cs.CLOct 14, 2026
SaveEvolve Together, Fall into Place: How Social Scaffolding Shapes Self-Evolving Multi-Agent Systems
Organizations: Chung-Ang University
Abstract
Recent work lets LLM agents evolve their own harness, up to the team itself. Yet in human groups, the same abilities achieve different outcomes under different social conditions. Borrowing the notion of scaffolding from developmental psychology, we investigate whether the social scaffolding of a self-evolving multi-agent system (MAS), namely its team composition and structure of interdependence (e.g., collaboration, conformity, or competition), shapes how the team evolves. We present AGORA (Agents Growing through Organization and Relational Arrangements), a framework that varies the two axes independently across three evolution mechanisms. Across web search, codebase understanding, and document QA, we find that (1) evolution alone rarely helps, while the right scaffolding surpasses a strong agent evolved alone by up to 25 points; (2) it does so by moving authority to the most capable member while the others step back; and (3) this member, having grown within the team, outperforms the same model alone. Designing a self-evolving MAS thus requires designing its social scaffolding, which shapes not only how its agents learn, but who gets to lead.
Figures & tables
Figure 1: How different social scaffolding leads to different evolutionary outcomes in human groups.
Figure 2: Overview of AGORA. After completing a task as a team ( Team Session ), each agent updates itself ( L1 ) and writes memos about its peers, guided by social scaffolding ( L2 ). Based on these memos, team-level structural changes are proposed ( L3 ) and then retained or rejected by a held-out gate ( L3 Gate ). Throughout this process, we vary the social scaffolding and the evolution mechanism to examine how a self-evolving MAS evolves.
| Level | What changes | Who changes it | Takes effect | Gate |
| L1 (individual) | Skill document | Mechanism (SkillOpt) | Next iteration | L1 gate |
| Memory bank | Mechanism (MemRL) | Next iteration | None | |
| Prompt patch | Agent itself (Meta-Team) | Next task | None | |
| L2 (relational) | L2 memo | Each agent, about teammates it worked with | Next iteration | None |
| L3 (team) | Roles, agent prompts and configurations, constitution | Proposed by Tier 1 agents | Next iteration, unless reverted by the L3 gate | L3 gate |
Table 1: Levels of the AGORA operation. L1 and L3 evolve the agents and the team, while L2 records relations between agents only when the social scaffolding includes interdependence. Social scaffolding acts directly only on L2 and L3, and reaches L1 through the trajectories and L2 memos that every mechanism learns from.
| Meta-Team | SkillOpt | MemRL | |
| What evolves | Prompt patches and skills written by the agent | Skill document | Memory bank with Q-values |
| Injected into execution | Own patches and skill list | Current skill document | Up to 5 retrieved memories (success/failure scripts) |
| Evolution timing | Right after each task | End of iteration | End of iteration |
| Evolution signal | Agent’s self-reflection | Reflection on trajectories with grading results | Task success or failure as reward |
| L1 gate | None | Held-out binomial test | None |
| Model performing the update | Agent’s own model (weak or strong) | The strong model (fixed), with L2 notes | The strong model (fixed), with L2 notes |
Table 2: Evolution mechanisms. Batch sizes are 17 for GAIA and 20 for LoCoBench and OfficeQA Pro. All mechanisms share the same L2 and L3 and the same L3 gate.
| Intervention | Aspect | None | Collaboration | Conformity | Competition |
| Permission tiers | Chairman | Fixed | Fixed | Fixed | Rotates to top scorer |
| Tier 1 roles | All roles | All roles | Chairman only | Chairman only | |
| L2 memo | Sharing | Disabled | Active | When stuck; deferential | Minimal; scores shown |
| Finalization | Disabled | After peer review | Immediately | After peer review | |
| L3 team config | Proposals | All roles | All roles | Chairman only | Chairman only |
| Peer review | No | Majority | No | Majority |
Table 3: Structures of interdependence. ‘Majority’ peer review passes once a majority of participating teammates have responded, and Tier 2 roles can only suggest L3 changes to the chairman. Details are given in Section 3.3.2 .
| GAIA | LoCoBench | OfficeQA Pro | |||||||||||
| Team | Structure | Pre | Meta-Team | SkillOpt | MemRL | Pre | Meta-Team | SkillOpt | MemRL | Pre | Meta-Team | SkillOpt | MemRL |
| solo_weak | None | 20.8 1.4 | 20.8 5.2 | 20.8 5.2 | 20.0 0.0 | 54.5 2.0 | 50.8 1.2 | 49.5 3.6 | 46.8 3.8 | 2.7 0.9 | 2.7 0.9 | 5.9 0.9 | 2.2 2.5 |
| solo_strong | None | 51.7 2.9 | 52.5 5.0 | 49.2 1.4 | 53.3 5.2 | 73.1 1.5 | 76.4 3.1 | 74.4 2.5 | 76.8 1.7 | 28.5 0.9 | 31.2 4.9 | 34.4 0.9 | 29.0 8.4 |
| onlyweak | None | 21.7 5.2 | 21.7 9.5 | 20.0 4.3 | 18.3 6.3 | 52.9 3.5 | 43.4 5.3 | 48.5 5.6 | 43.4 4.0 | 1.1 0.9 | 0.5 0.9 | 0.5 0.9 | 2.2 0.9 |
| Collab. | 25.0 4.3 | 19.2 6.3 | 24.2 1.4 | 59.3 4.1 | 54.9 6.1 | 46.8 5.1 | 0.0 0.0 | 0.5 0.9 | 1.1 0.9 | ||||
| Conform. | 25.0 2.5 | 20.8 1.4 | 17.5 2.5 | 54.2 2.3 | 45.8 6.7 | 39.1 5.9 | 1.1 0.9 | 0.0 0.0 | 0.5 0.9 | ||||
Table 4: Main results (avg@3 success rate, %) on the test set. Cell color shows the change from Pre : rose for gains and blue for drops ( Pre : same team before evolution), darker for larger changes. Bold : best structure of interdependence for each team, mechanism, and benchmark. Values are mean SD over three rollouts.
| GAIA weakstrong | LoCoBench weakstrong | LoCoBench onlystrong | |||||||||||||
| chair: planner (W) | chair: planner (S) | ||||||||||||||
| Run | Chair | Agents | Calls | TO | avg@3 | Run | Agents | Calls | TO | avg@3 | Run | Agents | Calls | TO | avg@3 |
| Solo strong, Pre | – | 1.00 | 22.6 | 0.0 | 51.7 | Solo strong, Pre | 1.00 | 24.3 | 12.8 | 73.1 | Team, Pre | 3.20 | 73.3 | 23.2 | 63.0 |
| Team, Pre | plan_agent (W) | 2.76 | 46.8 | 7.5 | 48.3 | Team, Pre | 4.18 | 80.0 | 5.4 | 80.5 | SO, none | 1.40 | 27.0 | 12.8 | 76.8 |
| SO, none | plan_agent (W) | 2.97 | 39.6 | 3.3 | 51.7 | SO, conform. | 4.62 | 107.3 | 5.1 | 79.1 | SO, collab. | 1.00 | 19.7 | 5.1 | 82.8 |
| SO, compet. | web_agent (S) | 1.38 | 19.4 | 0.0 | 74.2 | SO, compet. | 4.55 | 91.5 | 3.4 | 82.5 | SO, conform. | 1.00 | 18.3 | 4.4 | 87.2 |
Table 5: Process metrics per session in the final evaluation for the runs discussed in Section 6 (GAIA , LoCoBench sessions per run; all scores in Table 4 ). Chair : the role holding the chair and its model; on LoCoBench, the chair stays the planner in every run shown. Agents : roles started in the session; Calls : LLM calls; TO : share of sessions stopped by the 900-second limit (%). Bold: best avg@3 in each block. (SO: SkillOpt, MT: Meta-Team, MR: MemRL; W: weak, S: strong)
| GAIA | LoCoBench | |||
| Solo | Team | Solo | Team | |
| avg@3 | 49.2 | 74.2 | 74.4 | 87.2 |
| Skills accepted | 0/4 | 2/4 | 2/4 | 2/4 |
| Solved only here | 5 | 35 | 8 | 46 |
Table 6: Team-grown agent ( Team ) vs. the same model alone ( Solo ) under SkillOpt in GAIA weakstrong (competition) and LoCoBench onlystrong (conformity). Skills accepted: this agent’s skills that passed the L1 gate, out of its candidates that reached the gate. Solved only here: (task, rollout) pairs only this agent solved.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
| Benchmark | Train | Val | Test | Success criterion | Judge |
| GAIA | 17 | 20 | 40 | Exact match of the final answer | – |
| LoCoBench | 80 | 20 | 99 | LCBS 3.0 (out of 5) | – |
| OfficeQA Pro | 51 | 20 | 62 | Judged correct | gpt-5-mini |
Table 7: Benchmark splits and scoring. The validation set (20 tasks) is fixed first; the rest is split train : test = 4 : 5 (test 40, seed 42). Gates always use the validation set and the final evaluation always uses the test set.
| Benchmark | Role | Tools | Model in weakstrong |
| GAIA | plan_agent (chairman) | – | weak |
| web_agent | web_search, web_fetch, bash, read_file | strong | |
| file_agent | bash, read_file, write_file | weak | |
| answer_agent | bash, read_file | weak | |
| LoCoBench | planner (chairman) | read_file, bash | weak |
| developer | read_file, write_file, bash | strong |
Table 8: Roles, tools, and model assignment in weakstrong for each benchmark. Weak: Qwen3-30B-A3B-Instruct-2507; strong: Qwen3.6-35B-A3B. All roles also have the messaging tools; the chairman additionally has the role-management (start_agent / stop_agent) and answer-submission (finalize_task) tools. In weakstrong , only the bottleneck role is strong; OfficeQA runs this role (reader) as two identical instances, both strong.
| Meta-Team | AGORA | |
| Goal | Improve one team | Compare social conditions |
| L1 update | Agent’s own edits | Meta-Team, SkillOpt, or MemRL |
| L2 | Single fixed protocol | Set by structure; off under None |
| L3 proposal | Chairman alone; others suggest | Tier 1 proposes |
| L3 decision | Applied by the chairman | Rule set by structure |
| Chairman | Fixed by default | Fixed, or rotates under Competition |
Table 9: What AGORA inherits from and changes in Meta-Team. † Except under the Meta-Team mechanism, which keeps the original sequential execution and single training pass (Table 2 ).
| Stage | Model |
| Task execution | the role’s own model (weak or strong) |
| L1, L2, and L3 reflection in a session | the agent’s own model |
| L3 decision at the end of an iteration | the current chairman’s model |
| Gate sessions | as in task execution |
| SkillOpt analysis and edits | the strong model (fixed) |
| MemRL memory writing | the strong model (fixed) |
Table 10: Model used at each stage. “Fixed” means the same model regardless of team composition.
| Reference | L1: accept if candidate | L3: revert if changed team | |
| Successes (/20) | Rate | (/20) | (/20) |
| 0 | 0.00 | 2 | – |
| 1 | 0.05 | 2 | – |
| 2 | 0.10 | 4 | 0 |
| 3 | 0.15 | 5 | 1 |
| 4 | 0.20 | 6 | 2 |
Table 11: Gate thresholds for and . The reference is the stored skill score for L1 and the current team’s score for L3. L1 accepts a candidate skill only if it is clearly better (about two or more tasks above the reference); L3 reverts the changes only if the changed team is clearly worse (two or more tasks below). “–”: never reverted.
| None | Collaboration | Conformity | Competition | |
| L2 sharing | Disabled | Active | Only when blocked | Minimal |
| L2 note commit | – | Stage challenge commit | Direct write | Stage challenge commit |
| Comparative scores shown | – | – | – | Yes |
| Tier 1 | All roles | All roles | Chairman | Top scorer |
| L3 proposal right | All roles | All roles | Chairman | Tier 1 |
| L3 challenge tool | No | Yes | No | Yes |
Table 12: Full specification of the structure of interdependence. Scoring and revert rules are identical across conditions. Gate sessions run under the same condition, tiers, chairman, and constitution as training sessions.
| New chairman | Runner-up | Batch success (%) | avg@3 | |||||||
| Bench. | Mechanism | Team | Step | Chairman change | Model | score | score ( ) | before after | (%) | |
| GAIA | SkillOpt | onlyweak | 1 | plan_agent verifier_agent † | W | 0.67 | 3 | 0.20 (20) | 11.8 32.4 | 28.3 |
| GAIA | SkillOpt | weakstrong | 0 | plan_agent web_agent | S | 0.56 | 16 | 0.50 (20) | 58.8 88.2 | 74.2 |
| GAIA | SkillOpt | onlystrong | 0 | plan_agent file_agent | S | 1.00 | 1 | 0.75 (20) | 64.7 47.1 | 51.7 |
| GAIA | MemRL | onlyweak | 2 | plan_agent answer_agent | W | 48 | (49) | 23.5 41.2 | 13.3 | |
| GAIA | MemRL | onlyweak | 3 | answer_agent file_agent | W | 0.023 | 7 | (63) | 41.2 – | |
Table 13: Chairman changes under competition. Model : the model of the new chairman (W: weak, S: strong). The score that decides the ranking is the skill’s gate score for SkillOpt ( : val tasks), the mean of the memory bank for MemRL ( : memories), and the success rate on the role’s batch tasks for Meta-Team ( : tasks). Batch success is measured in the step of the change and averaged over the later steps; – marks a change at the last step. LoCoBench and OfficeQA Pro batches differ across steps, so their values are not directly comparable. In weakstrong teams, the change helps when the strong role takes the chair (GAIA: 74.2 and 63.3, the best structure) and hurts when a weak role does (OfficeQA Pro: 15.1, 10.2, and 14.5, at or near the bottom). No other competition run changed its chairman. † Role added by L3 in an earlier step.
| Metric | Definition |
| Challenge rate | Challenges to a teammate’s note or reflection, divided by the number of proposals |
| Withdrawal rate | Notes withdrawn by their author, divided by the number of notes |
| Reflection adoption rate | Reflections applied, divided by reflections proposed |
| Skip rate | Share of sessions in which the agent skips L2 or L3 reflection |
| Team-improvement suggestions | Suggestions sent to the chairman by Tier 2 agents |
| Teammate-model updates | Updates to teammate profiles and collaboration notes |
Table 14: Reflection metrics, counted per session and averaged per iteration.
| Metric | Definition |
| Roles added and removed | Per iteration and cumulative, counting only changes that pass the L3 gate |
| Final team size | Number of roles after the last iteration |
| Proposals and adoption rate | Proposals made, proposals kept by the decision rule, and proposals applied after the gate |
| Change types | Shares of role additions, constitution patches, and patches to existing roles |
| Gate margin | Validation successes of the current and modified teams and the -value of each L3 gate |
| Usage of added roles | Share of later sessions in which an added role is actually called |
Table 15: Team-structure metrics for L3.
| Metric | Definition | What it indicates |
| Messages | Messages sent per session | Volume of communication |
| Active roles | Roles called per session | Team size in use |
| Speaker concentration | Gini coefficient of each role’s share of messages | Hierarchical vs. flat |
| Chairman share | Share of messages between the chairman and a teammate, among all messages | Hub vs. mesh |
| Coordination overhead | Calls to wait for replies or check teammates’ status, divided by calls to work tools (file, shell, web search) | Waiting relative to working |
| Communication failures | Sends blocked as repeats and sends to roles that do not exist | Confusion as the team changes |
Table 16: Collaboration metrics, computed per session.
| Metric | Definition |
| Batch success rate | Success rate of each training batch |
| Timeout rate | Sessions that hit the 900-second limit, divided by all sessions |
| Session duration | Median and 90th percentile of session time |
| Cost per session | Messages, LLM calls, and input and output tokens |
| Abnormal endings | Sessions in which the chairman exits without submitting an answer, and sessions whose last message is used as a fallback answer |
| Context trimming | Times per session that older tool outputs are shortened (Appendix F ) |
Table 17: Efficiency and failure metrics.
| Solo | onlyweak | weakstrong | onlystrong | |||||||||||||||
| Mechanism | Step | Weak | Strong | None | Collab. | Conform. | Compet. | None | Collab. | Conform. | Compet. | None | Collab. | Conform. | Compet. | |||
| GAIA (17 problems per step, the same at every step) | ||||||||||||||||||
| SkillOpt | S0 | 29.4 | 70.6 | 5.9 | 35.3 | 11.8 | 17.6 | 58.8 | 41.2 | 47.1 | 58.8 | 58.8 | 70.6 | 82.4 | 64.7 | |||
| S1 | 35.3 | 58.8 | 29.4 | 17.6 | 23.5 | 11.8 | 41.2 | 41.2 | 58.8 | 94.1 | 70.6 | 76.5 | 82.4 | 35.3 | ||||
| S2 | 29.4 | 70.6 | 29.4 | 23.5 | 23.5 | 23.5 | 58.8 | 58.8 | 58.8 | 88.2 | 70.6 | 76.5 | 70.6 | 47.1 | ||||
| S3 | 41.2 | 70.6 | 35.3 | 29.4 | 11.8 | 41.2 | 52.9 | 70.6 | 52.9 | 82.4 | 64.7 | 82.4 | 88.2 | 58.8 | ||||
Table 18: Training batch success rate (%) at each step (S0–S3); Meta-Team runs a single step. On GAIA every step solves the same 17 train problems in a new order, so steps are comparable; on LoCoBench and OfficeQA Pro each step solves different problems, and step 2 of OfficeQA Pro has only 11 problems (Appendix A ).
| GAIA | LoCoBench | OfficeQA Pro | |||||||||||||
| Team | Structure | TO | Time | Calls | Msgs | TO | Time | Calls | Msgs | TO | Time | Calls | Msgs | ||
| onlyweak | Pre | 11.7 | 257 | 45.8 | 15.4 | 17.8 | 530 | 106.6 | 26.9 | 79.6 | 902 | 96.5 | 17.2 | ||
| None | 8.6 | 271 | 66.7 | 13.1 | 23.4 | 581 | 112.6 | 27.0 | 53.2 | 728 | 90.5 | 19.9 | |||
| Collab. | 5.0 | 226 | 60.3 | 9.1 | 19.1 | 575 | 99.3 | 26.4 | 46.8 | 773 | 79.1 | 14.8 | |||
| Conform. | 8.3 | 259 | 48.2 | 8.2 | 26.1 | 619 | 127.4 | 33.5 | 86.4 | 908 | 77.0 | 14.7 | |||
| Compet. | 5.3 | 168 | 75.0 | 18.7 | 35.5 | 738 | 142.3 | 40.6 | 41.7 | 692 | 106.8 | 27.3 | |||
Table 19: Efficiency in the final evaluation. TO : sessions stopped by the 900-second limit (%); Time : median session time (s); Calls : LLM calls per session; Msgs : messages sent per session. Evolved rows average the Meta-Team, SkillOpt, and MemRL runs of each structure; Pre is the same team before evolution.
| Team size | TO | Messages | Calls | ||
| GAIA | 36 | 0.59 | 0.46 | 0.06 | 0.28 |
| LoCoBench | 36 | 0.45 | 0.83 | 0.65 | 0.67 |
| OfficeQA Pro | 36 | 0.17 | 0.22 | 0.00 | 0.16 |
| Pooled | 108 | 0.32 | 0.42 | 0.36 | 0.33 |
Table 20: Spearman correlation between the change in avg@3 from Pre and the change in each process metric, over all evolved runs (36 per benchmark). Bold : .
| Best structure vs. none | None vs. Pre | |||||||
| Bench. | Mechanism | Team | Best structure | / | / | |||
| GAIA | Meta-Team | onlyweak | Compet. | +17/ 10 | 0.248 | +14/ 14 | 1.000 | |
| GAIA | Meta-Team | weakstrong | Compet. | +26/ 12 | 0.034 | +23/ 19 | 0.644 | |
| GAIA | Meta-Team | onlystrong | Collab. | +15/ 18 | 0.728 | +28/ 8 | 0.001 | |
| GAIA | SkillOpt | onlyweak | Compet. | +17/ 7 | 0.064 | +10/ 12 | 0.832 | |
| GAIA | SkillOpt | weakstrong | Compet. | +34/ 7 | <0.001 | +21/ 17 | 0.627 | |
Table 21: Task-level paired comparison on the 120 (GAIA), 297 (LoCoBench), and 186 (OfficeQA Pro) (task, rollout) pairs of the test set, which share seeds across runs. / : pairs solved only by the first / only by the second run; : exact McNemar test. Best structure : the best of collaboration, conformity, and competition. Bold : .
| Role | Sessions | Input tokens | Output tokens | Tool calls |
| web_agent | 78–80 | 50–78 | 28–52 | 33–54 |
| plan_agent (chairman) | 100 | 15–32 | 39–58 | 30–42 |
| answer_agent | 97–99 | 5–15 | 9–12 | 12–22 |
| file_agent | 2–9 | 0–2 | 0–4 | 0–3 |
Table 22: Workload of each GAIA role (%): share of sessions it joins and of the team’s tokens and tool calls, given as the range over the three teams before evolution.
| Strong role | Pre |
| web_agent (main weakstrong) | 48.3 1.4 |
| plan_agent | 28.3 2.9 |
| answer_agent | 25.0 2.5 |
| web + plan (2:2) | 52.5 6.6 |
| web + plan + answer (3:1) | 51.7 3.8 |
Table 23: Which role is strong matters (GAIA, no evolution, avg@3, %; mean SD over three rollouts). Each row makes the listed role(s) strong and the rest weak.
| 1:3 main (strong: web) | 2:2 (strong: web + plan) | 3:1 (strong: web + plan + answer) | ||||||||||
| Structure | Pre | Meta-Team | SkillOpt | MemRL | Pre | Meta-Team | SkillOpt | MemRL | Pre | Meta-Team | SkillOpt | MemRL |
| None | 48.3 1.4 | 51.7 2.9 | 51.7 2.9 | 50.0 4.3 | 52.5 6.6 | 56.7 1.4 | 49.2 5.2 | 57.5 2.5 | 51.7 3.8 | 58.3 2.9 | 55.8 8.0 | 56.7 7.6 |
| Collab. | 43.3 5.2 | 62.5 4.3 | 50.8 1.4 | 60.8 3.8 | 57.5 2.5 | 61.7 3.8 | 63.3 7.6 | 56.7 3.8 | 54.2 5.8 | |||
| Conform. | 47.5 0.0 | 50.8 3.8 | 58.3 5.8 | 58.3 1.4 | 55.0 6.6 | 58.3 5.2 | 54.2 3.8 | 48.3 2.9 | 55.0 2.5 | |||
| Compet. | 63.3 7.6 | 74.2 3.8 | 50.8 5.8 | 44.2 1.4 | 54.2 7.2 | 57.5 2.5 | 49.2 5.2 | 55.0 2.5 | 60.0 6.6 | |||
Table 24: Team composition on GAIA (avg@3, %, mean SD over three rollouts): the ratio of strong to weak roles in the weakstrong team. Strong roles are added in order of measured workload (web plan answer). Colors show the change from Pre of the same team as in Table 4 ; bold marks the best structure of interdependence in each column.
| Main gap (weak = Qwen3-30B-A3B) | Larger gap (weak = Qwen3.5-2B) | ||||||||
| Team | Structure | Pre | Meta-Team | SkillOpt | MemRL | Pre | Meta-Team | SkillOpt | MemRL |
| solo_weak | None | 20.8 1.4 | 20.8 5.2 | 20.8 5.2 | 20.0 0.0 | 16.7 5.8 | 14.2 5.8 | 15.0 6.6 | 8.3 3.8 |
| weakstrong | None | 48.3 1.4 | 51.7 2.9 | 51.7 2.9 | 50.0 4.3 | 3.3 3.8 | 4.2 1.4 | 5.0 5.0 | 5.8 2.9 |
| Collab. | 43.3 5.2 | 62.5 4.3 | 50.8 1.4 | 4.2 3.8 | 1.7 2.9 | 5.0 2.5 | |||
| Conform. | 47.5 0.0 | 50.8 3.8 | 58.3 5.8 | 3.3 1.4 | 5.8 6.3 | 1.7 1.4 | |||
| Compet. | 63.3 7.6 | 74.2 3.8 | 50.8 5.8 | 3.3 1.4 | 0.0 0.0 | 65.8 6.3 | |||
Table 25: Larger weak–strong gap on GAIA (avg@3, %, mean SD over three rollouts). The strong model is Qwen3.6-35B-A3B in both settings; only the weak model changes. Colors show the change from Pre of the same team as in Table 4 ; bold marks the best structure of interdependence in each column.
| Qwen (main) | Gemma-4 | ||||
| Team | Structure | Pre | SkillOpt | Pre | SkillOpt |
| solo_weak | None | 20.8 1.4 | 20.8 5.2 | 27.5 0.0 | 28.3 3.8 |
| solo_strong | None | 51.7 2.9 | 49.2 1.4 | 44.2 5.2 | 43.3 3.8 |
| onlyweak | None | 21.7 5.2 | 20.0 4.3 | 20.0 7.5 | 19.2 3.8 |
| Collab. | 19.2 6.3 | 20.8 7.6 | |||
| Conform. | 20.8 1.4 | 21.7 6.3 | |||
Table 26: Different model family on GAIA (avg@3, %, mean SD over three rollouts), SkillOpt only. Gemma-4: weak = E4B, strong = 26B-A4B; the SkillOpt optimizer is the family’s strong model. Colors show the change from Pre of the same team as in Table 4 ; bold marks the best structure of interdependence in each column.
| Qwen (main) | Mistral | ||||
| Team | Structure | Pre | SkillOpt | Pre | SkillOpt |
| solo_weak | None | 54.5 2.0 | 49.5 3.6 | 58.2 1.5 | 67.3 3.5 |
| solo_strong | None | 73.1 1.5 | 74.4 2.5 | 77.4 3.5 | 80.5 3.5 |
| onlyweak | None | 52.9 3.5 | 48.5 5.6 | 68.0 3.2 | 71.4 2.1 |
| Collab. | 54.9 6.1 | 70.0 3.1 | |||
| Conform. | 45.8 6.7 | 63.6 2.0 | |||
Table 27: Different model family on LoCoBench (avg@3, %, mean SD over three rollouts), SkillOpt only. Mistral: weak = Ministral-3-8B, strong = Mistral-Small-4-119B-A6B (NVFP4); teams that use only the strong model serve it on two GPUs. Colors show the change from Pre of the same team as in Table 4 ; bold marks the best structure of interdependence.
More in this topic
Multi-Agent LLM Systems
Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.
See all 1,040 papers