Research map / Large language models / Language, safety and interpretability
Language Model Networks: Supervision-Efficient Learning through Dense Communication
Shiguang Wu, Yaqing Wang, Quanming Yao
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 LMNet dense communication instead of prompting the same model | Qwen2.5-0.5B · General benchmarks average | relative improvement: +30.5% | Improved |
| SFT on the same training data as LMNet (ablation) | Qwen2.5-0.5B · General benchmarks average | relative improvement: -5.7% | Hurt |
| add Self-Consistency with 16 samples (ablation) | Qwen2.5-0.5B · General benchmarks average | relative improvement: +3.4% | No clear effect |
| add Self-Refine with 16 iterations (ablation) | Qwen2.5-0.5B · General benchmarks average | relative improvement: +0.5% | No clear effect |
| LMNet instead of prompting | Qwen2.5-0.5B · MATH | accuracy: 19.5 -> 38.8 | Improved |
| LMNet instead of prompting | Qwen2.5-0.5B · GPQA | accuracy: 24.8 -> 25.6 | No clear effect |
| replace SFT with LMNet | Qwen2.5-1.5B · MMLU limited data | Delta Acc: 35.02 -> 37.28 | Improved |
| use LMNet instead of latent reasoning COCONUT | Qwen2.5-1.5B · GSM8K limited data | accuracy: 48.63 -> 72.67 | Improved |
| COCONUT latent reasoning instead of plain prompting | Qwen2.5-0.5B · MMLU limited data | Delta Acc: 22.50 -> 19.22 | Hurt |
| increase vertex model size to 3B and 7B (ablation) | LMNet · MMLU and GSM8K | accuracy: MMLU 7B: 74.2 -> 76.0 vs Prompt | Improved |
| vary LMNet depth/width (ablation) | LMNet GPT2-M · E2E | BLEU: 1/1: 69.1, 1/2/1: 70.5, 1/4/4/1: 68.7 | No clear effect |
| LMNet edge-only tuning instead of LoRA | GPT2-M · E2E | BLEU: 68.9 -> 70.5 | Improved |
Closest papers by meaning
- Offline Training of Language Model Agents with Functions as Learnable Weights (ICML 2024 · similarity 0.88)
- Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch (ICML 2024 · similarity 0.88)
- Communicating Activations Between Language Model Agents (ICML 2025 · similarity 0.87)
- No Need to Talk: Asynchronous Mixture of Language Models (ICLR 2025 · similarity 0.87)
- Skip a Layer or Loop It? Learning Program-of-Layers in LLMs (ICML 2026 · similarity 0.87)
- KVComm: Enabling Efficient LLM Communication through Selective KV Sharing (ICLR 2026 · 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.
