LLMs · via HF Papers
TTR-Compressor rethinks sequence compression as test-time regression
The paper compresses token sequences by fitting a small regression at inference time and using its coefficients as the summary, aiming to beat pooling and learned projections in efficient language and multimodal models.