The authors systematically evaluated safety incompatibility in cross-model KV cache reuse across model pairs (base/instruction-tuned from Llama-3 and Qwen3 families, and LoRA variants) and reuse configurations (full, K-only, V-only).
Cross-model KV cache reuse introduces safety incompatibility in language models
The authors show that reusing value tensors causes larger safety deviations than reusing keys, and propose VReparo to mitigate the issue.
Chinese Tech
Zhiwei Lin · Jiahao Liu · Yan Liu · Chengpeng Wang
National University of Singapore · Ant Group
Research Digest··2 min read
The authors systematically evaluate safety incompatibility in cross-model KV cache reuse across model pairs and reuse configurations.
Why this paper
From Ant Group and National University of Singapore · Released code
In one line
Cross-model KV cache reuse can substantially alter LLM safety, especially through reused values, while affine value alignment reduces harmful behavior without losing most latency gains.
What it released
Code
What we could check
- ✓Code link in the paper (github.com)
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§