Two agents passed reasoning state instead of text, and the receiver cut generated tokens by up to 74.8 percent
AspireAI community · arXiv (Wang, Wang, Seneviratne, Chen, Xia, Wang & Halgamuge)
Latent communication between LLM agents has two halves that usually get separated: a way to keep reasoning continuous, and an interface that lets another model use that intermediate computation. J-Comm attacks both. GEAR, a training-free Geometric Agreement for Latent Reasoning, reweights token candidates over the model's existing embeddings so the geometric consensus between candidates improves monotonically at each step. The H+J interface then gives the receiver two views - continuous inputs derived from feedback distributions, plus intermediate-state cues in vocabulary-indexed Jacobian-lens coordinates that the receiver reconstructs with its own dictionary and injects during prefill.
No key-value caches are shared and the two models need not have identical widths. In Qwen3-8B to Qwen3-4B communication, J-Comm scores highest on mean across seven benchmarks: 94.80 percent on MATH500, 54.55 percent on GPQA-Diamond, 73.33 percent on AIME2024+2025 and 62.93 percent success on ALFWorld. Against TextMAS it improves performance by 1.29 to 8.59 percentage points and reduces generated text tokens by 19.2 to 74.8 percent.
The honest caveat is that the headline benchmark deltas are single-digit points, and the token saving is where the larger effect sits.
https://arxiv.org/abs/2610.32871
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.