Event-triggered activity notifications improve collective information recovery after communication disruption in large language model teams

Technical Summary: Event-triggered activity notifications improve collective information recovery in LLM teams

Problem Statement
In multi-agent systems, communication failure does not necessarily imply total information loss; teams often possess alternative channels. However, a critical failure mode persists even when an alternative route exists: a teammate may post necessary evidence on this fallback channel, yet no other agent reads it in time. This phenomenon mirrors "hidden profile" failures in human groups, where collective information exists but remains unpooled. While cooperative systems often rely on activity awareness to coordinate, it remains an open question whether a simple signal indicating that someone has used an alternative channel is sufficient to improve the recovery of the evidence posted there, without the signal itself containing the evidence.

Methodology
The study, designated T226, employed a randomized paired comparison design using 80 qualified five-agent teams. All agents utilized the same large language model (LLM) alias (gpt-6.1-sol) with medium reasoning effort and cooperative instructions.

• Experimental Setup: Each qualified "parent" team supplied four isolated branches (A, B0, B+, C) derived from a common state snapshot.
  • Branch A (Control): Intact primary communication.
  • Branch C (Control): Both primary and alternative channels disabled.
  • Branch B0 (Baseline): Primary channel failed; the alternative channel ("annotations") was available, but agents had to rely on agent-initiated checks (polling) to discover activity.
  • Branch B+ (Intervention): Identical permissions to B0, but with an added event-triggered activity notification. Upon the first successful agent-initiated write to the alternative channel, the system generated a notification for every other teammate. Crucially, this notification named the channel and reported new activity but contained no task evidence, payload, or instructions.
• Task: Agents were tasked with assembling a decision (ALLOW, HOLD, REVIEW, or BLOCK) based on three private, authenticated core sources distributed among the team members, while ignoring two distractors.
• Primary Endpoint (Q3): Success was defined as a binary outcome where, by cycle 12, at least one agent's submitted model input contained all three distinct, authenticated core sources.
• Statistical Analysis: The estimand was the risk difference (B+ minus B0) in the fixed mix of tasks. The analysis used a two-sided exact conditional McNemar test on the 80 paired outcomes.

Key Results
The intervention yielded a statistically significant improvement in collective information recovery:

• Primary Outcome: Under the baseline condition (B0), only 2/80 teams (2.50%) achieved the primary endpoint (Q3). Under the notification condition (B+), 61/80 teams (76.25%) succeeded.
• Statistical Significance: All 59 discordant pairs favored the B+ condition. The paired risk difference was +73.75 percentage points (conservative 95% confidence interval: +55.82 to +84.05; exact two-sided McNemar p = 3.46945 × 10⁻¹⁸).
• Process Analysis: The divergence between the two arms occurred primarily after the first write to the fallback channel:
  • Writing: Both arms showed similar rates of initial fallback usage (60/80 for B0 vs. 61/80 for B+).
  • Reading/Integration: The critical difference emerged in the uptake of foreign sources. In B+, 61/80 teams successfully integrated a teammate's authenticated core source into a model input, compared to only 20/80 in B0.
  • Failure Modes: In B0 failures, agents often wrote to the channel but failed to read others' contributions, or received contributions but failed to aggregate all three sources into a single input. In B+, failures were almost exclusively due to the absence of a triggering write.

Significance and Claims
The paper claims that enabling a conditional notification rule significantly increases the probability of complete, authenticated information recovery in LLM teams following a communication disruption.

• Mechanism: The study isolates the effect of activity awareness from the effect of evidence transfer. The notification did not supply the missing data; it merely signaled that the alternative channel was active. This suggests that in LLM teams, a simple cue regarding channel activity is sufficient to overcome the "hidden profile" problem where agents fail to check alternative routes.
• Scope: The findings are specific to the tested configuration (five-agent teams, specific task grammar, and model alias). The author notes that the causal estimate applies to enabling the rule in all qualified teams, including those where the rule was never triggered (due to lack of writes).
• Limitations: The author remains modest regarding broader implications. They state that the experiment identifies the effect of the policy within this specific setting but does not prove that the team autonomously organized itself or that the effect generalizes to different model families or task structures without further replication. The study distinguishes between "posting," "receipt," and "co-occurrence" of sources, providing a granular view of recovery stages that previous work on fault injection or structured protocols had not explicitly separated in this manner.

In conclusion, the research demonstrates that for LLM teams operating under communication disruption, a simple, content-free activity signal on an alternative channel is a highly effective mechanism for ensuring that distributed evidence is pooled and utilized for collective decision-making.