Research map / Multi-agent systems and games / LLM agents

PRIME: Planning and Retrieval-Integrated Memory for Enhanced Reasoning

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PRIME, a dual-system multi-agent framework where a fast System 1 agent answers first and a reflection agent triggers a planning/retrieval/hypothesis-testing System 2 pipeline when uncertainty is detected, evaluated with LLaMA 3 models on medical and multi-hop QA.

Problem

LLMs' fast intuitive answers are prone to hallucination and errors, while always-on deliberate multi-step reasoning is costly; the paper targets knowledge-intensive multi-hop and medical reasoning with adaptive compute.

Approach

Combines a subquestion-decomposing fast answer with a self-reflection gate that selectively invokes a modular System 2 agent pipeline (planning, search, reading, hypothesis, integration, decision), mirroring dual-process cognition.

Key results

• On MedQA/MedMCQA/MMLU-Medical average, PRIME with LLaMA 3.3 70B reaches 86.39 vs. 83.57 for GPT-4o and 81.10 for GPT-4 (reported as competitive/outperforming on MedQA and MMLU-Medical).
• With LLaMA 3.1 8B, PRIME averages 75.99 vs. 71.47 for Search-O1 and 71.21 for i-MedRAG.
• On Musique, 2Wiki, and HotpotQA with LLaMA 3.3 70B, PRIME reaches 35.17/48.81 (EM/F1), 68.84/79.81, and 46.51/60.68, exceeding Search-O1 on all three.

Surprising findings

• On Amboss, System 2 underperforms System 1 on Easy questions (77.78% vs 92.68%), suggesting deliberation can hurt on simple items.
• Reported that the model is often anchored to its initial hypotheses, limiting belief revision.

Limitations

• Performance depends on the Reflection Agent correctly detecting uncertainty; missed triggers leave errors uncorrected.
• Model is often anchored to initial hypotheses and struggles to revise beliefs against contradictory evidence.
• Multi-agent pipeline adds latency and complexity; evaluated mainly on QA-style tasks.

Experimental setup

Datasets: MedQA MedMCQA MMLU-Medical Musique 2Wiki HotpotQA Amboss

Benchmarks: MedQA MedMCQA MMLU-Medical Musique 2Wiki HotpotQA Amboss

Metrics: accuracy exact match F1 average accuracy over tokens

Baselines: CoT Self-Consistency MedRAG i-MedRAG Search-O1 Meditron 70B Mixtral 8x7B GPT-3.5 GPT-4 GPT-4o-mini GPT-4o Naive RAG

Builds on: Chain-of-Thought Self-Consistency Search-O1 i-MedRAG ReAct

Closest papers by meaning

• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.87)
• Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.87)
• Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · similarity 0.86)
• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.86)
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.86)
• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.86)