Research map / Large language models / Instruction and fine-tuning
TokMem: Tokenized Procedural Memory for Large Language Models
Zijun Wu, Yongchang Hao, Lili Mou
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 |
|---|---|---|---|
| use TokMem instead of LoRA fine-tuning | Qwen 2.5 0.5B · SNI 1000 tasks | ROUGE-L: 43.2 -> 50.0 | Improved |
| use TokMem instead of LoRA fine-tuning | Llama 3.1 8B · SNI 1000 tasks | ROUGE-L: 61.6 -> 64.8 | Improved |
| use TokMem routing instead of Sentence-BERT RAG retrieval | Qwen 2.5 0.5B · SNI 1000 tasks routing | accuracy: 79.7 -> 94.7 | Improved |
| add replay memory to fine-tuning | LoRA fine-tuning · Qwen 2.5 0.5B SNI 1000 tasks | ROUGE-L: 43.2 -> 46.7 | Improved |
| decouple memory token into address and steering embeddings (ablation) | TokMem · SNI average, Qwen 0.5B vs Llama 3B | ROUGE-L: 50.7 -> 51.1 (Qwen); 62.9 -> 62.6 (Llama 3B) | No clear effect |
| add renormalization of new memory embeddings (ablation) | TokMem · continual tool learning | retention of old tools | Improved |
| add brief adaptation of backbone on auxiliary tools (ablation) | TokMem · APIGen tool selection avg, Llama 3.2 1B | F1: 86.2 -> 98.4 | Improved |
| add brief adaptation of backbone on auxiliary tools (ablation) | TokMem · APIGen tool selection avg, Llama 3.1 8B | F1: 82.8 -> 99.1 | Improved |
| use TokMem (w/ adapt) instead of LoRA fine-tuning | Llama 3.2 1B · APIGen tool selection avg | F1: 9.0 -> 98.4 | Improved |
| use TokMem (w/ adapt) instead of LoRA fine-tuning | Llama 3.2 3B · APIGen argument avg | F1: 85.8 -> 86.3 | No clear effect |
| use TokMem instead of RAG | Llama 3.2 3B · APIGen tool selection avg | F1: 72.8 -> 99.2 | Improved |
| infix memory placement instead of prefix (ablation) | TokMem-1 · Fanfics 1024 tokens | perplexity: 3.81 -> 3.28 | Improved |
Closest papers by meaning
- M$^\star$: Every Task Deserves Its Own Memory Harness (NeurIPS 2026 · similarity 0.88)
- Procedural Memory Distillation: Online Reflection for Self-Improving Language Models (NeurIPS 2026 · similarity 0.88)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.87)
- Minerva: A Programmable Memory Test Benchmark for Language Models (ICML 2025 · similarity 0.87)
- MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning (NeurIPS 2026 · similarity 0.87)
- MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs (NeurIPS 2025 · similarity 0.87)
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.
