PyTorchKR Papers

최신 논문

총 524,001편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.03618v1cs.AI

Strategist: Evolving Reasoning with Agents

A model is only half of an AI agent system. The other half is the inference strategy: the reasoning procedure that turns model calls into a response. Today that second half is written by hand, and the experience of building one strategy never carries to the next. We introduce Strategist, a meta-agent that evolves inference strategies across tasks. Each strategy is written as a blueprint,executable code that other strategies can reuse, and together the blueprints form a library. Given a task and a small development set, Strategist retrieves related blueprints, writes candidate strategies, tests them on real samples, and revises them from the results. Every candidate then enters the library, so later tasks build on earlier discoveries. On 30 benchmarks, Strategist beats the strongest baseline on 24 and matches its average accuracy (74.7% vs.74.0%) at 39% lower inference cost. Its strategies also transfer to unseen tasks and other models without further search. The part of an AI system that is not improved on its own now does.

arXiv:2611.03617v1cs.CL

Tracing the Scientific Discovery Capacity of LLMs to Their Training Data

Large language models (LLMs) show growing promise for scientific discovery, yet basic questions remain about what enables reliable discovery. How can we trace the scientific discovery capacity of LLMs to their training data? We study this question as one of missing probability mass and develop a theoretical account in which monofacts-the facts observed once during training-set the discovery capacity. In materials science and regulatory genomics, we directly manipulate monofact rates in training sets and find that more monofacts consistently yield more discoveries across two open-weight model families, sampling strategies, and reasoning budgets. Further experiments show the relationship cannot be explained by training-set diversity alone and can transfer across scientific domains when paired with in-domain data. Together, our results identify monofact rate in training data as a key factor governing discoveries in language models.

arXiv:2611.03614v1cs.LG

Benchmark Prediction with Sparse and Unpaired Cached Responses

Evaluating a generative model on all queries from all available benchmark tasks is prohibitively expensive. As such, models are often evaluated on different subsets of tasks and/or different subsets of queries from a given task. The resulting collection of responses and the corresponding (model, task) score matrix are sparse and unpaired. In this paper, we introduce the Product Kernel Perspective Space (PKPS), a collection of low-dimensional representations of models based on their embedded responses. PKPS generalizes existing methods for representing black-box generative models by comparing responses to similar -- rather than just identical -- queries and is therefore able to utilize sparse and unpaired information. Theoretically, we show that benchmark prediction using PKPS is query-efficient relative to methods that only use information from dense and paired portions of the evaluation data. Empirically, we show that PKPS-based methods outperform baselines on benchmark prediction across two large benchmark suites.

arXiv:2611.03612v1cs.LG

Efficient Closure Learning in Dynamical Systems: A Comparison of Data-Assimilated and Embedded Approaches

Increased computing power has transformed computational science and improved predictive capabilities for complex systems, but many phenomena still exceed current simulation capabilities or require rapid evaluations for fast design iterations. Closure models address this challenge by approximating unresolved effects in reduced or incomplete dynamical systems. Data-driven closures offer a flexible mechanism for improving model accuracy and applicability, but approaches that embed the closure within a computational model for construction and training phases can be computationally expensive and cause numerical instabilities in the computational model during training. To avoid the challenges posed by embedded approaches, staged approaches to closure-model learning have been developed. The first stage uses data assimilation to infer state and closure trajectories from observations. The second stage trains a machine learning closure model offline using the assimilated data, thereby avoiding repeated simulator calls during training. In this work, we investigate the cost and accuracy tradeoffs between embedded and staged data-driven closure modeling approaches for a synthetic epidemiological SIQR model with a nonlinear state-dependent quarantine closure. The results illustrate a practical accuracy-efficiency trade-off: for this exemplar, embedded training provides the strongest average accuracy, while the staged approach yields comparable qualitative behavior with a moderate increase in error, substantially fewer model evaluations, and improved optimization robustness.

arXiv:2611.03605v1cs.CL

ufakzeka-kulak: Building and Evaluating a 31.6M-Parameter Turkish Speech Recognizer from Scratch

ufakzeka-kulak is an open Turkish speech-to-text model of 31.6M parameters, a FastConformer encoder with a token-and-duration transducer (TDT) and a CTC head, trained from scratch and released under Apache-2.0. It writes cased, punctuated Turkish with numbers as words. Its 4,977 training hours come from sources whose licences allow commercial use; 281 hours are human-transcribed, and for the other 4,696 the training text is an uploader subtitle or a machine transcript, kept only where a second, independent reading of the audio agrees with it. The recipe, checkpoint, labelling and export choices were made on development data, and the test sets were scored once, in a release gate fixed before training. With greedy TDT decoding the normalised word error rate (WER) is 0.0921 (95 percent interval 0.0792 to 0.1055) on FLEURS Turkish test and 0.0916 (0.0873 to 0.0958) on Common Voice 27.0 Turkish test, lower than Whisper base and Whisper small on both sets with paired intervals wholly below zero; Whisper large-v3 (greedy) scores 0.0604 and 0.0767. On conversational speech the WER is 0.3807. We also report punctuation and casing, degraded audio, and how closely each shipped export reproduces the trained model.

arXiv:2611.03602v1cs.CV

Reviso: Not All Visual Evidence Deserves Equal Attention in Video Captioning

Multimodal large language models (MLLMs) can generate detailed video captions, yet they may overemphasize minor details while overlooking essential context and key events. We argue that not all visual evidence deserves equal attention, as the importance of visual information depends on the content of each video. Motivated by this observation, we propose Reviso, an agentic video captioning framework that learns what to observe and how to gather relevant visual evidence. From a low-frame-rate global preview, Reviso constructs a hierarchical Context--Focus--Points (CFP) observation plan that identifies the video context, prioritizes key aspects, and defines concrete observation targets and corresponding tools. Guided by this plan, Reviso performs goal-directed active perception by selectively revisiting temporal segments and combining tools for fine-grained inspection. It interprets the acquired evidence, updates its plan as needed, and integrates global and local information to generate the final caption. To learn this observation strategy, we develop a verification-guided pipeline that constructs high-quality training trajectories through factual verification and teacher-guided refinement. We further train Reviso using Group Relative Policy Optimization (GRPO), incorporating a key-fact reward for caption quality, a Plan Alignment Reward (PAR) for identifying important observation targets, and an Unproductive Tool-Call Penalty (UTCP) that adjusts observation-cost penalties based on caption quality. Experiments across four video captioning benchmarks show that Reviso outperforms the evaluated open-source baselines, while its generated captions better support downstream video question answering. These findings highlight the effectiveness and potential of priority-guided active observation for video captioning.

arXiv:2611.03600v1cs.AI

Four Bridges: Measuring Deception Propensity in a Multi-Agent Game

AI agents are increasingly deployed in settings where they interact and coordinate with other agents, yet honesty is mostly measured in single-turn, single-agent benchmarks that pressure the model to deceive. We study deception that emerges in multi-agent interaction without any instruction to deceive. We introduce Kradle, a platform for evaluating AI models, and Four Bridges, an information asymmetry game in which four agents distribute across four rooms: one is the death room, the others hold food worth more if not shared. One agent, the informed agent, knows which room is the death room and can withhold or disclose that information. We evaluate eight frontier models as the informed agent. Deceptive behavior appears in every model and in 73% of all runs (6% to 100% per model), and does not decrease with capability tier within a model family. When the informed agent does disclose, the warning is often challenged, and agents that challenge it enter the death room twice as often as those that do not. Telling all agents that one of them knows the death room collapses deception in six of eight models; telling the informed agent that the room is permanently fatal moves only one. On an open-weight model, an honesty steering vector reduces deception to 0% and adversarial honesty training to 6%. The platform and game provide a testbed for measuring deceptive behavior and comparing interventions.

arXiv:2611.03599v1cs.LG

Should We Combine Our Data? Algorithms and Limits for Privately Assessing Data Sharing

In this paper, we study when two parties should pool their data to train a machine learning model. Two agents, Alice and Bob, each hold a sensitive labeled dataset and want to train a classifier from a given hypothesis class $\mathcal{H}$. Even when each party's dataset can be perfectly classified on its own, the two parties may follow conflicting labeling standards or have selection biases that affect their features and label sets, such that no single classifier in $\mathcal{H}$ fits their combined data. Before committing to data sharing, Alice and Bob therefore want to test whether a single classifier in $\mathcal{H}$ can jointly fit both datasets---without revealing their individual data points to one another. We formalize this problem from a learning-theoretic viewpoint via $\operatorname{opt}_\mathcal{H}$, the minimum \emph{number of mistakes} made by any hypothesis in $\mathcal{H}$ on their pooled data. Through interactive communication, the agents must distinguish $\operatorname{opt}_\mathcal{H}=0$ (compatible datasets) from $\operatorname{opt}_\mathcal{H}\geq t$ (conflicting datasets) under pure $\varepsilon$-two-party differential privacy. A trusted curator could do this with $t=O(1/\varepsilon)$, and we ask how small $t$ can be without one. For thresholds on the real line over a domain of size $m$, we show that the optimal $t$ is $\tilde{\Theta}(\log\log m/\varepsilon)$, exponentially smaller than the $\Theta(\log m/\varepsilon)$ samples needed to privately learn thresholds, even with a curator. For halfspaces in $\mathbb{R}^d$, the dependence on $m$ jumps to $\log(m)$: we give a protocol with $t=\tilde{O}(d^3\log m/\varepsilon)$ and show that some domains require $t=\Omega(d\log(m/d)/\varepsilon)$.

arXiv:2611.03595v1cs.LG

DynaSolver: A Neural Operator for Autoregressive 4D Flow Prediction with Dynamic Boundaries

Neural operators provide efficient surrogates for computational fluid dynamics (CFD), but most are designed for static geometries and are not equipped to model unsteady flows induced by continuously moving boundaries. We study dynamic-boundary 4D CFD prediction for articulated moving geometries and introduce RoboCFD-4D, a temporally resolved robotic CFD benchmark spanning diverse morphologies and flow conditions. To address the computational and autoregressive challenges of this setting, we propose DynaSolver, a neural operator that explicitly conditions on boundary evolution and causal flow history. DynaSolver employs an efficient physical-token formulation that amortizes point-to-token projection across the operator stack, substantially reducing repeated point-space spatio-temporal computation while retaining accurate point-level reconstruction. We further introduce Replay-Augmented Parallel Training, which exposes the model to autoregressively generated histories while preserving parallel dense supervision, narrowing the gap between ground-truth-conditioned training and autoregressive rollout. Experiments on RoboCFD-4D show that DynaSolver achieves the best full-rollout accuracy among a broad range of learned CFD surrogates, with its advantage increasing as rollout proceeds, indicating substantially reduced autoregressive error accumulation. Project page: https://made-gpt.github.io/DynaSolver-ProjectPage.

arXiv:2611.03592v1cs.LG

Evaluating Uncertainty Calibration in Probabilistic Time Series Foundation Models

Time series foundation models (TSFMs) are increasingly used for probabilistic forecasting across diverse domains, yet the reliability of their uncertainty estimates remains poorly understood. In this work, we study uncertainty calibration in probabilistic TSFMs from a broader methodological and empirical perspective. We introduce a taxonomy of how TSFMs represent predictive uncertainty and develop a fine-grained evaluation framework for calibration in multi-horizon forecasting that distinguishes pooled, horizon-wise, and stronger conditional notions of calibration. Using this framework, we conduct a comprehensive empirical evaluation of popular TSFMs across diverse datasets and show that current models exhibit distinct and systematic patterns of miscalibration that standard metrics often obscure. We further analyze common post-hoc recalibration methods and characterize which forms of calibration they can improve and which tend to persist. Our results provide practical guidance for evaluating and improving uncertainty calibration in TSFMs and highlight important limitations of current practice.

arXiv:2611.03590v1cs.LG

A Scaling Theory of Multi-Epoch Pretraining: Late-Time Overfitting and Ensembling

In the near future language model pretraining will become bound by total data rather than total compute. Such a regime requires revisiting scaling laws and scaling law theory. Motivated by prior works that have found utility in ensembling, we study how the number of training epochs and ensemble size jointly shape generalization when training data are fixed. In language-model experiments on FineWeb, logit averaging reduces late-training deterioration, delays optimal stopping, and can outperform wider or deeper single models at matched parameter-token budgets over a common training horizon. To analyze this interaction between data reuse and ensembling, we derive a dynamical mean-field theory (DMFT) for multi-epoch stochastic gradient descent in power-law random-feature ensembles, tracking error correlations across epochs and models. Ensembling reduces the prediction variance responsible for late-time growth in expected test error, changing the optimal stopping time. We use these stopping times to derive compute-allocation laws at fixed data and individual model sizes. Under this allocation, ensemble size grows with compute at least as fast as training time per member, while excess test loss above the limiting loss decreases as a power of compute.

arXiv:2611.03588v1cs.CV

3D-BBDA: Backpropagation-Free Black-Box Domain Adaptation for Robust 3D Point Cloud Recognition

Robust 3D point-cloud recognition is critical for robotics, mapping, and urban sensing, yet point-cloud classifiers often suffer performance drops under domain shifts caused by sensor changes, environmental variation, and acquisition artifacts. Many existing adaptation methods are difficult to deploy in practice because they require access to source data or the internal parameters of the source model, which may be restricted by privacy, licensing, or security constraints. We present 3D-BBDA, a backpropagation-free black-box domain adaptation framework that adapts a source-trained predictor to unlabeled target data using only black-box outputs. Our method performs pseudo-label guided prototype-based alignment and employs a weighted prediction module to fuse source logits with alignment logits, reducing sensitivity to noisy pseudo-labels. To stabilize prototype memory when pseudo-label coverage is incomplete, we construct a Gaussian-perturbed prototype prior from unlabeled target predictions and use it as a fallback for missing or unreliable class prototypes during alignment. Experiments on four corrupted 3D benchmarks demonstrate consistent improvements over strong baselines. On ModelNet-40C, 3D-BBDA achieves mean errors of 32.08% (PointNet) and 29.62% (GDANet), outperforming Source-only (35.88% / 34.03%) and BFTT3D (33.78% / 32.36%). 3D-BBDA requires only a single forward pass with no gradient updates making it suitable for real-time, privacy-preserving deployment in transferable 3D recognition pipelines. Project page is available at https://visdomlab.github.io/3D_BBDA

arXiv:2611.03586v1cs.AI

FedReady: Bounded Agents for Federation Formation

Federated learning (FL) runtimes standardize aggregation and deployment, but each new study still requires federation formation: selecting task-eligible sites and preparing their local data and task code for federated training or analysis. This preparation recurs when the target changes, while raw records and site-private mappings are excluded from cross-site FL messages. We present FedReady, a system that performs federation formation with bounded LLM agents. Agents propose shared contracts, local adapters, and task code; the runtime owns workflow state and accepts artifacts only after checks tied to their digests. Candidate site discovery prioritizes recall; local eligibility verification checks task support before site admission. Once participating sites and artifacts are fixed, conventional federated training or analysis proceeds without readiness agents. Across 39 simulated retinal sites and four tasks, discovery reaches recall 1.00 at precision 0.65; local eligibility verification removes all 17 false positives without losing an eligible site-task pair. The participating sites supply 45,575 task-record instances and complete four 100-round jobs, including one with a runtime-generated contract. Across ten patient cohorts, preparation supports seven patient populations and holds endpoints at three sources as unsupported. Comparison with manually prepared reference populations and analysis (the manual reference) gives a maximum survival-curve difference of 0.001133. We also analyze 19 development failures, targeted classification ablations, and standalone and heterogeneous agent configurations. These results establish the feasibility of task-specific data preparation before ordinary FL training or analysis.

arXiv:2611.03575v1cs.CL

VeriHarm-Bench: Measuring Harmful Success in LLM Coding Agents Verifiably

Large language model (LLM) coding agents increasingly operate with shell access, allowing them to write scripts, move files, alter permissions, and modify system configurations. However, an agent may achieve a user's goal while causing unintended harm, such as deleting unrelated files, weakening access controls, or corrupting data, in response to benign inputs without explicit jailbreaking. This failure mode is particularly important because apparent task success may reassure users that an agent behaved correctly, causing harmful side effects to go unnoticed. We call this harmful success: the agent successfully completes the intended benign task while simultaneously causing unintended harm. Despite its importance, harmful success is difficult to measure because conventional task-success metrics overlook harmful behaviors, while existing safety evaluations are often designed solely to assess harmful outcomes. We introduce VeriHarm, an automatic framework for constructing programmatically verifiable benchmarks that directly measure harmful success. VeriHarm minimally perturbs coding tasks to elicit unintended harmful behavior while maintaining the benignity of the task and preserving the original task objective. It then synthesizes executable harm tests that, together with the original task-correctness tests, evaluate both task accuracy and harmful outcomes. Frontier agents, including GPT-5.6-Sol and Claude-Opus-5, achieve an average task accuracy of 95% on our curated VeriHarm-Bench, yet still struggle to complete tasks safely: their safe success rate, the fraction of tasks completed successfully without triggering the designated harm, averages only 37%. These results suggest that successful task completion can frequently coexist with unintended harm, and that task-success metrics alone may substantially understate the potential safety risks of coding agents.

arXiv:2611.03574v1cs.LG

BISECT: Bilevel Spectral Decomposition for Low-Rank Continual Learning

Exemplar-free continual learning on pre-trained Vision Transformers increasingly relies on Low-Rank Adaptation (LoRA), yet prevailing methods govern the low-rank update space through strict subspace isolation. By projecting each incoming task into the approximate null space of historical token activations, isolation-based learners equate any cross-task feature overlap with catastrophic interference. We show that this orthogonality premise induces a structural bottleneck we term \emph{spectral starvation}: when down-projection bases are locked before task optimization, the gradient norm reaching the trainable up-projection is upper-bounded by the energy of current-task features captured by the chosen basis. Because natural visual tasks share low- and mid-level covariance structure, historical null tails contain barely $5.5\%$ of incoming task variance, starving new-class plasticity while blocking positive consolidation along shared directions. We propose BISECT (Bilevel Spectral Decomposition), a continual LoRA framework that partitions a fixed rank-$r$ budget into two orthogonal functional blocks: an invariant carrier spanned by the dominant modes of the joint historical-and-current Gram matrix, and a novelty complement extracted via generalized Rayleigh whitening followed by orthogonal complement projection. After single-task training, BISECT folds both blocks back into the frozen backbone using a closed-form precision gate that scales each rank-1 invariant atom by its current-to-joint Rayleigh ratio while preserving novelty atoms at unit gain. Across eight incremental settings on ImageNet-R, ImageNet-A, CIFAR-100, CUB-200, and DomainNet, BISECT outperforms prompt, adapter, and LoRA baselines with zero extra inference parameters or FLOPs, and remains complementary to sparse analytic expansion.

arXiv:2611.03557v1cs.CV

PlantVLM-Edge: Quantized Vision-Language Models for On-Device Crop Health Diagnosis

Timely and accurate crop health diagnosis is essential for sustainable agricultural production, yet accurate diagnosis often requires expert knowledge that is not readily available. Recent vision-language models (VLMs) have demonstrated remarkable capabilities in visual understanding, reasoning, and natural-language interaction, creating new opportunities for intelligent crop health assessment. However, deploying VLMs on resource-constrained devices remains challenging due to their substantial computational and memory requirements. This paper presents PlantVLM-Edge, a quantized vision-language framework for conversational crop health diagnosis that can run entirely on resource-constrained edge devices. PlantVLM-Edge combines agricultural domain adaptation, post-training quantization, and on-device serving with an offline chat interface, allowing growers to ask open-ended questions and receive diagnoses, symptom explanations, and management guidance without cloud connectivity. The framework is backbone-agnostic: we validate it with two different VLM types with varying scale and architecture (Qwen3-VL-2B and LLaVA-1.5-7B), four quantization schemes, and two diagnostic domains (plant diseases and insect pests), and we deploy all variants on an NVIDIA Jetson AGX Orin. Eight-bit quantization affects answer quality minimally relative to the full-precision model, and 4-bit activation-aware weight quantization (AWQ) retains 98.6% of full-precision answer quality on both domains while reducing the weight memory of the 2B backbone by 49%. In its most compact form, PlantVLM-Edge requires 2.01 GiB of weight memory, returns its first token in about 0.12 s, and completes a 64-token diagnostic response in 0.78 s. These results demonstrate that timely and accurate crop health diagnosis is practical on edge devices, providing a foundation for next-generation crop monitoring and decision support systems.

arXiv:2611.03551v1cs.CL

Can LLM Agents Be Forward-Deployed Engineers? Requirement Discovery and Delivery for Enterprise Text-to-SQL

LLM coding agents can increasingly automate model post-training, but whether they can act as forward-deployed engineer (FDE) agents, discovering customer requirements and turning them into deployment decisions, remains unclear. We study this on a real enterprise text-to-SQL engagement: a simulated customer, grounded in the real use case's requirements, starts with an ambiguous request. An agent must discover acceptance criteria, data permissions, and hardware constraints in the discovery phase, then plan and execute a post-training recipe in the delivery phase. Across 15 models, the strongest discover over 80% of withheld requirements and open-weight models discover 40-50%; in the delivery phase, only 25% of runs across eight coding-agent configurations pass all acceptance criteria. We identify four bottlenecks in this engagement: (1) incomplete discovery, with 40% of withheld requirements missing or misrepresented; (2) unreliable execution; (3) underused compute budgets and a lack of systematic comparisons across base models and post-training methods; and (4) benchmark accuracy does not guarantee production readiness. We suggest human review of discovered requirements, monitoring of the agent's plan during execution, and an explicit experiment-environment specification covering hardware, budget scope, and production checks.

arXiv:2611.03542v1cs.LG

Variational Functional Neural Networks (vfNN)

We present a variational functional neural network (vfNN) algorithm that optimizes physically motivated continuum functionals in integral form (e.g., Lagrangian, Hamiltonian, or free energy) to find the states of a system. That way, vfNN bypasses solving the equations of motion (a.k.a. the Euler-Lagrange equations), which are in the form of differential equations. To this end, our approach is fundamentally different from conventional physics-informed neural networks (PINN) that solve differential equations by minimizing a phenomenological loss function involving higher order derivatives and ad hoc weight factors. In contrast, vfNN merely evaluates lower order derivatives of the generating functional and replaces the weight factors with rigorous physical scales. This also fundamentally changes the loss landscape, and the desired solutions may become unstable. For this emergent nonconvex optimization problem, we develop a quasi-Newton optimization algorithm with simulated annealing that converges to the solutions of the corresponding Euler-Lagrange equations of the functional regardless of their stability properties. Our simulations using vfNN give accurate results for benchmark systems governed by linear or nonlinear ordinary differential equations.

arXiv:2611.03534v1cs.AI

An Agentic Framework for Automated Power System Studies and Grid Connection Assessment

Connecting new generation and demand to the grid depends on two scarce skills: checking applications against a large, changing body of regulation, and running the network studies that decide where and how a project connects. We present Grid Agents, an AI system that brings both together, and the methodology we use to evaluate it. The Grid Ontology is its knowledge system. It represents the regulation, standards and engineering knowledge that grid work depends on, makes it searchable, and evolves as the agents learn from use. Every answer drawn from it rests on evidence the agents cite. Grid Studio is an agent system specialised for power-system engineering. It works in its own engineering environment with domain tools and a study framework, draws on the Grid Ontology, and relies on established solvers for every number it reports. In both, engineers make the decisions. We bring eight published power-system benchmark families, previously scattered across separate papers, into one suite of 519 tasks, and we verify not only each answer but the work behind it, by replaying the agent's tool calls. The Grid Ontology cites the right regulatory evidence more often than a general coding agent or retrieval pipelines, leads on LegalBench-RAG and ranks first on the public leaderboards of the Enterprise RAG Challenge and ViDoRe V3. Grid Studio scores highest on the 519 power-system tasks and is the fastest of all the systems we compare, both by a statistically significant margin. With an optimiser in the loop, its reinforcement packages cost 33-67% less on average than those the agent designs alone.

arXiv:2611.03528v1cs.CV

Contrastive Adapter Training and Consensus Knowledge Distillation for Multi-Source-Free Domain Adaptation in Skin Cancer Diagnosis

Skin cancer diagnosis, particularly differentiating melanoma from benign nevi, is challenging due to the visual similarity between lesions. Although convolutional neural networks (CNNs) and vision transformers (ViTs) have shown promising performance, their effectiveness often deteriorates on data from heterogeneous clinical sources. Conventional domain adaptation methods address domain shift but require access to source data during adaptation, which is often infeasible due to privacy regulations. Multi-source-free unsupervised domain adaptation (MSFDA) addresses this limitation by leveraging multiple labeled source domains to generalize to an unlabeled target domain without requiring access to source data. However, existing MSFDA methods rely on full backbone fine-tuning, leading to catastrophic forgetting and overfitting on small clinical datasets, and address domain shift at the aggregation stage without establishing a shared domain-invariant feature space. Furthermore, their reliance on hard pseudo-labels or confidence-weighted aggregation introduces noisy supervision signals under domain shift. To address these limitations, we propose CAT-CKD, consisting of two components: (1) contrastive adapter training (CAT), which trains lightweight ConvPass adapters within a frozen ViT backbone using supervised contrastive learning (SCL) to establish a shared domain-invariant feature space before source-specific model training, and (2) consensus knowledge distillation (CKD), which aggregates logits from multiple source models into a consensus supervisory signal and adapts a student model on unlabeled target data using KL divergence. Experiments on five publicly available skin lesion datasets show that CAT-CKD achieves an average AUROC of 86.1%, outperforming existing MSFDA methods while requiring only 4.3M trainable parameters. The code is available at https://github.com/A-Abedi/CAT_CKD.