List your company

Summarization on GovReport (ROUGE)

46ROUGE ScoreSMD
23.22429.13735.0540.963Feb 20, 2025May 4, 2025Jul 17, 2025Sep 28, 2025Dec 11, 2025Feb 22, 2026May 7, 2026

Evaluation results

MethodLinks
SMDModel=Qwen2.5-7B, Rollout Method=SMD, Toks. saving=39.4%2026.0546
GRPO DenseModel=Qwen2.5-7B, Rollout Method=Dense2026.0545.2
Sparse-RLModel=Qwen2.5-7B, Rollout Method=Sparse-RL, Toks. saving=39.4%2026.0544.1
GRPO DenseModel=Qwen3-4B-Instruct, Rollout Method=Dense2026.0542.5
Dense BaselineKV cache=100%2026.0542.5
SMDModel=Qwen3-4B-Instruct, Rollout Method=SMD, Toks. saving=50.0%2026.0542.4
SMD SnapKVKV cache=50% drop2026.0542.4
BaseModel=Qwen2.5-7B, Rollout Method=–2026.0541.5
SMDModel=Qwen2.5-3B, Rollout Method=SMD, Toks. saving=42.0%2026.0541
GRPO DenseModel=Qwen2.5-3B, Rollout Method=Dense2026.0540.2
Sparse-RLModel=Qwen3-4B-Instruct, Rollout Method=Sparse-RL, Toks. saving=50.0%2026.0539.8
Sparse-RLKV cache=50% drop2026.0539.8
Sparse-RLModel=Qwen2.5-3B, Rollout Method=Sparse-RL, Toks. saving=42.0%2026.0539.5
GRPO w/ SnapKVModel=Qwen2.5-7B, Rollout Method=w/ SnapKV2026.0538.6
BaseModel=Qwen3-4B-Instruct, Rollout Method=–2026.0538
BaseModel=Qwen2.5-3B, Rollout Method=–2026.0536.4
GRPO w/ SnapKVModel=Qwen3-4B-Instruct, Rollout Method=w/ SnapKV2026.0535.2
SMDModel=Qwen2.5-1.5B, Rollout Method=SMD, Toks. saving=43.3%2026.0534.8
GRPO w/ SnapKVModel=Qwen2.5-3B, Rollout Method=w/ SnapKV2026.0534.8
GRPO DenseModel=Qwen2.5-1.5B, Rollout Method=Dense2026.0534.2
Sparse-RLModel=Qwen2.5-1.5B, Rollout Method=Sparse-RL, Toks. saving=43.3%2026.0533.5
LIFTTarget Model=Llama-3-8B-Instruct, Synthetic Task Generator=Qwen2.5-72B-Instruct2025.0230.9
GRPO w/ SnapKVModel=Qwen2.5-1.5B, Rollout Method=w/ SnapKV2026.0530
ICLEvaluation Mode=Truncated In-Context Learning, Target Model=Llama-3-8B-Instruct2025.0229.1
SMDModel=Llama-3.2-1B-Instruct, Rollout Method=SMD, Toks. saving=45.0%2026.0528.8
GRPO DenseModel=Llama-3.2-1B-Instruct, Rollout Method=Dense2026.0528.5
BaseModel=Qwen2.5-1.5B, Rollout Method=–2026.0528.1
Sparse-RLModel=Llama-3.2-1B-Instruct, Rollout Method=Sparse-RL, Toks. saving=45.0%2026.0526.5
BaseModel=Llama-3.2-1B-Instruct, Rollout Method=–2026.0524.2
GRPO w/ SnapKVModel=Llama-3.2-1B-Instruct, Rollout Method=w/ SnapKV2026.0524.1