Training Language Models to Reason Efficiently (2025 · 74 citations)
Large reasoning models solve hard problems by producing very long chains of thought. This makes serving them expensive and slow, because attention cost grows quadratically and the KV cache grows linearly with length. Arora and Zanette propose teaching such models to reach correct answers with fewer tokens. They use…
ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL (2024 · 54 citations)
Most reinforcement learning for LLMs optimizes one response at a time. As a result, agents cannot learn to gather information over several turns, credit early actions for delayed outcomes, or build on earlier steps. The authors introduce ArCHer (Actor-Critic Framework with a Hierarchical Structure), which splits…
Fast Best-of-N Decoding via Speculative Rejection (2024 · 50 citations)
Best-of-N aligns an LLM at inference time without retraining. It samples many candidate answers and keeps the one a reward model rates highest. It can rival post-training methods such as RLHF or DPO, but competitive N values run into the thousands, which would need many accelerators. The authors observe that reward…
Can Large Reasoning Models Self-Train? (2025 · 16 citations)
The paper asks whether a reasoning model can keep improving through reinforcement learning when no ground-truth answers are available and the training signal comes only from the model itself.

All works

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

Training Language Models to Reason EfficientlyConference on Neural Information Processing Systems (NeurIPS)Daman Arora, Andrea Zanette2025 · 74 citations
ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RLarXiv:2402.19446Yifei Zhou, Andrea Zanette, Jiayi Pan, et al.2024 · 54 citations
Fast Best-of-N Decoding via Speculative RejectionConference on Neural Information Processing Systems (NeurIPS)Hanshi Sun, Momin Haider, Ruiqi Zhang, et al.2024 · 50 citations
Can Large Reasoning Models Self-Train?arXiv:2505.21444Sheikh Shafayat, Fahim Tajwar, Ruslan Salakhutdinov, et al.2025 · 16 citations
Maximum Likelihood Reinforcement Learning ↗arXiv (Cornell University) · no summary yetFahim Tajwar et al.2026 · 4 citations
Tail-Likelihood Reinforcement Learning ↗arXiv (Cornell University) · no summary yetShrinivas Ramasubramanian et al.2026 · 1 citation
Harness Learning Enables Generalizable Test-Time Adaptation ↗arXiv (Cornell University) · no summary yetAlvin Zhang et al.2026

Co-authors here

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

Where their works here were published, the most works first, with the editions of those that have a page.

Papers of our library cite their works 194 times (of the 149,439 papers whose bibliographies we read); 85 citing papers with a page here →.

Updated 3 November 2026