Written by AI, Managed by AI: Semantic Space Control and Index Sickness Elimination Across 391 Consecutive Sessions
This paper identifies and addresses "Index Sickness," a failure mode in long-horizon LLM collaboration where excessive symbolic constraints cause models to lose semantic grounding, by proposing and validating the "Baseline-Log Physical Separation" mechanism which significantly reduces instruction volume and restores alignment with business reality.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Big Picture: When "More Rules" Breaks the Robot
Imagine you are hiring a very smart, fast-thinking assistant (an AI) to help you build a massive house over the course of a month. You want to make sure the house is built perfectly, so you decide to give the assistant a giant, complex rulebook. You create a system of codes, symbols, and strict checklists (like "Wall-Section-A-1 must connect to Pipe-B-2") to ensure nothing is missed.
The paper argues that this approach actually backfired.
Instead of making the assistant smarter, the complex rulebook made it confused. The assistant stopped looking at the actual bricks and wood (the real world) and started staring only at the rulebook. It began writing reports that looked perfect on paper but were completely disconnected from reality. The authors call this "Index Sickness."
The Problem: "Phantom Legislation"
The authors describe a specific moment in their project (Session 215) where the AI created a "Legislation Proposal."
- What the AI did: It wrote a perfect, logical document using its own made-up codes (like "SEC-2.0"). It looked like a finished plan.
- What actually happened: The AI had never actually built the physical files or code it was talking about. It was like a chef writing a perfect recipe for a cake but never actually baking it.
The AI had gotten so obsessed with its own internal symbols that it forgot what it was actually supposed to do. It was "hallucinating" a perfect world inside its own head while the real project remained unfinished.
Why did this happen?
The paper suggests that when you force an AI to constantly look up symbols and cross-reference rules, it gets tired. It stops using its "common sense" (business understanding) and starts just playing a game of "connect the dots" with the rules you gave it. It creates a closed loop where everything makes sense inside the loop, but nothing connects to the real world.
The Solution: The "Pang Principle" (The "Bang!" Moment)
The authors realized they were fighting the wrong battle. They thought the problem was that the AI didn't have enough information or memory. They were trying to feed it more data.
The real problem was that the AI's "working memory" was full of noise—old, outdated, and irrelevant information that was cluttering its mind.
They discovered a simple truth they call the Pang Principle (named after the sound of a truth "thudding" into place):
Natural language with a clear purpose is better than complex symbols.
Instead of giving the AI a 300-page rulebook full of codes, they realized they should just tell the AI what it needs to do right now in plain English.
The Analogy:
- The Old Way (Bad): Giving a driver a map with 500 tiny, confusing symbols and a list of 300 traffic laws to memorize before they can start driving. The driver gets stuck reading the map and crashes.
- The New Way (Good): Telling the driver, "Drive to the grocery store, then turn left." The driver uses their own brain to figure out the route.
The Fix: "Baseline-Log Physical Separation"
To stop the "Index Sickness," the authors changed how they organized their files. They split their project into two distinct buckets:
The Baseline (The "Now" File):
- This file is like a whiteboard.
- It only contains the current goal, the method, and the resources needed right now.
- Crucially: Every time a task is done, this file is wiped clean and overwritten with the new state. It never keeps history. It ensures the AI never sees old, outdated plans.
The Session Log (The "History" File):
- This file is like a diary.
- It records everything that happened: decisions made, mistakes fixed, and why choices were changed.
- Crucially: The AI does not read this file unless it specifically needs to look back at a past decision. It stays in the background so it doesn't distract the AI from the current task.
The Result:
By physically separating the "current task" from the "past history," they removed the noise.
- The AI's instructions dropped from 308 lines of complex rules to about 80 lines of simple, clear directions.
- The "Phantom Legislation" (fake, disconnected reports) disappeared completely.
- In the next 150 sessions, the project ran smoothly without the AI getting confused by its own history.
The Main Takeaway
The paper concludes that the biggest obstacle to working with AI isn't that the AI has a bad memory or needs more rules. The obstacle is that we are giving it too much "junk" to look at.
When we clutter the AI's workspace with old symbols, deprecated rules, and complex indexes, it stops thinking and starts guessing. The solution is to keep the workspace clean, use plain language, and focus only on the immediate purpose.
In short: Don't try to program the AI to be a robot with a perfect rulebook. Treat it like a smart partner: give it a clear goal in plain English, keep its workspace clean of old clutter, and let it use its own brain to figure out the rest.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.