OpenAI says GPT-6.1 Sol scores 86.6% in translating COBOL

The papers collectively establish that reward collapse in reasoning reinforcement learning stems from multiple identifiable sources: finite-sample diversity reduction (Cornell University), verifier reward hacking (Yonsei University, University of Washington, FuriosaAI), KL penalty interference (Alibaba Group), and vanishing safety feedback in group-relative methods (Baidu Inc. with multiple collaborators).

The accumulated papers collectively establish that the diffusion latent space is directly shaped by the design of the training objective, which can be strategically decoupled from the standard denoising task. Most recently, Tsinghua University and Microsoft Research found that a robust prediction principle applied to corrupted latents can replace the diffusion loss, while the University of Michigan, Alibaba Group, UC Berkeley and Nanjing University introduced Swift, using a frozen, undertrained denoiser to bias the regression targets.

The accumulated papers establish that offline experience, contextual distillation, and hindsight replay can robustly improve language model agents, but they also document that passing self-generated tests does not guarantee semantic correctness (UC Berkeley and Northeastern) and that co-evolved skill banks can erode gains over time (Alibaba Group). The line of inquiry has moved most recently toward closing this verification gap through verifiable runtimes for reversible tools (Amazon), falsifiable failure-diagnosis frameworks (UBC and MBZUAI), and reward models that forecast future action sequences (UIUC and Meta AI).