## Technical Architecture Diagram: LLM Continuous Thought Approaches ### Overview The image is a technical diagram illustrating two categories of "Continuous Thought" processing in Large Language Models (LLMs): **Intrinsic Continuous Thought** (top section) and **Auxiliary Continuous Thought** (bottom section). It compares six distinct architectural approaches (COCONUT, CODI, an unnamed LLM variant, LightThinker, CCoT, SoftCoT) with labeled components, color-coded token types, and trainable/frozen status indicators. The diagram uses dashed boxes for hidden states/loss and icons (flame = trainable, snowflake = frozen) to distinguish design choices. ### Components/Axes Since this is a diagram (not a chart), components are organized by section and sub-diagram: - **Top Section (Intrinsic Continuous Thought)**: Four left-to-right sub-diagrams: 1. COCONUT: "Large Language Model" (brown box) with input/output tokens. 2. CODI: "Latent Reasoning Model" (brown box) with input/output tokens and "Hidden States" (pink dashed box). 3. Unnamed LLM Variant: "Large Language Model" (brown box) with input/output tokens and "Hidden States" (pink dashed box). 4. LightThinker: "Large Language Model" (brown box) with input/output tokens and "Mask" (pink dashed boxes) on output tokens. - **Bottom Section (Auxiliary Continuous Thought)**: Two left-to-right sub-diagrams: 1. CCoT: "LLM Layer i" (pink box, labeled "i=L"), "Scorer" (white box), "CCoT φ" (brown box), "Decoder ψ" (brown box) with input/output tokens and "Loss" (pink dashed box). 2. SoftCoT: "Large Language Model" (brown box, frozen), "Assistant LLM" (brown box, frozen), "Linear Layer" (white box, trainable) with input/output tokens. - **Legend (Bottom-Right)**: Color-coded tokens and status icons: - Gray square: Prompt - Green square: Explicit Tokens - Blue square: Continuous Tokens - Pink dashed box: Hidden States - Flame icon: Trainable - Snowflake icon: Frozen - Pink dashed box with "Loss": Loss ### Detailed Analysis #### Intrinsic Continuous Thought (Top Section) 1. **COCONUT**: - Input: Gray (Prompt), blue (Continuous Tokens), green (Explicit Tokens) tokens. - Model: "Large Language Model" (brown box, trainable: flame icon). - Output: Blue (Continuous Tokens), green (Explicit Tokens) tokens. - Flow: Input tokens feed directly into the trainable LLM, which generates continuous and explicit output tokens. 2. **CODI**: - Input: Gray (Prompt), blue (Continuous Tokens), green (Explicit Tokens) tokens. - Model: "Latent Reasoning Model" (brown box, trainable: flame icon) with "Hidden States" (pink dashed box) connected to the model. - Output: Blue (Continuous Tokens), green (Explicit Tokens) tokens. - Flow: Input tokens feed into the latent reasoning model, which uses hidden states to generate continuous and explicit output tokens. 3. **Unnamed LLM Variant**: - Input: Gray (Prompt), blue (Continuous Tokens), green (Explicit Tokens) tokens. - Model: "Large Language Model" (brown box, trainable: flame icon) with "Hidden States" (pink dashed box) connected to the model. - Output: Blue (Continuous Tokens), green (Explicit Tokens) tokens. - Flow: Input tokens feed into the LLM, which uses hidden states to generate continuous and explicit output tokens. 4. **LightThinker**: - Input: Gray (Prompt), blue (Continuous Tokens), green (Explicit Tokens) tokens. - Model: "Large Language Model" (brown box, trainable: flame icon). - Output: Blue (Continuous Tokens), green (Explicit Tokens) tokens, with "Mask" (pink dashed boxes) applied to some output tokens. - Flow: Input tokens feed into the LLM, which generates masked continuous and explicit output tokens. #### Auxiliary Continuous Thought (Bottom Section) 1. **CCoT**: - Input: Gray (Prompt), blue (Continuous Tokens), green (Explicit Tokens) tokens. - Components: "LLM Layer i" (pink box, i=L), "Scorer" (white box, trainable: flame icon), "CCoT φ" (brown box, trainable: flame icon), "Decoder ψ" (brown box, trainable: flame icon). - Output: Blue (Continuous Tokens), green (Explicit Tokens) tokens. - Special: "Loss" (pink dashed box) connected to "LLM Layer i". - Flow: Input tokens feed into "LLM Layer i", which connects to the Scorer, CCoT φ, and Decoder ψ (all trainable) to generate output tokens; loss is computed from LLM Layer i. 2. **SoftCoT**: - Input: Gray (Prompt), blue (Continuous Tokens), green (Explicit Tokens) tokens. - Components: "Large Language Model" (brown box, frozen: snowflake icon), "Assistant LLM" (brown box, frozen: snowflake icon), "Linear Layer" (white box, trainable: flame icon). - Output: Blue (Continuous Tokens), green (Explicit Tokens) tokens. - Flow: Input tokens feed into the frozen LLM and Assistant LLM, which connect to the trainable Linear Layer to generate output tokens. ### Key Observations - **Categorization**: The diagram splits continuous thought approaches into **intrinsic** (integrated into the LLM) and **auxiliary** (separate external components). - **Trainable vs Frozen**: Intrinsic approaches (COCONUT, CODI, LightThinker) and CCoT use fully trainable components; SoftCoT uses frozen LLMs with a trainable Linear Layer. - **Special Elements**: Hidden states (CODI, unnamed LLM) imply latent reasoning; masks (LightThinker) suggest selective token processing; scorers/decoders (CCoT) indicate structured output generation; loss (CCoT) shows training signal integration. - **Token Consistency**: All approaches use three token types (Prompt, Explicit, Continuous) as defined in the legend. ### Interpretation This diagram clarifies tradeoffs in LLM continuous thought design: - **Intrinsic Approaches**: Embed reasoning directly into the LLM, simplifying architecture but requiring full model training. Hidden states (CODI) enable latent reasoning, while masks (LightThinker) allow selective token processing for efficiency. - **Auxiliary Approaches**: Use modular external components to augment LLMs. CCoT integrates training loss for structured output, while SoftCoT leverages pre-trained frozen LLMs with a trainable linear layer to reduce training costs. - **Practical Impact**: Researchers can choose intrinsic methods for end-to-end reasoning or auxiliary methods for modular, efficient augmentation. The diagram highlights how each approach modifies token flow, model training, and reasoning logic to enhance LLM capabilities.