ACE-Compression
A capability-centered evaluation framework for large language model compression. It decomposes the evaluation of compressed models from task accuracy into four deployment-relevant capability dimensions: procedural reasoning, knowledge grounding, social cognition, and persona fidelity. The framework is used to reveal the non-uniform effects of compression on a model's capability structure.
Timeline (last 90 days)
ACE-Compression is a capability-centric model compression evaluation framework aimed at assessing which core capabilities of a compressed model are retained or degraded, rather than focusing only on task accuracy.
ACE-Compression decomposes evaluation into four dimensions directly relevant to real deployment scenarios: procedural reasoning and constraint execution, knowledge grounding and truthfulness, social cognition and emotional understanding, and persona fidelity and stylistic consistency.
The research team evaluated over 35 compressed checkpoints across six open-weight models, covering the two mainstream method categories of quantization and pruning.
All Facts (3)
ACE-Compression is a capability-centric model compression evaluation framework aimed at assessing which core capabilities of a compressed model are retained or degraded, rather than focusing only on task accuracy.
50%UnverifiedACE-Compression decomposes evaluation into four dimensions directly relevant to real deployment scenarios: procedural reasoning and constraint execution, knowledge grounding and truthfulness, social cognition and emotional understanding, and persona fidelity and stylistic consistency.
50%UnverifiedThe research team evaluated over 35 compressed checkpoints across six open-weight models, covering the two mainstream method categories of quantization and pruning.
50%