Time-Series Forecasting
A Standard Benchmark and Protocol for Genuine Zero-Shot Transfer to Truly Unseen Domains in Structured Perception Tasks
Generated automatically from the limitations stated in 19 papers (CVPR, NeurIPS, ECCV), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current zero-shot claims in structured prediction are not actually zero-shot: evaluations use datasets sharing variable structure, sensor type, or category overlap with training data; some protocols leak a fraction of test samples or target scenes to select prompts or instructions; and others substitute qualitative visualizations for scores because no clean mapping exists. The result is that reported "zero-shot" numbers cannot be compared across papers, and the field cannot measure how far true transfer to genuinely unseen conditions actually is. Until evaluation is tightened, progress on cross-domain generalization is unmeasurable and likely overstated.
Why it matters
Trustworthy zero-shot numbers would let the field distinguish real cross-domain generalization from prior bias and leakage, and would give a reproducible target for methods aimed at closing the zero-shot-to-few-shot gap.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Audit and formalize: build a leakage taxonomy over existing zero-shot evaluation protocols (dataset-family overlap, prompt/instruction selection from test data, train-set contamination with target categories or scenes), and re-score published setups under a strict definition — measuring the gap between reported "zero-shot" accuracy and strict zero-shot accuracy on the same tasks.
- 2
Construct a held-out transfer suite: for time-series, point clouds, and segmentation respectively, designate target benchmarks whose variable structure, sensor domain, and categories are disjoint from any plausible training corpus, with fixed category mappings and quantitative metrics (accuracy/AP/mIoU per class, with rare classes broken out). Measure where zero-shot methods actually land versus few-shot baselines given equal total data.
- 3
Characterize failure structure: measure per-class and per-condition error as a function of distance between target domain and nearest training distribution (e.g., domain-similarity metrics), testing whether zero-shot degradation is predictable — this determines whether the wall is fundamental (prior-dominated predictions) or an artifact of evaluation.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Published authors may dispute the leakage taxonomy or argue their relaxed settings are the practically relevant ones, leaving the community without consensus on the strict protocol. Alternatively, the audit may simply confirm a well-known gap without yielding insight beyond "zero-shot is hard."
Sub-problems
Bridging the Zero-Shot Transfer Gap in Real-World 3D Point Cloud Perception
3D point cloud models currently fail to generalize reliably to unseen real-world scenes, exhibiting severe performance drops on benchmarks like ScanObjectNN, ScanNetV2, and SemanticKITTI compared to supervised baselines. Current zero-shot 3D methods rely heavily on 2D vision-language priors (like CLIP) via multi-view projection or learned embeddings, but fail under sensor noise, occlusion, density variations, and tail categories. Consequently, 3D perception pipelines cannot be deployed open-vocabulary in novel indoor or outdoor environments without costly in-domain annotation.
A Systematic Robustness and Protocol Benchmark for Zero-Shot Segmentation Across Domains and Granularities
Current zero-shot segmentation claims remain fragmented across narrow test settings (e.g., single tasks like RGB-D COD or standard natural-image benchmarks) without a unified assessment across granularity scales and domain shifts. Inconsistent protocols—such as unseen-pixel leakage in training images, pseudo-label contamination, and unmapped evaluation categories—prevent direct comparisons between strict zero-shot and open-vocabulary methods. Consequently, the research community cannot determine whether reported generalization failures (such as steep performance drops on subparts or medical images) stem from model architectural limits or disparate evaluation setups.
Evidence
Each paper's own statement of the limitation, verbatim.
- Routing Channel-Patch Dependencies in Time Series Forecasting with Graph Spectral DecompositionICLR 2026
Zero-shot evaluation is restricted to frequency-aligned ETT datasets; broader cross-domain transfer is explicitly left to future work
- Instruction-based Time Series EditingKDD 2026
Zero-shot generalization to entirely unseen conditions fails (near-zero RaTS); it requires a few-shot tuning procedure assuming users can supply example pairs from the unseen condition
- In-Context Learning of Temporal Point Processes with Foundation Inference ModelsICLR 2026
Zero-shot accuracy is biased toward the Hawkes-process prior, leading to failure on specific real-world patterns like consistently alternating marks in the TAXI dataset.
- BiPA: Bilevel Prompt Adaptation for Underwater Instance SegmentationCVPR 2026
Zero-shot generalization to unseen categories is supported only by qualitative visual comparisons, not quantitative scores, because UIIS and UIIS10K lack a reliable category mapping
- unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary ReasoningICML 2025
Zero-shot capability is bounded by natural-image objectness priors; domain-gap data like medical images are expected to fail
- Towards Learning to Complete Anything in LidarICML 2025
Zero-shot recognition is weak, producing a large gap to fully supervised PaSCo (13.12 vs 26.49 PQ† on SemanticKITTI), limited by the underlying vision foundation models and worse on rare classes (pedestrian, cyclist)
- Online Segment Any 3D Thing as Instance TrackingNeurIPS 2025
Gains over the closest baseline shrink in zero-shot transfer (+0.9 AP on SceneNN, +1.9 AP on 3RScan with SAM), and offline methods remain far ahead on ScanNet (Oneformer3D 59.3 AP vs 43.4 online)
- Occlusion-aware Text-Image-Point Cloud Pretraining for Open-World 3D Object RecognitionCVPR 2025
Zero-shot accuracy on real data remains moderate (63.5% ScanObjectNN, 49.0% ScanNetV2), with several classes near 0% (e.g., cabinet, picture, counter)
Show all 19 papers
- Exploiting Language Power for Time Series Forecasting with Exogenous VariablesTheWebConference 2025
Zero-shot evaluation limited to within-ETT-family transfers (same datasets with same variable structure), so cross-domain generalization is untested
- VSCode: General Visual Salient and Camouflaged Object Detection with 2D Prompt LearningCVPR 2024
Zero-shot generalization demonstrated only for one unseen task (RGB-D COD); generalization to arbitrary unseen domain-task combinations is untested
- UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series ForecastingTheWebConference 2024
Zero-shot transfer requires an instruction-selection protocol that uses 0.5% of test samples and forecasting loss to pick the instruction, so it is not a fully zero-shot procedure
- UniTS: A Unified Multi-Task Time Series ModelNeurIPS 2024
Zero-shot transfer to new domains is significantly less effective than few-shot adaptation and remains an open challenge.
- SceneVerse: Scaling 3D Vision-Language Learning for Grounded Scene UnderstandingECCV 2024
True zero-shot transfer to entirely unseen benchmarks is weak (32.4 on Nr3D vs 58.7 trained from scratch); large gains only appear in the zero-shot-text setting where target-domain scenes are seen during training
- SPIN: Hierarchical Segmentation with Subpart Granularity in Natural ImagesECCV 2024
All zero-shot models perform poorly at subpart granularity (best 14 mIoU vs 86 for objects), and even SAM with ground-truth box prompts drops ~20pp per granularity level, showing the task is far from solved
- PartGLEE: A Foundation Model for Recognizing and Parsing Any ObjectsECCV 2024
Pseudo-label and object-data additions expose models to PartImageNet-like categories in some comparisons, complicating zero-shot interpretation
- From Similarity to Superiority: Channel Clustering for Time Series ForecastingNeurIPS 2024
Zero-shot evaluation is limited to closely related ETT datasets (same sensor type, different region/granularity); cross-domain transfer to unrelated datasets is untested
- ULIP: Learning a Unified Representation of Language, Images, and Point Clouds for 3D UnderstandingCVPR 2023
Zero-shot accuracy on ModelNet40 ALL is only ~60% and ~50% on ScanObjectNN, leaving a large gap to supervised performance
- Semantic-Promoted Debiasing and Background Disambiguation for Zero-Shot Instance SegmentationCVPR 2023
Strict zero-shot setting excludes training images containing unseen-class pixels, unlike open-vocabulary methods that can exploit captions on unseen images
- CLIP2Point: Transfer CLIP to Point Cloud Classification with Image-Depth Pre-TrainingICCV 2023
Zero-shot performance on real-world data is unstable and low (authors note PointCLIP's ScanObjectNN results appear unreliable and theirs remain modest)
Nearest existing work
- ZeroDiff: Zero-Shot Time Series Reconstruction via Informed-Prior DiffusionICML 2026
- Zero-shot Forecasting by Simulation AloneICLR 2026
- VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series ForecastersICML 2025
- Progressive Ensemble Networks for Zero-Shot RecognitionCVPR 2019
- Unified Training of Universal Time Series Forecasting TransformersICML 2024
- Unified Transferability Metrics for Time Series Foundation ModelsNeurIPS 2025
- Unsupervised Domain Adaptation for Zero-Shot LearningICCV 2015
- Not All Tasks Are Born Equal: Understanding Zero-Shot GeneralizationICLR 2023
- Zero-Shot Learning Through Cross-Modal TransferNeurIPS 2013
- Zero-Shot Task TransferCVPR 2019
- TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context LearningNeurIPS 2025
- Zero-Shot Learning via Semantic Similarity EmbeddingICCV 2015
- Revisiting LLMs as Zero-Shot Time Series Forecasters: Small Noise Can Break Large ModelsACL 2025
- Conformal Prediction for Zero-Shot ModelsCVPR 2025
- A Generalization Theory for Zero-Shot PredictionICML 2025