Arxiv Paper Tracker

2. Competition and Cooperation of LLM Agents (arXiv:2604.00487) - Utility: 0.87

• Title: Competition and Cooperation of LLM Agents in Games
• Authors: Jiayi Yao et al.
• Key Finding: LLM agents tend to cooperate rather than converge to Nash equilibria
• Games Tested: Network resource allocation game, Cournot competition game
• Insight: Fairness reasoning is central to cooperative behavior in non-zero-sum contexts
• Practical Use: Understanding strategic behavior patterns in competitive multi-agent settings
• URL: https://arxiv.org/abs/2604.00487

Memory Management

3. Novel Memory Forgetting Techniques (arXiv:2604.02280) - Utility: 0.88

• Title: Novel Memory Forgetting Techniques for Autonomous AI Agents: Balancing Relevance and Efficiency
• Authors: Sunil Tiwari
• Problem: Long-horizon agents suffer from temporal decay and false memory propagation
• Solution: Adaptive budgeted forgetting framework with:
  • Relevance-guided scoring
  • Bounded optimization
  • Integration of recency, frequency, and semantic alignment
• Results:
  • Long-horizon F1 > 0.583 baseline
  • Higher retention consistency
  • Reduced false memory behavior
  • No increase in context usage
• Practical Use: Prevent memory bloat in extended conversational agents
• URL: https://arxiv.org/abs/2604.02280

Agent Architecture

4. The Self Driving Portfolio (arXiv:2604.02279) - Utility: 0.85

• Title: Agentic Architecture for Institutional Asset Management
• Authors: Andrew Ang
• Scale: ~50 specialized agents, 20+ portfolio construction methods
• Architecture Components:
  • Capital market assumptions agents
  • Portfolio construction agents (competing methods)
  • Critique and voting system
  • Researcher agent (proposes new methods)
  • Meta-agent (rewrites code/prompts based on past performance)
• Governance: Investment Policy Statement constrains autonomous agents
• Practical Use: Reference architecture for complex multi-agent systems with self-improvement
• URL: https://arxiv.org/abs/2604.02279

Evaluation Methods

5. User Turn Generation Probe (arXiv:2604.02315) - Utility: 0.86

• Title: User Turn Generation as a Probe of Interaction Awareness in Language Models
• Authors: Sarath Shekkizhar
• Key Concept: Interaction awareness - ability to generate grounded follow-ups
• Method: Let model generate under user role after seeing assistant response
• Finding: Interaction awareness is decoupled from task accuracy
• Tested Models: 11 open-weight LLMs (Qwen3.5, gpt-oss, GLM) across 5 datasets
• Insight:
  • Qwen3.5 family: GSM8K accuracy scales from 41% to 96.8%, but follow-up rates near zero under deterministic generation
  • Higher temperature sampling reveals latent awareness (22% follow-up rate)
• Practical Use: Evaluate conversational AI beyond task accuracy
• URL: https://arxiv.org/abs/2604.02315

Related Resources

Benchmarks

• MultiAgentBench (arXiv:2503.01935): Comprehensive multi-agent evaluation with milestone-based KPIs
• MAFBench (arXiv:2602.03128): Unified framework-level evaluation suite
• Theoretical Production Benchmark v2.0: Evaluating sustained conceptual coherence

Memory Systems

• A-MEM (arXiv:2502.12110): Zettelkasten-inspired agentic memory
  • Dynamic indexing and linking
  • 6-fold improvement in multi-hop reasoning
  • 85-93% reduction in token usage
• A-MemGuard: Proactive defense against memory poisoning attacks

Paper Collections

• VoltAgent/awesome-ai-agent-papers: Curated 2026 papers on multi-agent coordination, memory/RAG, tooling, evaluation, security

Trends (April 2026)

1. Multi-agent coordination is mainstream - Multiple papers on topology optimization, cooperation, competition
2. Memory management maturing - From simple storage to adaptive forgetting, dynamic organization
3. Production architectures emerging - Real-world applications showing complete system designs
4. Evaluation expanding - New dimensions beyond accuracy: interaction awareness, cooperation quality, conceptual coherence
5. Self-improvement mechanisms - Meta-agents that rewrite code/prompts based on performance

Quick Reference Commands

# Search for recent agent papers
web_search query:"arXiv LLM agents multi-agent coordination memory RAG 2026"

# Fetch specific paper
web_fetch url:"https://arxiv.org/abs/<paper_id>"

# Check awesome collection
web_fetch url:"https://github.com/VoltAgent/awesome-ai-agent-papers"

Last updated: 2026-04-03 22:00 (Asia/Shanghai) Papers analyzed: 188 submissions from Fri, 3 Apr 2026 High-utility count: 5 papers (utility >= 0.85)

Activation Keywords

• arxiv-paper-tracker
• arxiv-paper-tracker 技能
• arxiv-paper-tracker skill

Tools Used

• read - Read documentation and references
• web_search - Search for related information
• web_fetch - Fetch paper or documentation

Instructions for Agents

Follow these steps when applying this skill:

Step 1: Multi-agent coordination is mainstream

Step 2: Memory management maturing

Step 3: Production architectures emerging

Step 4: Evaluation expanding

Step 5: Self-improvement mechanisms

Examples

Example 1: Basic Application

User: I need to apply arXiv Paper Tracker to my analysis.

Agent: I'll help you apply arxiv-paper-tracker. First, let me understand your specific use case...

Context: Apply the methodology

Example 2: Advanced Scenario

User: Complex analysis scenario

Agent: Based on the methodology, I'll guide you through the advanced application...

Example 2: Advanced Application

User: What are the key considerations for arxiv-paper-tracker?

Agent: Let me search for the latest research and best practices...