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

XBRIDGE: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 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 XBridge, a decode-free heterogeneous LLM communication protocol that pairs token-level lexical anchor mapping with a gated cross-attention latent enrichment bridge, training only the bridge while sender and receiver stay frozen.

Problem

Cross-architecture latent communication between LLMs with different tokenizers and hidden spaces loses entity identity (rare-token compression collapse), while text-based communication is slow and lossy.

Approach

Combine a deterministic, training-free mapping of sender context tokens into receiver vocabulary embeddings (discrete entity anchors) with a learned gated cross-attention bridge that lets the receiver query sender hidden states at four layers; the receiver's self-attention binds the contextual signal to the anchored entities.

Key results

• XBridge averages 63.2 F1 vs 41.8 for NLComm-hetero on Llama->Qwen, outperforming NLComm on all 7 tasks across three model pairs.
• XBridge latency is 0.15 s/sample vs 1.70 s for NLComm on HotpotQA (about 11x lower).
• Removing LAM drops HotpotQA F1 from 78.8 to 30.3 (near NoComm 24.6); removing LEB gives 56.5.

Surprising findings

• Bridge-only (LEB without entity anchors) falls to roughly NoComm level, showing continuous bridges alone lose entity identity.
• Under same-architecture settings, the XBridge bridge outperforms KVComm on 6 of 7 tasks, but KVComm wins on MultiFieldQA.

Limitations

• Evaluated only one-directional, single-exchange communication; bidirectional and multi-turn dialogue not tested.
• Multi-agent composition tested only with two independently trained bridges on HotpotQA.
• Balanced training set (587 samples) becomes a bottleneck for larger senders such as 14B.

Experimental setup

Datasets: HotpotQA MuSiQue QASPER 2WikiMultihopQA MultiFieldQA-en Countries Tipsheets

Benchmarks: HotpotQA MuSiQue QASPER 2WikiMultihopQA MultiFieldQA-en Countries Tipsheets

Metrics: F1 latency (s/sample) entity rank cosine similarity

Baselines: NLComm KVComm FullComm NoComm C2C

Builds on: Flamingo KVComm NLComm C2C

Closest papers by meaning

• KVComm: Enabling Efficient LLM Communication through Selective KV Sharing (ICLR 2026 · similarity 0.87)
• Latent Collaboration in Multi-Agent Systems (ICML 2026 · similarity 0.86)
• Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression (ICML 2025 · similarity 0.86)
• Let Models Speak Ciphers: Multiagent Debate through Embeddings (ICLR 2024 · similarity 0.85)
• KVCOMM: Online Cross-context KV-cache Communication for Efficient LLM-based Multi-agent Systems (NeurIPS 2025 · similarity 0.85)
• Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.85)