Anthropic publishes its own ETH refutation as academics organise review of…

Across recent work, self-improving agents can learn from self-generated experience: Meta AI's retrospective self-distillation, Tsinghua's self-written long-context tasks, and UIUC/Berkeley/Google's repeated data-selection loops all report gains, but new benchmarks from over 50 experts, CUHK-Shenzhen and ByteDance, and UT Dallas et al. show that models still misjudge their own capabilities, fail to detect and recover from errors in long agentic tasks, and can regress under repeated training.

Recent work on training data subset selection converges on a central tension: per-point value scores such as Data Shapley or influence are fragile, since fixed top-k selection provably under-covers latent data modes, and Université de Montréal, Mila, TTIC and Google DeepMind show that even targeting full-training fidelity faces statistical limits, often requiring impractically large validation sets. The newest line of inquiry, from Beijing Jiaotong University, Hong Kong Polytechnic University and Huawei Noah's Ark Lab, shifts from static subsets to closed-loop control, using output entropy to dynamically reweight replay data during fine-tuning and thereby counter catastrophic forgetting.

The accumulated papers establish that explicit credit assignment at intermediate turns, tokens, or reasoning steps consistently improves long-horizon language-agent training over outcome-only reinforcement learning. Recent work has moved toward isolating each decision's effect via counterfactual rollouts (University of Sheffield, CMU, University of Michigan), self-assigned credit from summaries (Meta, Duke, NYU), or causal measures of step influence (HIT Shenzhen, University of Aberdeen, University of Cambridge), while also revealing failure modes like silent forks where outcome-based RL provides no gradient (NYU, CMU).