Matt Shumer's Reflection Llama-3.1 70B launches with new 'reflection-tuning' technique

A 70-billion-parameter model is uploaded to Hugging Face with a new training method called reflection-tuning.

Matt Shumer today released Reflection 70B, a 70-billion-parameter language model hosted on Hugging Face. The model is built on Meta’s Llama 3.1 70B Instruct and employs a new training method called reflection-tuning. According to the model card, this technique allows the model to output reasoning inside tags, catch errors with tags, and then deliver a final answer inside tags. The model card acknowledges an earlier upload issue and asks users to try again if they had poor results. The release page promises a dataset and a report next week, alongside a larger 405B model that Shumer says will be “the top-performing LLM in the world, including closed-source models.”

The record

Hugging Face: mattshumer/Reflection-Llama-3.1-70B

The room reactsas it happened

Matt Shumer@mattshumer

Reflection-Tuning teaches a LLM to detect mistakes in its reasoning and correct course.

One year later — open only if you can handle spoilers

Over the following week, independent evaluations could not reproduce Reflection 70B's benchmark scores, and evidence emerged that the demo API was proxying Anthropic's Claude. The episode became a cautionary tale about hype and reproducibility in open-source AI, though Shumer later released a corrected model with more modest claims.

Replay thisPost on X Reddit HN LinkedIn