Research map / Architectures and efficiency / Quantization
Compute Where it Counts: Self Optimizing Language Models
Yash Akhauri, Mohamed S. Abdelfattah
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 |
|---|---|---|---|
| learn per-token compute allocation with a GRPO-trained policy instead of a fixed static schedule | SOL-J-FL-T16 on frozen LLM · WikiText-style perplexity at matched net keep-rate (Llama-3.2-1B/3B, Llama-3.1-8B-Instruct) | perplexity | Improved |
| grow the action space from 8 to 27 to 1560 joint actions (ablation) | SOL-J policy · perplexity vs net keep-rate | perplexity | Improved |
| lengthen control horizon T from 4 to 16 to 64 (KV-refresh period) (ablation) | SOL-J-FL · perplexity vs net keep-rate | perplexity | Improved |
| use SOL controller versus fixed schedule on downstream tasks at matched budget | Llama-3.2-3B · ARC-Easy, PIQA, WinoGrande average | average accuracy | Improved |
| use SOL controller versus fixed schedule on downstream tasks at matched budget | Llama-3.1-8B-Instruct · MMLU (3 subjects, continuation format) | average accuracy: up to 7.3% | Improved |
| use SOL token-sparsity policy versus fixed schedule | DeepSeek-R1-Distill-Llama-8B · GSM8K-CoT-Llama 8-shot | accuracy | Improved |
| request very aggressive prune keep-rate (rho below about 0.5) | SOL budget adherence · requested vs realized budget | budget adherence | Hurt |
Closest papers by meaning
- Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for LLM Problem-Solving (ICLR 2025 · similarity 0.91)
- TokenPowerBench: Benchmarking the Power Consumption of LLM Inference (AAAI 2026 · similarity 0.91)
- Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.90)
- Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads (ICML 2024 · similarity 0.90)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.90)
- xLSTM 7B: A Recurrent LLM for Fast and Efficient Inference (ICML 2025 · similarity 0.90)
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.
