Reinforcement Learning via Self-Distillation (2026 · 74 citations)
Policy-gradient RL with verifiable rewards, such as GRPO, gets only one scalar score per attempt. That makes it hard to tell which tokens caused a failure, and learning stops when every rollout in a group gets the same score. The authors point out that many environments also return text explaining the outcome, such…
Self-Distillation Enables Continual Learning (2026 · 62 citations)
Foundation models usually stop learning after deployment. Fine-tuning them on new data with supervised fine-tuning (SFT) tends to erase earlier abilities, because SFT trains on a fixed external data distribution, which makes it off-policy. On-policy reinforcement learning forgets less, but it requires a reward…
Learning How Hard to Think: Input-Adaptive Allocation of LM Computation (2024 · 22 citations)
Costly decoding strategies such as best-of-k sampling with reranking, or calling a stronger model, improve LM outputs, but they are usually applied uniformly to every query.

All works

4 works of theirs in our library, the most cited first; 3 with a page here, the others link their original until we summarise them.

Reinforcement Learning via Self-DistillationarXiv:2601.20802Jonas Hubotter, Frederike Lubeck, Lejs Behric, et al.2026 · 74 citations
Self-Distillation Enables Continual LearningarXiv:2601.19897Idan Shenfeld, Mehul Damani, Jonas Hubotter, et al.2026 · 62 citations
Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty ↗arXiv (Cornell University) · no summary yetMehul Damani et al.2025 · 24 citations
Learning How Hard to Think: Input-Adaptive Allocation of LM ComputationarXiv:2410.04707Mehul Damani, Idan Shenfeld, Andi Peng, et al.2024 · 22 citations

Co-authors here

6 co-authors with a page, the most papers together first.

Papers of our library cite their works 182 times (of the 149,439 papers whose bibliographies we read); 38 citing papers with a page here →.

Updated 3 November 2026