{
  "id": "2405.14838",
  "title": "From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step",
  "related_terms": [
    "implicit CoT",
    "latent reasoning",
    "latent chain-of-thought",
    "latent CoT"
  ],
  "authors": "Yuntian Deng, Yejin Choi, Stuart Shieber",
  "first_public": "2024-05-23",
  "venue": "arXiv",
  "arxiv": "2405.14838",
  "paper_url": "https://arxiv.org/abs/2405.14838",
  "code_url": "https://github.com/da03/Internalize_CoT_Step_by_Step",
  "lens": [
    "internalized"
  ],
  "representation": [
    "hidden-state"
  ],
  "computation": [
    "depth-wise"
  ],
  "training": [
    "curriculum",
    "supervised"
  ],
  "questions": [
    "capability",
    "efficiency"
  ],
  "evidence": [
    "empirical"
  ],
  "modality": [
    "language",
    "symbolic"
  ],
  "summary": "Uses a curriculum that progressively removes intermediate reasoning steps during fine-tuning, teaching a model to internalize an explicit chain of thought.",
  "provenance": "Awesome-Latent-CoT",
  "author_names": [
    "Yuntian Deng",
    "Yejin Choi",
    "Stuart Shieber"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2405.14838/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2405-14838",
  "citation_date": "2024-05-23",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2405.14838/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2405.14838/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2405.14838/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2405.14838/citation.csl.json",
  "demos": [
    {
      "id": "multiplication",
      "title": "Multiplication without written steps",
      "label": "Multiplication",
      "description": "Compare GPT-2 models using implicit CoT, explicit CoT, or no CoT on two numbers of up to 20 digits each.",
      "limitations": "The models can make mistakes, including on valid inputs. The implicit model writes answer digits, not intermediate steps.",
      "model": "GPT-2",
      "url": "https://huggingface.co/spaces/yuntian-deng/gpt2-multiplication",
      "model_url": "https://huggingface.co/yuntian-deng/gpt2-implicit-cot-multiplication-20-digits",
      "topic_ids": [
        "internalized"
      ],
      "paper_relationships": [
        {
          "paper_id": "2405.14838",
          "relationship": "implementation"
        },
        {
          "paper_id": "2311.01460",
          "relationship": "related_method"
        }
      ],
      "relationship": "implementation"
    },
    {
      "id": "math-word-problems",
      "title": "Math word problems without written steps",
      "label": "Math word problems",
      "description": "Try a Mistral-7B model trained with stepwise internalization on grade-school math word problems.",
      "limitations": "Answers can be wrong. This model is trained for grade-school math word problems, not general conversation. It still generates the answer token by token.",
      "model": "Mistral-7B",
      "url": "https://huggingface.co/spaces/yuntian-deng/implicit-cot-math",
      "model_url": "https://huggingface.co/yuntian-deng/implicit-cot-math-mistral7b",
      "topic_ids": [
        "internalized"
      ],
      "paper_relationships": [
        {
          "paper_id": "2405.14838",
          "relationship": "implementation"
        },
        {
          "paper_id": "2311.01460",
          "relationship": "related_method"
        }
      ],
      "relationship": "implementation"
    }
  ],
  "method_comparisons": [
    {
      "title": "Distillation and stepwise internalization",
      "text": "Deng et al. (2023) introduced implicit CoT as reasoning through hidden states rather than written steps; distillation is the paper's implementation of that proposal. It learns to predict a teacher's intermediate hidden states, which help a student produce the answer. Stepwise Internalization instead repeatedly shortens the written solution and fine-tunes the same model on what remains. Both use worked solutions during training; neither requires writing the removed steps at inference. Their reported experiments focus on multiplication and grade-school math, not a guarantee of general reasoning ability.",
      "source_ids": [
        "2311.01460",
        "2405.14838"
      ],
      "source_urls": [
        "https://arxiv.org/html/2311.01460v1#S2.SS3",
        "https://arxiv.org/html/2311.01460v1#S3",
        "https://arxiv.org/html/2405.14838v1#S3"
      ],
      "source_labels": [
        "Implicit CoT proposal, §2.3",
        "Distillation implementation, §3",
        "Stepwise internalization, §3"
      ],
      "topic_id": "internalized",
      "topic_label": "Internalized reasoning",
      "topic_url": "https://latent-implicit-thinking.github.io/topics/internalized-reasoning/"
    },
    {
      "title": "From stepwise internalization to Coconut",
      "text": "Coconut explicitly credits Deng et al. (2024) as the inspiration for its multistage training curriculum. Stepwise internalization gradually removes written CoT tokens until the model answers directly. Coconut instead replaces removed steps with continuous hidden-state inputs, retaining a sequence of internal reasoning steps before the answer. It also follows the earlier work in resetting the optimizer between training stages.",
      "source_ids": [
        "2405.14838",
        "2412.06769"
      ],
      "source_urls": [
        "https://arxiv.org/html/2405.14838v1#S3",
        "https://arxiv.org/html/2412.06769v3#S3"
      ],
      "source_labels": [
        "Deng et al. (2024), §3",
        "Coconut, §3"
      ],
      "topic_id": "continuous",
      "topic_label": "Continuous-state reasoning",
      "topic_url": "https://latent-implicit-thinking.github.io/topics/continuous-states/"
    },
    {
      "title": "Changing generation order or removing written steps",
      "text": "Diffusion of Thought changes how a chain of thought is generated: it refines the solution through denoising rather than committing to words strictly from left to right. This changes the generation schedule, not necessarily the visibility of the reasoning. The resulting chain can still be text, unlike methods that only emit the final answer.",
      "source_ids": [
        "2402.07754",
        "2405.14838"
      ],
      "source_urls": [
        "https://proceedings.neurips.cc/paper_files/paper/2024/hash/be30024e7fa2c29cac7a6dafcbb8571f-Abstract-Conference.html",
        "https://arxiv.org/html/2405.14838v1#S3"
      ],
      "source_labels": [
        "Diffusion of Thought",
        "Deng et al. (2024), §3"
      ],
      "topic_id": "beyond-sequential",
      "topic_label": "Alternative generation methods",
      "topic_url": "https://latent-implicit-thinking.github.io/topics/alternative-reasoning-schedules/"
    }
  ]
}