Haven团队叶新武、唐相儒等联合斯坦福大学丛乐、普林斯顿大学王梦迪最新提出的LatentChem，想做的就是一件事： 把化学推理从"文本表面"挪到"模型内部"。 LatentChem团队 投稿 量子位 | 公众号 QbitAI AI做科学推理，可能不该总靠"把步骤写出来"。 过去几年，大模型一旦进入"推理模式"，几乎都会走同一条路线： 先输出一大段思维链，再给出最终答案。 这套方法在数学题、代码题、复杂问答里很常见，也确实有效。但到了化学场景，它未必还是最顺手的方式。 看起来像是在认真思考，结果却经常"说一套，做一套"。 LatentChem论文给出的解释很直接： 化学推理本身，更像是在连续、结构化的空间中进行搜索、调整和更新； 而自然语言token，本质上是离散的。 模型不一定要把每一步都翻译成文字，也可以先在连续隐空间里完成多步计算，最后再输出自然语言。 这不是"取消推理"，而是换了一种推理介质。 模型前面能写得头头是道。 用文字描述化学，为什么不够好？ 电子效应、位阻、官能团、反应位点，说得都很专业。 做过分子优化、分子编辑、反应预测的人，大多见过这种情况： 但到了最后，生成出来的SMILES或分子结 ...

Core Viewpoint - The article introduces SIM-CoT (Supervised Implicit Chain-of-Thought), a new advancement in implicit reasoning that addresses the core issue of latent state collapse when scaling implicit tokens, leading to a loss of reasoning semantics [2][9]. Group 1: SIM-CoT Overview - SIM-CoT employs a plug-and-play step-level supervision module that stabilizes optimization and prevents collapse by aligning each latent token with corresponding reasoning steps during training [2][10]. - The method allows for interpretable implicit reasoning, enabling the decoding of latent tokens into human-readable intermediate reasoning steps [2][10]. Group 2: Performance Improvements - During inference, SIM-CoT incurs zero additional overhead, yet it shows significant performance improvements: +2.1% over supervised CoT and +8.2% over Coconut on GPT-2, with stable gains of +1.5% to +9.0% on larger LLaMA models [3][18]. - In the GSM8k-Aug dataset, SIM-CoT improved accuracy from 36.6% to 44.8% (+8.2) while maintaining lower token usage, achieving 2.3× token efficiency [18]. - On out-of-domain datasets like GSM-Hard, MultiArith, and SVAMP, SIM-CoT's average accuracy increased from 42.6% to 46.9% (+4.3), demonstrating robust latent space reasoning [19]. Group 3: Stability and Efficiency - SIM-CoT maintains stability even with increased implicit tokens, addressing issues like latent instability and semantic homogenization that typically arise in implicit CoT methods [9][14]. - The auxiliary decoder used during training is removed during inference, ensuring that SIM-CoT's reasoning efficiency remains comparable to other implicit methods while still providing a speed advantage over explicit CoT [21]. Group 4: Experimental Validation - The authors conducted systematic evaluations of SIM-CoT, confirming that it is more accurate, stable, and token-efficient compared to existing methods [17]. - The framework was validated across various models, including GPT-2 and LLaMA 1B/3B/8B, consistently showing effective performance improvements [22].