Records found: 5
'Early Experience uses agent-generated future states as supervision to train language agents without rewards or large demo sets, outperforming imitation learning across eight benchmarks and improving RL initialization.'
'LIMI uses 78 curated, tool-grounded trajectories to fine-tune GLM models, hitting 73.5% on AgencyBench and outperforming large-sample baselines by a wide margin.'
'Google shows active learning can fine-tune LLMs with under 500 targeted labels, matching or improving model quality while reducing labeling needs by orders of magnitude.'
DSRL introduces a novel method to adapt diffusion-based robotic policies via latent-space reinforcement learning, significantly boosting real-world task performance without modifying base models.
Researchers reveal that training large language models with just one example using 1-shot reinforcement learning significantly enhances their math reasoning abilities, matching results from large datasets.