LZ Penalty: An information-theoretic repetition penalty for autoregressive language models.

TMLR Paper5520 Authors

Abstract: We introduce the Lempel-Ziv (LZ) penalty, a penalty specialized for reducing degenerate repetitions in autoregressive language models without loss of capability. The penalty is based on the codelengths in the LZ77 universal lossless compression algorithm. Through the lens of the prediction-compression duality, decoding with the LZ penalty has the interpretation of sampling from the residual distribution after removing the information that is highly compressible. We demonstrate the LZ penalty enables open-source reasoning models to operate with greedy decoding without loss of capability and without instances of degenerate repetition. In contrast, both the industry-standard frequency penalty and repetition penalty are ineffective, incurring degenerate repetition rates of up to 4% or more.

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

(1) Fix PDF rendering issue!
(2) Add Limitations subsection
(3) Add conclusion
(4) Add additional reference as suggested by reviewer.

Various other line-by-line improvements and clarifications suggested by reviewer.

Assigned Action Editor: ~Tim_Genewein1

Submission Number: 5520