MoRSE: Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts

David van Dijk; Peiwen Li; Rex Ying; Shiyang Zhang; Sizhuang He; Yangtian Zhang

REVIEW 3 major objections 5 minor 48 references

MoRSE: Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read MoRSE claims that conditioning each agent's parameters on both role and subtask, not just prompts, is the key to multi-agent LLM performance on complex code tasks.

desk verdict Solid empirical paper; the headline variance-reduction theorem has a real proof gap that needs correcting or downgrading before this goes out. read the letter →

arxiv 2608.09251 v1 pith:NPNGCJ6M submitted 2026-08-10 cs.MA cs.AIcs.CLcs.LG

Peiwen Li , Shiyang Zhang , Yangtian Zhang , Sizhuang He , David van Dijk , Rex Ying This is my paper · ORCID

classification cs.MAcs.AIcs.CLcs.LG

keywords multi-agentsystemslargelanguagemodelsmixtureofexpertslow-rankadaptationreinforcementlearningcreditassignmentcodegenerationtaskdecomposition

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large language model multi-agent systems currently differentiate agents mostly by prompts, and the paper's motivating diagnosis is that this leaves outputs redundant and capabilities underspecialized. MoRSE aims to show that spelling out each agent's responsibility as a (role, subtask) pair, and then giving agents parameter-level specialization through a prototype-routed mixture of role and subtask LoRA experts, substantially outperforms prompt-level role differentiation. A hierarchical group-relative policy optimization (HGRPO) with two-layer credit assignment is introduced to train experts and router stably under sparse step rewards. The paper reports best held-out whole-task and step-wise scores over all baselines on two code-generation benchmarks across three backbones, with trained specialization transferring to held-out task categories and domains. If correct, the result identifies parameter-level heterogeneity, not agent count or prompt design alone, as the key lever in multi-agent LLM systems.

What carries the argument

The load-bearing object is a dual-factorized mixture of LoRA experts: a small pool of role experts (one per role such as execute or merge) plus a shared pool of subtask experts, selected per agent call by a prototype-based semantic router that scores subtask embeddings against learnable prototypes and picks the top-$K$ subtask experts. The active set is composed as a LoRA combination on a frozen backbone, so each (role, subtask) call obtains its own low-rank adaptation subspace at near single-model cost. Training is carried by HGRPO, whose two-layer credit assignment computes a within-route advantage for the expert update and a cross-route advantage for the router update; this conditional-baseline structure removes the cross-route component of reward variance from expert gradients. The task-oriented DAG decomposition supplies the (role, subtask) labels and rule-based step rewards that condition and supervise this machinery.

What would settle it

Train MoRSE exactly as reported on SRDD with Qwen3-4B, then rerun the same HGRPO objective with each node's verifier reward replaced by a shuffled version of the same reward values. If whole-task and step-wise scores still reach the Table 2 row-D levels, the claim that valid step-level credit drives the gains is false; if performance drops to the MoRSE-base level, the verifier-based credit is confirmed as load-bearing.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that (role, subtask)-conditional specialization at both the task-structure and parameter levels yields measurable gains over untrained structural decomposition and over prompt/role-only multi-agent baselines. The central claim is that router and expert quality have to be decoupled: HGRPO assigns a within-route advantage to the LoRA experts and a cross-route advantage to the router, and Proposition 3.1 states this strictly reduces the LoRA-expert gradient variance relative to a single shared GRPO advantage. The supporting subspace analysis argues that different roles and subtasks occupy divergent rank-$\rho$ update directions, so no single shared LoRA can cover them, and the paper formalizes this as a coverage limit (Theorem L.2). Empirically, MoRSE attains the best held-out test scores over all baselines in both whole-task and step-wise performance on SRDD and SciCode across three backbones, and the trained specialization outperforms an untrained variant and a fixed-LoRA fine-tuning baseline on held-out categories and domains.

Load-bearing premise

The rule-based verifier is assumed to provide valid step-level rewards, and the planner's DAG decomposition is assumed accurate enough for subtask-level credit assignment to propagate correct signals; if either signal is noisy or misaligned with true subtask quality, the HGRPO training signal is corrupted and the reported gains may not transfer.

Editorial extensions

If this is right

Multi-agent LLM systems can shift from prompt-level role descriptions to actual parameter-level specialization while keeping the backbone frozen and training only small adapters.
Learned subtask experts compose for held-out task categories and domains, suggesting that specialization transfers as reusable skills rather than memorized training examples.
HGRPO makes routed-mixture training stable under sparse step rewards, removing a known fragility of standard GRPO when experts and router are co-optimized.
Because gains appear at the step level as well as end-to-end, per-subtask credit assignment is meaningful enough to drive process-level improvement.
The framework's cost profile stays close to a single model, since only LoRA experts and router parameters are trained on the shared backbone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the LoRA subspace-coverage limit generalizes, then any setting where agent subgroups need divergent low-rank updates should favor conditional mixtures of adapters over one shared adapter; this predicts similar gains outside code generation, for example in multi-step planning or tool use.
HGRPO's router-expert credit split is a structural rather than temporal decomposition, so it could be composed with rollout-time hierarchies to assign credit along both axes at once.
A testable corollary of the variance formula is that HGRPO's advantage over standard GRPO should grow as cross-route reward variance increases; experiments that manipulate router-induced reward spread could verify this directly.
Because the whole pipeline leans on rule-based step rewards, the magnitude of the reported gains should depend on verifier quality; comparing training under shuffled versus faithful verifier rewards would isolate that dependence.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

MoRSE is a strong systems paper whose empirical contribution is real and whose headline theoretical claim does not hold up as stated. The stress-test note on Proposition 3.1 is correct: the proof drops a cross term it is not entitled to drop. The "provable variance reduction" language needs to be fixed or walked back. The paper would still be worth publishing, because the empirical case for HGRPO does not rest on the theorem.

What is actually new: the (role, subtask) factorization of LoRA experts with a prototype-based router, and the two-layer credit assignment that gives experts and router separate advantages. The pieces are known (DAG decomposition, GRPO, LoRA-MoE), but the combination is a real step beyond prompt-level role differentiation, and the system delivers the gains it claims on held-out splits of SRDD and SciCode across three backbones. The evaluation is the best part: budget-matched ablations, including a total-parameter-matched fixed-LoRA control, category/domain-disjoint OOD splits, a router-utilization analysis, and an execution-verified evaluation whose checks come from a model family disjoint from the trained backbone. The limitations appendix is unusually honest about single-seed results and the partial reward/ECI alignment.

The soft spot that matters is the math in Appendix G. The proof writes u−μ̄ = (u−μ_b) + (μ_b−μ̄), calls the two terms uncorrelated, and concludes Var[(u−μ̄)∇] = Var[(u−μ_b)∇] + σ²_C E[||∇||²]. Orthogonality of the reward deviations gives E[AC]=0, but the gradient cross term is E[AC||∇ log p||²], which does not factor. The score identity only guarantees E[∇ log p | b] = 0, not E[(u−μ_b)||∇ log p||² | b] = 0. When reward correlates with gradient magnitude, the cross term is nonzero and Eq. (10) can fail with either sign. Corollary G.1 inherits the gap.

Happily, this does not sink the paper. The C-vs-D ablation and the Appendix H training dynamics show standard GRPO destabilizing with MoLE while HGRPO stays controlled; that is direct empirical evidence for the credit-assignment design. The fix is straightforward: prove a correct variance statement under explicit assumptions (e.g., conditional independence of reward and gradient norm within routes) or downgrade the claim to an empirical finding. The subspace-coverage theorem in Appendix L checks out, and the rest of the math is fine. The remaining concerns are minor: single-seed runs, and trust in the rule-based verifier's step rewards, which is reasonable for SciCode's unit tests and coarser for SRDD, though the independent execution-verified evaluation mitigates the latter.

Verdict: this deserves real peer review, and it would make a good reading-group paper — the proof gap is exactly the kind of thing worth dissecting. I would cite it for the MoLE+HGRPO design and the careful empirical protocol, and I would send it to referees with a request that the theoretical claim be corrected or softened.

Referee Report

3 major / 5 minor

Summary. The paper introduces MoRSE, a multi-agent LLM framework that combines (i) a task-oriented DAG decomposition assigning each agent an explicit (role, subtask) pair, (ii) a mixture of role and subtask LoRA experts selected by a prototype-based semantic router, and (iii) a hierarchical group-relative policy optimization (HGRPO) with within-route and cross-route advantages. The central claims are that the task-structure DAG alone improves held-out inference, that the LoRA expert/router module adds further parameter-level specialization, and that HGRPO provides a provable variance reduction for the expert updates while stabilizing co-optimization. The method is evaluated on SRDD and SciCode across three backbones, with budget-matched ablations, OOD splits, and an execution-verified evaluation protocol.

Significance. If the empirical results hold, the paper makes a meaningful contribution: it moves multi-agent LLM specialization beyond prompt-level role differentiation to parameter-level (role, subtask) adaptation, and it provides a careful evaluation design. The execution-verified independent SRDD evaluation, the budget-matched trained baselines, and the OOD category/domain splits are genuinely useful strengths, as are the three-backbone comparisons and the explicit ablation map in Appendix J. However, the paper's headline theoretical contribution, Proposition 3.1, is not supported by the proof as written, and this undermines one of the three stated pillars of the method. The empirical case for HGRPO also rests on single-seed runs, so the quantitative strength of the claims is weaker than the text suggests. The core empirical idea remains defensible and the theoretical overreach is fixable, but the current version overclaims.

major comments (3)

[§3.3, Eq. (10); Appendix G] The variance-reduction claim in Proposition 3.1 and Corollary G.1 does not follow from the proof. The proof treats the LoRA gradient as (u−μ_b)∇log p and the single-baseline gradient as (u−μ̄)∇log p, then asserts that the cross term between u−μ_b and μ_b−μ̄ vanishes because these two reward deviations are uncorrelated. Uncorrelatedness gives E[(u−μ_b)(μ_b−μ̄)]=0, but the gradient cross term is E[(u−μ_b)(μ_b−μ̄)∥∇Φ log p∥²], which does not factor as E[(u−μ_b)(μ_b−μ̄)]·E[∥∇Φ log p∥²]. Conditioning on the route b reduces the issue to E_b[(μ_b−μ̄) E[(u−μ_b)∥∇Φ log p∥² | b]], and the score-function identity only guarantees E[∇Φ log p | b]=0, not E[(u−μ_b)∥∇Φ log p∥² | b]=0. A concrete counterexample exists: for a Gaussian policy with mean f(x) and reward u=y², E[(u−μ_b)(y−f)² | b] is nonzero, so the claimed variance reduction can fail and can even reverse. Corollary G.1 inherits the same flaw, and it is compounded by the fact that Algorithm 1 computes the within-route mean μhat_b from the same M candidates that include the current sample, so the baseline is not independent of the action and the unbiasedness argument is only heuristic. The C-vs-D ablation (Table 2) and the training dynamics in Appendix H are legitimate empirical evidence, but they do not establish the 'provably reduces variance' statement. The authors should either prove the result under explicit additional assumptions (for example, conditional independence of the score norm and reward deviation given the route) or remove the formal guarantee and present HGRPO as a heuristic whose benefits are empirical.
[Appendix L, §L.1–L.3 and Figure 2] There is a mismatch between the object measured empirically and the object used in Theorem L.2. The diagnostic in App. L.1 computes rank-ρ subspaces of input activations to each target layer (A^(L) ∈ R^{Np×d}, the last-token input activation), whereas Assumption L.1 and Theorem L.2 concern 'group-optimal rank-ρ LoRA update subspaces' Va, Vb. The paper states in the 'Connection' paragraph that no bridging assumption is needed because Δ is 'exactly the rank-ρ subspace divergence reported in Figure 2(a,b)', but Figure 2 measures activation subspaces, not weight-update subspaces. Activation subspace divergence and update subspace divergence are different quantities; a low-rank adapter could in principle compensate for divergent input subspaces or fail despite similar input subspaces. The theorem should be restated as a conditional result on update-subspace divergence, and the activation diagnostic should be presented as suggestive evidence rather than as an exact measurement of the theorem's premise.
[§4, Appendix C; Tables 1, 2, 7, 8] All headline results are reported from a single seed, and the text states this only in the limitations appendix. With stochastic decoding at temperature 0.2, the differences in Table 1 and Table 2 have no confidence intervals, and some reported OOD deltas (e.g., Table 8 Mean Step Pass deltas of −0.63 and +0.51) are clearly within plausible run-to-run noise. The paper should either provide multiple seeds with standard deviations or, at minimum, a paired bootstrap or Wilcoxon analysis over benchmark instances, for the key comparisons (MoRSE vs. MoRSE base, MoRSE vs. fixed-LoRA, and HGRPO vs. standard GRPO). Without this, the quantitative claim that MoRSE 'attains the column-best score on every reported test cell' is not yet statistically supported.

minor comments (5)

[§4.1, Table 1] The sentence 'On the All set, the untrained MoRSEbase achieves the best overall performance comparing with all baselines' is contradicted by the Gemma-4-31B rows, where the Single base model has ECI Mean 0.763 vs. 0.734 and ECI Product 0.338 vs. 0.316. Please qualify this claim or state explicitly that 'best overall' refers to a majority of metrics rather than all metrics.
[§3.2, Eq. (8)] The router loss uses log πψ(E_s,i | s_i), where E_s,i is a top-K subset of subtask experts, but Eq. (5) only defines a per-expert softmax πψ(k | s_i). The paper should specify the distribution over subsets used for stochastic top-K sampling and the exact log-probability formula in Eq. (8).
[Appendix G, Corollary G.1] The bridge argument says both normalized estimators are divided by the same route-constant scale σ(b)+ε, but σ(b) is itself estimated from the same M candidates that define u(b,m), so it is not a fixed constant given the data. The finite-sample effect should be stated explicitly and not folded into the same conditional-independence argument.
[Figure 6] The left panel's axis label appears garbled in the preprint rendering ('010 2'); please correct the typesetting of the symlog axis.
[Table 1, SciCode Problem Pass, Gemma-4-31B] On the Test set, the Problem Pass value for MoRSE is 20.00, tied with Single, MacNet, and AFlow. The abstract's 'improvements in both whole-task and step-wise performance' should be nuanced to account for ties on this cell, or the claim should focus on the metrics where the gain is unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central empirical claims are tested against external held-out benchmarks, the variance-reduction claim is a self-contained proof statement, and the disclosed training-reward/evaluation overlap is acknowledged and independently verified.

full rationale

I walked the derivation chain and found no load-bearing step that reduces, by the paper's own equations or by self-citation, to its own inputs. The main empirical claims (Section 4, Tables 1-8) are evaluated on external benchmarks SRDD and SciCode with held-out test splits and category/domain-disjoint OOD splits, against budget-matched baselines sharing the same backbones, prompts, and decoding configurations; the untrained MoRSE base provides a no-training framework comparison, and trained MoRSE is compared with fixed-LoRA and standard-GRPO controls under matched activated/total parameter and rollout budgets (Appendix J, Tables 3-5). No reported score is a fitted parameter renamed as a prediction. HGRPO's variance-reduction claim (Proposition 3.1, Eq. 10) is presented as a mathematical identity derived from score-function identities and conditional baselines, with a proof in Appendix G; even if a critic challenges the proof's correctness, that is a separate mathematical concern and not circularity, because the claim does not encode its own conclusion by definition or via self-citation. The only point where training and evaluation signals are related is the SRDD reward/ECI overlap, and the paper explicitly discloses this in Limitations and Appendix I, states that the weights and implementations differ and that the training reward never observes the evaluation scorer, and provides an execution-verified independent evaluation in Appendix J (Table 6) using acceptance checks and a grader disjoint from the trained backbone and evaluation embedder. That is a disclosed and mitigated alignment concern, not a circular derivation. Theorem L.2's coverage-limit theorem takes the empirically measured subspace divergence as an explicit assumption (Assumption L.1) and derives a bound internal to that stated quantity. No load-bearing self-citation chain or ansatz-smuggling via citation was found. Verdict: no significant circularity.

Assumptions & free parameters 8 free parameters · 4 assumptions · 3 invented entities

The central claim rests on standard ML design choices (LoRA rank, router top-K, reward weights) and on domain assumptions about the verifier and planner reliability. No new physical or conceptual entity is postulated outside the trainable components.

free parameters (8)

Number of subtask experts Ks = 4
Fixed in advance, not tuned on validation/test (Appendix J). The number of experts is a design choice that affects routing capacity.
Top-K subtask experts per call = 2
Router selects K=2 subtask experts at each agent call (Appendix E).
LoRA rank rho = 8
Rank of each LoRA adapter for q/v/o projections in last 8 blocks (Appendix E).
LoRA scaling alpha = 16
Scaling factor in LoRA composition (Appendix E).
Router projection dimension D = 128
Dimension of projected subtask embedding and prototypes (Appendix E).
Router gradient scale alpha_pi = 0.15
Relative weight of router gradient vs LoRA gradient (Appendix E).
Reward weights w_exec, w_comp, w_cons = 0.5, 0.5, 1.0
Fixed a priori from design intent; affect the training signal (Appendix E).
Within-group std skip threshold = 5e-4
Updates skipped when within-group reward std is below threshold (Appendix E).

assumptions (4)

domain assumption The frozen backbone LLM can express specialized behavior via LoRA adapters on the last 8 layers
MoLE attaches LoRA pairs to q/v/o projections; the diagnosis in Sec 2.1 and Thm L.2 assume the specialization directions align with activation subspaces.
domain assumption The rule-based verifier provides a valid reward signal for each subtask artifact
HGRPO optimizes step rewards {u_i}; if verifier rewards are noisy or misaligned with real subtask quality, credit assignment is corrupted (Sec 3.1, Appendix E).
domain assumption The LLM planner's DAG decomposition accurately represents task dependencies and subtask identities
The method's per-node labels and topology come from a prompt-based planner; incorrect decomposition propagates wrong credit (Sec 3.1).
standard math Standard statistical identities used in the variance reduction proof (score function, law of total variance, Ky Fan trace maximum)
Prop 3.1 and Thm L.2 rely on classical results; the proof sketch cites Greensmith et al. [7] and Ky Fan.

invented entities (3)

Role expert pool Phi_r (one LoRA expert per role)
purpose: Provides role-specific parameter specialization for each agent call
Trainable LoRA adapters internal to the method; existence is only supported by the paper's experiments.
Subtask expert pool Phi_s (shared LoRA experts selected by router)
purpose: Provides subtask-specific specialization across open-ended subtask space
Same as above; no external falsifiable handle.
Prototype-based semantic router pi_psi
purpose: Selects top-K subtask experts per agent call based on semantic similarity to learned prototypes
Trainable router; evidence is internal to the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MoRSE: Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts." pith.science (2026). https://pith.science/paper/NPNGCJ6M

@misc{pith2026260809251,
  author       = {Pith},
  title        = {Pith review of: MoRSE: Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPNGCJ6M}},
  note         = {Machine review of arXiv:2608.09251}
}

read the original abstract

Large language model-based multi-agent systems have recently shown strong potential for complex, long-horizon tasks. However, existing methods mainly rely on coarse prompt-level differentiation without parameter adaptation for diverse subtasks, resulting in insufficient inter-agent heterogeneity and limited specialized capability that bottleneck performance on tasks with complex requirements. To address this, we introduce a Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts (MoRSE) that distinguishes agents with (role, subtask)-conditional specialization at both the task structure and parameter levels. To make agents' responsibility explicit at the task structure level, we formulate a task-oriented multi-agent system that decomposes each task into a dependency-aware Directed Acyclic Graph of subtasks and assigns each agent a specific (role, subtask), introducing task-level specialization across collaborating agents. Additionally, to address the diverse role and subtask parameter adaptation demands, we propose a dynamic Mixture of (role, subtask) LoRA Experts module with a prototype-based semantic router for subtasks, augmenting agents with parameter-level specialization on a shared LLM substrate cost-effectively. Then, to co-optimize experts and router stably under sparse task rewards, we further propose a hierarchical group-relative policy optimization with two-layer credit assignment that isolates expert updates from the cross-route variance introduced by routing decisions, disentangling expert quality from routing quality. Experiments on code-generation benchmarks across three backbones demonstrate the effectiveness of our approach, with improvements in both whole-task and step-wise performance, and the gains from trained specialization generalize across held-out task categories and domains.

Figures

Figures reproduced from arXiv: 2608.09251 by the authors.

**Figure 1.** Heterogeneity deficit in role-prompted, graph-structured MAS. Each thin line is one SRDD instance, drawn from its prompt-level Sim (top row) to its output-level Sim (bottom row); bold lines show per-method median trajectories. Smaller Sim indicates lower redundancy. The legend reports the median shift ∆ = Simoutput − Simprompt: positive (+) means the pipeline amplifies agent-output redundancy, negative (-) means it … view at source ↗

**Figure 2.** Role and subtask mismatch in the base LM (Qwen3-4B, SRDD). (a) Pairwise rank-8 subspace divergence between two roles (execute and merge) per LoRA-injected layer; the dashed line shows the label-shuffled binary baseline. Higher = more divergent subspaces. (b) Same diagnostic with K=8 subtask clusters; shaded bands show label-shuffled baselines. (c) Heatmap of pairwise cluster divergence at layer 31 (representative); … view at source ↗

**Figure 3.** Overview of the MoRSE framework: (1) task-oriented MAS that decomposes [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Output Sim. Parameter-level heterogeneity gain from MoLE. Sec. 2.1’s prelim diagnosed a heterogeneity deficit at the prompt level (role-prompted MAS); the role-subtask DAG framework alone mitigates it to some extent [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: OOD generalization experiments for two backbones. Bold numbers above D are ∆ vs. A. Per-metric tables are in Appendix K [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: HGRPO training dynamics on Qwen3-4B (SRDD). Left: per-epoch GRPO loss (updated steps only, symlog scale). Right: step-pass rate per epoch. Dataset splits. For SRDD we sample from the 1,200 examples with a 3:2 train/test ratio for in-distribution evaluation. The out-of-…

**Figure 7.** Figure 7: SRDD subspace divergence across all 8 LoRA layers. Each sub-panel is a cluster-pair heatmap (1 − ∥V ⊤ a Vb∥ 2 F /ρ) at one layer; off-diagonal structure is nearly constant across panels. Subspace and divergence measure. For each layer L and group g, restrict A(L) to ro…

**Figure 8.** Figure 8: SciCode robustness. Repeating the SRDD diagnostic on SciCode: real-vs-shuffle gap 0.29 with non-overlapping ±σ-bands across layers (left), and the same off-diagonal cluster-pair structure at the representative layer (right). Assumption L.1 (Group subspace divergence). …

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 16 canonical work pages

[1]

Deepseekmoe: Towards ultimate expert specializa- tion in mixture-of-experts language models

Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Yu Wu, et al. Deepseekmoe: Towards ultimate expert specializa- tion in mixture-of-experts language models. InProceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers), pages 1280–1297, 2024

work page 2024
[2]

Loramoe: Alleviate world knowledge forgetting in large language models via moe-style plugin.arXiv preprint arXiv:2312.09979, 2023

Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. Loramoe: Alleviate world knowledge forgetting in large language models via moe-style plugin.arXiv preprint arXiv:2312.09979, 2023

arXiv 2023
[3]

Improving factuality and reasoning in language models through multiagent debate, 2023.URL https://arxiv

Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate, 2023.URL https://arxiv. org/abs/2305.14325, 3, 2023

arXiv 2023
[4]

Group-in-group policy optimization for LLM agent training

Lang Feng, Zhenghai Xue, Tingcong Liu, and Bo An. Group-in-group policy optimization for LLM agent training. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. URLhttps://openreview.net/forum?id=QXEhBMNrCW

2026
[5]

Mixture-of-loras: An efficient multitask tuning method for large language models

Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Yu Han, and Hao Wang. Mixture-of-loras: An efficient multitask tuning method for large language models. InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11371–11380, 2024

work page 2024
[6]

Higher layers need more lora experts

Chongyang Gao, Kezhen Chen, Jinmeng Rao, Baochen Sun, Ruibo Liu, Daiyi Peng, Yawen Zhang, Xiaoyuan Guo, Jie Yang, and VS Subrahmanian. Higher layers need more lora experts. arXiv preprint arXiv:2402.08562, 2024

arXiv 2024
[7]

Variance reduction techniques for gradient estimates in reinforcement learning.Journal of Machine Learning Research, 5(Nov): 1471–1530, 2004

Evan Greensmith, Peter L Bartlett, and Jonathan Baxter. Variance reduction techniques for gradient estimates in reinforcement learning.Journal of Machine Learning Research, 5(Nov): 1471–1530, 2004

work page 2004
[8]

Chawla, Olaf Wiest, and Xiangliang Zhang

Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V . Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges, 2024. URLhttps://arxiv.org/abs/2402.01680

arXiv 2024

Show all 48 references

[9]

Grassmann discriminant analysis: a unifying view on subspace- based learning

Jihun Hamm and Daniel D Lee. Grassmann discriminant analysis: a unifying view on subspace- based learning. InProceedings of the 25th international conference on Machine learning, pages 376–383, 2008

2008
[10]

Moragent: Parameter efficient agent tuning with mixture-of-roles.arXiv preprint arXiv:2512.21708, 2025

Jing Han, Binwei Yan, Tianyu Guo, Zheyuan Bai, Mengyu Zheng, Hanting Chen, and Ying Nie. Moragent: Parameter efficient agent tuning with mixture-of-roles.arXiv preprint arXiv:2512.21708, 2025

2025
[11]

Hierarchy-of-groups policy optimization for long-horizon agentic tasks.arXiv preprint arXiv:2602.22817, 2026

Shuo He, Lang Feng, Qi Wei, Xin Cheng, Lei Feng, and Bo An. Hierarchy-of-groups policy optimization for long-horizon agentic tasks.arXiv preprint arXiv:2602.22817, 2026

2026
[12]

Metagpt: Meta programming for a multi-agent collaborative framework

Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, et al. Metagpt: Meta programming for a multi-agent collaborative framework. InThe twelfth international conference on learning representa...

2023
[13]

Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. URL https://arxiv.org/abs/2106.09685

2021 arXiv
[14]

Moe-grpo: Optimizing mixture-of-experts via reinforcement learning in vision-language models.arXiv preprint arXiv:2603.24984, 2026

Dohwan Ko, Jinyoung Park, Seoung Choi, Sanghyeok Lee, Seohyun Lee, and Hyunwoo J Kim. Moe-grpo: Optimizing mixture-of-experts via reinforcement learning in vision-language models.arXiv preprint arXiv:2603.24984, 2026

2026
[15]

Similarity of neural network representations revisited

Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. InInternational conference on machine learning, pages 3519–3529. PMlR, 2019

2019
[16]

Execution-grounded credit assignment for grpo in code generation.arXiv preprint arXiv:2603.16158, 2026

Abhijit Kumar, Natalya Kumar, and Shikhar Gupta. Execution-grounded credit assignment for grpo in code generation.arXiv preprint arXiv:2603.16158, 2026

2026
[17]

Hierarchical mixture of experts: Generalizable learning for high-level synthesis

Weikai Li, Ding Wang, Zijian Ding, Atefeh Sohrabizadeh, Zongyue Qin, Jason Cong, and Yizhou Sun. Hierarchical mixture of experts: Generalizable learning for high-level synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18476–18484, 2025

2025
[18]

Beyond entangled planning: Task-decoupled planning for long-horizon agents.arXiv preprint arXiv:2601.07577, 2026

Yunfan Li, Bingbing Xu, Xueyun Tian, Xiucheng Xu, and Huawei Shen. Beyond entangled planning: Task-decoupled planning for long-horizon agents.arXiv preprint arXiv:2601.07577, 2026

2026
[19]

Marft: Multi-agent reinforcement fine-tuning.arXiv preprint arXiv:2504.16129, 2025

Junwei Liao, Muning Wen, Jun Wang, and Weinan Zhang. Marft: Multi-agent reinforcement fine-tuning.arXiv preprint arXiv:2504.16129, 2025

2025 arXiv
[20]

Balancing the experts: Unlocking loRA-moe for GRPO via mechanism-aware rewards

Changlian Ma, Zizheng Huang, Xiangyu Zeng, Yi Wang, Cheng Liang, Kun Tian, Xinhai Zhao, and Limin Wang. Balancing the experts: Unlocking loRA-moe for GRPO via mechanism-aware rewards. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://open...

2026
[21]

Stabilizing moe reinforcement learning by aligning training and inference routers.arXiv preprint arXiv:2510.11370, 2025

Wenhan Ma, Hailin Zhang, Liang Zhao, Yifan Song, Yudong Wang, Zhifang Sui, and Fuli Luo. Stabilizing moe reinforcement learning by aligning training and inference routers.arXiv preprint arXiv:2510.11370, 2025

2025
[22]

Locating and editing factual associations in gpt.Advances in neural information processing systems, 35:17359–17372, 2022

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt.Advances in neural information processing systems, 35:17359–17372, 2022

2022
[23]

Grpo-λ: Credit assignment improves llm reasoning.arXiv preprint arXiv:2510.00194, 2025

Prasanna Parthasarathi, Mathieu Reymond, Boxing Chen, Yufei Cui, and Sarath Chandar. Grpo-λ: Credit assignment improves llm reasoning.arXiv preprint arXiv:2510.00194, 2025

2025
[24]

Chatdev: Communicative agents for software development, 2024

Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. Chatdev: Communicative agents for software development, 2024. URL https://arxiv.org/abs/ 2307.07924

2024 arXiv
[25]

Scaling large language model-based multi-agent collabora- tion.arXiv preprint arXiv:2406.07155, 2024

Chen Qian, Zihao Xie, Yifei Wang, Wei Liu, Kunlun Zhu, Hanchen Xia, Yufan Dang, Zhuoyun Du, Weize Chen, Cheng Yang, et al. Scaling large language model-based multi-agent collabora- tion.arXiv preprint arXiv:2406.07155, 2024

2024 arXiv
[26]

Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability.Advances in neural information processing systems, 30, 2017

Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability.Advances in neural information processing systems, 30, 2017

2017
[27]

Proximal policy optimization algorithms, 2017

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. URLhttps://arxiv.org/abs/1707.06347

2017 arXiv
[28]

Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

2024 arXiv
[29]

Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017

Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017. URLhttps://arxiv.org/abs/1701.06538. 11

2017 arXiv
[30]

Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning,
[31]

Bert rediscovers the classical nlp pipeline

Ian Tenney, Dipanjan Das, and Ellie Pavlick. Bert rediscovers the classical nlp pipeline. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 4593–4601, 2019

2019
[32]

Scicode: A research coding benchmark curated by scientists.Advances in Neural Information Processing Systems, 37:30624–30650, 2024

Minyang Tian, Luyu Gao, Shizhuo D Zhang, Xinan Chen, Cunwei Fan, Xuefei Guo, Roland Haas, Pan Ji, Kittithat Krongchon, Yao Li, et al. Scicode: A research coding benchmark curated by scientists.Advances in Neural Information Processing Systems, 37:30624–30650, 2024

2024
[33]

Parameter-efficient sparsity crafting from dense to mixture-of-experts for instruction tuning on general tasks

Haoyuan Wu, Haisheng Zheng, Zhuolun He, and Bei Yu. Parameter-efficient sparsity crafting from dense to mixture-of-experts for instruction tuning on general tasks. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 737–749, 2024

2024
[34]

Gap: Graph-based agent planning with parallel tool use and reinforcement learning.arXiv preprint arXiv:2510.25320, 2025

Jiaqi Wu, Qinlao Zhao, Zefeng Chen, Kai Qin, Yifei Zhao, Xueqian Wang, and Yuhang Yao. Gap: Graph-based agent planning with parallel tool use and reinforcement learning.arXiv preprint arXiv:2510.25320, 2025

2025
[35]

Autogen: Enabling next-gen llm applications via multi-agent conversation framework.arXiv preprint arXiv:2308.08155, 3(4), 2023

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework.arXiv preprint arXiv:2308.08155, 3(4), 2023

2023 arXiv
[36]

Phase- aware mixture of experts for agentic reinforcement learning.arXiv preprint arXiv:2602.17038, 2026

Shengtian Yang, Yu Li, Shuo He, Yewen Li, Qingpeng Cai, Peng Jiang, and Lei Feng. Phase- aware mixture of experts for agentic reinforcement learning.arXiv preprint arXiv:2602.17038, 2026

2026 arXiv
[37]

Understanding agent scaling in llm-based multi-agent systems via diversity.arXiv preprint arXiv:2602.03794, 2026

Yingxuan Yang, Chengrui Qu, Muning Wen, Laixi Shi, Ying Wen, Weinan Zhang, Adam Wierman, and Shangding Gu. Understanding agent scaling in llm-based multi-agent systems via diversity.arXiv preprint arXiv:2602.03794, 2026

2026
[38]

The expressive power of low-rank adaptation

Yuchen Zeng and Kangwook Lee. The expressive power of low-rank adaptation. InInternational Conference on Learning Representations, volume 2024, pages 5078–5123, 2024

2024
[39]

Towards stable and effective reinforcement learning for mixture-of-experts

Di Zhang, Xun Wu, Shaohan Huang, Lingjie Jiang, Yaru Hao, Li Dong, Zewen Chi, Zhifang Sui, and Furu Wei. Towards stable and effective reinforcement learning for mixture-of-experts. arXiv preprint arXiv:2510.23027, 2025

2025
[40]

Aflow: Automating agentic workflow generation

Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xiong-Hui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, et al. Aflow: Automating agentic workflow generation. InThe Thirteenth International Conference on Learning Representations, 2024

2024
[41]

MARTI: A framework for multi-agent LLM systems reinforced training and inference

Kaiyan Zhang, Kai Tian, Runze Liu, Sihang Zeng, Xuekai Zhu, Guoli Jia, Yuchen Fan, Xingtai Lv, Yuxin Zuo, Che Jiang, Yuru wang, Jianyu Wang, Ermo Hua, Xinwei Long, Junqi Gao, Youbang Sun, Zhiyuan Ma, Ganqu Cui, Ning Ding, Biqing Qi, and Bowen Zhou. MARTI: A framework for multi...

2026
[42]

Stackplanner: A centralized hierarchical multi- agent system with task-experience memory management.arXiv preprint arXiv:2601.05890, 2026

Ruizhe Zhang, Xinke Jiang, Zhibang Yang, Zhixin Zhang, Jiaran Gao, Yuzhen Xiao, Hongbin Lai, Xu Chu, Junfeng Zhao, and Yasha Wang. Stackplanner: A centralized hierarchical multi- agent system with task-experience memory management.arXiv preprint arXiv:2601.05890, 2026

2026 arXiv
[43]

Stronger-mas: Multi-agent reinforcement learning for collaborative llms.arXiv preprint arXiv:2510.11062, 2025

Yujie Zhao, Lanxiang Hu, Yang Wang, Minmin Hou, Hao Zhang, Ke Ding, and Jishen Zhao. Stronger-mas: Multi-agent reinforcement learning for collaborative llms.arXiv preprint arXiv:2510.11062, 2025. 12 A Notation We summarize the main notation used throughout Sections 2–3. Symbol...

2025
[45]

MoRSE w/o MoLE & HGRPO (MoRSEbase) untrained reference 90.00 0.765 0.370 0.00 0.000 0.000 B

MacNet prior MAS baseline 66.25 0.616 0.186 −23.75−0.149−0.184 A. MoRSE w/o MoLE & HGRPO (MoRSEbase) untrained reference 90.00 0.765 0.370 0.00 0.000 0.000 B. MoRSE w/o MoLE standard-training baseline88.75 0.765 0.374 −1.25 0.000 +0.004 D.MoRSE ours 96.25 0.783 0.392 +6.25 +0....
[46]

MoRSE w/o MoLE & HGRPO (MoRSEbase) untrained reference 72.50 0.676 0.242 0.00 0.000 0.000 B

MacNet prior MAS baseline 71.25 0.630 0.184 −1.25−0.046−0.058 A. MoRSE w/o MoLE & HGRPO (MoRSEbase) untrained reference 72.50 0.676 0.242 0.00 0.000 0.000 B. MoRSE w/o MoLE standard-training baseline76.25 0.702 0.285 +3.75 +0.026 +0.043 D.MoRSE ours 80.00 0.722 0.310 +7.50 +0....
[47]

MoRSE w/o MoLE & HGRPO (MoRSEbase) untrained reference 17.50 20.07 0.00 0.00 0.00 0.00 B

MacNet prior MAS baseline 8.75 9.79 0.00 −8.75−10.28 0.00 A. MoRSE w/o MoLE & HGRPO (MoRSEbase) untrained reference 17.50 20.07 0.00 0.00 0.00 0.00 B. MoRSE w/o MoLE standard-training baseline22.22 16.67 0.00 +4.72−3.40 0.00 D.MoRSE ours 24.69 19.44 0.00 +7.19 −0.63 0.00 Llama...
[48]

MoRSE w/o MoLE & HGRPO (MoRSEbase) untrained reference 7.50 9.06 0.00 0.00 0.00 0.00 B

MacNet prior MAS baseline 8.75 9.06 0.00 +1.25 0.00 0.00 A. MoRSE w/o MoLE & HGRPO (MoRSEbase) untrained reference 7.50 9.06 0.00 0.00 0.00 0.00 B. MoRSE w/o MoLE standard-training baseline16.05 9.26 0.00 +8.55 +0.20 0.00 D.MoRSE ours 18.52 9.57 0.00 +11.02 +0.51 0.00 L Role-S...
[2017]

URLhttps://arxiv.org/abs/1703.05175

arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Deepseekmoe: Towards ultimate expert specializa- tion in mixture-of-experts language models

Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Yu Wu, et al. Deepseekmoe: Towards ultimate expert specializa- tion in mixture-of-experts language models. InProceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers), pages 1280–1297, 2024

work page 2024

[2] [2]

Loramoe: Alleviate world knowledge forgetting in large language models via moe-style plugin.arXiv preprint arXiv:2312.09979, 2023

Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. Loramoe: Alleviate world knowledge forgetting in large language models via moe-style plugin.arXiv preprint arXiv:2312.09979, 2023

arXiv 2023

[3] [3]

Improving factuality and reasoning in language models through multiagent debate, 2023.URL https://arxiv

Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate, 2023.URL https://arxiv. org/abs/2305.14325, 3, 2023

arXiv 2023

[4] [4]

Group-in-group policy optimization for LLM agent training

Lang Feng, Zhenghai Xue, Tingcong Liu, and Bo An. Group-in-group policy optimization for LLM agent training. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. URLhttps://openreview.net/forum?id=QXEhBMNrCW

2026

[5] [5]

Mixture-of-loras: An efficient multitask tuning method for large language models

Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Yu Han, and Hao Wang. Mixture-of-loras: An efficient multitask tuning method for large language models. InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11371–11380, 2024

work page 2024

[6] [6]

Higher layers need more lora experts

Chongyang Gao, Kezhen Chen, Jinmeng Rao, Baochen Sun, Ruibo Liu, Daiyi Peng, Yawen Zhang, Xiaoyuan Guo, Jie Yang, and VS Subrahmanian. Higher layers need more lora experts. arXiv preprint arXiv:2402.08562, 2024

arXiv 2024

[7] [7]

Variance reduction techniques for gradient estimates in reinforcement learning.Journal of Machine Learning Research, 5(Nov): 1471–1530, 2004

Evan Greensmith, Peter L Bartlett, and Jonathan Baxter. Variance reduction techniques for gradient estimates in reinforcement learning.Journal of Machine Learning Research, 5(Nov): 1471–1530, 2004

work page 2004

[8] [8]

Chawla, Olaf Wiest, and Xiangliang Zhang

Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V . Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges, 2024. URLhttps://arxiv.org/abs/2402.01680

arXiv 2024