Activation Transplantation in Neural Models

• Activation transplantation is a method that leverages internal activation statistics to localize and transfer behavior between model configurations.
• The ARM approach uses role-conditioned neuron selection and conflict-aware transplantation to merge expert capabilities without updating the model during inference.
• ConTrans refines and remaps concept vectors from a source model to steer the target model’s behavior at runtime, improving alignment on tasks like truthfulness and toxicity.

Searching arXiv for the cited and closely related papers. Activation transplantation denotes a class of representation-engineering interventions in which activation statistics or activation-derived directions are used to transfer behavior from one model configuration to another. In current usage, the term covers at least two distinct mechanisms. In "ARM: Role-Conditioned Neuron Transplantation for Training-Free Generalist LLM Agent Merging" (Feng et al., 12 Jan 2026), it refers to activation-guided, role-conditioned selection and conflict-aware transplantation of MLP neurons from benchmark-specific expert agents into a merged backbone. In "ConTrans: Weak-to-Strong Alignment Engineering via Concept Transplantation" (Dong et al., 2024), it refers to refining a concept vector in a source LLM, reformulating it into a target model’s feature space, and adding that direction to the target residual stream at inference time. This suggests that activation transplantation is best understood as a family of methods that exploit internal representations to localize and transfer behaviorally salient structure.

1. Terminological scope and conceptual distinctions

The term has a narrower meaning in ARM than in ConTrans. ARM does not substitute activations at inference time; it uses activations measured on a calibration set to identify role-salient neurons and then edits the backbone’s parameters at those neurons, specifically rows, columns, and biases in the MLP projections. The intervention is therefore a static parameter-space transplantation rather than runtime manipulation of activations (Feng et al., 12 Jan 2026).

ConTrans uses the term more literally at the representation level. It extracts layerwise concept vectors from the hidden states of a source model, adapts those vectors to the target model via an affine map, and transplants them by adding the reformulated concept vector into the target residual stream during inference. No parameter updates of the target are required (Dong et al., 2024).

A common misconception is to treat these procedures as variants of the same operator. They are related by their reliance on internal activations, but they differ in intervention site, timing, and objective. ARM is an offline, training-free model-merging repair method for multi-turn agents; ConTrans is an inference-time steering method for weak-to-strong alignment transfer. Another misconception is to conflate ARM with activation substitution. ARM explicitly distinguishes itself from dynamic activation injection and frames its method as closer to representation surgery: trace signals to find the circuits that matter, then copy over the corresponding MLP channels from a donor to repair specific capabilities (Feng et al., 12 Jan 2026).

2. Role-conditioned neuron transplantation in ARM

ARM is organized as a 3-step framework: constructing merged backbones, selection based on role-conditioned activation analysis, and neuron transplantation for fine-grained refinements. Given expert checkpoints $M^{\mathrm{exp}}_{b_i}$ that share architecture and tokenizer, ARM first constructs a small pool of candidate merged backbones using training-free weight-merging operators $G$, including uniform averaging or model soups, task arithmetic, and TIES. For each operator $g \in G$,

$$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$

The goal is to choose an initialization that preserves role-critical behavior before any local repair (Feng et al., 12 Jan 2026).

Role conditioning is central. In multi-turn agents, roles are benchmark-critical output spans that must satisfy deterministic schemas or tool interfaces: tool-call spans in $\tau$-bench, final-answer JSON schema spans in OfficeBench, and action schema and argument spans in AgentBench OS and WebShop. ARM builds a disjoint calibration set of 699 tasks and 1240 trajectories for forward-pass activation tracing only. For each model $M$, layer $\ell$, and neuron $j$, role-conditioned saliency is defined as the expected mean absolute activation over role tokens,

$$s_{\ell,j}(M; b_i, r) = \mathbb{E}_{x \sim D(b_i)} \left[ \mathrm{mean}_{t \in T^{(b_i,r)}(x)} |z_{\ell,j}(t)| \right].$$

Top-$k$ selection per layer yields the role-salient set $G$0, and these sets are aggregated across layers (Feng et al., 12 Jan 2026).

Backbone selection uses the Activation-Overlap Score, a Jaccard index between the role-salient neuron sets of a candidate backbone and its corresponding expert:

$G$1

ARM chooses the operator

$G$2

and sets $G$3. The stated purpose is to preserve experts’ role-salient circuits and mitigate the high variability of weight merges in agent settings (Feng et al., 12 Jan 2026).

The final step is conflict-aware neuron transplantation. ARM identifies weak benchmarks $G$4 on a dev set where the selected backbone underperforms the donor expert. For a benchmark $G$5, it defines protected backbone neurons salient for other benchmarks,

$G$6

and transplants only donor neurons that do not collide with this protected set:

$G$7

For an MLP neuron $G$8, ARM writes donor parameters into the backbone by copying row $G$9 of $g \in G$0, entry $g \in G$1 of $g \in G$2, and column $g \in G$3 of $g \in G$4; for gated MLPs such as SwiGLU, the same index is used consistently across gate or up projections and the down projection. No attention heads or layer norms are transplanted; edits are confined to MLP neurons (Feng et al., 12 Jan 2026).

3. Concept transplantation in ConTrans

ConTrans presents activation transplantation as concept transplantation for weak-to-strong alignment engineering. The method transfers an aligned concept encoded in the hidden states of a smaller, weak-yet-aligned source model into a larger, strong-yet-unaligned target model. It proceeds in three steps: concept refinement in the source, concept reformulation via an affine map into the target feature space, and concept transplantation into the target residual stream (Dong et al., 2024).

Concept refinement uses paired positive and negative examples of a target concept. For a source model $g \in G$5 with layerwise hidden states $g \in G$6, ConTrans uses the last-token hidden state to represent the sequence. The per-layer concept vector is the mean-difference direction

$g \in G$7

The paper states that this contrast-based refinement cancels out low-level style or syntax while preserving the semantic direction of the concept. Although alternatives such as linear probes, PCA components, or DAS rotations are mentioned, ConTrans itself uses the mean-difference method (Dong et al., 2024).

Because source and target models can have different hidden dimensions and geometries, ConTrans learns a layerwise linear projection $g \in G$8 from paired hidden states collected on the same text in both models. The affine mapping is written as $g \in G$9 with no bias, and the reformulated concept is

$$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$0

The map is estimated by least squares,

$$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$1

with closed-form solution

$$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$2

or equivalently through an SVD-based expression for numerical stability. ConTrans trains these maps using base-model hidden states on 2,000 Wikipedia-derived samples from WikiSplit (Dong et al., 2024).

Transplantation itself is an additive write to the target residual stream. In a pre-norm transformer block, ConTrans injects the reformulated direction as

$$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$3

where $$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$4 is a scalar steering strength. The vector can be applied at the last-token position, matching extraction, or broadcast across positions; the paper uses the hidden-state tensor $$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$5 and adds $$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$6. The method focuses on one concept at a time, and multi-concept superposition is identified as an open question rather than a resolved capability (Dong et al., 2024).

4. Intervention sites, conditioning signals, and representational assumptions

The two methods differ most sharply in where the transplant lands and what signal defines saliency. ARM uses parser-detected role spans as conditioning variables. Saliency is measured only on tokens that directly express role-critical structures such as tool calls, JSON outputs, or action schemas. The stated effect is to reduce cross-benchmark entanglement so that fewer neurons appear salient for every task, enabling clean and localized edits (Feng et al., 12 Jan 2026).

ConTrans instead defines saliency through concept polarity. Positive and negative examples instantiate opposites such as honest versus dishonest prompts, non-toxic versus toxic prompts, or scenarios corresponding to one emotion versus the others. Its conditioning signal is therefore semantic rather than schema-based. The transplant does not identify a sparse donor channel set but rather a direction in hidden-state space that is mapped into the target and added during inference (Dong et al., 2024).

Method	Activations used for	Actual intervention
ARM	Identify role-salient neurons on calibration trajectories and score merged backbones via AOS	Write donor MLP parameters into selected backbone neurons
ConTrans	Refine layerwise concept vectors and fit source-to-target affine maps from paired hidden states	Add $$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$7 to the target residual stream at inference

ARM assumes homologous experts with the same architecture and tokenizer, and heterogeneous model families are explicitly out of scope. ConTrans allows $$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$8 and supports inter-family transfers through the learned affine map, but its reported gains are strongest when source and target share architecture and pretraining scale; more distant bases yield weaker benefits (Feng et al., 12 Jan 2026, Dong et al., 2024).

These differences imply distinct representational assumptions. ARM assumes that task-critical circuits can be isolated as role-salient MLP channels and repaired by copying donor parameters. ConTrans assumes that aligned concepts are directions present in a shared low-rank subspace across models and can be activated by residual-stream steering. This suggests that the two techniques occupy complementary positions within representation engineering rather than competing on a single axis.

5. Empirical behavior and reported effects

ARM is evaluated on in-domain benchmarks including $$M^{(0,g)} = g(\{M^{\mathrm{exp}}_{b_i}\}_{i=1..N}).$$9-bench, OfficeBench, and AgentBench, with DB-bench and AlfWorld as out-of-domain tests. On the Qwen3-8B expert pool, ARM achieves Avg 44.6, exceeding the BEST-of-Three oracle at 44.2 and outperforming training-free baselines such as AIM at 39.5 and TIES at 36.2. The same setting reports DB-bench 47.7 and AlfWorld 48.0, with AlfWorld the best among the listed baselines. On the Qwen2.5-7B expert pool, ARM achieves Avg 45.7 versus BEST-of-Three at 41.2 and AIM at 36.9, with strong gains on AlfWorld at 68.0 and DB-bench at 51.3. Generalization summaries improve from WS 19.1 to 28.5 and RHM 0.84 to 0.98 for Qwen3-8B, and from WS 16.4 to 22.0 and RHM 0.95 to 1.04 for Qwen2.5-7B (Feng et al., 12 Jan 2026).

The paper also reports mechanistic and efficiency findings. Role conditioning reduces cross-benchmark overlap of top-10% salient neurons from 61% to 41% for Qwen3-8B and from 50% to 43% for Qwen2.5-7B. Conflict-aware protection stabilizes performance when sweeping $\tau$0, whereas unprotected variants degrade faster as $\tau$1 grows. Forward-pass activation tracing for backbone selection costs about 0.5 GPU-hour per backbone-benchmark pair on a single H20, about 12 GPU-hours total for six candidates across four benchmarks; transplantation runs in under 20 minutes; transplant sets are typically about 2–3% of MLP neurons per repaired benchmark; and activation statistics storage is under 500MB compressed. Reported repaired failures include reducing invalid final JSON episodes in OfficeBench from 8.5% to 4.7%, avoiding error loops after tool failures in $\tau$2-bench, and correcting subtle command choices such as echo -n versus echo in OS tasks (Feng et al., 12 Jan 2026).

ConTrans is evaluated on emotions, truthfulness, and toxicity. For truthfulness on TruthfulQA MC1, base models begin around 17–22% accuracy, and transplanting $\tau$3 refined from a 7B instruct model improves base-model accuracy by about 12.8% for 7B, 15.3% for 13B, and 13.3% for 70B on average across LLaMA, LLaMA 2, and Code LLaMA families. Many target base models surpass their corresponding-size instruction-tuned models in MC1 accuracy after ConTrans, including reported 13B and 70B cases. On ToxiGen, base-model toxicity rates of about 80–92% drop markedly when transplanting $\tau$4 from a 7B instruct source; average toxic percentages fall below about 47% across families. Emotion transplantation from LLaMA-7B reliably biases LLaMA-13B and LLaMA-65B toward the target emotion on negative scenarios, with accuracy increasing as $\tau$5 grows (Dong et al., 2024).

Several sensitivity results further characterize the method. As few as 20 positive or negative pairs already yield strong emotion results; emotions and truthfulness typically use $\tau$6; toxicity tunes $\tau$7 by grid search in $\tau$8 to avoid incoherence. Analyses on the Pythia series indicate that emotion transplantation becomes effective only once emotion-relevant geometry emerges at roughly 410M parameters and above, while Amber-7B checkpoints show stronger transplantation benefits as pretraining data increases. The paper interprets these trends as support for the claim that ConTrans steers pre-existing conceptual directions rather than creating new ones (Dong et al., 2024).

6. Limitations, safety considerations, and unresolved questions

ARM’s limitations follow from its merging setting and its reliance on parser-defined roles. It requires homologous experts with the same architecture and tokenizer. If experts are highly dissimilar, AOS may be low, backbone selection becomes harder, and transplant overlap may grow; the stated mitigations are to keep $\tau$9 small, strengthen conflict-aware protection, or expand the backbone pool $M$0. Role leakage is another failure mode: if role spans are misidentified by parsers, saliency estimates may degrade. The paper also notes safety concerns, since transplanting tool-use behaviors can propagate brittle formatting assumptions; recommended safeguards include strict validators, schema checks, and controlled temperature during evaluation (Feng et al., 12 Jan 2026).

ConTrans’s limitations concern superposition, steering strength, and model mismatch. The method is presented as a single-concept intervention; multi-concept superposition remains insufficiently characterized. Large $M$1 can drive toxicity near zero but harm coherence, so validation-based tuning and perplexity checks on OpenWebText are used to balance safety and fluency. Transfers are strongest when source and target share architecture and data scale, and weaker when source and target bases are more distant. The paper also explicitly distinguishes value alignment from capability transfer: ConTrans targets concepts such as honesty, fairness, and emotions, and it reports no evidence of capability gains from a single direction edit (Dong et al., 2024).

A broader conceptual issue is that the phrase activation transplantation can obscure the difference between static and dynamic interventions. In ARM, activations are an analysis signal for selecting sparse MLP parameter edits; in ConTrans, activations are the object being directly written into the residual stream. This distinction matters for runtime cost, stability, and interpretability. ARM emphasizes static, lightweight, training-free edits without runtime overhead, whereas ConTrans emphasizes transparent, parameter-free steering with inference-time control. The shared idea is not a common implementation, but the use of internal representations as a locus for targeted transfer.

7. Relation to neighboring methodologies

Both papers position activation transplantation relative to adjacent areas of model editing and control. ARM connects its conflict-aware top-$M$2 selection to mixture-of-experts and gating, describing the method as analogous to selecting a sparse expert subnetwork per role and using protection to avoid overwrites of neurons used elsewhere. It also places backbone selection in relation to linear mode connectivity and task vectors: rather than relying only on global linear combinations that may cause mode hopping, ARM first chooses a backbone in a favorable region of connectivity and then applies local sparse edits. Its parameter-localized MLP transplantation is explicitly compared to representation surgery, with ROME and MEMIT offered as reference points for the idea that small localized edits can adjust specific behaviors while preserving global performance (Feng et al., 12 Jan 2026).

ConTrans is compared to activation addition or steering, editing methods such as ROME and MEMIT, and training-based alignment approaches including LoRA, fine-tuning, DPO, RLHF, and RLAIF. The paper’s specific claim is that ConTrans extends same-model activation addition to cross-model transfer through the learned affine map, enabling weak-to-strong alignment transfer without retraining. Relative to editing methods, it is parameter-free and operates at the representation level rather than modifying stored factual associations. Relative to training-based methods, it requires a few hundred labeled examples, a one-time affine fit on about 2,000 WikiSplit sequences, and no parameter updates of the target (Dong et al., 2024).

Taken together, these comparisons locate activation transplantation within a broader shift toward interventions on internal representations rather than end-to-end optimization. One branch uses activation traces to choose where to edit parameters; the other uses activation-derived directions to steer a model online. This suggests that the concept is less a single algorithm than a methodological pattern: identify a behaviorally meaningful substructure in representation space, adapt it to the target setting, and intervene at the most local site that preserves the desired behavior with minimal collateral change.