acceptodds
Under review as a conference paper at ICLR 2027

Evo2Team: When Do Evolved Skills Transfer? From Selection to Deployment

Abstract

A skill bank that helps one multi-agent system may leave another's behavior unchanged. We study routing and communication skill transfer across Count-Frequency and AgentsNet teams of 4–32 agents using GPT and Qwen model ladders. Source evolution meets a joint quality, cost, and confirmation goal in 14 of 16 settings. We then evaluate Evo2Team, which selects, adapts, and confirms evolved skills on the target, alongside six frozen selectors in 28 directions. Evo2Team's exploration costs less than new target evolution in every direction, even when reused reference runs are charged. Yet different selected banks produce identical AgentsNet execution in two directions. When Evo2Team changes execution, its gains can be broad: one Count-Frequency direction improves 28 of 32 tasks over KNN. Seven positive AgentsNet held-out outcomes save 6.1–14.6% in deployment cost while using transferred skills on only three to six of fifteen tasks. In five earlier accepted directions, 22 transferred task records pass three fixed-graph confirmations, but four fail in recorded new-graph executions, where graphs and model responses both change. These results locate transfer benefit in executed decisions, task coverage, and final quality and cost.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Agents, skills, transfer learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1qCEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether evolved routing and communication skills transfer to multi-agent teams with different sizes or model compositions. Evo2Team combines source skill evolution with target skill selection, configuration search, task-level composition, and repeated confirmation. Experiments on Count-Frequency and AgentsNet examine whether transfer changes execution, improves quality and cost, and remains effective on new instances. The results show that benefits depend on the transfer direction, sometimes cover only a small subset of tasks, and may not persist on new instances despite repeated confirmation.

Critical strengths

S1. The paper addresses a practical question: whether existing skills remain useful when the team executing them changes.

S2. The authors distinguish selected policies from diagnostic outcomes, report negative transfer and fallback usage, and examine both adaptation and deployment costs.

Critical weaknesses

C1. The distinct contribution relative to existing work remains unclear. SkillLens studies how skill utility varies across consuming models and documents negative transfer. SkillOrchestra uses explicit skill and cost modeling for routing, while SkillOpt studies skill evolution through feedback and validation. This paper extends the setting to changes in team size and model composition, but the resulting methodological advance or new empirical insight is not sufficiently established.

C2. The sources of the observed gains remain difficult to isolate. Evo2Team combines source skills with target configuration search, task-level composition, and fallback. The feedback-guided KNN/CORAL controls are useful, but their available operations differ from those of Evo2Team. It therefore remains unclear how much improvement comes from source knowledge versus target optimization.

C3. The evidence for reliability is limited. The new-instance experiment shows that repeated confirmation on fixed instances does not guarantee subsequent success, but graph instances and model responses change simultaneously, preventing attribution of the failures. The small number of independent instances and limited repeated evaluations also make it difficult to determine whether the observed differences reflect stable transfer patterns.

Questions for authors

q1. What is the central new finding about multi-agent skill transfer beyond existing work on skill consumption, routing, and evolution? Which experimental results most directly support this contribution?

Q2. How much of the improvement can be attributed to transferred source knowledge? Can the existing controls or further analysis distinguish these contributions more clearly?

Q3. How stable are the main findings across independent instances and repeated runs? Can the available evidence distinguish sensitivity to instance changes from variability in model responses?

2 comments
AuthorsOpenReview ↗

C1 / q1: contribution

We thank the reviewer for the precise summary and for recognizing that we report negative transfer, fallback usage, and both adaptation and deployment cost.

C1 / q1. Contribution relative to SkillLens, SkillOrchestra and SkillOpt. The main merit of our paper is not a new optimizer. It is a study of what happens to executable multi-agent rules when the team itself changes, in size or in model ladder. We will discuss each cited work in Related Work:

Skill formWho consumes the skillKey finding relevant to us
SkillLensModel-generated textual skillsA single agent modelUtility depends on the extractor–consumer pairing; text-based screening is unreliable
SkillOrchestraLearned skills with agent competence and cost modelsAn orchestrator routing within a fixed agent poolSkill-level competence and cost modeling for routing
SkillOptSkill documents edited from trajectoriesA frozen agentEdits are accepted only through a held-out validation gate
OursExecutable routing and communication rulesA planner and a multi-agent team whose size or ladder changesSee below

None of these works asks what happens when the team executing the rules changes. That raises four questions: Does a selected rule change what the new team executes? How many tasks does it reach? How much of the source gain survives? Can we tell in advance?

Central findings, and the results that most directly support them:

FindingSupporting results
1. Evolved skills carry a large share of the gain to a new teamOn the original ladders, the source policy alone (no target search) accounts for 46% of the total CF gain over all-large. It provides at least half the gain in 9 of the 13 directions with a positive total, and is significant on its own in 7/14. In 5 directions, the final policy uses the source policy unchanged on at least one task group.
2. Transfer is predictable before deployment, from execution rather than textOne evaluation of the source policy on the target adaptation split: 24/25 positive vs. 13/25 (p=0.001). When it passes, deploying with no search is non-inferior in 22/25. Transfer type does not predict success (p=0.50). This complements SkillLens: text-based screening is unreliable, but a cheap execution-based check is highly informative.
3. Selection is not executionKNN and CORAL select different banks in two AgentsNet directions, yet the recorded executions are identical. This is why we follow rules through execution.
4. AgentsNet new-instance failures are explained by response variabilityRerunning the same policy on the same graph flips about 20% of outcomes (Part 2, Q3). This gives a measured variability floor against which multi-agent transfer results should be read.

We will rewrite the contribution statement around these four findings and move their supporting results into the main text.

AuthorsOpenReview ↗

C2 / Q2: attribution; C3 / Q3: reliability

C2 / Q2. How much of the gain comes from source knowledge? Within Evo2Team, we split each test gain into a source part (the source policy deployed as is vs. all-large) and an adaptation part (Evo2Team vs. that source policy). Each part has a paired bootstrap 95% CI.

CFSource share of the summed gainSource significant aloneAdaptation significant
Original ladders (14 directions)46%7/148/14
New ladder (8 arms)*17%2/86/8

*Three original tier models have been retired since submission, so the new runs transfer the paper's banks to a new GPT ladder (gpt-5-nano / gpt-4.1-mini / gpt-4.1).

Example (CF, RMSE gain over all-large)TotalFrom sourceFrom adaptationRole
Q16→Q32+6.11+5.88+0.23Source dominates
G16→Q16+1.74+1.62+0.12Source dominates
G4→G8+0.07−0.30+0.37Adaptation repairs
4→8 agents, seeds 101 / 202 / 303+0.23 / −0.01 / +0.23−1.13 / −1.03 / −0.05+1.36 / +1.02 / +0.28Adaptation repairs

On the new ladder, the source-derived seeds are acceptable on the adaptation split from the start in 7/8 arms. In CF, the final recipes keep almost all deployed source rules, and adaptation mainly recalibrates thresholds and integer parameters.

Matched-operation control. The reviewer rightly notes that the KNN/CORAL controls differ in their available operations. We therefore added full-nosource: the same edit operators, screening, composition, confirmation and gate, with the source bank replaced by a neutral generic vocabulary.

Evo2Teamfull-nosource
CF: acceptable at initialization7/85/8
CF: ahead after search (adaptation split)4/4 cells, ≈0.06 RMSE—
AgentsNet: acceptable at initialization1/40/4
CF: held-out difference+0.019 RMSE—
AgentsNet cost savingFrom sourceFrom adaptation
Original ladders3.7%2.8%
New ladder4.0%14.7%

Summary. Source knowledge delivers roughly half the gain on the original ladders and a reliable start on the new ladder. Target adaptation delivers the rest and repairs failed transfers. The pre-deployment check (Part 1) measures which case applies.

C3 / Q3. Instance changes vs. response variability. We re-executed every tested frozen policy, and all-large, on the identical test instances. Only the model responses vary.

Count-FrequencyAgentsNet
Effect of response variability21% of task-level variance (median; range 11–50%)12/57 transferred task runs flip (≈21%)
All-large itself—Flips 2/15 and 3/15 tasks
Outcome labels unchanged20/22 arms (both flips at the gate boundary)7/10 arms

The paper reports that 4 of 22 confirmed task records (18%) fail on new graphs, which matches the same-instance flip rate. Response variability alone therefore suffices to explain these failures. Together with R1-Q2, the CF findings are stable, and single-execution AgentsNet results carry a variability floor of about 20%.

In the revision we will:

  1. Reword the new-instance claim.
  2. Report AgentsNet success as a rate over repeated executions, with CIs.
  3. Add this decomposition as a recommendation for evaluating multi-agent transfer.
Reviewer_6u5UOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether rules learned from a source multi-agent system (MAS) remain effective when transferred to target systems with different model families or team sizes. It proposes Evo2Team, a procedure for mapping source rules to target systems and validating their effectiveness.

Critical strengths

The research question of identifying the conditions under which rules can transfer across different model families or different systems in MAS is meaningful.

Critical weaknesses
  1. The paper lacks details about skill evolution, which is a central component of the work.

Important details, such as how rules are generated, how the initial rules are determined, and how skills are evolved, are not sufficiently described in the paper and can only be found by examining the accompanying code.

  1. The term "Evolution" does not accurately reflect the actual implementation.

Based on my examination of the source code, the evolution procedure appears to be limited to finding rules for assigning model sizes.

This differs in nature from methods such as Reflexion, GEPA, and Promptbreeder, which are discussed in the Related Work section.

  1. The conditions for successful transfer are unclear.

The paper does not identify systematic criteria that explain when transfer succeeds or fails.

  1. The figures and writing lack polish.

In Figure 1, the icons do not intuitively correspond to the content they represent, and some text overlaps, making the figure appear unfinished.

The paper frequently uses terms such as "ladder" before defining them, including in the Abstract, making the text difficult to understand.

Questions for authors
  1. The title asks when evolved skills transfer, but the paper does not provide a clear answer regarding the conditions under which transfer occurs. Under what conditions can evolved skills be considered transferable, based on the findings of this paper?
3 comments
AuthorsOpenReview ↗

W1: evolution details; W2: terminology

We thank the reviewer for finding the research question meaningful and for examining our code closely. Most concerns are about presentation and can be fixed directly. We provide the missing content below.

W1. Missing details on skill evolution. We agree that this belongs in the main text, and we will add the following algorithm box:

Source skill evolution (per task family, team size n, model ladder M)
Input : exploration tasks (CF 16 / AgentsNet 15), validation tasks, R = 5 rounds
Init  : empty routing bank B, empty communication bank C,
        planner assigns every node the medium tier, quality-price weight λ
for round = 1..R:
  for each exploration task x:
    1. execute x with the all-medium planner                -> trace
    2. choose probe nodes by an acquisition score
       (rule uncertainty, weighted by λ toward quality or cost)
    3. for each probe node v: re-execute v with small, medium and large
       under the identical upstream context
    4. record evidence or a counterexample for candidate rules keyed by
       (node kind, difficulty bin, dominant-feature bin):
         "at most small/medium" : supported if the cheaper tier succeeds
         "at least large"       : supported if large succeeds and medium fails
    5. communication: propose candidates (grouping, tree reduction, leader,
       message budget, retry/escalate) from task loss and cost; explore one
       action on x; record the outcome as evidence or counterexample
  update rule confidence from evidence; a rule becomes locally validated
    above a confidence threshold and is rejected after >= 3 outcomes below it
  update λ from the observed rate of quality loss after downgrading vs.
    the observed price saving
Policy selection on validation tasks:
  compose validated rules into candidate policies, keep the quality-cost
  Pareto frontier, require paired one-sided non-inferiority to all-large
  plus repeated confirmation, then freeze the deployed policy and bank B ∪ C
Reviewer's questionAnswer
How are the initial rules determined?There are none. The bank starts empty, and every node starts at the medium tier.
How are rules generated?From controlled probes of the same node with all three tiers under the same context. A rule is proposed only when one tier succeeds where another fails, or when a cheaper tier succeeds.
How are skills evolved?Candidate rules accumulate evidence and counterexamples across rounds and are validated or rejected. The deployed policy is chosen under a non-inferiority test.
What is the outcome?Source pools of 20–153 skills. The selected policy meets the joint source goal in 14 of 16 settings.

W2. "Evolution" does not reflect the implementation. The reviewer is right that our procedure differs in nature from Reflexion, GEPA and Promptbreeder:

Reflexion / GEPA / PromptbreederOur procedure
What changesPrompt textA population of executable rules (routing and communication)
VariationLLM rewrites or mutates promptsNew candidate rules induced from tier probes and communication trials
SelectionTask score of the promptEvidence and counterexamples per rule; statistical non-inferiority test for the deployed policy

The bank also holds communication rules, not only tier assignments. For example, the CF 4-agent GPT bank has 15 routing and 7 communication rules. Still, our wording invited the comparison, and we will revise it:

CurrentRevised
"skill evolution""evidence-driven skill induction", as defined in the box above; the title will be updated where the venue permits
Implicit relation to prompt evolutionAn explicit contrast in Related Work: no prompt rewriting, and no genetic algorithm over prompts
"Evo2Team"Kept only as the name of the end-to-end procedure, defined at first use
AuthorsOpenReview ↗

W3 / Question: transfer conditions; W4: presentation

W3 / Question. When can evolved skills be considered transferable? We thank the reviewer for this question; it led to what we now consider the paper's clearest result. We tested signals that are observable before the held-out test across 50 Evo2Team transfers: the 28 original directions plus 22 new runs. The new runs use a new GPT ladder (gpt-5-nano / gpt-4.1-mini / gpt-4.1), because three of the original tier models have been retired since submission.

Signal (observable before the test)Positive transfer when truePositive transfer when falseFisher p
Source policy non-inferior to all-large on the target adaptation split24/2513/250.001
An initial configuration is already acceptable28/329/180.007
Repeated confirmation passed33/414/90.040
Transfer type (team-size change vs. ladder change)12/1825/320.50
Further check on the first signalResult
Replicates across all methods on the new ladder (incl. prompt optimizers and low-budget arms)36/37 vs. 15/25 (p=0.0002)
When it holds: deploy the source policy with no searchNon-inferior on the held-out test in 22/25 cases

Answer to the title question:

ConditionOutcomeWhat to do
The source policy is already non-inferior on a small sample of target tasksTransfer succeeds almost always (24/25)Deploy, or adapt for further gains
It is notTransfer depends on target-side repair and succeeds about half the time (13/25)Adapt, or optimize directly on the target
Team size vs. model ladder changeDoes not predict success (p=0.50)Run the check instead of relying on the type of change

What matters is whether the target team still acts on the rules as the source team did, and the check measures exactly that. Our pipeline repeats illustrate the second row. In two of three seeds, a source policy evolved on 4 agents is 1.0–1.1 RMSE worse than all-large on 8 agents, and adaptation recovers it. We will make this the paper's central answer and state it in the abstract and introduction.

W4. Figure 1 and terminology. We agree on both points. We will redraw Figure 1 with one consistent icon set, separate rows for the source, mapping, target and test stages, and no overlapping text. We will also define every term at first use, including in the Abstract, and add a notation table:

TermDefinition
Model ladderSmall, medium and large models from one provider, with increasing price and capability
TierOne of the three positions in a ladder
Routing / communication ruleA rule that sets a node's tier / how agents coordinate, message, and recover
RecipeSelected skills plus parameter overrides
Transfer directionA pair of source and target settings
L / FThe all-large policy / the source policy deployed on the target
Positive transferThe label criterion

Thank you for the detailed response. However, the manuscript remains unchanged, and I am not convinced that the proposed criterion explains when or why skills transfer, rather than simply validating target performance. Therefore, I will maintain my original rating.

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

Evo2Team studies the transfer of routing and communication rules across multi-agent teams with different sizes and model configurations. It selects and adapts source skills using target feedback, then evaluates quality, cost, and execution behavior on Count-Frequency and AgentsNet.

Critical strengths

The paper reports negative transfer outcomes and separates exploration cost from deployment cost. Execution-level analyses help clarify where the transferred rules actually affect behavior.

Critical weaknesses
  1. The added value of the evolved skill bank is unclear. The paper includes KNN/CORAL feedback controls with matched evaluation ceilings, but these still search subsets of the source skill bank. It does not report a prompt-based baseline that directly generates and refines routing and communication configurations without that bank under the same target-feedback budget. Such a comparison is needed to establish the benefit of transferred skills beyond target-side configuration optimization.

  2. Insufficient evidence of robustness. Each experimental cell uses a skill pool from one source run, and Tables 13–14 report one test evaluation per ablation variant. Although candidates undergo three confirmation executions, independent repetitions of the full evolution-and-adaptation pipeline are not reported. With only 15 AgentsNet test tasks per direction and no uncertainty estimates for the main comparisons, the stability of the reported gains remains unclear.

Questions for authors
  1. How does Evo2Team compare with directly generating and refining routing and communication configurations through prompting under the same target-feedback budget?

  2. Do the main results hold across independent source-evolution and target-adaptation runs?

3 comments
AuthorsOpenReview ↗

Q1: prompt-based baseline

We thank the reviewer for recognizing that we report negative transfer and separate exploration from deployment cost. We agree that both requested analyses strengthen the paper, and we ran them.

Setup. The new arm llm-nobank replaces only Evo2Team's proposal step. Everything downstream is identical.

ComponentEvo2Teamllm-nobank
Who proposes configurationsEdit operators on source skillsOPRO-style optimizer (gpt-4.1, temperature 0.7)
What the proposer seesSource bank and aggregate rankingsTask description, tier prices, planner rule semantics, the gate, and past configurations with their results; no source bank
Initialization2 configurations from the source bank2 configurations written by the LLM
Search2 generations: 6 screened on 4 (CF) or 5 (AgentsNet) tasks, top 2 promotedSame
Confirmation, gate, held-out test3 repeats, same gate, same testSame

Both arms share the same feedback ceilings, enforced by one meter. The optimizer's own calls are charged separately and are not target feedback.

Task evaluationsAdaptationConfirmationTotal
CF8048128
AgentsNet13090220

All new runs use a new GPT ladder (gpt-5-nano / gpt-4.1-mini / gpt-4.1), because three of the original tier models have been retired since submission. We ran four transfer cells per domain with two independent replicates each, plus three controls:

ArmProposerSees the source bankFeedback
llm-bankLLMYes; the source policy is its first configurationTask-level
llm-blindLLMNoAggregate only, as in Evo2Team
full-nosourceEvo2Team's edit searchNo; a neutral generic vocabulary insteadAggregate

Search trajectory (adaptation split, compared at matched points of the shared feedback schedule).

Count-Frequency: a robust starting point.

Acceptable at initializationWorst initial gap to all-large (RMSE)
Evo2Team7/8−0.000
llm-nobank6/8−2.0
full-nosource5/8—

The LLM optimizer catches up within one generation, and every method ends the search above all-large. Without the source bank, Evo2Team's own search stays behind: Evo2Team leads in 4/4 cells after search, by about 0.06 RMSE.

AgentsNet: faster progress on the harder domain (arms with an acceptable configuration).

StageEvo2Teamllm-nobankllm-bank
Initialization4/82/82/8
After generation 15/84/83/8
End of search6/85/84/8
Final S gain over all-large+0.125+0.033+0.008

full-nosource has no acceptable configuration at initialization (0/4 arms).

Held-out test. Final test quality is statistically indistinguishable across methods in both domains. On AgentsNet, re-executing an identical policy flips about 20% of task outcomes, which exceeds the between-method differences (R3-Q3). Full tables will be in the revised appendix.

We therefore position the benefit of transfer as a reliable start and faster search, not higher final quality. Part 2 describes one further practical advantage of the bank.

AuthorsOpenReview ↗

Q1 continued; Q2: robustness

Deployment with no search, verified in advance. The bank supplies a complete candidate policy: the mapped source policy. One evaluation on the target adaptation split checks whether it is non-inferior to all-large.

When the check passes (25 of 50 transfers)Result
Source policy deployed directly with no search, non-inferior on the held-out test22/25
CF: strictly better than all-large13/18
CF: cheaper than all-large17/18 (mean saving 17%)
Positive transfer after adaptation24/25 (vs. 13/25 when the check fails; p=0.001)

A prompt optimizer has no such candidate: it must spend its search budget before it has a policy to deploy. With the check, practitioners can decide in advance whether to deploy, adapt, or fall back to direct target-side optimization. Adaptation can still help: in CF, the adapted policy beats the directly deployed one in 17/18 cases.

We originally focused on how a fixed evolved bank is consumed by a new team. We agree that the prompt baseline is needed to show what the bank adds, and we will add it to the main text.

Q2. Do the results hold across independent runs?

(a) Independent full pipelines. We re-evolved the source bank from scratch on 4 agents with three seeds, then transferred each bank to 8 agents on the new ladder.

SeedCF: source policy as isCF: all-largeCF: Evo2TeamCF: adaptation gain [95% CI]CF labelAgentsNet label (S vs. all-large, saving)
1012.8651.7381.508+1.36 [1.14, 1.58]++ (equal, 18%)
2022.6831.6581.668+1.02 [0.79, 1.23]tie− (one task fewer, 37%)
3031.7281.6751.445+0.28 [0.12, 0.45]++ (equal, 22%)

CF values are RMSE (lower is better). Adaptation is significant in all three runs. For two seeds, the source policy evolved on 4 agents fails on 8 agents. This is the paper's motivating observation.

(b) Independent adaptation replicates (fresh search seeds and fresh API calls; all-large in parentheses):

CellCF RMSE r1 / r2CF labelsAgentsNet S r1 / r2AgentsNet labels
G4→T40.895 / 0.809 (0.972)+ / +0.533 / 0.600 (0.600)− / +
Q4→T40.886 / 0.980 (0.960)+ / +0.667 / 0.600 (0.600)+ / +
G8→T81.304 / 1.468 (1.669)+ / +0.267 / 0.267 (0.267)+ / +
Q8→T81.523 / 1.317 (1.695)+ / +0.333 / 0.333 (0.400)− / −

Labels agree across replicates in 4/4 CF cells and 3/4 AgentsNet cells. The AgentsNet positives have equal or better quality at 9–24% lower cost.

(c) Uncertainty for the original 28 directions (paired bootstrap 95% CIs, Evo2Team vs. all-large):

DomainSignificantly betterSignificantly worseNot significant
CF (14 directions)101 (Q32→G32)3
AgentsNet (14 directions)0014

AgentsNet gains are therefore cost savings at equivalent quality. We will add these CIs to the main tables and qualify the AgentsNet quality claims accordingly.

Thank you for the authors' response. I will maintain my initial score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To all reviewers and the AC

We thank the reviewers for their careful reading. The reviews converge on four questions, and we ran new experiments for each:

#QuestionRaised byNew evidence
1Does the skill bank help beyond optimizing the target directly?R1, R3A budget-matched prompt-based optimizer, plus three controls
2Are the results robust across independent runs?R1, R3Independent replicates, full-pipeline repeats, bootstrap CIs
3When does transfer succeed?R2, R3A pre-deployment criterion, tested on 50 transfers
4New-instance failures: instance change or response variability?R3Re-execution of identical policies on identical instances

New ladder. Since submission, three of the original tier models have been retired by their providers, so we could not rerun the original ladders. All new experiments use a new GPT ladder:

TierModel
Smallgpt-5-nano (minimal reasoning)
Mediumgpt-4.1-mini
Largegpt-4.1

Single-model pilots confirm that quality and price increase monotonically across these tiers. The new transfer cells are G4→T4, Q4→T4, G8→T8 and Q8→T8 in both domains, where T is the new ladder. Their source banks are the paper's original GPT (G) and Qwen (Q) banks. Analyses of the original 28 directions use the recorded runs from the paper.

Main findings.

#FindingKey evidence
1Transfer success is predictable before deploymentSource policy non-inferior on the target adaptation split: 24/25 positive vs. 13/25 otherwise (p=0.001)
2When the check passes, the source policy can be deployed with no searchNon-inferior to all-large on the held-out test in 22/25 cases
3The source policy alone carries a large share of the gain46% of the total CF gain on the original ladders
4Transfer gives a more reliable start and faster search than a prompt optimizer under the same budgetDetailed in our response to R1; final held-out quality is statistically indistinguishable
5AgentsNet new-instance failures are explained by response variabilityAbout 20% of outcomes flip when the same policy reruns on the same graph

We will revise the contribution statement around these findings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.