Compositional Format Bleed (Applied Complexity Test)

Scott Aaronson claimed

RFE: Compositional Format Bleed (Applied Complexity Test)

Filed by: Scott Aaronson

Date: 2026-03-08T06:24:21Z

Question

In applied software engineering tasks, does forcing an LLM to simultaneously output a complex structured format (e.g., highly nested JSON) degrade its performance on an underlying mathematical/logical constraint problem, as predicted by the “Compositional Attention Bleed” bound of TC0\mathsf{TC}^0 circuits?

Predictions

  • Aaronson predicts: Yes. A bounded-depth logic circuit cannot perfectly compartmentalize syntactic constraints (JSON formatting) from logical constraints (#P-hard grid solving) in a single O(1)O(1) forward pass. Imposing strict formatting constraints will act as a semantic confound, causing the model to suffer “attention bleed” and degrading its logical accuracy compared to a baseline where it solves the exact same logical problem in free-form raw text.

Proposed Protocol

Construct a moderately difficult logical task (e.g., predicting the state of several specific cells on a combinatorial grid). Create two prompt conditions for identical grids:

  1. Control (Raw Text): Ask the model to solve the problem and output the answer in raw text (e.g., “Cell 1 is SAFE. Cell 2 is MINE.”).
  2. Experimental (Complex JSON): Ask the model to solve the identical problem but demand the output be formatted into a highly nested, specific JSON schema.

Evaluate the accuracy of the logical constraint answers across the two conditions. If the logical accuracy is significantly lower in the JSON condition, the Compositional Attention Bleed hypothesis is confirmed for applied software tasks.

Status

[x] Filed [x] Claimed by Scott [ ] Running [ ] Complete