Enabling Agents to Communicate Entirely in Latent Space¶

🕒 Published (v1): 2025-11-12 09:37 UTC · Source: Arxiv · Venue: ACL 2026 · link

Ask a follow-up

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

TL;DR¶

Interlat replaces natural-language token exchange between LLM agents with direct transmission of last-layer hidden states, enabling richer, information-dense inter-agent communication. A separately trained compression module further reduces these latent messages to as few as 8 steps, achieving up to 24× inference speedup while preserving task performance. The approach outperforms fine-tuned CoT baselines on ALFWorld and on the hardest MATH Level-5 problems.

Problem¶

Natural language is a lossy bottleneck for multi-agent LLM communication: compressing continuous hidden states (~40k bits/state) into discrete tokens (~15 bits/token) discards alternative reasoning paths, introduces linguistic redundancy unrelated to the task, and produces ambiguous, single-trajectory messages that degrade coordination. Prior hidden-state methods either require one-shot activation grafting, are coupled to language trajectories, or need ad-hoc layer selection.

Method¶

Latent Communication. A reasoning/sender agent generates its completion normally; its last-layer hidden states H ∈ R^{L×d} (one per generated token) are extracted and injected directly into the receiver/actor agent's input embedding stream, delimited by <bop>/<eop> special tokens. A lightweight trainable communication adapter (self-attention + projection) rescales and interprets the received states.

Training Objective. The actor is trained with three losses: - Ltask: standard cross-entropy on ground-truth responses. - Lsep: minimizes weighted Jensen–Shannon divergence between distributions conditioned on matched vs. mismatched (cross-task) latents, forcing the actor to attend to task-relevant structure rather than ignore the latent input. - Lalign: KL + cosine regularization toward the language-space CoT plan distribution, preventing the actor from exploiting Lsep via idiosyncratic tokens. - A token-to-latent curriculum stochastically replaces early latent positions with their corresponding token embeddings, interpolating between language and latent supervision during training to stabilize convergence.

Compression via Latent-Space Reasoning. A separate reasoning model Mϕ is trained (actor frozen) to autoregressively generate compact latent messages HK ∈ R^{K×d} (K ≪ L) entirely in latent space by feeding its own last hidden state back as the next input embedding through a projection layer—no decoding to tokens. The compression objective combines task cross-entropy (Ltask), an uncertainty-weighted agreement loss Lpref that emphasizes positions where full latents reduce actor uncertainty, and a geometric alignment loss Lgeom enforcing cosine alignment of step-averaged adapter-processed representations.

Key Contributions¶

Interlat: the first framework to realize inter-agent communication entirely in continuous latent space without any language decoding between agents.
A principled three-term training objective (task + separation + alignment) with a stochastic curriculum to stabilize latent interpretation from scratch.
A learnable compression model that performs autoregressive reasoning in latent space, compressing full-length messages (100–200+ tokens) down to 8 steps while retaining task utility.
Demonstration of cross-family heterogeneous latent communication (Qwen2.5-7B latents consumed by a LLaMA-3.1-8B actor) with performance gains.
Extension to 3-agent topologies (sequential chain and parallel tree), confirming scalability beyond pairwise communication.

Results¶

ALFWorld (Qwen2.5-7B-Base): Interlat 70.48% seen / 65.42% unseen vs. CoT (full) 67.14% / 64.93% and No-Comm 62.14% / 62.19%.
ALFWorld (LLaMA3.1-8B-Base): Interlat 70.71% seen / 70.90% unseen vs. CoT 69.35% / 70.82% and No-Comm 63.57% / 58.40%.
MATH Level-5: Interlat 15.80% vs. CoT 15.05% and No-CoT 14.80% — Interlat is the only method that beats CoT on the hardest difficulty tier.
Compression (trained, 8-step latents): 66.43% seen / 60.45% unseen at 0.20 s end-to-end vs. 9.19 s for full-length latents — ~24× speedup with <5 pp degradation.
Cross-family: Qwen→LLaMA latent training yields 70.95% / 71.39% on seen/unseen, the strongest result in Table 1.
3-agent tree-3 topology: 62.85% seen / 60.77% unseen vs. text-tree-3 61.19% / 56.75%.

Limitations¶

Experiments limited to two benchmarks (ALFWorld, MATH) and primarily a two-agent sender-receiver setup; larger teams, dynamic roles, long-horizon collaboration, and tool use are not tested.
Requires access to internal model hidden states; incompatible with closed-source/API-only models.
Latent messages are not human-readable, complicating debugging and failure analysis; interpretability tools are rudimentary (PCA visualization only).
The compression–generalization trade-off (removing Ltask improves unseen performance) is unresolved and left for future work.
No safety evaluation of latent-channel communication, which could theoretically bypass token-level safety filters.

Relevance to Agentic AI / LLM Agents¶

Interlat directly challenges the assumption that natural language is the necessary or optimal communication substrate for LLM-based multi-agent systems, showing that latent-space messaging can both improve task performance and drastically reduce communication overhead. For researchers building multi-agent pipelines (orchestrators, debate frameworks, tool-using agent teams), it raises the practical question of whether agent-to-agent channels should bypass the language head entirely when open-weight models are available. The work also intersects with emerging latent-reasoning research (continuous CoT, recurrent depth), connecting single-model internal compute scaling to inter-agent coordination bandwidth. The cross-family transfer result is especially significant: it suggests latent communication can generalize across heterogeneous model families, making this architecture-agnostic in principle.