Causal Audit Finds Some AI Agent Cache Relays Transfer Nothing

Researchers swapped multi-agent AI systems' shared memory caches with blanks, noise, and scrambled versions to test whether the content matters, per an August causal audit.

The stakes are real for anyone crediting multi-agent gains to shared latent thought. On GSM8K and ARC-Challenge, a scrambled cache performed within 2.8 points of the real one across three Qwen3 model sizes. Wiping the cache entirely cost up to 14.7 points in one test case.

Document QA broke the other way. There, the receiving agent needs a specific fact from the sender's text. Real caches solved the task 100% of the time, against 23-25% for swapped ones, replicated across three model families and five checkpoints.

The gap tracks what each task requires. Math and trivia questions don't hinge on one agent handing another a private fact, so a wrong cache barely hurts. Retrieval-style QA does hinge on that fact, so gutting the cache guts the answer.

Checked against three released architectures, LatentMAS's cache relay hit the ceiling for real transfer. KVComm's partial layer-subset design showed some. C2C's projector-based relay showed none detected.

Papers crediting multi-agent reasoning gains to shared latent thoughts are likely pointing at the wrong mechanism. On math and trivia benchmarks, a scrambled cache does almost as well as the real one. Whatever drives the improvement, it isn't the specific content being relayed.