raymin0223/mixture_of_recursions

• highreadme#1

  Add an explicit introductory sentence to the README to clarify the repo's purpose.

  Why:

  COPY-PASTE FIX

  Copy

  This repository provides the official PyTorch implementation of the Mixture-of-Recursions (MoR) model, as presented in our NeurIPS 2025 paper, focusing on adaptive token-level computation for efficient LLMs.

  OpenDoneDismiss
• mediumreadme#2

  Introduce a 'Key Features' section to highlight the implementation's specific contributions.

  Why:

  COPY-PASTE FIX

  Copy

  ## ✨ Key Features of this Repository
  - **Official PyTorch Implementation:** Reproduce the results from our NeurIPS 2025 paper.
  - **Mixture-of-Recursions Model:** Explore dynamic recursive depths for adaptive token-level computation.
  - **KV Cache Handling:** See our novel solution for the missing Key-Value cache problem in early-exiting.
  - **Router Mechanism:** Understand how tokens are dynamically routed through the model.

  OpenDoneDismiss
• lowreadme#3

  Add a brief comparison section to the README.

  Why:

  COPY-PASTE FIX

  Copy

  ## 🆚 Comparison to Existing Adaptive Computation Methods
  Mixture-of-Recursions distinguishes itself from traditional early-exiting methods by directly addressing the Key-Value (KV) cache problem, which often limits the practical applicability of dynamic token-level computation. Unlike methods that approximate or recompute KV pairs, MoR learns dynamic recursive depths to adaptively manage computation while maintaining KV cache integrity for subsequent tokens.

  OpenDoneDismiss

Category GEO backends resolved for this scan: google/gemini-2.5-flash, deepseek/deepseek-v4-flash

• CATEGORY QUERY

  How can I make large language models run faster and more cost-effectively with adaptive computation?

  you: not recommended
  AI recommended (in order):

  1. DeepSpeed (microsoft/DeepSpeed)
  2. Fairseq (facebookresearch/fairseq)
  3. Hugging Face Transformers (huggingface/transformers)
  4. OpenAI Triton (openai/triton)
  5. PyTorch FSDP (pytorch/pytorch)
  6. TensorRT
  7. ONNX Runtime (microsoft/onnxruntime)

  AI recommended 7 alternatives but never named raymin0223/mixture_of_recursions. This is the gap to close.

  Show full AI answer
• CATEGORY QUERY

  What are approaches for implementing early exit mechanisms in LLMs to save computational resources?

  you: not recommended
  AI recommended (in order):

  1. Google's Switch Transformers
  2. Fairseq
  3. DeepSpeed
  4. BranchyNet
  5. Hugging Face Transformers
  6. PyTorch
  7. TensorFlow
  8. Adaptive Computation Time (ACT)
  9. DistilBERT
  10. TinyBERT
  11. PaddlePaddle
  12. PaddleSlim
  13. ONNX Runtime
  14. TensorFlow Serving
  15. TorchServe

  AI recommended 15 alternatives but never named raymin0223/mixture_of_recursions. This is the gap to close.

  Show full AI answer

• Metadata completeness

  pass

• README presence

  pass

• Compared to common alternatives in this category, what is the core differentiator of raymin0223/mixture_of_recursions?

  pass
  ⚠AI did not name raymin0223/mixture_of_recursions — likely talking about a different project

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?

• If a team adopts raymin0223/mixture_of_recursions in production, what risks or prerequisites should they evaluate first?

  pass
  ✓AI named raymin0223/mixture_of_recursions explicitly

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?

• In one sentence, what problem does the repo raymin0223/mixture_of_recursions solve, and who is the primary audience?

  pass
  ✓AI named raymin0223/mixture_of_recursions explicitly

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?