, and a clean final report.It is removed, along with the workflow, from the orchestrator’s training input.Source: §B.3 · p. 17

C.1 EVALUATION ROLES AND REPORTING UNIT

  • In homogeneous MASS, the same current model provides workflow feedback and teacher-ranking judgments, while external panel judgments are withheld from agents and are not SFT targets.Workspace pairs usually receive six judgments, treated as repeated assessments rather than independent task executions.Source: §C.1 · p. 18, Table 4 · p. 18

C.2 FIRST-CYCLE TASK-SOLVING RESULTS

  • The teacher comparison pairs student rollouts with ranked eligible teachers; its 108 pairs use selected workflow-prompted references, and none of the test trajectories enters SFT or validation.The 17.3% share measures a gap to those selected references, not a bare-prompt model comparison.Source: §C.2 · p. 18, Table 6 · p. 19

C.3 TRANSFER TO WORKFLOW GENERATION

  • Workflows optimized by L(1) outperform those optimized by L(0) in 64% of comparisons on average across 11 iterations and 12 tasks when both workflow sets are executed by L(0).This comparison holds execution weights fixed to assess workflow quality.Source: §C.3 · pp. 18–19

C.4 GENERIC-DIRECTIVE CONTROL

  • Under the matched explicit team directive, M+ records 352 wins, 306 losses, and two ties against the base, a 53.5% share.The four shared training tasks alone show a 47.1% share for M+ versus 40.8% for M; these four-task results are not comparable with the 11-task aggregate.Source: §C.4 · pp. 19–20

C.5 PUBLIC BENCHMARKS

Public-benchmark results vary by benchmark and metric. MLR-Bench’s higher aggregate score accompanies more delivered papers, while the reported DSBench and software-task scores do not show a uniform increase across generations.

C.5 PUBLIC BENCHMARKS

  • ScienceAgentBench uses a 102-task verified split and three trials per model; reference programs reached 80.4% in a local environment check, limiting comparison with other environments.The authors report three-trial means descriptively, without a significance claim.Source: §C.5 · p. 20
  • MLR-Bench aggregate scores are higher for L(2), alongside delivery increasing from 31/60 to 50/60; mean scores among delivered papers are 3.07, 3.00, and 2.88 for L(0), L(1), and L(2).The conditional means are from different completed-task sets, and the paired task-mean tests do not establish a consistent gain across research tasks.Source: §C.5 · pp. 20–21, Table 7 · p. 20
  • The reported AstaBench rubric scores are 0.053 ± 0.002, 0.055 ± 0.004, and 0.067 ± 0.017 for L(0), L(1), and L(2), with report deliveries of 60/120, 69/120, and 91/120.Mean scores among delivered reports are 0.105, 0.095, and 0.088, respectively, but these refer to different completed-task sets.Source: §C.5 · pp. 20–21, Table 7 · p. 20
  • For L(0), L(1), and L(2), DSBench scores are 0.473 ± 0.029, 0.465 ± 0.011, and 0.482 ± 0.025, respectively.Source: §C.5 · pp. 20–21, Table 7 · p. 20
  • Terminal-Bench 2.0 pass rates are 37.5 ± 2.2, 35.3 ± 2.1, and 35.2 ± 1.7 for L(0), L(1), and L(2), while SWE-bench Verified rates are 62.6 ± 1.4%, 61.8 ± 1.9%, and 61.1 ± 0.6%.Source: §C.5 · pp. 21–22

D.1 DATASETS AND EVALUATION SCOPE

These sections document the evaluation and token-accounting choices, compare generated teacher outputs, and describe workflow-search conditions and measures.

D.1 DATASETS AND EVALUATION SCOPE

  • M+ uses 119 episodes from eight tasks, but its broader task coverage and different sampler mean it is not a matched-data control.Figure 6 pools evaluation over four shared training tasks and test task 60 rather than reporting task 60 alone.Source: Table 8 · p. 22, §D.1 · p. 22

D.2 STORED TOKENS, PROCESSED TOKENS, AND INTERPOLATION

  • The 8.3- and 6.3-point gaps for M and M+ use a linearly interpolated S/S+/S++ reference, not single-agent models trained at those exact budgets.The M+ versus S++ comparison uses scores against the base model, not a direct head-to-head comparison.Source: §D.2 · p. 23

D.3 TEACHER-OUTPUT QUALITY AND GENERATED TRACE LENGTH

  • 141,844 versus 59,355 generated tokens per eligible episode (2.39×) means multi-agent execution averaged about 2.4 times the generated output length.Generated trace length sums completion tokens across the orchestrator and all workers; it differs from stored or processed supervised tokens.Source: §D.3 · p. 23
  • 85.1% pooled win rate (79.3% with equal task weighting) favors multi-agent outputs in saved comparisons, but does not isolate delegation or equalize generation budgets.The comparable-length result uses only ten pairs; the comparisons also retain differences in workflow mixtures, verification instructions, and successful-completion selection.Source: Table 9 · p. 23, §D.3 · p. 24

E.1 MODEL ROLES AND SEARCH HISTORY

  • The six fixed-weight workflow-search ablations keep executor L(0) and vary evaluator/optimizer roles or the workflow-comparison history.The history variants include retained workflows, successive-workflow differences, and a fixed bare-workspace reference.Source: §E.1 · p. 24

E.2 COVERAGE, WIN RATES, AND EDIT SIZE

  • Coverage tracks whether a task has ever achieved a unanimous win, while the win-rate curve evaluates the current iterate across pooled judge–seed assessments.Error bars describe variation among six judge–seed rates over the same executions, not independent search runs.Source: §E.2 · p. 24
  • Edit size has a weak association with judged-quality changes, with large reversals in every populated bin; each proposed workflow is executed once.The analysis therefore combines prompt changes with execution variability.Source: §E.2 · p. 24

E.3 EXECUTION VARIABILITY UNDER AN UNCHANGED WORKFLOW

Identical workflow and prompt text produced sharply different judgments on two observed executions, illustrating that execution variability contributes to aggregate search trajectories.

E.3 EXECUTION VARIABILITY UNDER AN UNCHANGED WORKFLOW

  • Six wins versus six losses arose from identical workflow and rendered-prompt texts on task 53, despite the same recorded generation seed.The later execution halted under loop detection; the two observations do not estimate workflow reliability or identify the source of runtime variation.Source: §E.3 · pp. 24–25, Table 11 · p. 25
  • The later run repeatedly polled the same background-shell log without changing tool arguments, illustrating why edit size alone cannot explain every regression.The episode is one observed execution, not evidence that the workflow is generally unreliable.Source: §E.3 · pp. 24–25

F.1 WORKFLOWS AND REPRESENTATIONS

The appendix examines retained workflow components and their task information and redundancy, then audits what coordination behaviors appear in first-cycle student traces and training exposure.

F.1 WORKFLOWS AND REPRESENTATIONS

  • The retained-path analysis tracks role, instruction, contract, and a shared ordered hop at iterations 3, 6, 9, and 11; initial workflows lacking contracts and hops are excluded.The retained path can stay at one workflow while later proposals are executed.Source: §F.1 · p. 25, Figure 8 · p. 25

F.2 ESTIMATORS AND PERMUTATION ADJUSTMENT

  • Task information is an embedding-based proxy relating workflow components to task descriptions, while task-centered total correlation is not an exact conditional mutual-information estimate.Permutation adjustment estimates excess over a shuffled baseline and does not guarantee an unbiased information estimate.Source: §F.2 · p. 26

F.3 VALUES AND INTERPRETATION

  • From iteration 3 to 11, roles and instructions lose task information while contracts and hops gain it; unconditional redundancy falls in all four configurations.Task-centered redundancy falls in three configurations, with permutation-adjusted MPNet as the exception.Source: §F.3 · p. 26, Table 12 · p. 27, Table 13 · p. 27
  • The workflow statistics come from 43–51 agents across 8–9 tasks and describe retained-path text, not execution messages or a causal learning mechanism.Agents sharing workflow task descriptions and routing are not independent task replications, and the redundancy decrease does not hold for every proposed iterate.Source: §F.3 · p. 26

G.1 SCOPE AND DEFINITIONS

  • The audit covers first-cycle M+ bare-prompt rollouts against base references and counts successful explicit tool operations as delegation, handoff, worker revision, or main edit.Reads through imports or shell commands and writes inside shell scripts are not counted.Source: §G.1 · pp. 26–27

G.2 BASE CAPABILITY AND SUPERVISED EXPOSURE

  • The audit finds retained orchestrator assignments in all 114 teachers, retained code handoffs in 98, and retained worker revisions in 57.Sampled counts are 112, 79, and 37 episodes respectively; conservative matching and incomplete trace coverage make these lower bounds on observed exposure.Source: §G.2 · p. 28
  • The observations support increased bare-prompt use of coordination already available in the base, but do not identify which training examples caused student actions.They also do not isolate optimized workflow text from the directive, data selection, or sampler.Source: §G.2 · pp. 27–28

G.3 EXAMPLES AND LIMITS

The examples show artifact-dependent implementation, revision after evaluation, and a case where coordination improves workspace judgments while changing the scientific meaning of success.

G.3 EXAMPLES AND LIMITS

  • 2,934.9 versus 571.4 average expanded nodes favored Manhattan distance over the learned A* heuristic, and the student retained this unfavorable result after repair and review.The student corrected a normalization-interface mismatch and requested an independent critic before reporting the result.Source: §G.3 · pp. 28–29
  • Task 305’s student assembled a worker-dependent pipeline and won all six comparisons with its base reference, though the trace does not certify every simulation choice or metric.Figure 9 depicts artifact dependencies, with arrows requiring a producer write and return followed by a consumer read.Source: §G.3 · p. 28, Figure 9 · p. 29
  • A simulator revision produced success on all three task types by changing force-driven motion to direct goal repositioning and removing simulation steps, altering the meaning of success.The episode still received five favorable judgments out of six, showing workspace preference does not guarantee preservation of the scientific task.Source: §G.3 · p. 29

H REPRESENTATIVE TASK SPECIFICATIONS

  • The representative bare prompts specify a computational study, required outputs, and shared execution and logging instructions; a bare-prompt episode still receives the full task specification.The examples cover finance, pharmacy, and robotics tasks.Source: H REPRESENTATIVE TASK SPECIFICATIONS · p. 29

H.1 FINANCE: TASK 51

  • Task 51 requests a cross-asset macro nowcasting and trading study with time-based splits, an event-time embargo, transaction costs, and robustness checks.Its deliverables include a conference-style report, plots, reproducible Python code, metrics, model comparisons, and an ablation log.Source: §H.1 · pp. 29–30

H.2 PHARMACY: TASK 202

The appendix specifies a geometric GNN task for protein-interface prediction and a PyBullet task for generating and validating manipulation-planning datasets. It also gives saved workflow examples with explicit agent handoffs and validation.

H.2 PHARMACY: TASK 202

  • Task 202 prototypes a geometric GNN for protein-interface prediction using public PDB-derived structures, with generalization across families as a target.The prompt calls for AUROC/AUPRC reporting and ablations comparing distances, distances plus angles, and a full equivariant model.Source: §H.2 · pp. 30–31

H.3 ROBOTICS: TASK 307

  • Task 307 generates synthetic manipulation-planning instructions and action sequences in PyBullet, then measures dataset diversity and plan executability.The requested deliverables include scene and instruction statistics and executable-plan rates by scene complexity.Source: §H.3 · p. 31

I COMPLETE WORKFLOW EXAMPLES

  • The saved workflow examples connect task-specific specialist agents through explicit handoffs and validation, but their presence does not establish that every run performs checks correctly.Task 51 links data, features, modeling, backtesting, auditing, and reporting; Task 202 links structure acquisition, graph construction, training, and evaluation, with a validator returning failed-agent and violation details.Source: I COMPLETE WORKFLOW EXAMPLES · p. 32

I.1 COMPLETE CROWN WORKFLOW: TASK 51, VERSION 10

The saved designs specify multi-agent workflows for finance and protein-interface modeling, using specialist agents, structured outputs, validation gates, and recall loops. Both designs direct the orchestrator to address failures before reporting results.

I.1 COMPLETE CROWN WORKFLOW: TASK 51, VERSION 10

  • Task 51’s seven-hop workflow runs data alignment, feature auditing, modeling, backtesting, auditing, reconciliation, and reporting, with explicit recall routing when checks fail.The orchestrator tracks pipeline state and retries failed work up to the specified maximum.Source: §I.1 · pp. 32–33
  • The data pipeline requires non-overlapping temporal splits, alignment diagnostics, and sufficient pre- and post-2020 coverage before downstream analysis.It also sets thresholds for alignment gaps and calendar coverage and requests self-correction or recall when those checks fail.Source: §I.1 · p. 33
  • Task 51’s feature and modeling agents enforce leakage controls, embargo-aware validation, regime gating, and distinct IC and RankIC calculations.The workflow defines output contracts and validation reports for the feature and modeling stages.Source: §I.1 · pp. 33–34
  • Its backtest, auditor, and report stages check costs, subperiod coverage, ablation consistency, metric alignment, and required plots before synthesis.The auditor can issue targeted recalls, while the report stage must match reported metrics to the results data.Source: §I.1 · pp. 34–36

I.2 COMPLETE CROWN WORKFLOW: TASK 202, VERSION 8

  • Task 202’s orchestrator divides work into data acquisition, graph construction, parallel model training, validation, and programmatic reporting, recalling agents for specific failures.Its instructions include stall prevention and require validation and cross-agent consistency checks before termination.Source: §I.2 · pp. 36, 38
  • Task 202 assigns specialists to real PDB data and labeling, family-separated graphs, three geometric model variants, training, and validation.The training instructions require family-level metrics and pinned dependencies; the validator checks data integrity, equivariance, and report-metric alignment.Source: §I.2 · pp. 37–38

J COMPARISON WITH RELATED METHODS

The comparison distinguishes MASS from prior search-and-distillation and inference-only approaches by its use of optimized multi-agent workflows and shared-weight training. The reported comparison concerns mechanisms, not matched benchmark performance, and the recursive loop has no guarantee of improvement in every future cycle.

J COMPARISON WITH RELATED METHODS

  • Earlier methods combine search, policy learning, rationale generation, reward filtering, or interaction training; MASS applies this pattern to optimized multi-agent research workflows with shared weights for orchestrator and workers.The cited examples include Expert Iteration, STaR, ReSTEM, and SiriuS.Source: J COMPARISON WITH RELATED METHODS · p. 38
  • Unlike inference-only critique or communicating-agent approaches, MASS tests whether execution traces improve shared model weights and subsequent workflow search.The table compares reported mechanisms rather than matched performance benchmarks, and sharing one model across roles alone does not establish weight learning.Source: J COMPARISON WITH RELATED METHODS · p. 38
  • The recursive loop is empirical and guarantees no improvement in every future cycle; workflow feedback and teacher ranking use the current model.Checkpoint selection uses validation loss, while external judgments assess quality after execution.Source: J COMPARISON WITH RELATED METHODS · p. 38
Loading 2610.12176v1…