New Approaches to LLM Training and Reasoning
Researchers have published four papers addressing different aspects of large language model training and performance optimization.
According to arxiv.org, a team introduced Amortised Sequential Information Gathering (ASIG), a fine-tuning approach that incorporates Bayesian Experimental Design into LLM policies through a multi-turn extension of Group Relative Policy Optimisation with an Expected Information Gain reward. When evaluated on the 20 Questions task, ASIG “more than doubles the success rate of the 7B base model and reduces inference cost by over $25\times$” compared to a baseline called BED-LLM. The paper was accepted to FoGen 2026, an ICML workshop.
In another arxiv.org paper, researchers presented dOPSD (on-policy self-distillation for diffusion language models), which addresses challenges in training diffusion-based language models. According to the paper, dOPSD “derives the teacher’s privilege directly from the student’s own denoising trajectory, evaluating masked positions using later, more-decoded steps of that same trajectory rather than an external label.” The method reportedly improved both in-domain math reasoning and out-of-domain code generation on Dream and LLaDA benchmarks.
Additionally, arxiv.org reported on OpenTinker, an infrastructure for training LLM agents with multiple LoRA-backed policies over shared resources, and TaoSR-AGRL, an Adaptive Guided Reinforcement Learning framework deployed on Taobao Search serving “hundreds of millions of users,” according to the paper accepted to WWW 2026.