LatComm
A latent communication framework for multi-LLM collaboration that separates context encoding from question reading, enabling a context to be encoded once and reused multiple times, thereby improving inference efficiency and accuracy.
Timeline (last 90 days)
LatComm proposes a latent communication interface that, by separating encoding from reading, enables a context to be encoded once and reused multiple times.
On 12 multiple-choice question datasets, LatComm achieves an average accuracy up to 10.2 percentage points higher than the single-vector baseline.
LatComm reduces input by 57 times and speeds up by 3 times on the QuALITY dataset.
The LatComm system includes a frozen context encoder LLM and a frozen receiver LLM, which can be different models.
LatComm trains only the cross-attention reading module and does not fine-tune the encoder and receiver.
LatComm uses fixed-size data packets (a set of latent vectors) as the context representation.
LatComm's latent-space alignment may limit generalization across model families.
LatComm is currently evaluated mainly on multiple-choice questions, and its performance on open-ended generation tasks remains to be verified.
LatComm's code and models have not yet been released.
LatComm speeds up answering all questions about an article by 3 times on the QuALITY dataset.
All Facts (10)
LatComm proposes a latent communication interface that, by separating encoding from reading, enables a context to be encoded once and reused multiple times.
50%UnverifiedOn 12 multiple-choice question datasets, LatComm achieves an average accuracy up to 10.2 percentage points higher than the single-vector baseline.
50%UnverifiedLatComm reduces input by 57 times and speeds up by 3 times on the QuALITY dataset.
50%UnverifiedThe LatComm system includes a frozen context encoder LLM and a frozen receiver LLM, which can be different models.
50%UnverifiedLatComm trains only the cross-attention reading module and does not fine-tune the encoder and receiver.
50%UnverifiedLatComm uses fixed-size data packets (a set of latent vectors) as the context representation.
50%UnverifiedLatComm's latent-space alignment may limit generalization across model families.
50%UnverifiedLatComm is currently evaluated mainly on multiple-choice questions, and its performance on open-ended generation tasks remains to be verified.
50%UnverifiedLatComm's code and models have not yet been released.
50%UnverifiedLatComm speeds up answering all questions about an article by 3 times on the QuALITY dataset.
50%