Fudan University  ·  Shanghai Innovation Institute  ·  arXiv:2601.07309

Many expert agents, one generalist — with no training.

ARM — role-conditioned neuron transplantation that merges benchmark-specialist agents into a single deployable model, without any gradient optimization

ARM·Agent-Role Merging — specialists in, one generalist out

Keep every expert’s edge. Pay no training cost.

Zhuoka Feng∗  ·  Kang Chen∗  ·  Sihan Zhao∗∗  ·  Kai Xiong∗∗  ·  Yaoning Wang∗∗  ·  Minshen Yu  ·  Junjie Nian  ·  Changyi Xiao  ·  Yixin Cao†  ·  Yugang Jiang

Watch the film The idea Read the paper

The one idea

Specialists don’t transfer. Merge them — and keep each one’s edge.

Fine-tuned LLM agents are strong inside the environment they were trained on and brittle everywhere else: change the tool schema, the action interface, or the trajectory distribution and accuracy drops. Retraining one model across every environment is expensive and finicky. Model merging is the training-free alternative — fold several same-architecture experts into one model — but on multi-turn agents naïve merges are unstable, and small errors on role-critical spans (a malformed tool call, a broken action) cascade into long-horizon failure.

Bars comparing several training-free merge baselines across interactive agent benchmarks. No single operator is strong everywhere: each recipe that wins on one suite collapses on another, showing pronounced benchmark-dependent variance.
Fig. 1Training-free merge baselines are not uniformly strong across interactive agent environments — each recipe has a suite where it collapses. This motivates selecting a stable backbone before any fine-grained repair.
01

Specialists are brittle

An agent tuned for one benchmark degrades sharply on another. Maintaining a separate expert per environment — or retraining one model over all of them — is costly to build and to serve.

02

Merging is cheap but unstable

Weight-space merges need no training, yet different operators trade suites off against each other, and interference on role-critical tokens triggers cascading failures in multi-turn rollouts.

→

Fix it at the neuron level

Choose a stable backbone with an answer-free signal, then repair the weakest suite by grafting a few role-salient neurons — while protecting the neurons every other suite depends on.

How ARM works

Build backbones. Select by activation overlap. Transplant role-salient neurons.

ARM is a three-stage, gradient-free pipeline. It never optimizes weights end-to-end and never reads answer strings to choose a model: selection is driven purely by where each expert’s role-critical neurons fire.

The three-stage ARM pipeline. Stage 1: apply multiple training-free weight-space merge operators to benchmark-specialized experts to produce a pool of candidate merged backbones. Stage 2: a selector computes the Activation-Overlap Score from role-conditioned MLP activations on a small calibration set and picks the candidate that maximizes mean AOS. Stage 3: for the dev-diagnosed weak benchmark, transplant a small top-k percent of role-salient donor MLP neurons into the backbone while protecting neurons salient for other benchmarks.
Fig. 2Overview of Agent-Role Merging. (1) merge operators build a pool of candidate backbones; (2) the Activation-Overlap Score selects one, answer-free; (3) conflict-aware transplantation grafts role-salient donor neurons and protects the rest.
  1. pool

    Backbone pool construction

    Apply a set of standard training-free weight-space operators — uniform averaging, task arithmetic, TIES — to the same-architecture experts, producing a small pool of candidate merged backbones instead of committing to one recipe.

  2. AOS

    Backbone selection — answer-free

    On a lightweight calibration set, trace role-conditioned MLP activations and measure how well each candidate preserves an expert’s role-salient neurons. Pick the backbone that maximizes mean Activation-Overlap Score — no interactive evaluation, no answer strings.

    AOS(b) = |Sb ∩ Sbb||Sb|
  3. graft

    Conflict-aware neuron transplantation

    For the dev-diagnosed weak suite, transplant a small top-k% of the donor expert’s role-salient MLP neurons into the backbone — but only those not salient for any other benchmark. Protecting shared neurons keeps repair local and avoids negative transfer.

Every stage is training-free: no gradients, no router edits, no changes to quantization or decoding. ARM edits a single merged checkpoint at the neuron level. Selection is answer-string-free — it reads activation geometry, not outputs.

Why it works

Role-conditioned tracing localizes each suite — so repair stays surgical.

The transplant only helps if a suite’s salient neurons are actually distinct. ARM’s analysis shows that conditioning the trace on role-critical spans yields more benchmark-specific neuron sets than un-conditioned tracing — the sets overlap less, so a graft can lift the weak suite without disturbing the others.

①

AOS tracks capability

The answer-free overlap score correlates strongly with downstream accuracy across candidates (Pearson r≈0.84, Spearman ρ≈0.99) — enough signal to rank backbones without running the interactive benchmarks.

②

Role-conditioned = more specific

Conditioning activation tracing on role-critical tokens produces sharper, less-overlapping salient sets per benchmark — the precondition for conflict-aware, localized repair.

→

A few neurons, chosen well

Only a small top-k% of donor neurons are grafted. Sensitivity analysis shows a broad stable range — the method is not balanced on a knife-edge.

Scatter relating each candidate backbone's answer-free Activation-Overlap Score to its measured downstream accuracy; higher AOS trends with higher accuracy, supporting AOS as an answer-free selector.
Fig. 3The answer-free Activation-Overlap Score tracks downstream capability across candidate backbones — near-perfect rank correlation (Spearman ρ = 0.99) — so ARM can screen merges without any interactive evaluation.

The payoff

One checkpoint that’s best on average — and beats the oracle selector.

ARM merges three Simia benchmark-specialist experts into a single model and is evaluated across six aggregates — four in-domain suites (τ-bench, OfficeBench, WebShop, OS) and two held-out out-of-domain benchmarks (DB-bench, AlfWorld) — on both Qwen3-8B and Qwen2.5-7B expert pools.

44.6 / 45.7best-average merged generalist · Qwen3-8B / Qwen2.5-7B
+10.9%over the oracle expert selector · Qwen2.5-7B pool
6aggregates · 4 in-domain + 2 out-of-domain

Across both expert pools ARM is the strongest single merged model — average 44.6 (Qwen3-8B) and 45.7 (Qwen2.5-7B) — beating every training-free merge baseline and even the BEST-of-Three oracle expert selector (44.2 / 41.2), while preserving out-of-domain generalization to DB-bench and AlfWorld.

Best average, both families

On Qwen3-8B, ARM reaches 44.6 mean over six aggregates — above every weight-space and activation-aware baseline, and above the oracle expert selector (44.2). On Qwen2.5-7B it reaches 45.7 (+10.9% over the oracle).

Worst-suite robustness, kept OOD

The gains are balanced rather than lopsided: ARM lifts the weakest suites instead of trading one for another, and it holds out-of-domain performance on DB-bench and AlfWorld — the benchmarks no expert was trained on.

Sensitivity of ARM to the transplant fraction top-k percent on the two expert pools; performance is stable across a broad range of k rather than peaking at a single fragile setting.
Fig. 4Transplant-fraction sensitivity. ARM is stable across a broad range of top-k% — the repair does not hinge on a single fragile hyperparameter.

Targeted neuron edits, not retraining

ARM never runs gradient descent and never votes over answer strings. It screens merges by activation overlap and repairs the weakest suite by grafting a handful of role-salient neurons — evidence that role-critical circuits are a practical, training-free lever for robust multi-turn agent merging.

Built from 3 Simia experts · 2 model families · 6 aggregates Qwen3-8B · Qwen2.5-7B expert pools · in-domain: τ-bench · OfficeBench · WebShop · OS (AgentBench) · out-of-domain: DB-bench · AlfWorld

Watch

Many specialists, one generalist — a visual essay

A silent vector essay, drawn live in your browser — no video file. Three specialists become one generalist: build candidate backbones, select by activation overlap, transplant a few role-salient neurons, protect the rest.

Cite

Reference

Zhuoka Feng1,∗, Kang Chen1,∗, Sihan Zhao1,∗∗, Kai Xiong∗∗, Yaoning Wang1,∗∗, Minshen Yu1, Junjie Nian1, Changyi Xiao1, Yixin Cao1,2,†, Yugang Jiang1

1 Fudan University    2 Shanghai Innovation Institute    ∗ Co-first authors    ∗∗ Co-second authors    † Corresponding author

Correspondence: yxcao@fudan.edu.cn  ·  first-author contact 23307130211@m.fudan.edu.cn

Links: Paper (PDF)  ·  arXiv  ·  Project page

@misc{feng2026arm,
  title   = {ARM: Role-Conditioned Neuron Transplantation for
             Training-Free Generalist LLM Agent Merging},
  author  = {Feng, Zhuoka and Chen, Kang and Zhao, Sihan and Xiong, Kai and
             Wang, Yaoning and Yu, Minshen and Nian, Junjie and Xiao, Changyi and
             Cao, Yixin and Jiang, Yugang},
  year    = {2026},
  eprint  = {2601.07309},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  note    = {ARM (Agent-Role Merging)},
  url     = {https://arxiv.org/abs/2601.07309},
}

Many specialists. One generalist — merged, not trained.