Research map / Large language models / Agents, code and math
LightWM: Training-Free Hierarchical Working Memory for Small Language Model Agents
Ziyi Wang, Haonan Jin, Zian Wang, Wendong Wang, Lanshan Zhang
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 subtask-specific action templates (ablation) | LightWM · ALFWorld valid unseen, Qwen3-4B (FP16) | success rate: 0.821 -> 0.910 | Improved |
| add rule-constrained patch updates (ablation) | LightWM · ALFWorld valid unseen, Qwen3-4B (FP16) | success rate: 0.766 -> 0.910 | Improved |
| add action templates and patch constraints on top of HWM only (ablation) | LightWM · ALFWorld valid unseen, Qwen3-4B (FP16) | success rate: 0.697 -> 0.910 | Improved |
| add hierarchical goal-conditioned working memory (HWM only vs MemGPT) (ablation) | Qwen3-4B (FP16) · ALFWorld valid unseen, Qwen3-4B (FP16) | success rate: 0.317 -> 0.697 | Improved |
| add rule-constrained patch updates (output tokens per call) (ablation) | LightWM · ALFWorld valid unseen, Qwen3-4B (FP16) | output tokens per call: 139 -> 32 output tokens/call | Improved |
| use LightWM instead of ReAct | Qwen3-4B (FP16) · ALFWorld valid unseen | success rate: 0.062 -> 0.910 | Improved |
| use LightWM instead of ReAct | Gemini-2.5 Flash · ALFWorld valid unseen | success rate: 0.786 -> 0.945 | Improved |
| use LightWM instead of ReAct | Qwen3-8B (FP16) · ALFWorld valid unseen | success rate: 0.579 -> 0.828 | Improved |
| use Simply Summarize instead of ReAct | Qwen3-8B (FP16) · ALFWorld valid unseen | success rate: 0.579 -> 0.469 | Hurt |
| use summarization-based memory (Simply Summarize) instead of ReAct | Gemini-2.5 Flash · ALFWorld valid unseen | success rate: 0.786 -> 0.531 | Hurt |
| use LightWM instead of MemGPT | Qwen3-4B (FP16) · ALFWorld valid unseen | success rate: 0.317 -> 0.910 | Improved |
| use FP16 instead of FP8 inference | LightWM · ALFWorld valid unseen, Qwen3-4B | success rate: 0.821 -> 0.910 | Improved |
Closest papers by meaning
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.89)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.89)
- Minerva: A Programmable Memory Test Benchmark for Language Models (ICML 2025 · similarity 0.89)
- MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.89)
- A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.88)
- PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents (ICML 2026 · similarity 0.88)
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.
