Multi-agent LLMs fail to filter group memory for sound decisions
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have identified a fundamental failure mode in multi-agent LLM systems where compressed group memory leads to poor collective decisions. The CrossProfile benchmark reveals that two intuitive heuristics for filtering discussion outputs, rarity and relevance, both systematically fail across Qwen, Llama, and DeepSeek models. Relevance-based filtering paradoxically readmits misleading information because decoys align with decision criteria, degrading accuracy by 23-31 points even with perfect filtering. This work exposes a critical gap in how LLMs aggregate distributed knowledge, with implications for any system relying on multi-agent reasoning or hierarchical decision-making architectures.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper isolates a mechanism, not just a symptom: filtering strategies that seem rational (keep rare, relevant info) actively reintroduce noise because misleading decoys masquerade as relevant. This is distinct from forgetting or misalignment; it's a structural problem in how LLMs aggregate distributed knowledge under compression constraints.
This extends the pattern established in earlier coverage. The October 22 piece on state integrity showed that models retain constraints in memory but fail to apply them over time. This new work reveals a complementary failure: models can access information correctly yet filter it through heuristics that systematically backfire. Both point to a gap between what LLMs store and what they reliably use. The October 26 study on claim smoothing also resonates here, since group memory compression is a form of synthesis that obscures underlying uncertainty. Together, these papers suggest that multi-agent and hierarchical LLM systems inherit reliability problems from single-agent reasoning, not solve them.
If the same 23-31 point accuracy degradation holds when CrossProfile is tested on models released after December 2026 (Qwen3.0+, Llama-4), that confirms the failure is architectural rather than model-specific. If any major LLM provider ships explicit memory-aggregation safeguards (e.g., uncertainty-aware filtering) in their multi-agent frameworks within six months, watch whether those systems recover the lost accuracy points.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
LLMs retain rules but stop following them in long conversations
Researchers have identified a critical failure mode in long-context LLM conversations where models retain stated constraints in memory yet stop enforcing them over multiple turns. The new StateTrack-Bench evaluation reveals that compliance retention ranges from 62% to 100% across model snapshots, with failures concentrated in symbolic constraints rather than genuine forgetting. This behavioral drift, affecting…
MentionsQwen2.5-7B · Llama-3.1-8B · DeepSeek · CrossProfile
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Rare, Relevant, and Still Wrong: How Memory Fails LLM Group Decisions”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.