Xiaohongshu (Little Red Book) (China)

DepthGraft turns cross-layer KV sharing from a cache-saving inference trick into a structural regularizer during training: deep layers no longer compute their own KV but rebuild it from a stride-aligned shallow–middle source pair using learnable channel-wise coefficients. Gradients from deep layers then flow back along each sharing edge into the source KV projections, sharpening key retrieval directions while enriching the value content subspace. Across dense and MoE models from 0.5B to 16B parameters, downstream averages improve consistently, with a 7.0-point absolute gain on MMLU-Pro for the 16B MoE and roughly one third less KV cache.

DepthGraft turns cross-layer KV sharing from a cache-saving inference trick into a structural regularizer during training: deep layers no longer compute their own KV but rebuild it from a stride-aligned shallow–middle source pair using learnable channel-wise coefficients. Gradients from deep layers then flow back along each sharing edge into the source KV projections, sharpening key retrieval directions while enriching the value content subspace. Across dense and MoE models from 0.5B to 16B parameters, downstream averages improve consistently, with a 7.0-point absolute gain on MMLU-Pro for the 16B MoE and roughly one third less KV cache.

The work identifies a "Symmetric Conservatism" failure mode of GBPO in generative recommendation (symmetric update bounds suppress rare positive signals such as cold-start items, static negative-sample constraints fail to prevent diversity collapse, and group-normalized multi-objective rewards yield low-resolution training signals) and proposes SAGE, which uses a geometric-mean sequence-level importance ratio and a decoupled multi-objective advantage estimator to reduce token-level variance and mitigate reward collapse, plus asymmetric adaptive bounding that applies a positive Boost to successful slates and an entropy-aware penalty to low-diversity failures; on three Amazon Product Reviews categories and the large-scale RecIF-Bench, the native-text TextRec optimized with SAGE outperforms O

The work identifies a "Symmetric Conservatism" failure mode of GBPO in generative recommendation (symmetric update bounds suppress rare positive signals such as cold-start items, static negative-sample constraints fail to prevent diversity collapse, and group-normalized multi-objective rewards yield low-resolution training signals) and proposes SAGE, which uses a geometric-mean sequence-level importance ratio and a decoupled multi-objective advantage estimator to reduce token-level variance and mitigate reward collapse, plus asymmetric adaptive bounding that applies a positive Boost to successful slates and an entropy-aware penalty to low-diversity failures; on three Amazon Product Reviews categories and the large-scale RecIF-Bench, the native-text TextRec optimized with SAGE outperforms O