Test-Time Capability Transfer

• Test-Time Capability Transfer is a family of methods that makes task-relevant competence available during inference through temporary parameter updates, activation changes, prompts, tools, external code, or adaptive system structures.
• Applications include long-context adaptation, image denoising, cross-language programming, reasoning improvement, domain-robust segmentation, terrain-specific robot control, and co-evolving multi-agent coordination.
• The approach can improve accuracy, robustness, or efficiency without conventional retraining, but its benefits are often instance-specific and depend on reliable feedback, computational resources, and compatibility with capabilities already present in the model.

Test-Time Capability Transfer is a family of methods that acquire, adapt, preserve, or redeploy task-relevant competence during inference rather than relying exclusively on fixed pretrained parameters or conventional offline retraining. The transferred object may be a pair-specific translation prior, a context encoded in temporary weights, an activation-space reasoning direction, a procedural harness, an executable functional specification, a terrain-specific policy, or an evolving multi-agent state. Across these formulations, inference is treated as an active computational phase in which test inputs, auxiliary models, execution feedback, self-supervision, or generated curricula alter how an existing model or system solves the current instance.

1. Conceptual scope and taxonomy

Test-Time Capability Transfer encompasses methods that operate at different levels of the computational stack. Some optimize model parameters on the current observation; others modify activations, prompts, memories, communication graphs, or external procedures without changing the underlying model. The common property is that capability is made available to the current test instance through an inference-time mechanism rather than being fully fixed at training completion.

A useful distinction is between the capability being transferred and the carrier through which it is transferred:

Transfer carrier	Representative mechanism	Transferred object
Temporary parameters	ETT, test-time denoising, DTP	Context, image-pair translation behavior, or image-specific adaptation
Activation directions	Unlock	Reasoning or CoT behavior latent in another model or prompt
External procedures	AI4AI, PaT, CodeChemist	Rules, decomposition, verification, routing, or functional specifications
Structured inference state	TacoMAS	Prompt refinements, memories, tools, roles, and communication pathways
Adapted control policy	TTT-Parkour, SETA	Terrain-specific locomotion or OOD appearance-and-pose handling
Spectral or feature interfaces	TTAwPCA, T3s	Corruption robustness or target-domain style compatibility

The term does not imply that a model permanently acquires a general capability. In Deep Translation Prior (DTP), the optimized parameters encode a translation prior specific to one content/style pair (Kim et al., 2021). In the implicit-neural-representation style-transfer method, the optimized network supports many later queries for the same pair but does not generalize to unrelated images or styles (Kim et al., 2022). In PaT, the smaller generator does not permanently learn the larger planner’s reasoning ability; it receives decomposition artifacts for selected instances (Yoon et al., 8 May 2026). TacoMAS similarly performs within-instance capability adaptation through contextual memory and prompt refinement rather than explicit inter-agent parameter sharing (Xu et al., 10 May 2026).

The family can therefore be organized along two axes. The first is adaptation locus: parameters, activations, representations, prompts, tools, topology, or external code. The second is transfer persistence: ephemeral adaptation for one sample, temporary reuse across queries to one instance, reuse across a test set, or persistent offline modification. ETT resets its parameters after answering each test example (Zahirnia et al., 8 Jul 2025), whereas MITT uses generated reasoning prefixes for lightweight LoRA adaptation before evaluating target-language inference (Bajpai et al., 21 May 2025). AI4AI harnesses are built using validation data and then applied to hidden examples without modifying the target model (Qian et al., 12 Aug 2026).

2. Parameter and representation adaptation

Several methods transfer capability by changing the computational representation used by a frozen or partially trainable model.

Pair-specific neural optimization

DTP performs test-time training separately for each content/style pair. Given content image $I_{\mathcal C}$ and style image $I_{\mathcal S}$, a correspondence network and randomly initialized decoder are optimized using only that pair. ImageNet-pretrained VGG-19 supplies the initial feature representation, while the DTP-specific correspondence and generation networks are learned at test time (Kim et al., 2021).

The correspondence module constructs an all-pairs cosine-similarity matrix and converts it into a temperature-scaled soft warp. The warped style features and warped RGB image are combined through a decoder:

$$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$

with $\lambda_w=1/9$. Optimization uses contrastive content, patch-based style, cycle-consistency, and feature-moving-average mechanisms. The parameters become specialized to the current relationship between content and style. The method uses approximately 1,000 iterations, a learning rate of $10^{-4}$, and roughly 150 seconds per $256\times256$ image pair on an RTX 3090.

The test-time capability is therefore a structured translation prior rather than a universally reusable style-transfer network. DTP combines explicit content and style objectives associated with optimization-based methods with the inductive bias of a neural generator, analogous in spirit to Deep Image Prior. Its reported main setting achieves average PE $=0.93$ and SSIM $=0.92$, with evaluations on photorealistic benchmarks, CelebA-HQ, and FFHQ.

A related formulation optimizes an implicit neural representation (INR) on one content/style pair (Kim et al., 2022). The INR maps positional encodings and an interpolated latent code

$z'(\alpha)=\alpha z_c+(1-\alpha)z_s$

to RGB values. Training samples $\alpha\sim U(0,1)$, enabling the optimized MLP to support subsequent style intensities, spatially varying style fields, masks, gradients, arbitrary resolutions, and arbitrary aspect ratios without further optimization. This is within-instance capability transfer: the initial optimization produces a continuous function over coordinates and content/style codes, but the learned parameters remain tied to the specific pair.

Context stored in temporary weights

ETT extends short-context Transformers to long-context understanding by fine-tuning selected parameters on overlapping chunks of the current input (Zahirnia et al., 8 Jul 2025). A long context is divided into 512-token subsequences with 32-token overlap. The model is trained at test time using next-token prediction, typically for 10 epochs with Adam, learning rate $I_{\mathcal S}$0, and weight decay $I_{\mathcal S}$1.

The context is encoded into temporary parameters:

$I_{\mathcal S}$2

The adapted model answers using $I_{\mathcal S}$3, after which parameters are reset to $I_{\mathcal S}$4. Since each attention operation processes a fixed-size chunk, memory is constant with respect to total context length, while the number of chunks grows linearly. ETT extends GPT-Large from approximately 1k to 32k tokens and Phi-2 by up to $I_{\mathcal S}$5, with reported improvements of up to 30% in LongBench accuracy.

Selective tuning is central. The experimental ablation identifies the FFN up-projection, described as the key layer, as particularly effective. For GPT-Large, tuning this component yields a LongBench score of 12.57 compared with 11.30 for full fine-tuning; for Phi-2, the corresponding scores are 18.33 and 17.33. Updating deeper layers can reduce the trainable portion to about 15% of the full model with little or no loss in the reported setting.

Activation-space transfer

Unlock transfers reasoning behavior without parameter updates (Balasubramanian et al., 7 Apr 2026). It extracts a capability contrast from Source Locked and Source Unlocked variants:

$I_{\mathcal S}$6

The contrasts are aggregated by a mean or first principal component. A low-rank linear map aligns Source and Target activation subspaces. If $I_{\mathcal S}$7 and $I_{\mathcal S}$8 are the leading singular-vector bases, the alignment is learned through

$I_{\mathcal S}$9

with lifted operator

$$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$0

The transferred direction is injected into the Target residual stream at every layer using norm-preserving activation steering. No gradient descent, weight merging, correctness labels, or Source-model execution during generation is required.

The method’s Master Key Hypothesis proposes that capabilities may correspond to low-dimensional latent directions. Observed spectral entropies of approximately $$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$1–$$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$2 nats imply effective ranks of roughly $$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$3–$$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$4, despite hidden dimensions of at least 1024. These results support an operational low-dimensional transfer mechanism but do not establish that capabilities are literally single causal directions. Transfer is strongest when the Target already contains compatible latent capability; it cannot reliably create behavior absent from the Target’s representational basis.

3. Self-supervised and sequential test-time adaptation

A second group uses unlabeled test observations to optimize a model against auxiliary objectives selected or meta-learned to improve the primary task.

Meta-learned denoising adaptation

Test-Time Adaptation for Real Image Denoising via Meta-transfer Learning trains a network to improve denoising through self-supervised masked reconstruction (Gunawan et al., 2022). Stage 1 uses Meta-Auxiliary Learning on DIV2K with Gaussian, speckle, and salt-and-pepper corruption. Stage 2 performs MAML-based meta-transfer learning on SIDD and Poly. At test time, only the primary and auxiliary heads are adapted; the body and mask generator remain fixed.

The auxiliary objective reconstructs the observed noisy image on mask-selected regions:

$$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$5

The test image itself supplies the target, so no clean image is needed. Meta-training optimizes the initialization so that a few self-supervised updates improve clean-image denoising. With $$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$6 test-time updates and a learning rate of $$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$7, test PSNR improves from 39.2499 to 39.2653 and test SSIM from 0.9672 to 0.9685 on the reported setting. The immediate gain is modest, while MAXL initialization contributes approximately 1.1–1.6 dB over random initialization depending on the comparison.

The transferred capability is the ability to use a noisy observation to select informative regions, update a small parameter subset, and improve clean-image prediction. This differs from standard domain adaptation because adaptation occurs per image rather than through collective source-target distribution alignment.

Sequential OOD adaptation

SETA addresses open-world pose transfer by adapting two distinct capabilities sequentially: OOD appearance first, followed by OOD skeleton/posture (Chen et al., 2023). Appearance adaptation uses augmented views of an OOD person image and cross-reconstruction through the pretrained pose-transfer model. The appearance loss combines $$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$8 reconstruction, VGG perceptual loss, and attention loss.

After appearance adaptation, the model generates the person in a target posture. Skeleton adaptation uses pretrained person re-identification features and human-parsing masks. Global identity preservation is enforced through Re-ID feature consistency, while local body-part consistency uses Gram matrices:

$$I_{\mathcal C\leftarrow\mathcal S} = \lambda_w\mathcal F(r_{\mathcal C\leftarrow\mathcal S};\omega_g) + (1-\lambda_w)R_{\mathcal C\leftarrow\mathcal S},$$9

The ordered procedure

$\lambda_w=1/9$0

is essential to the method’s rationale. Skeleton adaptation performed before appearance adaptation would rely on unreliable generated images when the model cannot yet preserve unfamiliar clothing or identity. On SHHQ, DeepFashion2, and TikTok, SETA improves SSIM, LPIPS, and FID across pretrained ADGAN, CASD, and NTED systems. For NTED on TikTok, SSIM rises from 0.678 to 0.837, LPIPS falls from 0.259 to 0.097, and FID falls from 72.68 to 35.56.

Spectral robustness adaptation

TTAwPCA inserts a PCA module after an intermediate layer and adapts a small diagonal spectral filter using entropy minimization (Cordier et al., 2022). The PCA basis and source singular values are fixed before deployment. Only $\lambda_w=1/9$1 scalar filter parameters are learned at test time, with $\lambda_w=1/9$2 sufficient in the reported CIFAR-10-C and CIFAR-100-C experiments.

The adapted representation is

$\lambda_w=1/9$3

The method attenuates selected principal directions but cannot amplify weakened task-relevant components or restore information removed by corruption. At severity 5, mean CIFAR-10-C error falls from 43.53% without adaptation to 20.42% with ReLU-TTAwPCA and 20.35% with exponential-TTAwPCA. On CIFAR-100-C, the corresponding errors are 85.54%, 36.62%, and 37.89%, trailing TENT at 34.56%.

TTAwPCA demonstrates parameter-efficient robustness transfer rather than acquisition of a new semantic capability. It is most appropriate when corruption introduces removable nuisance variation in intermediate features. Its online results are less reliable, particularly for CIFAR-100-C, where long-running entropy adaptation can become unstable.

Feature-space style compatibility

T3s performs cross-organ pathology domain generalization through style/content decomposition and test-time style projection (Meng et al., 24 Mar 2025). A DINOv2-style pathology foundation encoder produces features decomposed into domain-specific style statistics and normalized content. Learnable style bases

$\lambda_w=1/9$4

are combined through similarity-weighted projection. CSDM encourages orthogonality among style bases, while mixup broadens the source-style space.

The transferred capability is source-trained segmentation applied to unseen organs after target style has been represented in the learned basis space. The method preserves target content while adapting style. On the reported cross-organ segmentation task, T3s achieves 77.68% IoU and 87.43% Dice, compared with 76.90% and 86.94% for TTDG. The foundation model provides the largest individual gain, while mixup and CSDM improve the complete system.

The supplied description leaves the precise low-rank adaptation and test-time optimization schedule under-specified. Consequently, T3s is most securely characterized as structured style projection with possible parameter-efficient adaptation, rather than as a fully specified generic online optimization algorithm.

4. Procedural and artifact-mediated transfer

Some methods preserve frozen model parameters and transfer capability through external artifacts, verification, decomposition, or executable procedures.

Cross-language functional transfer

CodeChemist transfers functional knowledge from a high-resource programming language to a low-resource language using generated tests (Wang et al., 1 Oct 2025). High-resource programs, typically Python, are generated and executed on diverse inputs. Majority voting over valid outputs constructs language-independent input/output test cases. Low-resource candidates are sampled at temperatures $\lambda_w=1/9$5, $\lambda_w=1/9$6, $\lambda_w=1/9$7, and $\lambda_w=1/9$8, producing ten candidates in total. The candidate with the highest test pass rate is selected.

The functional interface is

$\lambda_w=1/9$9

where $10^{-4}$0 and $10^{-4}$1 are programs in different languages and $10^{-4}$2 is the observed oracle output. The transfer object is therefore not source code, model parameters, or syntax, but a problem-specific behavioral specification.

On MultiPL-HumanEval, Qwen 1.5B Lua performance rises from 34.1 to 57.8, Qwen 1.5B Java from 43.5 to 69.6, and Qwen 1.5B C++ from 34.4 to 52.2. The strongest improvements occur for smaller models and larger gaps between Python and the target language. CodeChemist requires no model retraining, but it depends on executable infrastructure, sufficiently informative tests, and at least one target-language candidate capable of exploiting the transferred specification.

Planner-to-generator transfer

PaT transfers problem structure from a powerful planner to a cheaper generator (Yoon et al., 8 May 2026). The generator first attempts direct code generation using Best-of-$10^{-4}$3 with $10^{-4}$4. If all candidates fail strict execution-based verification, a larger planner decomposes the problem into helper-function specifications. The generator recursively implements and verifies these subproblems, after which the components are composed and re-evaluated.

The central transfer path is

$10^{-4}$5

The smaller model does not acquire the larger model’s general reasoning ability. Rather, the planner changes the granularity and representation of the task. On foundational benchmarks, homogeneous PaT improves average Pass@1 over FunCoder while reducing normalized cost. In the heterogeneous Qwen3 setting, an 8B generator with a 32B planner achieves 87.39% Pass@1 at 0.31 normalized cost, compared with 88.37% at cost 1.00 for homogeneous Qwen3-32B, representing an approximately 69% cost reduction.

PaT’s advantage depends on reliable verification and conditional escalation. Standard generation solves approximately 76% of foundational instances directly, so planning is invoked on roughly 24%. On xCodeEval, however, the 4B generator’s PaT cost is 17.93 normalized units compared with 12.95 for FunCoder, although Pass@1 rises from 29.0 to 34.2. Thus, test-time capability transfer can improve performance without reducing absolute cost when difficult instances dominate.

Strong-to-weak harness transfer

AI4AI at Test-Time studies builder models that construct inference-time harnesses for fixed weaker target models (Qian et al., 12 Aug 2026). A harness may contain deterministic rules, symbolic solvers, task routing, prompt templates, parsers, fallback logic, verification, and strict answer-format enforcement. The builder uses 5% validation data, iteratively refines the scaffold, and submits a final entry point for hidden evaluation.

For GPT-5.4-mini, the mean over scaffolded runs rises from 0.488 to 0.763, while the best GPT-5.5-built harness reaches 0.912. The best automated scaffold obtains approximately 1.00 on BigToM, 0.80 on Hi-ToM, 0.84 on MMToM-QA, and 0.88 on MuMA-ToM. The strongest associations with accuracy are deterministic offloading, benchmark-specific routing, polarity and negation logic, structured extraction, and strict output formatting. The paper reports a correlation of $10^{-4}$6 between deterministic fraction and accuracy, compared with approximately $10^{-4}$7 for code length.

This mechanism transfers cognitive structure rather than internal model knowledge. The builder externalizes unstable reasoning into code and procedures, while the target model handles residual cases. The largest gains occur for weaker targets with greater baseline headroom. Over-scaffolding can harm strong targets near ceiling; every builder regressed on at least one Gemini-3.5-flash benchmark in the matched analysis.

5. Environment-specific and system-level adaptation

Test-Time Capability Transfer also applies to embodied agents and multi-agent systems, where the transferable object includes control policies, memories, roles, and interaction structures.

Terrain-specific policy adaptation

TTT-Parkour transfers a pretrained humanoid locomotion capability to a particular unseen terrain through rapid real-to-sim-to-real adaptation (Zhu et al., 2 Feb 2026). A real terrain is captured with RGB-D, reconstructed as a metrically scaled collision mesh, aligned to gravity and traversal direction, and inserted into simulation. The pretrained policy is then fine-tuned with PPO on the reconstructed terrain before deployment.

The transferred capability includes dynamic locomotion, foothold selection, recovery behaviors, depth-based terrain interpretation, and safety-aware control. Test-time training supplies geometry-specific adaptation. The reconstructed mesh is essential because it provides the physical constraints that caused failure in the real environment.

The policy outputs 29-dimensional target joint positions, which are converted into torques through proportional-derivative control. Full fine-tuning is the most consistently effective of the tested strategies. On 13 target terrains, TTT-1 raises many pretrained success rates near 0–10% to approximately 99–100% in simulation. Scratch training on one target terrain fails on nearly every terrain despite 25,000 iterations, demonstrating that the pretrained dynamic capability is necessary.

The entire capture, reconstruction, and adaptation procedure takes less than approximately ten minutes on most tested terrains. Real-world performance is lower than simulation because of camera noise, actuator dynamics, terrain movement, and mesh inaccuracies. The method is terrain-specific and does not provide a formal guarantee of retention on the original terrain distribution.

Co-evolving multi-agent systems

TacoMAS adapts both agent capabilities and communication topology during inference (Xu et al., 10 May 2026). The MAS state is

$10^{-4}$8

where $10^{-4}$9 is the directed communication graph and $256\times256$0 contains role-specific prompts, contextual memory, and tool inventories.

A fast loop evaluates agent trajectories and contribution scores, then writes prompt edits, memory seeds, or tool-use changes. A slow loop, executed every $256\times256$1 fast rounds, performs agent birth and death operations and edge edits. The default configuration uses 10 fast rounds, $256\times256$2, five initial agents, a maximum of 20 agents, at most two birth/death pairs, and at most four edge edits per slow update.

The method’s fast capability update has a replicator-style abstraction:

$256\times256$3

In implementation, this corresponds to verbal reinforcement through state editing rather than literal neural parameter multiplication. A useful behavior can persist in an agent’s memory across rounds, influence other agents through messages, or motivate creation of a specialized agent. Explicit cross-agent memory copying, persistent cross-instance skill transfer, and gradient-based parameter updates are not demonstrated.

The two-timescale design reflects a stability trade-off. Capabilities must change rapidly in response to local failures, while topology must evolve more slowly to preserve coordination. TacoMAS achieves reported average improvement of 13.3% over the strongest baseline across Finance, BrowseComp-plus, PlanCraft, and Workbench. Its principal limitation is that scratch memory is cleared between instances, preventing persistent skill accumulation.

6. Test-time curriculum and reasoning transfer

A distinct line of work transfers reasoning behavior by constructing self-supervised curricula from the test set or by adapting the distribution of inference trajectories.

Multilingual initial-thought transfer

MITT transfers high-resource-language reasoning prefixes to improve test-time scaling in low-resource languages (Bajpai et al., 21 May 2025). The method collects 100 independent 32-token initial prefixes per question in English or in English, Italian, German, and Portuguese. These prefixes are used for LoRA-based causal-language-model adaptation of DeepSeek-R1-Distill-Qwen-7B. The base model is 4-bit quantized, computation uses bfloat16, and LoRA is applied to query and value projections with rank 8, $256\times256$4, and dropout 0.05.

The training objective is next-token prediction on model-generated prefixes. At target inference, prompts remain strictly monolingual; MITT does not paste English reasoning into each target prompt. It changes the model parameters so that target-language reasoning is more likely to begin in a stable and productive trajectory.

For Qwen-7B, Vietnamese accuracy at 8k tokens rises from 0.091 in the base condition to 0.162 with English-only prefix training and 0.209 with multilingual high-resource prefixes. Tagalog rises from 0.109 to 0.349 under English-only prefix training. The gains are not uniformly monotonic across budgets or source-language mixtures. MITT therefore supports transfer of an initial reasoning prior, but does not prove that the transferred object is an abstract language-independent reasoning capability rather than a mixture of structural, stylistic, and language-specific effects.

Test-time curriculum synthesis

TTCS uses two co-evolving policies: a question synthesizer and a reasoning solver (Yang et al., 30 Jan 2026). The synthesizer generates structurally isomorphic but surface-distinct variants of test questions. The solver evaluates these variants through multiple sampled responses. Questions near the solver’s uncertainty frontier receive higher capability reward:

$256\times256$5

Here $256\times256$6 is the self-consistency score. Questions that are too easy yield near-uniformly correct responses, while questions that are too difficult yield near-uniformly incorrect responses. Intermediate disagreement is intended to provide a stronger learning signal.

The synthesizer also receives penalties for similarity to the original question and redundancy within the generated group. The solver trains on both original test questions and synthetic variants, retaining examples satisfying

$256\times256$7

This co-evolution creates a moving curriculum: as the solver improves, the synthesizer is encouraged to produce more difficult variants.

On Qwen2.5-Math-1.5B, average performance across six mathematical benchmarks rises from 17.30 for the pretrained model to 41.49 with TTCS. On Qwen2.5-Math-7B, TTCS reaches 52.54 compared with 48.42 for TTRL. On Qwen3-4B-Base, it reaches 47.21 compared with 43.59 for TTRL. Training on MATH-500 improves AIME24 performance from 7.1 to 12.9 in the reported cross-benchmark transfer.

TTCS also reports transfer from mathematical adaptation to MMLU-Pro, SuperGPQA, and BBEH. These results support within-model and cross-benchmark transfer, while the stronger cross-domain interpretation remains limited by possible test-set overfitting, noisy pseudo-labels, and incomplete semantic validation of synthetic problems.

7. Evaluation principles, limitations, and unresolved questions

The empirical evidence across these methods supports several recurring principles.

Capability and representation must be distinguished. A model may already contain a latent capability that is difficult to elicit, as in Unlock, or it may require a new external procedure that changes the task representation, as in AI4AI and PaT. In TTAwPCA, semantic competence remains fixed while robustness is adapted. In TTT-Parkour, locomotion competence is pretrained while geometry-specific control is learned in simulation.

Test-time transfer is often instance-specific. DTP, INR style transfer, ETT, SETA, and TTT-Parkour adapt to a particular pair, context, posture distribution, or terrain. Their learned states are reset, discarded, or unsuitable for unrelated instances. TacoMAS explicitly clears scratch memory between queries. Such methods should not be conflated with persistent skill learning.

Verification and self-supervision determine reliability. CodeChemist and PaT exploit executable tests, which provide stronger feedback than confidence alone but remain vulnerable to sparse coverage, correlated errors, and false positives. AI4AI uses a validation proxy whose best score correlates with hidden accuracy at $256\times256$8 in the reported study. TTCS uses self-consistency but acknowledges that majority agreement is not correctness. Entropy-based TTAwPCA can produce overconfident incorrect predictions.

Adaptation capacity creates a trade-off. Small adapters and low-rank directions reduce cost and preserve the base model but may be underpowered. TTAwPCA uses 2,000 filter parameters and trails TENT on CIFAR-100-C. Denoising adaptation updates approximately 0.12 million head parameters and produces only a modest direct meta-test gain. Unlock deteriorates at excessively high alignment rank, while ETT finds that selective FFN tuning can outperform full fine-tuning.

Test-time computation is not necessarily cheaper. ETT has linear rather than quadratic dependence on total context length but performs repeated forward/backward passes. CodeChemist is substantially slower than vanilla inference. TacoMAS adds judges, reflections, topology updates, and repeated agent calls. TTT-Parkour reduces adaptation time relative to retraining but still requires terrain capture and simulation. Test-time capability transfer generally trades offline training or model size against inference-time latency, memory, execution, or orchestration cost.

Generalization claims remain bounded. The strongest results are tied to structured shifts: style correspondence, visual corruption, cross-organ style statistics, multilingual reasoning prefixes, executable code behavior, reconstructed terrain geometry, or recurring Theory-of-Mind structures. There is limited evidence for arbitrary semantic shifts, heterogeneous architectures, unrelated tasks, unseen scripts, persistent cross-instance learning, or capabilities absent from the target model.

The central unresolved question is whether these methods transfer genuine capabilities or merely improve deployment conditions. Activation steering may sharpen a pre-existing reasoning distribution rather than create mathematical reasoning. Harnesses may externalize computation rather than strengthen the target model. PaT transfers decomposition artifacts but not planner competence. ETT may memorize context in temporary weights without extending intrinsic general reasoning capacity. TTCS may learn useful procedures or may partially overfit self-generated curricula.

A precise operational definition is therefore preferable to a strong ontological one. Test-Time Capability Transfer is best understood as the inference-time movement of task-relevant computational competence across a boundary—between domains, languages, representations, models, agents, subtasks, or environmental configurations—using temporary optimization, activation intervention, self-supervision, external artifacts, execution feedback, or adaptive system structure. The transferred competence may be persistent only within the current instance, and successful transfer generally depends on compatibility between the pretrained capability and the adaptation interface.