Research map / Architectures and efficiency / Attention and transformers
Bottlenecked Transformers: Periodic KV Cache Consolidation for Generalised Reasoning
Adnan Oomerjee, Zafeirios Fountas, Haitham Bou Ammar, Jun Wang
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add Cache Processor KV rewrites to SFT backbone | Llama 3.2 1B · SVAMP | accuracy: 38 -> 44.6 | Improved |
| add Cache Processor KV rewrites to SFT backbone | Llama 3.2 3B · GSM8K | accuracy: 46.78 -> 51.33 | Improved |
| add Cache Processor KV rewrites to SFT backbone | Qwen 3 0.6B · MATH | accuracy: 26.68 -> 29.08 | Improved |
| add Cache Processor KV rewrites to SFT backbone | Llama 3.1 8B · LogiQA | accuracy: 20.74 -> 23.81 | Improved |
| add Cache Processor KV rewrites to SFT backbone | Llama 3.2 3B · LogiQA | accuracy: 22.12 -> 20.12 | Hurt |
| add Cache Processor KV rewrites to SFT backbone | Qwen 3 0.6B and Llama 3.1 8B · Gaokao-MathQA | accuracy | Hurt |
| add 16 pause tokens during SFT only (ablation) | SFT backbone · seven math benchmarks | accuracy: GSM8K Llama 3.1 8B 70.28 -> 69.22 | No clear effect |
| add 16-token latent rollout during SFT (ablation) | SFT backbone · seven math benchmarks | accuracy: GSM8K Llama 3.1 8B 70.28 -> 4.02 | Hurt |
| match total epochs: Cache Processor epoch versus extra SFT epoch (ablation) | Llama 3.2 1B · GSM8K, GSM-Hard, SVAMP, LogiQA | accuracy | Improved |
| match total epochs: Cache Processor epoch versus extra SFT epoch (ablation) | Llama 3.2 1B · MATH, TheoremQA | accuracy | Hurt |
| vary reconsolidation budget k from 16 to 256 (ablation) | Bottlenecked Transformer Llama 3.2 1B · seven benchmarks | accuracy: MATH 13.12 (k=16) -> 13.34 (k=256) | Mixed |
| vary recent-step window R from 16 to 96 (ablation) | Bottlenecked Transformer Llama 3.2 1B · seven benchmarks | accuracy | No clear effect |
Closest papers by meaning
- R-KV: Redundancy-aware KV Cache Compression for Reasoning Models (NeurIPS 2025 · similarity 0.90)
- Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains (NeurIPS 2025 · similarity 0.90)
- KaVa: Latent Reasoning via Compressed KV-Cache Distillation (ICLR 2026 · similarity 0.90)
- Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces (ICLR 2025 · similarity 0.89)
- Deliberation in Latent Space via Differentiable Cache Augmentation (ICML 2025 · similarity 0.89)
- Beyond Speedup - Utilizing KV Cache for Sampling and Reasoning (ICLR 2026 · similarity 0.89)
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.
