Research map / Large language models / AI and society
F.A.C.U.L.: Language-Based Interaction with AI Companions in Gaming
Wenya Wei, Sipeng Yang, Qixian Zhou, Ruochen Liu, Xuelei Zhang, Yifu Yuan, Yan Jiang, Yongle Luo, Hailong Wang, Tianzhou Wang, Peipei Jin, Wangtong Liu and 3 more
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 |
|---|---|---|---|
| Train BERT with fuzzy intent detection (entropy maximization on unsupported intents) (ablation) | BERT -> BERT-FID · instruction reasoning | accuracy: 0.714 -> 0.749 | Improved |
| Fine-tune Qwen2-1.5B on domain-specific command-action data | Qwen2-1.5B · instruction reasoning | accuracy: 0.642 -> 0.908 (+26.6%) | Improved |
| Domain-aligned 1.5B LLM vs larger general Qwen2-7B | LLM* · instruction reasoning | accuracy: 0.807 -> 0.908 | Improved |
| Route low-confidence commands from BERT-FID to LLM* instead of using LLM* alone | BERT-FID & LLM* · instruction reasoning | accuracy: 0.908 -> 0.886 | Hurt |
| Route low-confidence commands from BERT-FID to LLM* instead of using LLM* alone | BERT-FID & LLM* · instruction reasoning | QPS: 245 -> 813 | Improved |
| Use BERT-FID instead of original BERT as the router with LLM* (ablation) | BERT & LLM* · instruction reasoning | accuracy: 0.831 -> 0.886 | Improved |
| Fine-tune CLIP on in-house asset labels (ablation) | CLIP · entity retrieval Top 5 | F1: 0.429 -> 0.790 | Improved |
| Combine text similarity with fine-tuned CLIP instead of text similarity alone (ablation) | entity retrieval · Top 5 | F1: 0.888 -> 0.908 | Improved |
| Combine text similarity with fine-tuned CLIP instead of text similarity alone (ablation) | entity retrieval · Top 20 | F1: 0.941 -> 0.960 | Improved |
| Full F.A.C.U.L. (BERT-FID routing, LLM*, CLIP) vs fast BERT-FID plus text similarity | F.A.C.U.L. · end-to-end recognition | accuracy: 73.9% -> 87.2% | Improved |
| Full F.A.C.U.L. vs slow LLM* plus fine-tuned CLIP | F.A.C.U.L. · end-to-end recognition | accuracy: 89.2% -> 87.2%, QPS 210 -> 916 | Hurt |
Closest papers by meaning
- CollabBench: Benchmarking and Unleashing Collaborative Ability of LLMs with Diverse Players via Proactive Engagement (ICML 2026 · similarity 0.85)
- TowerMind: A Tower Defence Game Learning Environment and Benchmark for LLM as Agents (AAAI 2026 · similarity 0.85)
- Multi-agent KTO: Enhancing Strategic Interactions of Large Language Model in Language Game (NeurIPS 2025 · similarity 0.84)
- MARSHAL: Incentivizing Multi-Agent Reasoning via Self-Play with Strategic LLMs (ICLR 2026 · similarity 0.84)
- BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games (ICLR 2025 · similarity 0.83)
- LLM-PySC2: Starcraft II learning environment for Large Language Models (NeurIPS 2025 · similarity 0.83)
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.
