Research map / Reinforcement learning / Visual and representation RL
An Imitation Learning Approach for Cache Replacement
Evan Liu, Milad Hashemi, Kevin Swersky, Parthasarathy Ranganathan, Junwhan Ahn
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use end-to-end imitation learning from Belady's instead of LRU | Parrot · SPEC2006 and Web Search | cache hit rate: 16% higher cache hit rates than LRU averaged over all programs; 13.5% higher raw cache hit rate than LRU on Web Search | Improved |
| Use end-to-end imitation learning from Belady's instead of Glider | Parrot · SPEC2006 (13 programs) | normalized cache hit rate: 20% higher normalized cache hit rate averaged over all programs | Mixed |
| Use Parrot instead of LRU | Parrot · Web Search | normalized cache hit rate: 61% higher normalized cache hit rate | Improved |
| Use the byte embedder instead of separate address and PC embeddings (ablation) | Parrot · SPEC2006 (13 programs) | normalized cache hit rate: full-sized model performs better; byte model achieves an average of 8% higher normalized cache hit rate than Glider | Hurt |
| Add reuse-distance prediction as an auxiliary task (ablation) | Parrot · SPEC2006 (4 programs) and Web Search | normalized cache hit rate: 16.8% higher normalized cache hit rates | Improved |
| Use reuse-distance prediction to directly choose the line with the highest predicted distance rather than as an auxiliary loss (ablation) | Parrot · SPEC2006 (4 programs) and Web Search | cache hit rate | Hurt |
| Train on-policy with DAgger rather than on oracle-policy states (ablation) | Parrot · SPEC2006 (4 programs) and Web Search | normalized cache hit rate: 9.8% higher normalized cache hit rate overall | Mixed |
| Use ranking loss instead of log-likelihood loss (ablation) | Parrot · SPEC2006 (4 programs) and Web Search | normalized cache hit rate: 3.5% normalized cache hit rate improvement | Improved |
| Increase the attended history length up to 80 accesses (ablation) | Parrot · SPEC2006 | normalized cache hit rate: 80 | Improved |
| Increase history length beyond 80 accesses (ablation) | Parrot · SPEC2006 | normalized cache hit rate: 80 | No clear effect |
Closest papers by meaning
- Robustifying Learning-Augmented Caching Efficiently without Compromising 1-Consistency (NeurIPS 2025 · similarity 0.85)
- DropKV: Decoupling Residual-Output Perturbation for Near-Optimal KV-Cache Eviction (NeurIPS 2026 · similarity 0.85)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.84)
- MiniPIC: Flexible Position-Independent Caching in <100LOC (NeurIPS 2026 · similarity 0.83)
- Value-Aware Stochastic KV Cache Eviction for Reasoning Models (NeurIPS 2026 · similarity 0.83)
- Tvcache: A Tool-Value Cache for Post-Training LLM Agents (ICML 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
