Few-shot learning
PulseAugur coverage of Few-shot learning — every cluster mentioning Few-shot learning across labs, papers, and developer communities, ranked by signal.
11 day(s) with sentiment data
New theoretical frameworks will emerge to unify in-context learning with probabilistic and meta-learning concepts.
One cluster explicitly links in-context learning (ICL) to Bayesian inference and meta-learning, proposing a statistical theory to explain its mechanisms. This indicates a growing research effort to provide a more rigorous mathematical understanding of ICL, moving beyond empirical observations.
Few-shot learning frameworks will increasingly leverage LLM-generated synthetic data for training and evaluation.
Multiple recent clusters highlight the use of LLMs and in-context learning (ICL) to generate synthetic ground truth data for tasks like emotion classification and to improve grammatical error correction. This suggests a trend towards using LLMs not just for inference but also for data augmentation in few-shot learning scenarios, reducing reliance on expensive human labeling.
In-context learning is being adapted for complex model migration and cross-framework adaptation tasks.
A recent cluster details an agentic framework using ICL to automate the migration of deep learning models from PyTorch to JAX, achieving high numerical equivalence. This demonstrates that ICL is being applied to more complex engineering challenges beyond simple task adaptation, such as cross-framework compatibility.
-
LLM agents may rely on statistical extrapolation over reasoning in strategic tasks
A new research paper explores whether large language model (LLM) agents improve their decision-making through genuine reasoning or by extrapolating statistical patterns from interaction history. The study used multi-age…
-
Research: Few-shot degradation in LLMs is task-dependent, new metric shows
A new research paper investigates the phenomenon of "few-shot degradation" in language models, where providing examples can sometimes harm performance instead of improving it. The study, which tested 12 open-weight mode…
-
Gating mechanisms hinder State Space Models' in-context learning, research finds
A new research paper published on arXiv explores the role of gating mechanisms in State Space Models (SSMs), which are emerging as an alternative to Transformers for sequence modeling. The study reveals that these gatin…
-
New V-ICAL benchmark reveals significant limitations in video-based learning for multimodal agents
A new benchmark called V-ICAL has been introduced to evaluate how well multimodal agents can learn from video demonstrations in interactive environments. This benchmark, comprising 342 tasks across 37 environments, asse…
-
In-context learning emerges across diverse AI modalities, study finds
A new research paper proposes the Convergent Emergence Hypothesis, suggesting that few-shot in-context learning (ICL) capabilities, observed in large language models, may emerge broadly across different data modalities.…
-
Few-shot learning evaluation protocols may overestimate model performance
A new paper critically examines the assumptions behind few-shot learning evaluations, particularly the common practice of pre-training models on large auxiliary datasets. Researchers found that pre-training, even with c…
-
New framework aligns multimodal LLMs with reasoning paths beyond imitation
Researchers have developed a new framework for multimodal in-context learning (ICL) that aims to improve how large language models (LLMs) align their responses with the reasoning process required by complex multimodal i…
-
DearICL framework enhances LLM in-context learning with novel sample selection
Researchers have developed DearICL, a novel framework for selecting optimal demonstration examples for in-context learning (ICL) in large language models (LLMs). Unlike static, offline selection methods, DearICL treats …
-
New AI approach boosts information extraction for Industry 4.0 asset data
Researchers have developed AAS-RAIL, a novel retrieval-augmented in-context learning approach to improve information extraction for Asset Administration Shells (AAS) from PDF product datasheets. This method dynamically …
-
New paper unifies LLM training methods via Bayesian lens
A new paper proposes a unified Bayesian framework to understand various large language model training and evaluation paradigms, including supervised fine-tuning (SFT), in-context learning (ICL), and KL-regularized reinf…
-
Robots learn new tasks from demonstrations using In-Context Learning · 3 sources tracked
Robotics startups Skild AI and Generalist AI are pioneering In-Context Learning (ICL) for robots, enabling them to learn new tasks from demonstrations without fine-tuning. Skild AI's S1 model can process up to 10-minute…
-
New Context Compilation Architecture Boosts LLM In-Context Learning
Researchers have introduced a new Context Compilation Architecture (CCA) designed to improve how large language models handle in-context learning (ICL). The CCA aims to address the brittleness of current models in tasks…
-
New metrics reveal LLMs can fake in-context learning
Researchers have developed a new method to evaluate in-context learning (ICL) in large language models, specifically focusing on how fine-tuning affects this ability. The study introduces "In-Context Sensitivity" (ICS) …
-
Causal Foundation Models leverage in-context learning for causal inference
Researchers have introduced Causal Foundation Models (CFMs), which leverage pretrained neural networks to estimate causal effects on new datasets through in-context learning. This approach eliminates the need for fine-t…
-
New BIRD-History Benchmark Evaluates Text-to-SQL Systems Using Query Logs
Researchers have introduced BIRD-History, a new benchmark designed to evaluate text-to-SQL systems' ability to leverage historical query logs for understanding underspecified natural language questions. The benchmark in…
-
Landau theory explains criticality in linear in-context learning
Researchers have developed a Landau theory to explain the critical phenomenon observed in linear in-context learning (ICL). This theory frames the double-descent singularity, which occurs when pretraining samples approa…
-
AI's next frontier: Data-efficient learning beyond Scaling Laws
A researcher from BIMSA, Wang Yaqing, argues that the Scaling Law paradigm in AI is hitting its limits due to data scarcity and high computational costs. She proposes "Data-Efficient Agentic Learning" (DEAL) as the next…
-
Open-source LLMs show promise for predicting software vulnerability severity
Researchers have conducted an industrial case study on predicting CVSS v3.1 scores for software vulnerabilities using in-context learning with locally deployable, open-source Large Language Models (LLMs). The study comp…
-
Robots learn new tasks in seconds with GPT-3-like one-shot learning
Generalist AI has released GEN-1.5, a new foundation model for robots that enables one-shot learning. This model can learn new tasks by observing short demonstrations, similar to how GPT-3 learned from text examples. GE…
-
6 Architectural Shifts to Optimize LLM Pipelines for Cost and Latency
The article proposes six architectural shifts to optimize large language model (LLM) pipelines by reducing token costs and latency. It advocates for implementing strict retrieval-augmented generation (RAG) with vector d…