Track the frontier, from condensed matter to cosmology.
Search papers, labs, and topics across Lattice.
Affiliation:
4
0
8
Exposing the critic to both successful and failed same-prompt rollouts as contrastive foils resolves PPO's notoriously brittle credit assignment in multi-step reasoning—retaining its efficacy even when paired with a downscaled critic.
Stop wasting compute on full rollouts: ADWIN dynamically adapts on-policy distillation windows, slashing training costs by up to 4.1x without sacrificing accuracy on reasoning tasks.
Most RLVR datasets are just remixes of a few originals, and this paper shows how to trace them back to their source, revealing widespread data contamination.
Training tool-calling agents with just an 8B language model outperforms traditional methods that depend on expensive resources, reshaping the landscape of tool learning.