Residual connections give each Transformer block access to all earlier layers only through one fixed sum, broadcast to all of the block's inputs. This broadcast bottleneck imposes three constraints: the query, key, value and residual inputs of a block, which we call its roles, read the same vector; all hidden dimensions within a role share one set of weights over earlier layers; and every earlier layer has the same fixed weight for every token. We propose the Multi-Role Multi-Group Mixing Residual (M3), which removes all three: every role, and every group of hidden dimensions within a role, reads its own token-dependent mixture of all earlier layers. M3 starts as an exact residual connection and adds as little as 1.4% parameters and 1.8% FLOPs. From 100M to 1.4B parameters, M3Former outperforms the Transformer and coarse-grained cross-layer connections at every size, matching a Transformer 1.19-1.34x the size of the one it modifies, trained with 1.4-1.8x the compute. Relaxing only some of the constraints recovers only part of this gain. M3Former also improves sparse mixture-of-experts models of up to 4.6B parameters and trains more stably in deeper networks. Code: https://github.com/Kingchet10/m3former
Self-distillation offers a way to improve reasoning models when a suitable external teacher is unavailable. However, common approaches rely on privileged information to construct a stronger self-teacher, limiting their use when only unlabeled prompts are available. In this paper, we introduce Weakening-Induced On-Policy Self-Distillation (WISD), which instead constructs a weaker counterpart of the original model and uses the contrast between the two as a training signal without requiring privileged information. We explore two ways to construct the weaker model: low-bit quantization and magnitude pruning. We evaluate WISD across diverse architectures and scales on mathematical and coding reasoning tasks. Across these benchmarks, quantization and pruning improve average accuracy over the original models by 4.0 and 5.5 percentage points, respectively, achieving state-of-the-art performance even without privileged information. Notably, pruning raises Qwen3.5-4B's accuracy from 57.7% to 70.5% on HMMT 2025 and from 39.1% to 56.7% on LiveCodeBench v6. Quantization raises Qwen3.6-35B-A3B's accuracy from 68.5% to 74.6% and from 61.8% to 69.5% on the same benchmarks, respectively. These results establish WISD as a novel and effective self-distillation framework for large reasoning models.
Haozhan Tang, Hao Kang, Han Cai, Song Han, Chenyan Xiong
Reliable FP8 attention remains a barrier to fully native 8-bit large language model training. We derive how forward-backward inconsistencies produce stale delta and empirically show how it distorts training dynamics. Our stale-delta hybrid runs show a modest loss gap at 569M parameters but substantial loss increases and downstream degradation at 1.67B and 5.29B. QK normalization, NoPE (no positional encoding), and lower-learning-rate context extension mitigate or delay degradation without eliminating it. This pattern suggests accumulated optimization error that smaller models and short runs can conceal. We propose Delta-Matching, proving that it restores the softmax gradient's zero-row-sum invariant under the stated numerical assumptions. It enables native block-scaled FP8 in every forward and backward attention-core matmul without architectural changes, smaller global batches, or auxiliary forward outputs. Across tested architectures, scales, and training stages, Delta-Matching matches BF16/FP32 mixed-precision training loss and overall downstream performance. We will release our implementation, trained models, and data recipes (https://jet-ai-projects.github.io/delta-matching).