RESEARCH · RESEARCH · #2476
Paper provides causal evidence of hidden continuous communication between diffusion-LM agents
arXiv:2610.28905v1 ('Hidden Messages') reports experiments where a sender compresses its denoising state into four dense vectors consumed by a receiver; ablating or substituting those vectors causally affects the receiver's outputs. The authors find large content and channel effects (e.g., +63.6 pts on Tiny‑A2D, +100.0 pts on Dream‑7B; matched vs wrong‑example messages on QASC: 80.8% vs 46.8%) and show a single late transfer can outperform multiple recurrent exchanges.
KEY POINTS
- arXiv:2610.28905v1 ('Hidden Messages') reports experiments where a sender compresses its denoising state into four dense vectors consumed by a receiver; ablating or substituting those vectors causally affects the receiver's outputs.
- The authors find large content and channel effects (e.g., +63.6 pts on Tiny‑A2D, +100.0 pts on Dream‑7B; matched vs wrong‑example messages on QASC: 80.8% vs 46.8%) and show a single late transfer can outperform multiple recurrent exchanges.
- This provides causal evidence that continuous denoising states can serve as compact hidden channels between diffusion-LM agents and that timing of transfer can be more important than number of exchanges.
WHY IT MATTERS
This provides causal evidence that continuous denoising states can serve as compact hidden channels between diffusion-LM agents and that timing of transfer can be more important than number of exchanges.