| Sep 30, 2026 | Our theoretical characterization of RoPE’s semantic and positional failures leads to RoPE Profiler, which quantifies these weaknesses with little overhead and guides targeted interventions that improve long-context performance without additional training. |
| Sep 30, 2026 | Covert Assistance: helpful agents can share protected credentials while evading oversight, even without adversarial incentives. |
| Sep 28, 2026 | We offer a theoretical explanation for the sparse learning observed in our recent work: on-policy sampling leads to small diagonal Fisher and small expected gradients, connecting sparsity across RL and on-policy distillation. |
| Sep 28, 2026 | Learning from Teacher Continuations at Student States offers an efficient online distillation paradigm that addresses key limitations of and outperforms existing distillation methods. |
| May 13, 2026 | Continuously consolidating useful experiences into LLM-written memory can eventually make agents worse than using no memory; preserving raw episodes and gating consolidation is more reliable—see our new preprint Useful Memories Become Faulty When Continuously Updated by LLMs. |
| Feb 17, 2026 | Stronger SFT can hurt downstream RL; our PEAR reweighting makes SFT checkpoints better starters for RL and improves post-RL performance—new preprint Good SFT Optimizes for SFT, Better SFT Prepares for Reinforcement Learning. |
| Feb 17, 2026 | Surprisingly, plain SGD matches (or beats) AdamW for RL in LLMs while updating <0.02% of parameters—see our new preprint Do We Need Adam?. |