When two LLM agents exchange a message, something is lost. The question is whether that loss matters. A new paper from arXiv (arXiv:2607.14103) constructs the first systematic measurement framework for comparing text-based and latent communication channels between language model agents, and arrives at a carefully qualified negative result: the infrastructure for latent communication is sound, but text remains sufficient for every task type currently tested.
The motivating intuition is straightforward. When an LLM processes a rich, contextual input, it builds a high-dimensional internal representation. To communicate that to another agent, it must compress it into tokens. The receiving agent then reconstructs its own representation from those tokens. There is no guarantee the two representations are equivalent. The Platonic Representation Hypothesis adds a more optimistic framing: if diverse neural networks trained on sufficient data converge toward compatible representational geometries, then a direct latent channel bypassing text might preserve more of what the sender actually encoded. This paper tests that hypothesis rigorously.
Three Channels, One Framework
The authors construct three communication channels between a sender and receiver model (Llama 3.1 8B-Instruct and Mistral 7B):
- Dense latent: the full activation vector from the sender's residual stream, injected directly into the receiver
- SAE-sparse: a sparse code produced by a pre-trained Sparse Autoencoder, transferring a compressed representation
- Text: the standard approach, serialising the concept to natural language and re-encoding it
Channel fidelity is measured via linear probes trained on concept representations, cross-architecture retrieval accuracy after Procrustes alignment, and a zero-shot multiple-choice task protocol designed to reflect realistic inference conditions. The experimental structure is gated: two prerequisite checks must pass before the main claims are tested. Concepts must be linearly separable in activation space (confirmed at 100% probe accuracy), and SAE feature overlap must track semantic relatedness (confirmed, with a cosine similarity delta of +0.319 between related and unrelated pairs at layer 23).
Key Findings
The results are worth unpacking carefully because the headline numbers can mislead in either direction.
- The SAE-sparse channel achieves 99.4% probe accuracy at 28-fold compression over the dense-latent baseline. The text channel achieves 80.4%. On this metric, latent channels look substantially better.
- Text serialisation destroys 88% of SAE features. This is not attenuation: lost features have statistically indistinguishable activation magnitudes from surviving ones (2.36 vs 2.46). The representation moves to a different neighbourhood of the 131K-dimensional feature space entirely.
- Cross-architecture alignment works. Unaligned cosine similarity between Llama and Mistral concept vectors is -0.002, essentially random. After Procrustes alignment with 140 anchor concepts, top-1 retrieval reaches 92%, against a 0.87% chance baseline. CCA alignment in a 79-dimensional shared subspace achieves 100% top-1 retrieval at 80 anchors, confirming that cross-architecture semantic correspondence is concentrated in a low-rank subspace.
- Despite all of the above, the latent channel never exceeds the text channel on task-level concept identification. The 3-10 percentage point gap favours text on every tested task type, including cross-lingual variants designed to favour architecture-agnostic latent structure.
The resolution to this apparent contradiction lies in what the destroyed features actually encode. The augmentation experiment is decisive: when lost SAE features are re-injected into the receiver's representation, performance decreases. The features that text serialisation destroys carry surface form, specifically tokenisation artifacts, prompt-structural features, and positional encoding effects, not concept-discriminative semantics. The 88% feature loss is real but task-irrelevant for current benchmarks.
Methodological Strengths and Weaknesses
The paper's experimental design is careful in ways that matter. The distinction between attenuation and identity replacement in the feature survival analysis is not just semantic: it rules out a whole class of potential fixes (threshold tuning, magnitude rescaling) and points instead to a structural property of how bare concept names are processed versus concepts in context. The gated experimental structure also means the main claims rest on verified prerequisites rather than assumed ones.
The limitations are acknowledged honestly. Both models are 8B parameters; larger models may develop richer latent structures where contextual representations carry more task-relevant information, but pretrained SAEs for larger models are scarce. The concept dataset covers only 165 general-knowledge English concepts with 2-4 prompts each, which is thin for variance estimation and may not generalise to domain-specific vocabularies. Most critically, all task types are designed with text-expressible answers. This is precisely the condition under which text performs best, and the paper is transparent that the hypothesis cannot be properly tested until tasks exist that probe knowledge genuinely beyond textual expressibility.
The pilot experiments with Gemma 3 27B-IT reveal a practical obstacle that will matter for anyone attempting to deploy latent injection in real systems. Instruction-tuned models with structured prompt templates have attention sinks at BOS token positions where hidden-state norms reach 35 times the magnitude of regular positions by layer 20. Injecting sender vectors at these positions produces near-chance accuracy regardless of blending coefficient. Last-token injection at layer 20 achieves above-baseline performance, but only modestly. Template structure in instruction-tuned models couples semantic extraction tightly enough to format that tokens outside the expected structure are treated as positional noise.
Implications and Open Questions
The Platonic Representation Hypothesis receives meaningful empirical support here. The finding that base Mistral aligns slightly better than instruction-tuned Mistral (92% vs 88% top-1 at 140 anchors) is consistent with the hypothesis: alignment quality reflects pretraining convergence on shared data statistics, not instruction-tuning objectives. RLHF does not improve cross-architecture geometric compatibility, which is the right prediction if representational convergence is driven by exposure to the same statistical regularities in training data.
For the multi-agent systems community, the practical upshot is nuanced. Text communication is not being rescued from a serious information loss problem, at least not for current task types. The 88% feature destruction is real, but it destroys the wrong things. For researchers interested in latent communication as a future capability, the bottleneck is not alignment quality (CCA gets to 100% retrieval) or even the existence of shared geometric structure (confirmed). The bottleneck is identifying tasks where the contextual richness of a latent representation actually encodes something that text cannot express.
The authors flag cross-modal alignment with vision models as a natural next frontier, both as a practical application and as a stronger test of representational convergence. That is a reasonable direction. A more immediate challenge is constructing evaluation tasks where the answer genuinely depends on information that survives in latent space but not in text. Without such tasks, the hypothesis remains unrefuted rather than confirmed.
This is a well-executed study that resists the temptation to oversell its positive results. The convergence finding is real and the measurement framework is sound. What it cannot yet show is when latent communication would actually help. Full paper: arXiv:2607.14103.