Research map / Large language models / Reasoning and chain of thought
BOLT: Decision‑Aligned Distillation and Budget-Aware Routing for Constrained Multimodal QA on Robots
Tengjun Ni, Xin Yuan, Shenghong Li, Kai Wu, Ren Ping Liu, Wei Ni, Wenjie Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BOLT, a decision-aligned distillation and budget-aware routing framework for constrained multimodal question answering on robots.
Problem
Robotic systems need multimodal reasoning under tight latency, memory, and energy constraints, where standard instruction tuning and token-level distillation fail to deliver decision quality, reliability, and interpretability.
Approach
Treats multiple-choice prediction as a decision surface that is aligned during distillation (decision-aligned distillation), and adds budget-aware routing to meet resource constraints, instead of matching teacher tokens alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| option-level decision distillation instead of token-level KD (ablation) | Qwen2-VL-2B (teacher LLaVA-1.5-13B) · Robo2VLM-1 | accuracy: 37.58 -> 42.89 | Improved |
| option-level decision distillation instead of token-level KD (ablation) | Qwen2-VL-2B (teacher LLaVA-1.5-7B) · Robo2VLM-1 | accuracy: 33.91 -> 38.53 | Improved |
| option-level decision distillation instead of token-level KD (ablation) | PaliGemma2-3B (teacher LLaVA-1.5-13B) · Robo2VLM-1 | accuracy: 35.50 -> 39.64 | Improved |
| distill from LLaVA-1.5-13B with ODD | Qwen2-VL-2B · Robo2VLM-1 | accuracy: 28.66 -> 42.89 | Improved |
| add budget-aware test-time augmentation (bTTA) | ODD student · Robo2VLM-1 | accuracy: 42.89 -> 50.50 | Improved |
| add high-resolution re-evaluation only (ablation) | ODD student · Robo2VLM-1 | accuracy: 42.89 -> 46.64 | Improved |
| add type-matched retrieval exemplars only (ablation) | ODD student · Robo2VLM-1 | accuracy: 42.89 -> 44.31 | Improved |
| add question decomposition only (ablation) | ODD student · Robo2VLM-1 | accuracy: 42.89 -> 45.47 | Improved |
| combine HR, tmRAG and QD (ablation) | ODD student · Robo2VLM-1 | accuracy: 46.64 -> 50.50 | Improved |
| replace zero-shot teacher by distilled student with bTTA (deployment) | BOLT · GPU memory | MB: 26,878 -> 3,035 | Improved |
| use BOLT instead of 13B teacher | BOLT · latency per question | seconds: 22.30 -> 8.97 | Improved |
| use retrieved exemplars (tmRAG) | BOLT · Robo2VLM-1 contradiction rate | RCR: RCR 21.73% | Hurt |
Experimental setup
Baselines: standard instruction tuning token-level distillation
Builds on: knowledge distillation
Closest papers by meaning
- DeeR-VLA: Dynamic Inference of Multimodal Large Language Models for Efficient Robot Execution (NeurIPS 2024 · similarity 0.86)
- ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping (ICLR 2026 · similarity 0.85)
- MARS: Multi-Agent Adaptive Reasoning with Socratic Guidance for Automated Prompt Optimization (AAAI 2026 · similarity 0.85)
- A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning (ICLR 2026 · similarity 0.85)
- Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (AAAI 2026 · similarity 0.85)
- Consilience for Verifier-Free Test-Time Scaling (NeurIPS 2026 · similarity 0.85)
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.
