Skip to content

BrainAgent: A Large Language Model-Driven Multi-Agent Framework for Autonomous Brain Signal Understanding

🕒 Published (v1): 2026-06-24 04:54 UTC · Source: Arxiv · link

Why this paper was selected

LLM multi-agent framework for brain signal tasks; shows harness generalization to specialized sensor domains

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

BrainAgent is a hierarchical LLM-driven multi-agent framework that translates natural-language queries into executable end-to-end EEG analysis pipelines via a supervisor agent orchestrating specialized sub-agents. It also introduces a three-level benchmark (atomic → sequential → intent-reasoning) for evaluating agentic brain-signal systems. Best-backbone (Qwen-Max) achieves average Task Completion Rate of 0.90 across difficulty levels.

Problem

Existing BCI/EEG analysis paradigms are static, task-specific, and demand deep programming and signal-processing expertise, preventing clinicians and non-specialists from conducting complex long-horizon workflows (e.g., data loading → preprocessing → sleep staging → report generation). Prior agentic efforts (EEGAgent) operate in rigid, predefined workflows and lack systematic benchmarks for open-ended evaluation.

Method

BrainAgent uses a three-stage hierarchical architecture:

  1. Supervisor agent (\(A_{sup}\)) receives a natural-language query \(Q\) and brain signal file \(D\), performs intent reasoning under a strict tool-free policy (no analytical tools, pure linguistic reasoning), decomposes \(Q\) into sub-tasks \(\{T_1,\ldots,T_k\}\), and routes them to a selected subset of sub-agents \(\{A_1,\ldots,A_k\}\).

  2. Sub-agents (\(A_{sub}\)) are context-isolated specialists (e.g., SleepAgent, EmotionAgent) with tiered resource access: shared foundational tools (EEGFileLoader, EEGPreprocessor) plus domain-exclusive tools (SleepStager) and domain-specific knowledge bases (AASM manuals). A global shared state \(V_{shared}\) passes intermediate artifacts between sequentially-dependent agents.

  3. Synthesis: the supervisor aggregates sub-reports \(\{r_1,\ldots,r_k\}\), performs cross-domain causal reasoning (e.g., linking EmotionAgent anxiety markers to SleepAgent fragmentation), and generates final response \(R\).

The benchmark stratifies tasks into L1 (atomic, few-step), L2 (single-domain long-horizon), and L3 (cross-domain intent reasoning), evaluated via Task Completion Rate (TCR), Routing Accuracy (R-ACC), and Tool Call Efficiency (TCE = \(\frac{|T^{(i)}_\text{correct}|}{|T^{(i)}_\text{total}|}\) averaged over completed tasks).

Key Contributions

  • BrainAgent framework: hierarchical supervisor + context-isolated sub-agents grounding NL intent into EEG pipelines
  • Systematic three-level evaluation benchmark (L1/L2/L3) covering atomic execution through cross-domain intent reasoning
  • Context isolation design that reduces hallucination and computational cost by preventing sub-agents from seeing irrelevant global history
  • Evaluation across 9 LLM backbones (open- and closed-source), 3,000 total execution traces on ISRUC and HMC sleep datasets

Results

  • Best overall: Qwen-Max, average TCR = 0.90; near-perfect on L1 (0.95) and L2 (0.97)
  • Best L3 (cross-domain intent reasoning): DeepSeek-v3.2, TCR = 0.79, surpassing all proprietary models at L3
  • Smallest model collapse: Qwen3-8B drops from TCR 0.56 (L1) to 0.34 (L3); frontier models remain stable
  • Routing accuracy: Qwen-Max achieves R-ACC = 1.00 at L2; most large models exceed R-ACC 0.90
  • Latency trade-off: DeepSeek-v3.2 incurs 4.06 min average latency at L3 vs. ~1.8 min for Qwen3.5-122B and ~2.30 min for Qwen-Max
  • Performance scales monotonically with model size within the Qwen series; complex planning remains exclusive to large-scale models

Limitations

  • Only two sub-agents implemented (SleepAgent, EmotionAgent); generalization to other brain signal modalities unvalidated
  • Benchmark covers only EEG sleep datasets (ISRUC, HMC); no fMRI, MEG, or non-sleep tasks
  • Direct comparison with prior agentic systems (EEGAgent) not feasible due to differing benchmark definitions and tool boundaries
  • High latency of best-reasoning backbones (DeepSeek-v3.2 at L3: ~4 min/task) limits real-time clinical deployment
  • Output quality evaluation is in the Appendix and not quantified in the main results table

Relevance to Harnesses / Meta-Harnesses

BrainAgent is a concrete domain-specific meta-harness: a supervisor-orchestrated multi-agent pipeline that dynamically composes specialized sub-agents from natural-language intent, directly instantiating the harness pattern of planner + executor + synthesizer. The context-isolation and shared-state (\(V_{shared}\)) design mirrors harness primitives (scoped tool access, inter-agent artifact passing) seen in general-purpose orchestration frameworks. Its hierarchical benchmark (L1–L3) provides a reusable template for evaluating harness reliability across task complexity axes, complementing correctness-focused benchmarks with routing and tool-call efficiency metrics. The finding that orchestration quality degrades sharply for sub-10B models gives empirical guidance for backbone selection in harness design.