# kuleshov-group/bd3lms

[ICLR 2025 Oral] Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models

Repository: https://github.com/kuleshov-group/bd3lms
Canonical: https://ross.abutalabs.com/products/bd3lms
Homepage: https://m-arriola.com/bd3lms/
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2025-07-10T19:51:46+00:00

## Health v2 (maintenance only)
Score: 34/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 31, release rhythm 35, longevity 39
- inputs: {"age_days": 552, "days_push": 419, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1029, forks 78 (observed 2026-08-28T04:03:17.667607+00:00)

## What it is
BD3-LMs is a research implementation of Block Discrete Denoising Diffusion Language Models that interpolate between autoregressive and diffusion language models. It provides training and evaluation code, noise schedules, samplers supporting arbitrary-length generation, and baseline implementations of AR, SEDD, MDLM, and SSD-LM.

## Use cases
- train a block diffusion language model
- generate arbitrary-length text with diffusion models
- compare autoregressive vs diffusion language model baselines
- experiment with discrete denoising diffusion for text
- reproduce ICLR 2025 block diffusion results
- tune block size to trade off generation quality and speed
- evaluate likelihoods of diffusion language models

## When to choose
- you need flexible-length generation with KV caching and parallel sampling
- you want state-of-the-art diffusion language modeling likelihoods
- you are researching hybrid autoregressive-diffusion models
- you need baseline implementations of SEDD, MDLM, or AR models

## When to avoid
- you just need a production-ready LLM for inference
- you want a simple autoregressive model without diffusion
- you need a plug-and-play chatbot or API service
- you lack GPU resources for training language models

## Facets
- artifact type: library
- maturity: active
- function: llm-training, llm-inference, machine-learning, deep-learning, benchmarking
- domain: large-language-models, machine-learning, deep-learning
- platform: python
- tags: diffusion-language-models, block-diffusion, discrete-diffusion, iclr-2025, research-code, text-generation, kv-caching, huggingface, natural-language-processing, research, gpu, linux

## Member repositories
- kuleshov-group/bd3lms (main) score 34

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:17.667607+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T07:07:39.622198+00:00, confidence not recorded.
  - readme: https://github.com/kuleshov-group/bd3lms (fetched 2026-08-28T04:03:17.667607+00:00, sha 02a11d3cd57a)
  - homepage: https://m-arriola.com/bd3lms/ (fetched 2026-08-29T13:07:42.792373+00:00, sha 56df496f40b5)
- Data as of 2026-08-30T08:39:29.467469+00:00.