Tag
SALA is a Semantic-Aware Logical Alignment framework that improves demonstration selection for complex reasoning in in-context learning by automatically learning task-specific reasoning operations and using dynamic time warping for flexible alignment, outperforming existing methods.
This paper investigates the dissociation between attention-level proxies and behavioral performance in large language models under fine-tuning, revealing that attention sensitivity alone is unreliable for diagnosing in-context learning capabilities.
S1 is a universal robot foundation model that learns complex, long-duration tasks from a single video demonstration using in-context learning, enabling flexible, robust performance without extensive data collection.
ChorusTIC is a training-free foundation model for multivariate time series classification that uses in-context learning to handle heterogeneous channel configurations without target-task updates, demonstrating strong performance on standard benchmarks.
ICI-Time is a novel framework that reframes time series forecasting as a visual inpainting task, leveraging large vision models to enable adaptable forecasting without fine-tuning or architectural changes.
Zero-WAM is a causal video-action model that enables zero-shot robotic manipulation of unseen tasks by conditioning on in-context human video guidance, with the HumanGen dataset and a future-chunk prediction objective to improve generalization.
S1 is a new foundation model that learns from a single example and can perform 10-minute tasks from video prompts without fine-tuning, showcased in real-time operation.
Tydra is a hybrid Transformer-SSM architecture for tabular data that reduces inference time by 30% compared to TabPFN while maintaining similar predictive performance and outperforming Hydra.
RecPFN introduces a prior-fitted network for in-context learning in sequential recommendations, pretrained on synthetic clickstream data to achieve state-of-the-art zero-shot performance across benchmarks.
The paper proposes LLM-Detector, a framework that uses large language models with in-context learning to perform tabular anomaly detection without fine-tuning, demonstrating consistent improvements over existing methods on multiple datasets.
This paper introduces BayesPrompt, a Bayesian approach to optimize prompts for large language models, ensuring they are both efficient and human-readable, unlike traditional methods that produce unintelligible pseudoprompts.
This paper proposes the Structure-Internalized Rule Language Model (SIRLM) to address reasoning evidence perception drift in knowledge graph reasoning with LLMs, improving faithfulness and effectiveness by coupling structural and parametric knowledge, with experiments showing superiority over state-of-the-art methods.
This paper investigates using in-context learning with the TabPFN model to predict perceived neighborhood walkability based on built environment features for mobility-impaired older adults, achieving higher performance than baseline models and employing SHAP-IQ for interaction analysis.
This paper proposes FairTFM, a training strategy to incorporate fairness constraints into tabular foundation models, enabling fair predictions via in-context learning without task-specific retraining.
This paper develops a theoretical framework for in-context learning on partial orders, analyzing identifiability, teaching cost, and representation limits with exact completion trichotomies.
This paper introduces a simulation-aware large language model framework for analog IC layout refinement using in-context policy improvement, achieving better post-layout performance with fewer simulations compared to existing methods.
UI-Mate is a foundation GUI agent that uses environment-grounded training and in-context demonstrations to improve reliability on long-horizon office tasks, achieving state-of-the-art results on computer-use benchmarks.
The article introduces BDH-CQ, a 150M parameter recurrent model that combines in-context learning with latent reasoning, achieving 29.5% on ARC-AGI-1 at a cost of $0.0007 per task, setting a new standard for cost efficiency.
This paper introduces MAG, a manifold-guided framework for semi-supervised multi-modal in-context demonstration selection, leveraging unlabeled data to improve few-shot ICL for MLLMs. Experiments on eight benchmarks show consistent gains in label-scarce regimes.
This paper presents REAP, a two-stage pipeline for closed-book knowledge base construction from LLMs that combines relation-specific elicitation strategies with deterministic JSON parsing, achieving a macro-F1 of 0.62 on the AKBC Shared Task 2026 test set using Mistral-Small-24B-Instruct-2501.