基础模型适配

Production retrieval systems require three properties jointly: domain adaptation, preservation of general-retrieval quality, and compatibility with existing vector indices. Existing adaptation methods achieve at most two. Standard parameter-efficient fine-tuning (PEFT) adapts to a domain but breaks the index and degrades general retrieval; backward-compatible training (BCT) preserves the index approximately via distillation while sacrificing general retrieval under domain shift; frozen-backbone side networks (L2C) preserve general retrieval but still modify the deployed embedding via additive combining, so existing indices are not preserved. We introduce \textbf{CoSE} (Complementary Subspace Expansion), a side-branch LoRA whose frozen pathway is structurally untouched, so any index built from the base encoder remains valid --- the adapter is \emph{pluggable}, added on top of an already-deployed index without modifying existing entries. A joint training objective converts the frozen+LoRA concatenation into Pareto-efficient retrieval. Across four distinct domains (Korean broadcast, medical radiology, art paintings, remote sensing) and 3 seeds, CoSE at $6.4$M trainable parameters is the only configuration that simultaneously stays within $\pm 0.048$ of the frozen base on Flickr, COCO, and ImageNet at the deployment-default $\alpha{=}0.5$, is domain-competitive with standard LoRA (ties on RSICD, within $\sim 2\sigma$ on HAN and SemArt, $-0.05$ on ROCO), and is backward-compatible by construction. Deployment-time knobs --- $\alpha$ blending, Matryoshka truncation, partial branching --- expose further cost/quality trade-offs without retraining.

We propose a label-free approach to adapt powerful but generic vision foundation models to specialized scientific domains. Standard supervised fine-tuning is often ill-suited to these settings: labels are scarce, and task-specific training can collapse the model's generality and hurt robustness. We instead leverage metadata to adapt representations to new domains in a self-supervised manner. Our method, META-DINO, combines a standard self-supervised objective with flexible metadata guidance that handles both highly granular discrete metadata and continuous metadata. It encourages the representation to preserve informative factors while suppressing spurious ones. Across subcellular fluorescence microscopy, Earth observation, wildlife monitoring, and medical imaging, META-DINO consistently outperforms standard unsupervised domain adaptation and fully supervised adaptation. It also exceeds highly-specialized domain-specific state of the art, while using no task labels for backbone adaptation and only lightweight probes for supervision.

Fine-tuning large transformers routes gradients through every token position, paying the cost in activation memory that scales linearly with sequence length and dominates the GPU budget. Existing efficient methods reduce the parameter side---low-rank adapters, prefix tuning, weight-decomposed updates---but leave the activation footprint untouched. We present the Token Hypothesis, a memory-efficient fine-tuning procedure motivated by a Neural Tangent Kernel analysis of how fine-tuning signal is distributed across token positions. We prove that every position is a low-dimensional capacity bottleneck whose width does not grow with the model, that positions with aligned hidden-state geometry are interchangeable in bidirectional models, and that causal positions form a monotone capacity chain in which later positions strictly dominate earlier ones. The analysis pins down how many positions are needed and which ones, both computable from forward passes alone: any subset suffices in bidirectional models, while the last few positions suffice in causal ones. We translate this into a procedure that backpropagates through only a small subset of positions while running the full forward pass over the rest. Empirically, the procedure matches or exceeds full-sequence fine-tuning across six task families---commonsense reasoning, mathematical reasoning, visual question answering, video question answering, GLUE, and image classification---and five backbones spanning bidirectional and causal architectures. It reduces activation memory substantially, translating to up to ∼ 20% lower total GPU memory (∼ 33GB) on long-context tasks, and composes cleanly with parameter-efficient methods such as LoRA and DoRA.

Parameter-efficient fine-tuning (PEFT) and probing enable adaptation of foundation models using only a small number of trainable parameters, making it attractive for video understanding where annotation and computation are expensive. However, video PEFT has focused on adapting image-pretrained models, while standard PEFT methods can also be applied to video representations. These settings are rarely compared and both confine temporal reasoning to a single component of the model, leaving open how temporal context should be distributed across backbone, PEFT and probe. In this work we provide a systematic study of model adaptation strategies for video understanding. We evaluate methods across appearance-focused, motion-focused and spatially dense settings, with a particular focus on scenarios with limited data where parameter-efficiency is most beneficial. Our results provide new insights into PEFT and probing across settings and demonstrate the importance of temporal context allocation for effective video adaptation.

Tabular prior-fitted networks (PFNs) work well across tabular classification problems but can fail silently at deployment when distributions shift. We propose **CALD** (Calibrated Anchor-Loss Disagreement), a lightweight unsupervised monitor that triggers an alarm to warn when a PFN might fail. By finetuning the decoder to maximize disagreement with the test-set prediction while maintaining accuracy on an ID anchor set, **CALD** identifies harmful OOD batches through the resulting anchor loss, requiring no labels on test batches. We prove that the post-finetune anchor loss has a strictly lower expectation under harmful shift than under ID eval, yielding a one-sided test with a formal power guarantee. Across 14 datasets covering tabular and structured distribution shift, on both TabPFN and TabICL, **CALD** matches or outperforms baselines on harmful shifts and is markedly more resilient to benign-shift false alarms.

Rehearsal-free continual learning with parameter-efficient adapters can be cast as a sequence of task-vector write-in operations: for each new task, a low-rank adapter is learned and merged into a running model. We propose Proximity Regularized Merging (PRM), a minimal modification to sequential LoRA merging that adds a proximal penalty during task-vector training without changing the subsequent write-in rule. PRM acts as a robust task-vector regularizer: in the reported Base$\rightarrow$+Prox diagnostics, it improves AAA across multiple write-in rules, backbones, and class-incremental settings, while its fixed-coefficient variant remains competitive with strong coefficient-based baselines. Mechanistically, matched-prefix norm controls and proximal-strength sweeps show that proximal training shrinks the task-vector radius, lowers Fisher-weighted interference, broadens the coefficient plateau, and exposes a stability--plasticity trade-off. Together, these results suggest that the effectiveness of sequential LoRA merging depends not only on how much of a task vector is written in, but also on whether the task vector itself has been trained to be mergeable.

A representation that scrambles the true degrees of freedom of the world cannot support reliable planning or compositional generalization. We prove that LeJEPA is guaranteed to recover a linear representation of latent variables from nonlinear observations, a property known as *linear identifiability*. What kind of world and learning objective admit this guarantee? We consider World Models with three properties (independent, stationary, and additive-noise transitions) that generate positive pairs, and study representations trained with the LeJEPA objective. Our main result: if the latent distribution is Gaussian, the representation provably achieves linear identifiability. The key insight comes from a *spectral decomposition* of the representation with respect to the transition structure: each spectral component corresponds to a degree of nonlinearity, and higher degrees are strictly penalized by alignment, making the linear map the *unique* optimum. We then prove the converse: among all latent variable distributions satisfying these three properties, the Gaussian is the *unique* one that leads to linear identifiability. The guarantee degrades gracefully when the two objectives are only approximately satisfied, with an explicit bound validated empirically. Our theory turns an empirically successful recipe into a mathematical guarantee, providing the foundation for building World Models that provably recover the structure of the world.

Pretraining noise in foundation models has been shown to significantly impair the generalization performance of downstream tasks, leading to catastrophic inheritance. However, when downstream tasks arrive sequentially, how pretraining noise affects continual adaptation, which involves the forgetting of previously learned knowledge, remains underexplored. This paper is the first to study this problem. Using asymmetric label noise as a realistic setting, we find that pretraining noise persistently degrades performance on new tasks during continual adaptation and may further exacerbate forgetting of previously adapted tasks. Further analysis shows that the persistent effect of pretraining noise mainly manifests as inherited output contraction, where the logit gap between the true class and competing classes is compressed. To mitigate this issue, we propose Ambiguous Boundary Correction (ABC), a method that performs anchor-guided neighborhood reweighting to correct ambiguous regions induced by inherited output contraction. Experiments on synthetic and real-world noisy pretrained models show that ABC improves robustness to pretraining noise under both domain shifts and class shifts.

Clinical electroencephalogram (EEG) analysis rests on a hand-crafted feature catalog refined over decades, *e.g.,* band power, connectivity, complexity, and more. Modern EEG foundation models bypass this catalog, learn directly from raw signals via self-supervised pretraining, and match or outperform feature-engineered baselines on most clinical benchmarks. Whether the two representations align is an open question, which we decompose into three sub-questions: *what does the model learn*, *what does the model use*, and *how much can be explained*. We answer them with layer-wise ridge probing, LEACE-style cross-covariance subspace erasure, and a transparent classifier benchmarked against a random-feature baseline. The audit covers three foundation models (CSBrain, CBraMod, LaBraM), five clinical tasks (MDD, Stress, ISRUC-Sleep, TUSL, Siena), and a 6-family 63-feature lexicon. Of the $945$ (model, task, feature) units, $648$ ($68.6\%$) are representation-causal and $199$ ($21.1\%$) are encoded-only. Across tasks, $50$ features qualify as universal candidates with strong support (all three architectures RC) in two or more tasks. Frequency-domain features dominate, but the other five families each contribute substantial causal mass. Confirmed features recover, on average, $79.3\%$ of the foundation model's advantage over the random baseline, with a clean task gradient (MDD $\approx 0.99$ down to Stress $\approx 0.56$): tasks near ceiling are almost fully recovered by the lexicon, while harder tasks leave a non-trivial residual that pinpoints a concrete target for future concept discovery.

Vision-Language Models (VLMs) such as CLIP have become a standard backbone for open-vocabulary recognition, yet their zero-shot predictions remain vulnerable to distribution shifts encountered at deployment. Test-Time Adaptation (TTA) has recently been extended to CLIP as a lightweight solution, leading to a rapidly growing body of TTA4CLIP methods. However, empirical progress in this area has largely outpaced our understanding of what truly drives adaptation, where their gains originate, and under which shifts they remain reliable. In this paper, we take a step back from the pursuit of state-of-the-art accuracy and conduct a systematic controlled study of TTA4CLIP. We first organize existing methods into three unified paradigms according to what is updated at test time. We then introduce TTABC, an open-source TTA Benchmark for CLIP, which standardizes evaluation protocols and integrates more than 20 representative methods. Our controlled empirical analysis focuses on three key areas. First, we determine the driving factors in parameter-based methods, revealing that adaptation gains are primarily driven by test-time evidence and reliable proxies rather than heavy optimization. Second, we explore evidence utilization beyond heavy parameter tuning, showing that competitive and efficient performance can be achieved through cross- or current-sample evidence and lightweight prototype updates. Finally, we demonstrate that there is no silver bullet for TTA; no single adaptation paradigm is universally optimal, and the preferred paradigm depends on the nature of shift. We hope our benchmark and study provide a clearer understanding of the current TTA4CLIP landscape and establish a foundation for further research.

Time-series foundation models (TSFMs) are increasingly deployed as zero-shot forecasters in observability platforms, yet the operational concepts encoded within their internal representations remain largely un-audited. To address this gap, we introduce $\texttt{toto-interp}$, a control-first interpretability protocol that pairs linear probes with four rigorous baselines: raw-feature, shuffled-label, randomized-backbone, and supervised raw-window Fourier Neural Operator (FNO) controls. Applying this harness to TOTO on BOOM, we demonstrate that three structural telemetry concepts, such as cadence bucket ("frequency"), metric type, and domain, are linearly decodable, layer-localized, and specifically tied to pretraining. A budget-matched replication on MOMENT-base successfully recovers these same axes, confirming these representations are not isolated artifacts of the TOTO architecture. Furthermore, our matched-null and on-manifold interventions reveal a critical two-way dissociation between linear decodability and forecast-sensitivity. For instance, concepts like future burstiness are highly decodable yet completely inert under intervention, whereas other weakly decodable directions can severely disrupt forecasts off-manifold. By releasing the audit harness and detailing negative control cases (such as cardinality and shift risk), we demonstrate that linear decodability alone is insufficient for reliable post-training steering or deployment auditing in TSFMs.

A central question in genomic representation learning is whether DNA foundation models~(FMs) encode information beyond simple sequence composition, or whether their apparent representational power largely reflects signals already captured by classical compositional statistics. We address this by establishing $k$-mer frequency vectors as a rigorous biologically grounded baseline and introducing a geometric decomposition that partitions FM embeddings into a component linearly predictable from $k$-mer frequencies and an orthogonal residual representing genuinely non-compositional information. By evaluating each component separately as a downstream feature space, we directly measure what FMs encode beyond composition and whether that content is beneficial, neutral, or detrimental. Across 57 genomic classification datasets and 2 gene expression regression tasks, we compare five FMs (NTv3-650M, HyenaDNA, DNABERT-2, Caduceus-Ph, Evo2-1B) against $k$-mer features ($k \in \{4,5,6\}$) under a controlled frozen-embedding protocol with a fixed Random Forest classifier. NTv3 is the only model to consistently outperform $k$-mers (63.2\% of datasets, mean MCC $= 0.562$), with gains concentrated in splice-site and promoter tasks; HyenaDNA, DNABERT-2, and Evo2 are outperformed on 86\%, 95\%, and 93\% of datasets. Wilcoxon signed-rank tests with FDR correction confirm these patterns are systematic and robust to dataset heterogeneity. Our decomposition reveals that FM advantage originates exclusively from the non-compositional residual, and only in tasks with inherently positional regulatory signals. For HyenaDNA and DNABERT-2, the residual actively degrades classification. The fraction of embedding variance explained by $k$-mers ($R^2$) is uncorrelated with downstream quality, demonstrating that the quantity of non-compositional information is not a proxy for its utility. These results show that, in the frozen-embedding regime, current DNA FM representations largely recapitulate sequence composition, and provide a principled diagnostic for predicting which task conditions are likely to benefit from non-compositional structure.

Current foundation models of electroencephalography (EEG) sought to characterize the fluctuations in neural signals, yet this objective stands fundamentally at odds with the intrinsic physiological mechanisms governing brain activity. As with their LLM predecessors, these models pack millions of neurons, even those contributing negligibly small weights, yielding a computational footprint incompatible with resource-constrained portable brain-computer interfaces (BCIs). To overcome these challenges, we propose *Vibe-Spike*, an energy-preserving and biologically-realistic EEG foundation model operating in the regime of oscillatory neural synchronization. Drawing on principles from cognitive neuroscience, our foundation model is designed to recover the masked EEG signals through the transient spikes of spatially coupled neurons, from which self-organized cortical synchronization (aka. *brain rhythms*) emerges to reciprocally modulate ongoing neural firings. Under the hood, we introduce a macro-micro learning mechanism that bridges two scales of neural computation: lower layers capture neural oscillations via a spiking neural network (SNN), while upper layers model oscillatory synchronization across large neuronal populations, with cross-scale interactions learned end-to-end. Taken together, our *Vibe-Spike* sheds new light on learning intrinsic representation of functional neural fluctuations. We pre-train the model to reconstruct segmented EEG signals directly from the synchronized spiking representations using a purely self-supervised objective on a diverse corpus of **75** public datasets. To demonstrate its universal representational power, we conduct massive-scale evaluations across ten heterogeneous downstream datasets, including clinical pathology, sleep staging, and cognitive assessment. *Vibe-Spike* not only facilitates deployment on portable edge devices through its lightweight architecture and energy-preserving inference, but also provides profound neuroscience insights. By further capturing pre- and post-synaptic firing, *Vibe-Spike* unlocks a new pathway for characterizing effective connectivity and the oscillatory synchronization that underpins cognition.

Reinforcement Learning with Verifiable Rewards (RLVR), particularly GRPO, has become the standard for eliciting LLM reasoning. However, its efficiency in exploration and difficulty adaptation remains an open challenge. In this work, we identify an implicit advantage symmetry inherent in Group Relative Advantage Estimation (GRAE) as a structural property that provides a new perspective for understanding these bottlenecks. This symmetry induces two critical limitations: (i) at the group level, strict symmetry in weights between correct and incorrect trajectories leaves unsampled action logits unchanged, thereby hindering the exploration of novel correct solution. (ii) at the sample level, the algorithm implicitly prioritizes medium-difficulty samples, remaining agnostic to the non-stationary demands of difficulty focus. Through controlled experiments, we reveal that this symmetric property is sub-optimal, yielding two pivotal insights: (i) asymmetrically down-weighting the advantages of correct trajectories encourages essential exploration but risks instability; (ii) learning efficiency can be boosted by a curriculum-like transition—prioritizing simpler samples initially before gradually shifting to complex ones. Motivated by these findings, we propose Asymmetric GRAE (A-GRAE), which dynamically modulates exploration incentives and sample-difficulty focus. Experiments across seven benchmarks demonstrate that A-GRAE consistently improves GRPO and its variants across both LLMs and MLLMs.

Geometry estimation from perspective images has greatly advanced, maturing to the point where off-the-shelf foundation models are able to reconstruct 3D scene structure not only from multi-view imagery, but even from a single view. A natural extension is 3D reconstruction from panoramas, with the exciting prospect of recovering a full $360^\circ$ scene from a single panoramic image. In this work, we introduce PaGeR (Panoramic Geometry Reconstruction), a framework to lift powerful 3D foundation models, designed for perspective imagery, to the panorama domain. Our strategy is to start from a pretrained transformer for 3D reconstruction and turn it into a unified high-performance model that predicts scale-invariant depth, metric depth, sky masks and surface normals from both perspective and omnidirectional images, a single forward pass. By keeping architectural changes to a minimum and mixing perspective and panoramic images during training, Pager retains the rich 3D prior of the underlying foundation model while learning to also estimate geometrically consistent $360^\circ$ scenes from single panoramas. We extensively test our method in both indoor and outddor environments and find that it delivers state-of-the-art performance and excellent zero-shot performance across a wide range of scenes.

Fine-grained visual classification (FGVC) with Multimodal Large Language Models (MLLMs) remains challenging, as visually similar subordinate categories inherently induce classification uncertainty due to subtle inter-class differences. Existing reinforcement learning methods typically require the model to generate a single class name and optimize it with sparse correctness rewards. However, a seemingly incorrect prediction can still provide highly informative signals if it falls within the same local confusion set as the ground-truth class. By treating all non-exact matches as equally wrong, the rigid single-answer paradigm overlooks these valuable near-misses. Consequently, it fails to capture the model's intrinsic uncertainty and provides limited learning signals. To address these limitations, we propose List-GRPO, an uncertainty-aware listwise Group Relative Policy Optimization framework. Rather than enforcing a single prediction, our method prompts the model to adaptively generate a ranked candidate list, explicitly capturing its predictive uncertainty. Furthermore, to mitigate the advantage vanishing when online exploration completely misses the target, we introduce a dynamic ground-truth anchored response strategy. We synthesize an off-policy anchor by prepending the ground truth to the most frequent hard negatives mined from online samples. To prevent the anchored response from dominating learning, we isolate the advantage estimation of online samples from the anchored response and use a gating coefficient to modulate its contribution. Extensive experiments on multiple fine-grained benchmarks demonstrate that List-GRPO significantly improves Top-1 accuracy while maintaining high ground-truth coverage with only about three candidate predictions on average.

Recent long-tailed recognition methods increasingly adopt two-stage fine-tuning, where a generic representation is first learned and imbalance-aware adaptation is then performed in a lightweight second stage. However, existing second-stage designs often allocate objective emphasis using delayed empirical summaries, such as historical class-wise performance or validation feedback, and commonly restrict adaptation to classifier heads. These choices may make routing unreliable when empirical feedback is noisy or skewed, and may also limit the feature plasticity required for rare categories. In this paper, we propose U-MOF, an uncertainty-guided parameter-efficient multi-objective fine-tuning framework for long-tailed recognition. U-MOF introduces lightweight residual adapters into a shared-backbone expert committee, uses entropy-based committee statistics as diagnostic routing proxies to modulate tail-oriented and smoothing-oriented objectives, and applies conflict-aware orthogonal projection to coordinate heterogeneous objective gradients. Experiments on CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018 show that U-MOF improves rare-category and overall performance in controlled stage-two comparisons and remains competitive with strong published long-tailed recognition baselines, while preserving the efficiency advantages of decoupled adaptation. These results indicate that reliable stage-two long-tailed fine-tuning benefits not only from selecting appropriate objectives, but also from diagnosing when each objective should dominate and from retaining limited feature-level plasticity. An anonymized implementation is included in the Supplementary Material.

Properly tuned hyperparameters are critical for reinforcement learning algorithms to perform well, limiting their use in real world applications. AutoRL aims to to address this by nesting the reinforcement learning problem within an outer optimization loop, where a black-box optimizer tunes the reinforcement learning algorithm's hyperparameters leaving the AutoRL algorithm's hyperparameters (hyper-hyperparameters) fixed. We provide the first empirical study of an AutoRL algorithm's sensitivity with respect to its hyper-hyperparameters. Our results demonstrate that SMAC3 with Hyperband, used to tune PPO, is sensitive to its hyper-hyperparameters, albeit less so than PPO is to its hyperparameters, and this reduction comes at a significant computational cost. The per-environment tuned performance gains do not outperform random search, suggesting the reduction in sensitivity may not justify the substantial compute and engineering overhead AutoRL imposes.

CLIP has become a cornerstone of multimodal representation learning, yet improving its performance typically requires a prohibitively costly process of training from scratch on billions of samples. We ask a different question: *Can we improve the performance of open-weight CLIP models across various downstream tasks using only existing self-supervised datasets?* Unlike supervised fine-tuning, which adapts a pretrained model to a single downstream task, our setting seeks to improve general performance across various tasks. However, as both our experiments and prior studies reveal, simply applying standard training protocols starting from an open-weight CLIP model often fails, leading to performance degradation. In this paper, we introduce **TuneCLIP**, a self-supervised fine-tuning framework that overcomes the performance degradation. TuneCLIP has two key components: (1) a warm-up stage of recovering optimization statistics to reduce cold-start bias, inspired by theoretical analysis, and (2) a fine-tuning stage of optimizing a new contrastive loss to mitigate the penalization on false negative pairs. Our extensive experiments show that TuneCLIP consistently improves performance across model architectures and scales. Notably, it elevates leading open-weight models like SigLIP (ViT-B/16), achieving gains of up to +2.5% on ImageNet and related out-of-distribution benchmarks, and +1.2% on the highly competitive DataComp benchmark, setting a new strong baseline for efficient post-pretraining adaptation.

Step-size selection remains a central challenge in large-scale neural network optimization: conservative steps slow convergence, while aggressive steps can destabilize training. We propose Zero-and-First-Order optimization (ZFO), a lightweight framework that decouples direction selection from step-size selection. ZFO uses a trusted first-order optimizer to determine the direction and performs zeroth-order evaluations only along this one-dimensional subspace to choose how far to move. Using the current forward-backward pass and two additional fixed-batch function evaluations, ZFO constructs a local model of the objective along the proposed direction and selects a curvature-aware step within a bounded search interval. This yields an adaptive step-selection mechanism with lower cost than full line search. We provide theoretical guarantees showing that shared-sample evaluations produce reliable finite-difference curvature estimates, that the induced local model selects a near-optimal step along the search interval, and that ZFO converges to a neighborhood of a stationary point. Across multiple language models and datasets, ZFO improves accuracy and stability over standard first-order fine-tuning algorithms.