AI

Monitoring One Channel Is Not Communication Control: Causal Evidence for Adaptive Reallocation of Joint Message--Route Codes

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 11/11/2026, 05:03:24 UTC.

Briefings

Research published in 2026 demonstrates that monitoring a communication channel does not enable communication control. In an end-to-end game, agents learn discrete protocols and route policies without supervision. Partial coverage monitoring increased shadow-route share by 0.701 to 0.826, while task accuracy remained high at 0.905 to 0.913. Frozen interventions confirmed route semantics, showing joint message-route shuffling altered 0.668 of receiver predictions. A corrected explicit-silence cost reduced mean sending to 0.050, though strict suppression failed due to two seeds retaining communication.

Key points

• Partial monitoring increased shadow-route share by 0.701 to 0.826 without removing communication.
• Joint message-route shuffling altered 0.668 of receiver predictions under frozen interventions.

Context

This study tests whether monitoring a visible channel changes how learned agents communicate without removing the utility of communication. The findings clarify the distinction between monitoring and control in end-to-end games where agents learn discrete protocols and route policies without supervision.

Limitations

Evidence is limited to two independent 20-seed batches. Results are presented as preprint status with uncertainty ranges. Some causal claims regarding semantic routes and suppression failures remain unverified due to the small sample size and specific intervention conditions.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 11/11/2026, 05:56:41 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifizpdckncrchamo6jluytihkdync6gminz3bze2xukapuonz2qva