Last released Jun 25, 2026
Inference-time value steering for vLLM: dynamic abstention + value-filtered decoding.
Supported by