‘Reflection 70B’ AI model could be the answer to pesky LLM hallucinations

There’s yet another new artificial intelligence chatbot entering the already crowded space, but this one can apparently do what most can’t — learn from its mistakes.  

In a Sept. 5 post on X, HyperWrite AI CEO Matt Shumer announced the development of ‘Reflection 70B,’ claiming it to be “the world’s top open-source model.”

He added that the new AI was trained using “Reflection-Tuning,” which is a technique developed to enable LLMs to fix their own mistakes.

Reflection Llama-3.1 70B can “hold its own” against even the top closed-source models such as Anthropic’s Claude 3.5 Sonnet, and OpenAI’s GPT-4o in several benchmarks he claimed. Llama 3.1 is Meta’s open-source AI that was launched in July. 

He said that current AI models can often hallucinate but Reflection-Tuning enables them to recognize their mistakes and correct them before committing to an answer.

“Current LLMs have a tendency to hallucinate, and can’t recognize when they do so.”

An AI hallucination is a phenomenon when a generative AI chatbot perceives patterns or objects that are nonexistent or imperceptible to human observers, creating outputs that are inaccurate.

Reflection tuning is a technique used to improve AI models by having them analyze and learn from their own outputs. 

AI responses can be fed back into the AI where it can be asked to evaluate its own responses, identifying strengths, weaknesses, and areas for improvement, for example. 

The process is repeated many times, allowing the AI to continuously refine its capabilities with the goal of making it more self-aware of its outputs and better at critiquing and improving its own performance.

Shumer added that “with the right prompting, it’s an absolute beast for many use-cases,” providing a demo link for the new model. 

Related: Amazon to revamp Alexa with Anthropic’s Claude AI model: report

Microsoft-backed OpenAI released a research paper in 2023 with ideas on how to help prevent AI hallucinations.

One idea was “process supervision,” which involves training AI models to reward themselves for each individual, correct step of reasoning when they’re arriving at an answer, instead of just rewarding a correct final conclusion.

“Detecting and mitigating a model’s logical mistakes, or hallucinations, is a critical step towards building aligned AGI [artificial general intelligence],” Karl Cobbe, a researcher at OpenAI, told CNBC at the time. 

Magazine: AI drone ‘hellscape’ plan for Taiwan, LLMs too dumb to destroy humanity: AI Eye