Tag
BridgeVLA++ is a memory-augmented vision-language-action framework for 3D robot manipulation that builds on BridgeVLA to add spatio-temporal memory, achieving state-of-the-art results on memory-dependent manipulation benchmarks while preserving data efficiency and generalization.
Introduces Attention Head Reweighting (AHR), a data-efficient method for adapting LLMs to text classification tasks by learning a single scalar per attention head, drastically reducing trainable parameters while outperforming LoRA in limited data settings.
The article announces the first ChineseBabyLM Challenge at NLPCC 2026, which asks researchers to train language models from scratch on 100 million Chinese tokens and evaluate them on NLU, cognitive alignment, and Hanzi knowledge, promoting data-efficient and cognitively plausible modeling for Chinese.
Introduces a reinforcement learning with verifiable rewards recipe for data-efficient adaptation of audio-language models to code-switched ASR, achieving significant gains across 10 language pairs with minimal data.
Proposes GenDa, a unified framework for unsupervised reinforcement learning that addresses non-stationary skill semantics and brittle generalization via skill relabeling and a complementary information bottleneck, significantly improving data efficiency and generalizability.
The paper introduces OPDLM, a method that transforms autoregressive language models into diffusion language models via on-policy distillation, requiring 15x to 7000x fewer training tokens while retaining knowledge from the original model.
This paper presents a data-efficient anatomy-aware benchmark for cardiac pathology prediction on the ACDC MRI dataset, showing that under limited labels, anatomical representation matters more than model complexity.
This paper introduces BrainSimSiam, a lightweight self-supervised framework using siamese networks to learn robust fMRI representations from positive-only pairs, achieving strong performance on downstream tasks even with limited data.
This paper proposes a retrieval-based approach for multi-label legal annotation that uses frozen embedding models to retrieve labels via k-nearest neighbors, achieving competitive accuracy, high data efficiency, and eliminating label hallucination by design.
FrameSkip is a data-layer frame selection method that improves Vision-Language-Action (VLA) policy training by prioritizing high-importance frames based on action variation and visual-coherence metrics, achieving a macro-average success rate of 76.15% across three benchmarks while using only 20% of unique frames.
This paper introduces 'Hint Tuning,' a data-efficient method that reduces token usage in reasoning models by calibrating reasoning depth based on problem difficulty. It achieves significant token reduction (24–66%) on models like Qwen3-Thinking and DeepSeek-R1-Distill using only 1K self-annotated samples.