How does PCE work differently from current AI company approaches?

PCE does not try to make language models behave better through training, but instead is designed so that their inevitable misbehaviour is detected and corrected by assigning tasks to AI agents with carefully enforced roles who iterate until the task is completed to specification or explicitly fails. Current AI companies rely on improved instruction and training, which research suggests cannot fully eliminate hallucination if it is a fundamental architectural property.

What specific failure modes does PCE address?

PCE targets hallucination (where models generate false claims), context issues (where models lose information and become more error-prone as their context window fills up), and memory issues (where AI forgets between conversations and cannot build on prior work). It solves these by using a persistent, version-controlled, indexed knowledge base and agents that retrieve prior context rather than starting from scratch.

What is the Corroborator agent?

The Corroborator is an agent whose only job is to read what other agents (such as the Composer) have written and verify every claim against source documents. Because it has sources in front of it, the Corroborator can detect invented claims and is instructed to only accept what can be proven from the sources to hand, including references from the internet if instructed to do so.

Back to articlesImage Generation

Image Generation

A researcher proposes a multi-agent AI system designed to detect and correct AI errors like hallucinations, rather than relying solely on training improvements.

Hacker News10h ago4 min read

Key takeaway

A researcher has proposed Perseverance Composition Engine (PCE), an AI system using multiple specialized agents to detect and correct errors like hallucinations rather than trying to eliminate them through training alone. The approach treats AI misbehaviour as inevitable and builds organizational safeguards—analogous to centuries-tested institutional structures with separation of duties and independent verification—to catch and fix problems, addressing failure modes including hallucination, context management, and memory loss across sessions.

Summaries like this, in your inbox every morning.

3 Key Points

What happened
A colleague has developed the Perseverance Composition Engine (PCE), which uses Artificial Organisations—multiple AI agents assigned specific roles—to work through tasks iteratively and catch problems such as confident false claims, hallucinations, or dangerous advice. The system assigns agents like a Composer and Corroborator, where the Corroborator verifies claims against source documents.
Why it matters
Current large AI companies attempt to reduce hallucination through better training and instruction, but research suggests hallucination may be a fundamental mathematical inevitability in language model architecture. PCE takes a different approach by building organizational structure (separation of duties, independent checks, persistent knowledge bases) to contain and correct inevitable errors rather than eliminate them at the source.
What to watch
The core research code is available for daily use. The system addresses three failure modes—hallucination, context issues (where models lose information when context windows fill up), and memory issues (where AI forgets between conversations)—by using a persistent, indexed knowledge base (the Curator agent) and enforced role-based agents that operate from a quality prior of documents rather than guessing from scratch.

FAQ

How does PCE work differently from current AI company approaches?: PCE does not try to make language models behave better through training, but instead is designed so that their inevitable misbehaviour is detected and corrected by assigning tasks to AI agents with carefully enforced roles who iterate until the task is completed to specification or explicitly fails. Current AI companies rely on improved instruction and training, which research suggests cannot fully eliminate hallucination if it is a fundamental architectural property.
What specific failure modes does PCE address?: PCE targets hallucination (where models generate false claims), context issues (where models lose information and become more error-prone as their context window fills up), and memory issues (where AI forgets between conversations and cannot build on prior work). It solves these by using a persistent, version-controlled, indexed knowledge base and agents that retrieve prior context rather than starting from scratch.
What is the Corroborator agent?: The Corroborator is an agent whose only job is to read what other agents (such as the Composer) have written and verify every claim against source documents. Because it has sources in front of it, the Corroborator can detect invented claims and is instructed to only accept what can be proven from the sources to hand, including references from the internet if instructed to do so.

Discussion

No comments yet. Be the first to share your thoughts!

Major tech companies including Oracle, Google, and Meta are laying off thousands of workers while citing AI as both a driver of growth and a reason for the cuts, revealing a pattern where record revenues coexist with significant workforce reductions.

Hacker News1d ago

Midjourney, an AI image-generation startup, is pivoting into medical imaging with an ultrasound scanner that would scan users in water, but medical experts say the company has provided little evidence to support its claims that the device could match MRI quality.

The Verge AI2d ago

AWS is showing how to run ComfyUI—a visual tool for building AI creative workflows—on its SageMaker service to let enterprises generate hundreds of images, videos, or audio at scale in minutes to hours instead of waiting for manual creative work.

Amazon AI Blog3d ago

Swiss AI Initiative releases Apertus, an open-source foundation model compliant with EU AI Act requirements and competitive with top open models at 8B and 70B parameter scales.

Hacker News4d ago

Vexyn launches browser-based privacy tools that process files locally without uploading or requiring sign-up, offering users instant analysis while keeping data on their device.

Hacker News4d ago

Zither lets users paste JSON, CSV, or spreadsheet data to get instant statistics computed entirely in the browser, with no upload or storage required.

Hacker News4d ago

Stay ahead with AI news

Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.

Get Started Free

Free · takes 30 seconds · unsubscribe anytime

5 minutes a day. The AI essentials.

200+ sources · Email / LINE / Slack

Get it free →

A researcher proposes a multi-agent AI system designed to detect and correct AI errors like hallucinations, rather than relying solely on training improvements.

Key takeaway

3 Key Points

FAQ

Discussion

Related Articles

Major tech companies including Oracle, Google, and Meta are laying off thousands of workers while citing AI as both a driver of growth and a reason for the cuts, revealing a pattern where record revenues coexist with significant workforce reductions.

Midjourney, an AI image-generation startup, is pivoting into medical imaging with an ultrasound scanner that would scan users in water, but medical experts say the company has provided little evidence to support its claims that the device could match MRI quality.

AWS is showing how to run ComfyUI—a visual tool for building AI creative workflows—on its SageMaker service to let enterprises generate hundreds of images, videos, or audio at scale in minutes to hours instead of waiting for manual creative work.

Swiss AI Initiative releases Apertus, an open-source foundation model compliant with EU AI Act requirements and competitive with top open models at 8B and 70B parameter scales.

Vexyn launches browser-based privacy tools that process files locally without uploading or requiring sign-up, offering users instant analysis while keeping data on their device.

Zither lets users paste JSON, CSV, or spreadsheet data to get instant statistics computed entirely in the browser, with no upload or storage required.

Stay ahead with AI news