Research map / Vision-language and multimodal / Multimodal LLMs
RMultiplex200K: Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications
Sijia Chen, Bin Song
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 |
|---|---|---|---|
| fine-tune on RMultiplex200K | Qwen-2-VL-72B · RMultiplex200K testbed | accuracy (%): 21.3 -> 36.8 | Improved |
| add TC-Navigator process reward verifier (best-of-100) | fine-tuned Qwen-2-VL-72B · RMultiplex200K testbed | accuracy (%): 36.8 -> 46.8 | Improved |
| add TC-Navigator verifier with self-consistency | fine-tuned Llama-3.2-90B · RMultiplex200K testbed | accuracy (%): 39.2 -> 46.1 | Improved |
| use process reward model (TC) instead of outcome reward model (ablation) | fine-tuned Qwen-2-VL-72B · RMultiplex200K testbed | accuracy (%): 44.5 -> 46.8 | Improved |
| use PRM instead of ORM on problems with more reasoning steps (ablation) | PPO-reinforced Llama-3.2-11B · RMultiplex200K CSP category | accuracy (%): 31.9 -> 36.2 | Improved |
| use PRM instead of ORM on problems with fewer reasoning steps (ablation) | PPO-reinforced Llama-3.2-11B · RMultiplex200K SC category | accuracy (%): 35.8 -> 37.9 | Improved |
| add RA1 to TC-Navigator verifier (ablation) | TC-Navigator PPO · RMultiplex200K testbed | accuracy (%): 36.3 -> 37 | Improved |
| add RA2 to TC-Navigator verifier (ablation) | TC-Navigator PPO · RMultiplex200K testbed | accuracy (%): 35.9 -> 37 | Improved |
| reinforce generator with step-by-step PPO using TC reward | Llama-3.2-11B · RMultiplex200K testbed | accuracy (%): 34 -> 37 | Improved |
| combine PPO-reinforced generator with TC+SC verification | Llama-3.2-11B · RMultiplex200K testbed | accuracy (%): 37 -> 43.3 | Improved |
| train verifier on ApPA labels instead of MATH-Shepherd labels (ablation) | TC verifier Qwen-2-VL-72B · RMultiplex200K testbed | accuracy (%): over 6% | Improved |
| use TC verifier trained on RMultiplex200K for external benchmark | Qwen-2-VL-72B · TeleQnA | accuracy (%): 66.3 -> 76.8 | Improved |
Closest papers by meaning
- Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.87)
- Visual CoT: Advancing Multi-Modal Language Models with a Comprehensive Dataset and Benchmark for Chain-of-Thought Reasoning (NeurIPS 2024 · similarity 0.87)
- MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models (AAAI 2026 · similarity 0.86)
- MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.86)
- Point-RFT: Improving Multimodal Reasoning with Visually Grounded Reinforcement Finetuning (NeurIPS 2025 · similarity 0.86)
- VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models (ICLR 2026 · similarity 0.86)
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.
