AI 推理时偷偷把“没说出来”的思考喂回自己
大模型生成回答时，每走一步，只有最终选中的那个词会被送回模型底部，而它内部算出来的、更丰富的“隐藏状态”却被直接扔掉。这篇论文给这条反馈通道加了带宽：把上一层的隐藏状态和选中的词融合后，一起喂回模型，相当于让模型在“说出口”之外，还能把“没说出来”的思考也带进下一步。训练时用了一种分阶段的方法，先正常训练，后期才引入这种反馈，并混入少量更深的反馈来保持稳定。在 1B 参数的模型上训练到 400B token 后，它在语言理解、数学、代码和指令跟随上都更好，而且解码速度几乎没变，相当于用同样的算力达到了多训练 50% 数据的效果，还能用更短的推理链达到同样甚至更好的准确率。它不是你明天就能用上的功能，但指向一个趋势：AI 的“内心戏”不再只是副产品，而是可以被利用的资产。
📄 原文摘要(英文)
Autoregressive transformers compute along two axes: horizontally across generated tokens, and vertically through model depth. Dense attention gives each token broad horizontal access to the past, but the vertical feedback channel between decoding steps remains narrow: only the sampled token returns to the bottom of the stack, while the top-layer hidden state is discarded. We introduce the full-bandwidth transformer, which widens this channel with latent feedback: at each decoding step, the previous top-layer hidden state is fused with the sampled token embedding through a gated linear unit and fed back as the next input. Latent feedback lets non-verbalized computation re-enter the stack with a renewed depth budget, while preserving the standard transformer architecture, KV cache, and language-modeling objective. To train full-bandwidth transformers without losing parallel teacher forcing, we use a scheduled multi-pass objective that introduces latent feedback late in pretraining and mixes a small fraction of deeper feedback passes for stability. We train 1B-parameter full-bandwidth transformers up to 400B tokens and find that latent feedback improves validation loss, 5-shot language-model evaluation, math and coding generation, and instruction-tuned performance. With negligible per-token decoding overhead, full-bandwidth transformers match or approach standard transformers trained with roughly 1.5times more tokens, and manage to produce shorter reasoning traces at equal or better accuracy.