PaperlayerView exact revision on arXiv

Source-linked AI summary

EVOCHAMBER: Test-Time Co-evolution of Multi-Agent System at Individual, Team, and Population Scales

Yaolun Zhang, Tianyi Xu, Shengyu Dai, Zhenwen Shao, Qingyun Wu, Huazheng Wang

arXiv:2605.11136v1cs.AI

TL;DR

Prior test-time methods either isolate agents’ experiences or share them symmetrically, leaving continuous, gradient-free evolution of collaboration and knowledge flow across heterogeneous tasks unresolved. EVOCHAMBER evolves agents at individual, team, and population levels, using CODREAM for asymmetric transfer, and consistently outperforms baselines while producing stable niche specialists from identical agents.

  • Problem

    Prior methods do not address how multi-agent systems can continuously evolve at test time across heterogeneous task streams without gradient updates.

  • Method

    EVOCHAMBER evolves individual memories, niche-conditioned teams, and the agent population, with CODREAM routing distilled insights asymmetrically from strong to weak agents.

  • Results

    EVOCHAMBER consistently outperforms baselines across three heterogeneous task streams and two model families, reaching 63.9% on Hard Math with Qwen3-8B.

  • Takeaways & Limitations

    Identically initialized agents spontaneously form four to five stable niche specialists, a reproducible structural pattern unavailable to single-agent learners.

  • Takeaways & Limitations

    The framework is validated on only two model families, and its roughly 3.6× single-agent inference cost may be prohibitive in latency-sensitive settings.

Abstract

from arXiv · show

We argue that multi-agent test-time evolution is not single-agent evolution replicated N times. A single-agent learner can only evolve its own context and memory. A multi-agent system additionally evolves who collaborates, how they collaborate, and how knowledge flows across the population. These components have no single-agent counterpart and can produce phenomena such as emergent specialization. Yet prior test-time methods either confine experiences to individual agents, forfeiting cross-agent learning, or broadcast symmetrically to all agents, erasing the specialization that makes collaboration valuable. We present EVOCHAMBER, a training-free framework that instantiates test-time evolution at three levels over a coevolving agent pool. At its core is CODREAM (Collaborative Dreaming), a post-task protocol triggered on team failure or disagreement, in which agents collaboratively reflect, distill insights, and route them asymmetrically from strong to weak agents on the failed niche, preserving specialization while filling knowledge gaps. Team-level operators assemble niche-conditioned teams and select collaboration structures online. Population-level lifecycle operators fork, merge, prune, and seed agents under performance pressure. On three heterogeneous task streams with Qwen3-8B, EVOCHAMBER reaches 63.9% on competition math, 75.7% on code, and 87.1% on multi-domain reasoning, outperforming the best baseline by 32% relative on math and confirming asymmetric cross-agent transfer as the primary driver in ablation. Starting from several identically initialized agents, four to five stable niche specialists spontaneously emerge, a structural signature of multi-agent evolution that no single-agent learner can express. See our code at: https://github.com/Mercury7353/EvoChamber

1 Introduction

EVOCHAMBER frames multi-agent test-time evolution as a richer process spanning individual, team, and population state, then instantiates all three levels online without training. Its asymmetric cross-agent transfer and population dynamics improve heterogeneous-task performance while producing emergent niche specialization.

  • Motivation: Multi-agent evolution extends beyond individual context and memory to team collaboration, knowledge flow, and population membership, enabling emergent specialization.These team- and population-level components have no single-agent counterpart.
  • Motivation: Prior methods either isolate experiences within agents or broadcast them symmetrically, respectively sacrificing cross-agent learning or erasing specialization.Other co-improvement methods use RL fine-tuning or offline structure search with fixed roles within a single domain.
  • EVOCHAMBER: EVOCHAMBER is a training-free framework that evolves an agent pool at individual, team, and population levels online.It combines private experience and niche competence, niche-conditioned teams of three, online collaboration-structure selection, and CODREAM-based population evolution.
  • Results: 63.9% on Hard Math, 75.7% on Hard Code, and 87.1% on AFlow-Stream are achieved with Qwen3-8B, outperforming MemCollab by 32% relative on math.EVOCHAMBER also achieves a 5× improvement on CodeContests over a single agent.
  • Results: −10.8% is the single largest ablation drop when CODREAM is removed, identifying asymmetric cross-agent transfer as the primary driver.Disabling the team or population level produces level-specific performance drops.
  • Results: Four to five stable niche specialists spontaneously emerge from identically initialized agents, a reproducible pattern structurally impossible for a single-agent learner.The specialists emerge across the coevolving agent pool.

2 Related Work

Prior work improves multi-agent reasoning through fixed roles, individual memory, symmetric sharing, or gradient-based co-evolution, but does not provide EVOCHAMBER’s combination of persistent pool state and asymmetric, inference-time knowledge transfer. CODREAM addresses the specialization loss of symmetric memory by routing distilled insights to deficit agents.

  • Static multi-agent systems: Static multi-agent systems fix or dynamically group roles, while debate and test-time reasoning enhancements lack persistent state across tasks.Workflow and architecture discovery methods search offline rather than evolving agents with the task stream.
  • Individual agent memory: Individual memory methods improve agents separately but provide no mechanism for one agent’s learning to transfer to another.This limitation is especially important at low success rates, where individual memory accumulates slowly.
  • Symmetric shared memory: Symmetric shared memory broadcasts identical distilled insights to all agents, conflating domain-specific strategies and destroying specialization.MemCollab enables collective learning, whereas CODREAM uses asymmetric, gap-targeted distillation for deficit agents.
  • Gradient-based co-evolution: Gradient-based co-evolution methods require gradient updates on a static training distribution, whereas EVOCHAMBER evolves prompts at inference time.The paper positions EVOCHAMBER as achieving comparable qualitative co-evolution goals without gradient updates.

3 Method

EVOCHAMBER evolves a multi-agent state at individual, team, and population levels by selecting niche-conditioned teams, learning collaboration structures, routing knowledge after failures, and periodically reshaping the agent pool. Its state tracks private competence, pairwise synergy and overlap, leadership experience, and a mutable roster.

  • Per-task loop: Each task selects three roles—anchor, complement, and scout—executes a leader-chosen collaboration structure, propagates shared reward, and triggers CODREAM after failure or disagreement.Lifecycle operators fork, merge, prune, and seed agents every τ tasks.
  • Evolvable state: The multi-agent state maintains a pool, k-sized teams, collaboration structures, pairwise synergy, style overlap, and a mutable roster that persist across tasks.These quantities drive team formation and population adaptation beyond individual memories.
  • Team composition: Team composition combines niche competence, anchor synergy, stylistic distinctness, and under-exposure to preserve complementary skills while ensuring exploration.The anchor is the niche’s current best performer; the complement supplies missing capability, and the scout favors under-exposed agents.
  • Collaboration structure: LeadLearn retrieves similar team and task profiles to choose among voting, debate, generator-critic, and decompose structures, then records outcomes and reflections.The shared leadership bank accumulates structure-selection knowledge as the anchor rotates.
  • CODREAM: CODREAM fires when mean reward falls below θ or members disagree, using a five-phase reasoning loop to diagnose failures, contrast successful approaches, and imagine hypothetical strategies.It routes useful knowledge between existing agents without diluting specialization.
  • Population lifecycle: Every τ tasks, genesis, fork, merge, and prune reshape the roster to fill coverage gaps, create specialist headroom, combine redundancy, and remove persistently weak agents.With |P| > k, unused agents retain their state alongside newly seeded or modified specialists.

4 Experiments

Across three heterogeneous streams, EVOCHAMBER improves most on difficult tasks through asymmetric cross-agent transfer, while robustness and ablations identify CODREAM as the main contributor. Pool analysis shows dynamically emerging, niche-specific specialists and structured knowledge routing rather than fixed assignments.

  • Performance: 0.639 overall accuracy on Hard Math, a 32% relative improvement over MemCollab and double the single-agent baseline.Gains reach +0.160 on math_hard and +0.167 on AIME’24; majority voting falls to 0.067 on AIME.
  • Performance: 0.352 CodeContests accuracy represents a 5× improvement over a single agent, as debugging experience transfers from easier MBPP+ tasks.MBPP+ saturates near 0.85, while EvoMem and MemCollab score 0.027 and 0.084 on CodeContests.
  • Performance: 0.871 AFlow-Stream accuracy exceeds EvoMem at 0.840 and MemCollab at 0.832 across sequential six-domain adaptation.EVOCHAMBER wins or ties on five of six domains, with largest gains on MATH and MBPP.
  • Ablations: −0.108 is the largest ablation drop when CODREAM is removed, establishing asymmetric cross-agent transfer as the primary driver of collective learning.HotpotQA declines from 0.895 to 0.572 and DROP from 0.860 to 0.480.
  • Robustness: 0.655 and 0.662 accuracy under random Hard Math permutations exceed 0.639 under the default order, ruling out favorable curriculum effects.The gains therefore remain under task reordering rather than depending on a particular sequence.
  • Pool evolution: Niche-specific specialists emerge dynamically: dominant experts differ across AIME years, activate when their niches arrive, and exchange CODREAM knowledge through concentrated giver →recipient channels.The specialization index rises from ∼0.1 on Hard Math toward ∼0.3 on AIME’24 as task difficulty increases.

5 Conclusion · Appendix

EVOCHAMBER frames multi-agent test-time evolution as fundamentally distinct from replicating single-agent evolution, because teams and populations evolve collaboration and knowledge flow beyond individual context and memory. Its gradient-free three-level framework uses CODREAM for verified asymmetric transfer and produces spontaneous niche specialization while outperforming baselines.

  • 5 Conclusion: Multi-agent evolution extends beyond individual context and memory to who collaborates, how collaboration occurs, and how knowledge flows across the population.These team- and population-level components have no single-agent counterpart.
  • 5 Conclusion: These additional team and population dynamics produce emergent phenomena that no individual learner can express.
  • 5 Conclusion: EVOCHAMBER instantiates individual, team, and population evolution over a coevolving agent pool without gradient updates.Its core mechanism is CODREAM.
  • 5 Conclusion: CODREAM enables verified asymmetric knowledge transfer across the coevolving agent pool.
  • 5 Conclusion: Across three heterogeneous task streams and two model families, EVOCHAMBER consistently outperforms all baselines.
  • 5 Conclusion: N identical agents spontaneously differentiate into several stable niche specialists without being explicitly engineered to do so.The passage presents this spontaneous differentiation as the most striking emergent result.

A Limitations and Future Work

EVOCHAMBER’s evaluation is limited by narrow architectural coverage and substantially higher inference cost, despite being more accurate than SC with k=5 at 72% of SC’s token budget. Future work targets stronger backbones, longer streams, specialization stability, insight obsolescence, and formal role-conditioned credit attribution.

  • Limitations: EVOCHAMBER is validated on only two model families, limiting the demonstrated evidence for generalizability.The authors expect transfer because the mechanism uses prompts without architecture-specific components.
  • Limitations: 3.6× inference cost relative to a single agent may be prohibitive in latency-sensitive settings.EVOCHAMBER is nevertheless more accurate than SC with k=5 at 72% of SC’s token budget.
  • Future work: Stronger backbones and streams beyond 1000 tasks would enable studies of scaling limits, long-horizon specialization stability, and insight obsolescence.These extensions are proposed as future work to test how the system evolves over longer horizons.
  • Future work: Formalizing role-conditioned credit attribution beyond the current shared team reward is another future direction enabled by the three-level decomposition.This would extend the framework’s treatment of responsibility within collaborative teams.

B More Experiments … E.4 Evaluation Protocol

EVOCHAMBER’s experiments show reproducible specialization patterns despite seed-dependent specialist identities and minimal sensitivity to initial pool size. The framework combines niche-aware knowledge routing, explicit agent-lifecycle operators, defined collaboration metrics, and a standardized evaluation setup spanning related-work positioning through protocol details.

  • B More Experiments: Mean specialization index was 0.123 ± 0.008 and unique anchors were 38 ± 5 from an initial N=20 across three seeds, while specialist identities differed.The dominant-specialist pattern is seed-invariant, but which agent specializes in each niche is seed-dependent.
  • B More Experiments: The final-accuracy gap between N=3 and N=20 was only 0.011 absolute, with N=3 matching or slightly exceeding N=20 across all four AIME years.Both settings converged to comparable effective pools: N=3 grew to 8 active agents, while N=20 retained 9 routinely selected agents.
  • C Related Work Positioning Table: EVOCHAMBER was the only compared method satisfying all five criteria: training-free operation, an agent pool, cross-agent transfer, asymmetric transfer, and continuous task-stream evolution.The comparison table uses ✓, ×, ◦, and — to denote fully satisfied, unsatisfied, partially satisfied, and inapplicable criteria.
  • D Experience Archive Design Justification: The experience archive separates niche-specific lessons from cross-domain meta-insights, enabling targeted retrieval and transfer without polluting general metacognition with tactical noise.Niche lessons are retrieved by cosine similarity over task embeddings, while cross-domain insights support broader transfer; near-duplicates are deduplicated.
  • E.1 Operational Details: Style overlap ωij is cosine similarity between agents’ competence vectors, with missing subdomain entries treated as zero, while pair synergy σij(z) is mean reward on prior shared niche-z tasks.Synergy remains zero until a pair has co-participated in at least five niche-z tasks, limiting estimates based on small samples.
  • E.1 Operational Details: Lifecycle operators fork top-performing agents, merge highly similar mature pairs, prune sustained underperformers, and seed empty-memory agents when pool coverage or size falls short.Fork targets the top 10% by rolling reward; merge requires profile cosine similarity above 0.95 and at least 10 tasks; prune uses 0.8× the pool mean for 10 consecutive tasks.
  • E.2 Inference Configuration: All experiments use locally served Qwen3-8B with vLLM, thinking mode across streams, 4,096 solving tokens per agent, 8,192 for Hard Math, and 2,048 tokens per CODREAM phase.Inference uses two instances with round-robin load balancing, while retrieved insights are prepended to the system prompt.
  • E.3 Hyperparameters / E.4 Evaluation Protocol: A single hyperparameter configuration is used across streams and backbones, with fixed task order across methods and generic initial personas so domain knowledge emerges from experience.The protocol is designed to keep learning trajectories comparable and avoid supplying domain-specific knowledge at initialization.

F CODREAM Isolation Experiment

The controlled isolation shows that maintaining a multi-agent pool without cross-agent knowledge sharing provides no advantage over a single agent. Adding CODREAM produces the multi-agent gain through asymmetric transfer, with larger effects on dependent-reasoning domains.

  • Controlled isolation: 0.633: EVOCHAMBER without CODREAM matches the single-agent configuration exactly on the 30-task math subsequence.The comparison retains the 20-agent pool, team composition, and lifecycle operators while removing CODREAM.
  • Controlled isolation: 0.700: Adding CODREAM yields a +10.5% relative improvement, confirming asymmetric transfer as the mechanism behind the multi-agent advantage.This result is reported on the same 30-task AFlow math subsequence.
  • Controlled isolation: −0.108: Removing CODREAM lowers full 600-task AFlow-Stream overall accuracy, especially on HotpotQA and DROP.The larger effect is concentrated in dependent-reasoning domains where cross-agent coordination knowledge is most valuable.

G Hard Code Stream Per-Benchmark Breakdown

The hard code stream’s benchmark variation is concentrated in CodeContests: HumanEval saturates at ceiling for every method, while MBPP+ remains near 0.85 across multi-agent methods. EVOCHAMBER’s CodeContests advantage is attributed to experience-guided debugging and cross-difficulty transfer from MBPP+ failures.

  • HumanEval: HumanEval saturates at 1.000 for every method, so Table 3 omits its column.Overall is the micro-average over all 586 tasks, including HumanEval.
  • HumanEval: HumanEval reaches ceiling for all methods, with remaining run-to-run diversity within grading tolerance.Task-level inspection confirmed that every method solves every HumanEval problem correctly.
  • MBPP+ and CodeContests: MBPP+ clusters near 0.85 for all multi-agent methods, leaving CodeContests as the discriminating subset.This benchmark contrast explains why CodeContests provides the clearest separation among methods.
  • CodeContests: 0.352: EVOCHAMBER’s CodeContests accuracy is a 1.8× improvement over SC k=5 and 3.5× over AgentNet.The reported mechanism is experience-guided debugging using prior failure patterns and repair strategies for similar problem classes.
  • CodeContests: Cross-difficulty transfer lets debugging patterns learned on easier MBPP+ problems accumulate in agent profiles and improve later CodeContests corrections.Agents with relevant experience archives attempt more targeted repairs on subsequent CodeContests problems.

H Order and Execution Robustness: Setup

The robustness experiments keep the main Hard Math Stream configuration fixed while testing shuffled task orders and uniform voting. The setup also uses the default fixed-order SA score as a permutation-invariant reference.

  • Experimental setup: All robustness runs use the Qwen3-8B backbone, identical pool and team sizes, and the main Hard Math Stream code version.These settings match the Table 2 configuration and support the Table 5 robustness experiments.
  • Shuffle conditions: The default order presents domains in sequential blocks, whereas two seeded shuffle conditions reorder all 382 tasks across domains without changing the task set.Shuffling changes only task ordering while preserving the complete set of 382 tasks.
  • SA reference: SA uses the default fixed-order score and is statistically indistinguishable under task permutations because it does not accumulate experience across tasks.A shuffled run with seed 42 produced equivalent results.

I Per-Subset Regime Analysis

EVOCHAMBER’s gains follow a regime structure determined by per-agent success rates: diversity and dynamic team structures dominate at high accuracy, while CODREAM is most effective at intermediate accuracy. At very low accuracy it abstains, and its non-voting structures avoid self-consistency collapse when rare correct solutions exist.

  • Very high accuracy: At ≥80% accuracy, gains are modest (+0.01 to +0.07), with team diversity and dynamic leader-selected structures dominating cross-agent distillation.For GPT-4.1-mini on math_hard, base accuracy is 0.824.
  • Mid accuracy: At 40%–70% accuracy, CODREAM reaches its design sweet spot, with gains from +0.07 on Qwen3-8B math_hard to +0.20+ on GPT-4.1-mini AIME’22.This regime combines enough verified solutions for crystallization with wide gaps between struggling and successful agents, making cross-agent distillation meaningful.
  • Low accuracy: At 20%–40% accuracy, Qwen3-8B AIME gains from CODREAM are small and statistically noisy, while full EVOCHAMBER still improves AIME through team selection and lifecycle.Qwen3-8B AIME success rates are near 15–20%, compared with 30–40% for GPT-4.1-mini.
  • Very low accuracy: Below 15% accuracy, CODREAM abstains when no agent solves a task because verified solutions are unlikely to emerge.With 10% per-agent accuracy and a team of 3, at least one agent succeeds with probability 1 −(0.90)3 = 0.271, while two independent verified attempts are rarer.
  • Self-Consistency collapse: The empirical self-consistency result is 0.067 on AIME, while EVOCHAMBER uses debate, generator-critic, or decompose instead of voting when rare successes exist.With five independent agents at 20% accuracy, majority voting can override rare correct answers; EVOCHAMBER avoids that failure mode through leader-selected non-voting structures.

J Case Study: How EVOCHAMBER Learns Competition Mathematics … L.2 Code Insights, Hard Code Stream, Qwen3-8B

The case study shows EVOCHAMBER developing a concentrated expert core through CODREAM insight production and lifecycle specialization on hard mathematics, while adapting lifecycle behavior when AIME coverage gaps emerge. CODREAM yields actionable cross-task math and code techniques, but its insights degrade when no agent solves the underlying problem.

  • J Case Study: How EVOCHAMBER Learns Competition Mathematics: At task 11, CODREAM crystallizes three insights after a team score of 0/3, including an inclusion-exclusion technique with LCM adjustments.The first trigger involves agents 6f3dcc14, bb411e98, and 119b9e09.
  • J Case Study: How EVOCHAMBER Learns Competition Mathematics: When no team member solves an underlying problem, CODREAM produces generic reading-comprehension advice rather than targeted mathematical techniques.This failure mode occurs because crystallization lacks a successful trajectory to distill, although the verification gate still passes candidates.
  • J Case Study: How EVOCHAMBER Learns Competition Mathematics: 34 CODREAM triggers across 382 tasks produce 93 verified insights, while late-stream lifecycle behavior shifts from specialization to genesis and pruning during AIME.The AIME phase contains 5 genesis events for coverage gaps and 1 prune after six consecutive underperforming tasks.
  • J Case Study: How EVOCHAMBER Learns Competition Mathematics: 77% of verified math_hard insights come from the same three agents, which also account for all 43 specialize events.They produce 72 of 93 verified insights, while the top three agents overlap completely with the top forked agents.
  • K Lifecycle Operator Analysis: Fork, merge, and retire operators mainly preserve diversity and prune stagnation over long streams rather than accelerating early learning.Lifecycle contribution is negligible early and grows modestly in mid-to-late phases.
  • L CODREAM Insight Examples: CODREAM insights are generated during post-task reflection, verified by re-attempt, and routed to the appropriate experience archive.This protocol describes the provenance of representative insights from actual experimental runs.
  • L.1 Math Insights, Hard Math Stream, Qwen3-8B: Math examples cover modular arithmetic integration, trapezoid-height measurement, and constraint-graph reformulation for constraint-satisfaction problems.These examples emphasize structural modeling and careful geometric or combinatorial reasoning.
  • L.2 Code Insights, Hard Code Stream, Qwen3-8B: Code insights promote memoization with state compression, bounded arithmetic with saturation, and symbolic–numerical cross-validation as actionable cross-task patterns.Most insights are classified as cross-domain, while niche-specific routing and selective sharing use classification and cosine-similarity checks.
Loading 2605.11136v1…