Frozen Representations

An AI evaluation can be perfectly reproducible and still support the wrong claim. This risk is acute in closed-loop systems: policy determines visited states, observable components, and which failures leave a measurable trace. We propose a claim-safe protocol with three actions. Refuse: abstain when a clean reference stream or matched runtime comparison lacks support. Decompose: report protocol execution, operational false admission, and structural hypotheses separately rather than as one PASS/FAIL label. Refresh: treat distribution-shift alarms as requests to invalidate and recompute a reference map, not as fault evidence. We instantiate the protocol in an aggregate-only simulator with 24 policy components, three demand regimes, two fault-mask families, and independent development and heldout seeds. The preregistered heldout contains 1,440 cases and 21,600 partition rows. Only 55/72 regime-component units were reference-admitted and 54/55 remained runtime-admitted, making abstention part of the result. Stable false admission was 0/20 represented components, with a one-sided exact 95% upper bound of 0.1391 under a frozen 0.20 rule. Within admitted units, affected clean traffic outpredicted nominal fault-cell fraction: across 540 unit-arm rows nested in 20 component clusters, the cell-minus-traffic negative-log-likelihood difference was 0.1264 nats per row, with a 95% component-cluster interval of [0.0593, 0.1918]. A drift log shows why "null" must be reference-relative: clean fault-null streams triggered 15/15, 0/15, and 14/15 alarms across three regimes, while only the middle regime matched the frozen detector reference. Rather than a universal threshold, we contribute an executable contract linking observable support, statistical calibration, and justified claims.

Text-based person retrieval under a sim-to-real gap (synthetic training data, a real-image gallery) is usually tackled with costly fine-tuned cross-encoders. We ask whether a frozen-encoder system can compete. We present SCOUT, which casts cross-modal retrieval as prediction in embedding space. A trainable predictor maps the patch tokens of a frozen video encoder into the embedding space of a frozen text encoder under a bidirectional InfoNCE objective, and no encoder is fine-tuned in the base model. The video encoder is V-JEPA, the text encoder is EmbeddingGemma, and the predictor is initialized from a Qwen3.5-0.8B decoder. We make three findings. First, the best frozen text encoder is simply the one whose geometry best matches the video features. A training-free alignment score ranks three candidate text encoders in the same order as their retrieval accuracy on our held-out split (Spearman $ρ= 1.0$); a fourth, LLM-based encoder shows the rule is metric-dependent, holding for a neighborhood-overlap score ($ρ= 0.8$) but not for a linear probe ($ρ= -0.2$). Second, two precision-targeted levers, parameter-efficient ExPLoRA adaptation of the video encoder and a training-free attribute-decomposed reranker built on a vision-language model, improve the top-rank precision that otherwise limits the frozen system, adding 2.2 points of leaderboard R@1. Third, a local-versus-public calibration study explains which interventions transfer to the real domain. On AI City Challenge 2026 Track 4 the full retrieve-fuse-rerank system reaches 84.25 mAP@10 on the final leaderboard, while a single frozen model submitted alone reaches 60.63. Our trained components cost about 95 GPU-hours. CMP, the dataset authors' fine-tuned cross-encoder that trains for sixteen GPU-days, is one fusion member of the full system, not an alternative. Code and annotations: https://github.com/abtraore/SCOUT-ECCV

Black-box model attribution is increasingly relevant when large language models (LLMs) are served through relay and reseller APIs. A tempting low-cost signal is the prompt-token count returned by an OpenAI-compatible endpoint: two models that share a tokenizer and chat template may produce the same count sequence up to a fixed offset. Yet the validity of this signal for broader \emph{model-family} attribution has received little direct holdout testing. We conduct a frozen-threshold study over 24 labeled endpoint pairs, split evenly into a development set and an untouched holdout set, with three temporal repeats and 30 controlled texts per pair. We introduce a validity-gated result contract that distinguishes an observed dissimilarity from an uninformative measurement caused by missing usage data, rate limits, or endpoint policy. The resulting shift-invariant exact-match score perfectly separates the 12 development pairs, yielding a frozen threshold of 0.725. On holdout, however, only 6 of 12 pairs are eligible under the pre-specified three-repeat rule. Among eligible pairs, balanced accuracy is 0.75, sensitivity is 0.50 (95% Wilson interval 0.15--0.85), and specificity is 1.00 (0.342--1.00). Two same-family pairs---Qwen 3.8 and DeepSeek V4 variants---fall below the frozen threshold. Across 4,320 formal API calls, every log is replayable, while holdout contains 189 non-200 responses and 157 successful responses without prompt-token usage. The study therefore validates token-count consistency as a fingerprint of a shared \emph{tokenization stack}, but rejects its use as a standalone necessary test for model-family lineage.

Understanding motion in daily living requires context beyond kinematics, because similar inertial patterns during activities of daily living (ADLs) can reflect intentional stopping, object interaction, or pathological movement impairment. Egocentric vision provides task-related context that may help disambiguate these cases. We investigate this challenge through freezing of gait (FOG) detection in Parkinson's disease (PD), a symptom strongly influenced by contextual factors during ADLs. Using synchronized egocentric video, wearable IMUs, and expert-annotated FOG labels collected from 13 PD participants in their homes, we evaluate frozen representations from pretrained ego-video and time-series foundation models, alongside an IMU-based TCN trained from scratch, under leave-one-subject-out evaluation. The IMU-based TCN achieved the strongest event-detection performance, reaching 42.3 F1 and 83.0 AUROC, compared with 32.6 F1 and 77.2 AUROC for V-JEPA2 ego-video features. Although ego-video alone did not outperform IMU-based sensing, it showed above-chance discrimination, and qualitative analyses suggest that egocentric vision may capture FOG-relevant information independent of IMUs. Together, these results support the use of pretrained ego-video representations to add contextual information to wearable-sensor-based clinical motion understanding in daily living.

Modern sequence models, from Transformers to State Space Models, have enabled powerful generative modeling across diverse domains, yet they are typically trained to predict what happens while treating when it happens as a secondary concern. In data-mining settings where events are associated with explicit timing information, this separation can limit temporal reasoning, anomaly detection, and faithful reconstruction of event chronology. A common strategy is to treat timing as an auxiliary signal, training a separate timing model using representations learned solely for event prediction. However, this two-stage approach implicitly assumes that representations optimized for event prediction already contain sufficient temporal structure. We introduce ChronoSSM, an autoregressive State Space Model (SSM) that jointly models events and timestamps with a shared backbone trained using combined token and temporal generation objectives. We compare the joint regime, where temporal supervision updates the backbone, with the two-stage regime, where timing is learned only using the frozen event representations. Across four domains spanning dense and partial timestamp supervision, joint training consistently makes inter-arrival information more recoverable from frozen representations without any systematic degradation in content-generation quality overall. Our results show that temporal supervision can produce more temporally informative representations without materially degrading autoregressive event modeling.

The Independent Chip Model (ICM) converts tournament chips into reference prize equity, and policies are routinely constructed against those values. Because ICM reads only stack sizes, it omits action order, blind obligations, and seat rotation, and it does not price the elimination pressure a big stack puts on the short stacks it can bust. Those omissions can alter the successor-state contrasts that determine a move. We introduce Strategic-Continuation Optimization (SCO), a policy-construction method that enumerates current-hand outcomes, maps them to successor states, prices those states with continuation values computed from the finite tournament model, and optimizes and freezes the resulting current-hand policy. The fixed-ICM comparison policy changes one thing only: the same optimizer solves the same game with successor states priced by analytic ICM, so the two policies differ only through that pricing. We evaluate the resulting policies in a three-player jam/fold tournament with a $1M prize pool. Relative to the frozen strategic-continuation benchmark, analytic ICM has $9{,}036 mean absolute value error across all 2,838 state--seat entries. That value error rewrites the ranges it prices: measured against each decision point's own fixed-ICM jam range, SCO moves the jam frequency by an average of 14.08%. To price those different moves, we compare all 946 states and three policy owners while changing only the focal policy and holding both opponents and the continuation evaluator fixed. The policy produced by SCO earns $214.33 more prize equity per hand on average and is favored in 2,433 of 2,838 matched units. The ordering survives replacing the solver-built opponent with two LLMs and with a family of non-modeling threshold players. This value-to-policy-to-cost chain shows directly when ICM becomes an inadequate objective for tournament strategy construction.

Current LLM safety guardrails face a fundamental tension: fine-tuning distorts pre-trained representations while generative judges incur prohibitive inference costs. We ask a complementary question: how far can safety be achieved through pure geometric reasoning over frozen representations, and where does it fail? We present HoloAegis, a minimally parametric topological inference framework that decouples representation from reasoning: an un-fine-tuned encoder maps text to the unit sphere S^{d-1}, and all decisions reduce to Gibbs-Boltzmann free-energy differences over pre-computed anchor centroids. We contribute a boundary-mapping study rather than a leaderboard claim. On a frozen three-benchmark protocol, HoloAegis (3.2 MB) statistically matches WildGuard-7B (14 GB) on toxicity (0.96 vs. 0.96), exceeds it on harmful behaviors (0.99 vs. 0.79), and cedes oversafety detection (0.62 vs. 0.98) -- while ShieldGemma-2B fails on indirect harms (0.34). These failure modes are complementary and mechanistically traceable: potential-difference scoring senses manifold clustering, whereas policy-conditioned LLM judging requires explicit taxonomy matching. We restate our Topological Boundary Stability conjecture in ratio form and validate it via reference-set bootstrap: anchor banks reduce score variance 4-15x and boundary displacement to approximately 0.44 + 0.23 sqrt(k/K) of the full-space estimator. Per-domain analysis further reveals that geometric separability tracks within-domain semantic homogeneity. Our results chart where geometric guardrails substitute for, and where they must defer to, LLM judges.

Routine CT interpretation is inherently comprehensive, capturing incidental findings across the entire scan volume. 3D CT foundation models could assist this process by providing generalizable representations of anatomy and pathology. To evaluate their diagnostic breadth, we benchmark ten frozen CT encoders across three cohorts of thoracic CT scans, including an unseen internal clinical dataset, using $k$-nearest neighbors, zero-shot prompting, and linear probing. We find no universal state-of-the-art, with rankings fluctuating significantly depending on the evaluation context. While models combining fine-grained image tokenization with vision-language alignment generally perform best, a lightweight supervised encoder remains highly competitive, demonstrating that explicit labels can effectively substitute for scale. Crucially, rather than model architecture, we observe that the primary determinant of performance is a physical bottleneck: a finding's detectability scales with its contrast against surrounding tissue and its spatial extent. Through controlled within-organ comparisons, we empirically demonstrate that widespread or high-contrast abnormalities, such as devices and effusions, are reliably recovered. Conversely, small, low-contrast focal lesions remain a persistent challenge across all evaluated encoders. We attribute this to the inherent limitations of globally pooled embeddings, suggesting that accurately representing small, low-contrast structures will require region- or lesion-level pretraining.

Self-supervised pretraining objectives are spatially uniform: the teacher temperature and the per-patch loss weight are identical everywhere in the image, so a lesion a few patches wide contributes no more to the training signal than the surrounding parenchyma. Prior work biases the views toward annotated regions, which changes what the model sees but adds no pressure on the objective. We instead condition the targets of self-distillation, a method we call SALT (Spatially Adaptive Label-guided Temperature). Weak, box-derived labels, available only during pretraining, define a compact region on the encoder's patch grid, inside which the teacher's softmax temperature is sharpened and the masked-patch loss is up-weighted. The objectives, the masking policy and the centering statistics are otherwise unchanged, and at every downstream use the encoder is a plain feature extractor with no labels and no conditioning. We evaluate by freezing the encoder and training only a lightweight multi-depth CenterNet-style head, detecting lesions in 3D on four CT cohorts, and we isolate the mechanism against a backbone identical in architecture, pretraining data, schedule and label-guided cropping but with no target conditioning. We report patch-level separability, 3D detection stratified by cohort and by lesion size, box quality, and a detector-free probe in which a single frozen patch embedding re-identifies a lesion in a follow-up scan without registration, masks or fine-tuning. Because the conditioning is expressed through a spatial indicator rather than through label semantics, the formulation admits any weak spatial annotation; we instantiate and validate it for lesions.

Recent work has shown that a simple linear probe on frozen representations from modern vision foundation models (VFMs) can achieve state-of-the-art AIGI detection performance, substantially outperforming specialized detectors in challenging in-the-wild scenarios. This finding has established DINOv3 as the dominant foundation-model baseline for subsequent improvements. However, we find that the vision-language model Perception Encoder (PE) holds greater potential for AIGI detection, because its language-aligned representation preserves high-level provenance semantics. Specifically, PE exhibits stronger local provenance organization than DINOv3 in its frozen feature space. However, semantic-agnostic linear probing fails to exploit this structure, as PE-Linear still underperforms DINOv3-Linear by 4.1% on In-the-Wild. Based on this observation, we propose Semantic Prototype Calibration (SPC), which constructs category prototypes from forensic semantic information and calibrates them with supervised data. We apply SPC to PE and refer to the resulting detector as PE-SPC. Our analysis shows that this simple design achieves stronger generalization. Across cross-generator, post-processing, and in-the-wild benchmarks, PE-SPC surpasses the previous DINOv3 baseline and achieves new state-of-the-art results.

Tracking links observations of the same object through visual change, yet cannot by itself determine when the object is empty or filled, intact or cut. We formulate identity-conditioned state-moment retrieval: given a tracked-object history and alternative state descriptions, localize an interval in which each described state holds. Absolute image-text similarity scores descriptions independently; because every visible frame depicts the same target, shared object compatibility can obscure the state evidence needed to identify the target interval. The alternatives provide the missing reference: evidence for one state should be measured against the others. We introduce Déjà Cue, a training-free framework that turns these alternatives into a vocabulary-relative coordinate system. It subtracts their state-balanced centroid from each description, calibrates frame scores, and scans multiple durations within contiguous visible runs using a frozen encoder. On 78 VOST histories, holding the temporal scan fixed and changing only the query reference nearly doubles R@1 at tIoU 0.5 from 10.3% to 20.5% and raises Top-1 tIoU from 16.0% to 21.5%. Candidate-rank analyses show that vocabulary-relative queries rank useful intervals higher within the same candidate set. Related state descriptions can therefore serve as an object-specific, query-time coordinate system for reading frozen visual representations.

Contrastive vision-language models such as CLIP map semantically opposite phrases (e.g., "a dog" vs. "not a dog") to nearly identical embeddings, rendering them insensitive to negation. We attribute this failure to a phenomenon we call Representational Collapse: by tracking compositional divergence and visual alignment across the CLIP text encoder, we show that middle layers build compositional syntax, but the final layers collapse this structure as visual alignment rises, producing a syntax-blind final representation. To recover the lost negation signal without altering pretrained weights, we propose PeakPatch, a lightweight post-hoc correction system that intercepts the encoder at its compositional peak. An Embedding Correction Network (ECN) uses cross-attention to extract a negation-specific signal from the peak layer, anchored to a stable baseline, and predicts a deviation vector that re-injects the lost syntax into the final-layer embedding space. A complementary Score Correction Network (SCN) predicts bounded scalar score offsets for discriminative tasks. Both modules are trained jointly end-to-end while all CLIP parameters remain frozen, adding only 5.2M parameters (3.5% of the backbone) and preserving the standard cosine similarity interface. On NegBench, PeakPatch achieves 74.3% on COCO MCQ (+35.1 over CLIP, +17.8 over the best encoder fine-tuning method) and 65.5% on VOC MCQ, while outperforming all fine-tuning baselines on fully out-of-distribution negation retrieval despite training only 3.5% of the parameters. The corrected embeddings also transfer to text-to-image generation (+18.4 negation score) and generalize across ViT-B/32, ViT-L/14, and SigLIP backbones. Project URL: https://stevencylu.github.io/PeakPatch/.

We study how far a deliberately simple behavioral-cloning policy can progress in a visually rich first-person game before adding reinforcement learning or explicit memory. Cortex is a compact Quake policy with 10.98 million trainable parameters in a six-layer transformer over a frozen DINOv3 encoder. It is trained on the Quake subset of the public Pixels2Play corpus: 6,849 recordings (about 474.7 hours), represented as 17.09 million cached decision frames with keyboard and mouse actions. One sampled training epoch uses 517,048 four-frame windows and takes 3.3 minutes of policy-head optimization on one RTX 5080, excluding one-time feature extraction. We evaluate two independent batches of 20 stochastic, 120-second episodes on Quake E1M1. Cortex does not complete the level, but every episode reaches the opening door, button room, and gate descent; 19 of 20 episodes in each batch record at least one kill. Under the same time-controlled harness, released P2P-150M and NitroGen checkpoints remain shallower in five matched-duration episodes each. These comparisons are limited by small reference samples and different native interfaces. Ablations show that denser visual tokens improve combat and survival, while longer optimization and naive action history improve offline metrics without consistently improving play. The remaining failures are consistent with covariate shift and motivate targeted corrective data. We release the policy implementation, checkpoint, and a representative rollout.

Vision-based heading prediction is useful for UAV--UGV cooperation, but accurate prediction alone does not guarantee that every predicted heading should be issued directly as a control command. This paper investigates the decision problem of when and how a fixed vision-based heading predictor should be trusted for command issuance. A lightweight confidence-gated framework is proposed in which execution decisions are made using two interpretable reliability proxies derived from the perception stream: bounding-box area as a visibility-related proxy and short-window variation in predicted heading as a stability-related proxy. During low-confidence intervals, the framework compares the baseline freeze-HOLD policy with a bounded-blend fallback that updates the issued command conservatively. The method is evaluated on a real UAV--UGV dataset under clean and perturbed conditions. The results show that confidence gating creates a clear trade-off among execution rate, executed-frame accuracy, issued-command accuracy, and smoothness. The results further show that sparse execution can cause severe stale-command error under the baseline freeze-HOLD policy, whereas the bounded-blend fallback substantially improves command-level behavior under the same gate decisions. These findings highlight that reliable perception-driven autonomy depends not only on prediction accuracy, but also on decision-aware command issuance during low-confidence

Text-to-motion models are competent at the action a prompt names but unreliable at when each stroke lands: four punches alternating left and right rarely return four separable strokes. We introduce typed temporal events called Action Units (AUs) that make the individual stroke -- its body track, action class, time window, and impact timing -- an explicit conditioning signal. We ground a frozen text-to-motion backbone on the AU set through a lightweight gated adapter injecting two streams (per-stroke tokens and a per-frame phase channel), and at inference close residual timing errors with a training-free classifier gradient from a frozen frame-level detector. We measure per-stroke control on StrokeBench, whose prompts specify count, ordering, track, and core-frame placement, paired with an audited stroke corpus. AU grounding markedly raises the rate of correctly placed single strokes over the strongest prior interface, at the best motion quality among text-, interval-, and frame-level baselines. The prompted core frame emerges as a further steerable axis.

Fake speech detectors are increasingly challenged by the development of new and more accurate generative models. To cope with this problem, continual learning techniques are nowadays widely considered feasible strategies for updating models to new datasets, but they also lead to decreased performance on previously seen samples (catastrophic forgetting). In this work, we propose a forgetting-resilient solution based on the adoption of domain translators within a frozen detector, which remaps the new feature spaces into the original ones by means of a traceback translator network. Experimental results show that this strategy enables the achievement of high detection rates with respect to traditional retraining, while minimizing the computational effort and preserving the detection accuracy on previous data.

Cross-modal distillation from Vision Foundation Models (VFMs) to LiDAR backbones has recently emerged as a self-supervised pretraining strategy that reduces reliance on dense point-wise annotation for 3D scene understanding. However, existing distillation pipelines typically treat the VFM as a frozen feature source and train a heterogeneous 3D backbone to match fixed image embeddings, forcing the student to bridge both the modality gap and the cross-architecture gap between dense ViT token representations and sparse 3D encoders. We propose TOLiD, a self-supervised pretraining method for LiDAR representation learning that addresses this gap by coupling a LiDAR backbone with a student Vision Transformer (ViT) initialized from a frozen VFM teacher and applying supervision over compatible patch-token representations. TOLiD converts the set of point features within each image patch frustum into a token using Frustum Pooling followed by Frustum Attention, and performs token-level distillation with visibility masking. For LiDAR-only deployment, we lift token features back to per-point representations using masked bilinear sampling to avoid patches that have limited LiDAR points. We extensively evaluate TOLiD on five heterogeneous LiDAR datasets and four cross-sensor adaptation pairs, demonstrating improved transfer with frozen backbones and lightweight heads.

Training in reduced floating-point precision can silently halt learning: when a gradient-descent weight update falls below half the unit in the last place (ULP) of the weight, it rounds away and that coordinate freezes while its gradient is still nonzero. The freeze is deterministic, governed by a per-coordinate half-ULP condition, and predictable from a high-precision trajectory and the target mantissa length alone, without low-precision data. In a small GPT trained under the standard AdamW-plus-cosine recipe with bf16-equivalent stored weights, training proceeds normally and then permanently freezes just past mid-run, within four steps of the a-priori prediction. In a $124$-million-parameter GPT-2 transformer whose weights are constrained to the $8$-bit floating-point grid after every optimizer step, with no master weights, the dense weights freeze at initialization in both fp8 formats -- predicted \emph{a priori} from an fp32 reference -- and validation loss plateaus while full precision keeps improving. Stochastic rounding removes the persistent freeze, and the same reference predicts that too. The condition transfers across frozen-feature regression, a mantissa-truncation emulator spanning $128\times$ in precision, small networks, and a CNN on MNIST: a computable axis of low-precision training, not diffuse noise.

Few-shot segmentation asks a model to delineate a target class in a query image from only a handful of annotated examples, a setting most acute in remote sensing, where labels are scarce and the imagery departs sharply from the natural images on which vision backbones are pretrained. Prevailing approaches either train a segmenter on labelled episodes, which raises accuracy within the training distribution but binds the model to it, or reduce each class to a lossy summary of frozen features, a single prototype, a few cluster prototypes, or a discrete clustering, none of which preserves the internal structure of a multimodal class. We argue that a class is better described by a distribution than by a point, and that frozen self-supervised features already carry enough structure to estimate that distribution directly. We introduce FROST, a training-free few-shot segmenter that treats the reference foreground and background as two point clouds on the unit sphere of frozen DINOv3 features and labels each query token by a nonparametric density ratio, with a threshold the Bayes rule fixes at zero under equal priors. Because the variance of a density estimate shrinks as its sample grows, the decision sharpens as references accumulate, and every remaining quantity from the kernel bandwidth to the spatial gate is read from the support set rather than tuned. We develop FROST for overhead imagery, where a class is typically a scatter of many small and dissimilar instances that a density tracks but a lossy summary blurs. Across seventeen remote-sensing benchmarks FROST surpasses both training-free and learning-based methods, leading by 5.6 mIoU from a single annotated example and widening its lead as the support set grows, all while remaining among the smallest models compared. Code is available at https://github.com/jhpark-ai/FROST.

Electroencephalography (EEG) foundation models aim to learn generalizable representations from large-scale brain recordings. However, the role of temporal feature extractors and whether pretrained time-series foundation models (TSFMs) can be effectively transferred to this setting remains underexplored. We conduct a controlled comparison of three temporal feature extraction strategies, including a linear baseline, a convolutional encoder, and a frozen pretrained TSFM (MOMENT), within a unified EEG foundation model. We evaluate their impact on representation quality using two downstream tasks: motor imagery and emotion recognition. Results reveal different trends across the evaluated benchmarks. On the motor imagery dataset, simple temporal representations perform competitively, whereas the emotion dataset benefits from richer temporal modeling. Although not specifically adapted to EEG, the pretrained TSFM serves as an effective temporal feature extractor, suggesting that general-purpose time-series representations can be transferred as frozen temporal feature extractors within EEG foundation models.

Visual grounding often fails even when the target object is present in the proposal pool, because the language-side referent is bound to the wrong region. We study this binding failure under frozen perception and ask whether an explicit query--region alignment hook, together with a perception-grounded abstention mechanism, can improve faithful grounding without retraining the detector or the vision-language backbone. HKVLM freezes a language-aligned open-vocabulary detector for localization and learns a lightweight hook that maps referential query embeddings to detector proposals in a shared space; a verifier abstains when no region sufficiently supports the query. We prove an exact proposal-level diagnostic decomposition, $(1-\mathrm{SeeErr})(1-\mathrm{SayErr})$, separating proposal-coverage failures from conditional binding failures, and a monotonicity result that characterizes the faithfulness--recall trade-off induced by abstention. Across RefCOCO, RefCOCO+, RefCOCOg, and POPE, HKVLM improves over untrained and trained matched-perception binding controls and substantially reduces hallucination through abstention. Strong coordinate-decoding and end-to-end fine-tuned baselines remain much higher in raw grounding accuracy, and a reasoning-stress set exposes binding as the main current bottleneck. We therefore present HKVLM as a diagnostic and mechanism-level study of query--region binding under frozen perception, not as an absolute localization leader.

This technical report presents submission systems for Task 7(domain-incremental audio classification) of the DCASE 2026 Challenge. The main obstacle is that, the system is unable to access to past or future domain's data at once. We approached domain-incremental learning (DIL) as a frozen-feature replay problem. At each incremental stage, one or two compact experts are trained and then kept fixed; at the final stage, the penultimate features from all frozen experts are concatenated and used to train a lightweight per-class prototype classifier solely on cached features. This design prevents catastrophic forgetting by preserving each expert models at inference. To retain earlier-domain knowledge without storing raw audio, some experts were trained with DeepInversion-based generative replay. A cross-stage regression imputer was trained to fill the expert feature slots that did not yet exist at an ealier stage. We submit four fully DIL-compliant systems: three systems based on diverse frozen five-expert backbones and their cross-stack ensemble achieving 78.15% micro / 77.03% macro on the development set, outperforming every individual backbone on both evaluations.

We study whether pretrained video foundation models encode intuitive-physics information in their frozen representations, and how this information varies across model families, layers, and probe types. Using frozen-feature probing on IntPhys2 and Minimal Video Pairs (MVP), we compare predictive joint-embedding models (V-JEPA), masked reconstruction models (VideoMAE), and a diffusion-based video generator (LTX-Video). V-JEPA achieves the strongest overall results across benchmarks, especially with probes that model temporal dynamics, while VideoMAE remains competitive and LTX-Video recovers weaker but non-trivial signal. Layerwise analyses show that physics-relevant information is weakest in early layers and becomes most accessible at intermediate-to-late depth, and temporal controls show that disrupting frame order substantially reduces performance, especially on MVP. Together, these results suggest that intuitive-physics knowledge emerges reliably in pretrained video representations, but its accessibility depends strongly on pretraining paradigm, representational depth, and readout mechanism.

Texture segmentation stresses foundation segmentation because meaningful regions are defined by material or repeated appearance rather than object identity. Segment Anything Models (SAMs) often fail by default on such texture-defined partitions, but this failure is ambiguous: the texture evidence may be absent, missing from the proposal bank, or present but selected or assembled incorrectly by an object-centric readout. We ask what texture-relevant evidence is already preserved in frozen SAM before adaptation. We study two frozen evidence spaces: multiscale features, probed with a minimal clustering readout, and the automatic proposal bank, treated as evidence for a supervised consolidation readout. SAM is frozen throughout; we do not fine-tune the backbone or retrain the proposal generator. Across RWTD, STLD, an ADE20K-selected refined-crop complement, and a ControlNet-stitched PTD bridge archive, frozen SAM is not a texture segmenter by default, but its failures are not simple texture blindness. Coarse frozen features preserve texture organization, and proposal banks often contain texture-aligned masks or fragments. Natural scenes more often require assembly and commitment over fragments, while cleaner synthetic cases more often reduce to selecting an already coherent proposal. Default mask failure should therefore be decomposed into representation evidence, proposal-bank support, readout mismatch, and commitment failure.

Progress in genomic foundation models is difficult to assess due to fragmented benchmarks, incompatible evaluation protocols, and task-specific reporting. As a result, claims of superiority or generality across models are often not directly comparable. We introduce GENEB, a large-scale diagnostic benchmark that evaluates frozen representations from 40 genomic foundation models across 100 tasks spanning 13 functional categories under a unified probing-based protocol, including few-shot regimes. GENEB enables controlled comparison across model scale, architecture, tokenization, and pretraining data while explicitly exposing task-level trade-offs. Our analysis shows that aggregate leaderboards are unstable: model rankings vary sharply across task categories, scale provides only modest and inconsistent gains, and architectural and pretraining alignment frequently outweigh parameter count. These results highlight limitations of current evaluation practices and position GENEB as a reference framework for principled comparison and category-aware model selection in genomic machine learning.

Blind image quality assessment (BIQA) predicts perceived image quality without access to a pristine reference and is fundamental to applications such as image compression, transmission, and restoration. Recent BIQA methods increasingly rely on large vision-language models (VLMs). Although frozen VLMs provide an efficient alternative to computationally expensive full fine-tuning, it remains unclear how much performance is sacrificed by not adapting the backbone and, more importantly, under what conditions such adaptation is truly beneficial. Answering this question, however, is complicated by the widespread use of image-level splitting on synthetic-distortion benchmarks, where distorted versions of the same reference image can appear in both training and test partitions. This content overlap artificially inflates the apparent performance of frozen representations, masking their true generalization ability and potentially leading to incorrect conclusions about the value of backbone adaptation. We therefore address these two issues jointly. We develop an efficient BIQA framework that fuses a natural-scene-statistics descriptor with frozen SigLIP and CLIP-H embeddings through a lightweight regression head, and then apply parameter-efficient Low-Rank Adaptation (LoRA) to the SigLIP backbone, training only $0.23\%$ of its parameters. Evaluating both frozen and adapted models across six datasets under image-level and reference-level protocols, we find that image-level splitting inflates frozen-feature SROCC by up to $0.44$ and masks wide variation in true difficulty, which reference-level evaluation reveals. Under this content-independent protocol, LoRA adaptation recovers performance in proportion to the exposed difficulty, with the largest gains where frozen features generalize poorly (up to $+0.357$ SROCC on TID2013) and little benefit where they are already strong.

Vision--language--action (VLA) policies are trained to imitate actions; their loss never asks them to estimate reward, progress, or future success. Their frozen representations nevertheless carry such information, and it can be read out and used to guide action choice without retraining the policy. From mixed successful and failed manipulation trajectories on LIBERO-Goal, we recover Monte-Carlo outcome targets using lightweight linear probes on frozen features. The targets are consistently predictable from OpenVLA, Pi0.5, DINOv2, and CLIP features, and substantially less so from baselines built on progress, time-to-go, task identity, or proprioception. To rule out task and temporal shortcuts, we evaluate the probes under same-task, same-timestep matched comparisons: Pi0.5 probes still reach roughly 92% pairwise ordering accuracy, while label-shuffled controls stay at chance. Used as a test-time selector over sampled Pi0.5 action prefixes, the same probe turns this offline finding into behavior: on push-plate, success rises from 26.7% under greedy decoding to 44.3%, with a second positive case on wine-rack. The gains are not universal and require additional inference compute, but the underlying finding is clean: frozen VLAs already encode information about success that their imitation objective never explicitly demands.

Spatial intelligence requires visual representations that capture both semantic objects and geometric structure in the physical world. To support this, two major pre-training schemes are now widely used as foundation backbones: Vision-Language Models (VLMs), which use language supervision to align visual observations with semantic concepts, and Video Generation Models (VGMs), which learn from temporally evolving visual worlds. However, it still remains unclear which pre-training scheme provides a better representation substrate for spatial intelligence. In this paper, we present the first systematic frozen-feature probing study of VLMs and VGMs across three representative axes of spatial intelligence: semantic tagging, instance grouping, and 3D geometry prediction. Using the lightweight probe, our framework enables a controlled comparison of what information is already encoded in frozen representations from two model families. Experimental results reveal a clear complementarity: VLMs are stronger at semantic tagging and instance grouping, while VGMs provide more accessible signals for dense geometry and camera motion. Moreover, a naive fusion of the two already yields a representation that excels at both geometry and semantics, suggesting a promising direction for building stronger spatial-intelligence backbones by effectively integrating features from both model families. Our code is available at \href{https://github.com/om-ai-lab/Probing-VLM-VGM}{https://github.com/om-ai-lab/Probing-VLM-VGM}.

Frozen Vision Foundation Models (VFMs) with lightweight classification heads are increasingly used in medical imaging because they offer efficient and reproducible deployment. Yet noisy-label learning methods for this frozen-feature regime remain poorly understood, and most existing methods still rely on a small-loss assumption inherited from end-to-end training. We present a controlled benchmark of eight noisy-label methods across five medical datasets, three backbones, two noise types, and five noise rates (150 conditions, 6,000 training runs), evaluated with balanced accuracy. The benchmark shows that there is no universal winner: Friedman ranking over the 150 conditions yields $χ^2 = 333.2$ ($p = 4.77 \times 10^{-68}$), ELR wins the most conditions (49/150), while CUFIT attains the best mean rank (2.51). The practical cost of method choice grows sharply with noise severity, from 4.5pp on clean data to 18.8pp at asymmetric 40% noise. To explain these benchmark-level patterns, we revisit the small-loss assumption in a representative high-risk regime. Under frozen DINOv2 features, clean and noisy loss distributions overlap by 53--61%, and matched-rate clean-sample detection shows that prediction agreement is markedly more stable than loss ranking under asymmetric noise (3pp vs.\ 13pp precision drop). On ISIC2019 with asymmetric 40% noise, Co-Teaching reaches 68% overall accuracy while collapsing to 35.1% balanced accuracy with zero recall on three minority classes. Together, these results recast noisy-label learning for frozen VFMs as a regime-aware method-selection problem rather than a search for a single dominant algorithm. We conclude with evidence-based guidance and a low-regret feature-space selector for practical recommendation.

Understanding how linguistic structure emerges in language models is central to interpreting what these systems learn from data and how much supervision they truly require. In particular, semantic role understanding ("who did what to whom") is a core component of meaning representation, yet it remains unclear whether it arises from pre-training alone or depends on task-specific fine-tuning. We study whether semantic role understanding emerges during language model pre-training or requires task-specific fine-tuning. We freeze decoder-only transformers and train linear probes to extract semantic roles, using performance to infer whether role information is already encoded in pre-training or learned during adaptation. Across model scales, we find that frozen representations contain substantial semantic role information, with performance improving but not fully matching fine-tuned models. This indicates partial but incomplete emergence from pre-training alone. We show that semantic role structure emerges from language modeling objectives, but its internal implementation shifts toward more distributed representations as model scale increases.

Probing frozen vision transformers typically uses permutation-invariant aggregation (GAP or $\texttt{[CLS]}$), treating patch tokens as an unstructured set. Content-dependent probes such as self-attention are useful accuracy controls, but they do not expose a fixed token schedule or fixed position weights for auditing. We introduce $\textbf{SSMProbe}$, an explicitly inspectable probe that replaces invariant pooling with a Sinkhorn-learned evidence route followed by a diagonal S4 decoder. The S4 decoder is a linear time-invariant (LTI) system whose final state has fixed, position-dependent coefficients, so the probe-induced routed sequence can be audited as a concrete object rather than inferred only from accuracy. Our central measurement is the geometry of routed evidence: which patch tokens are moved to influential positions by this diagnostic, whether those tokens form spatially organized regions or random-like dispersed sets, and how the fixed S4 kernel weights them. Across MAE, BEiT, DINOv2, and supervised ViT, this route geometry separates MAE's dispersed, nearly random-like routes from the more spatially organized routes of BEiT, ViT, and DINOv2, with DINOv2 retaining a distinct strong $\texttt{[CLS]}$ profile. SSMProbe uses the mathematical transparency of state-space models to turn a frozen ViT readout into an auditable evidence-routing analysis.

We introduce a novel method that averages the logits of a frozen reference policy (e.g., SFT) and a trainable policy, and incorporate the method into Group Relative Policy Optimization (GRPO). In contrast to Reinforcement Learning with Verifiable Rewards (RLVR) methods, our proposal does not involve a Kullback Leibler (KL) regularization or critic; the trainable policy and the reference anchor are coupled through the logit averaging structure to leverage the reasoning expertise of the trainable policy while maintaining the formatting advantage of SFT. Our method is evaluated on MATH, cn-k12, and MMLU, and the results show a higher accuracy or at least comparable accuracy relative to the canonical KL-regularized GRPO.