Structured Reasoning Distillation with Policy Optimization and Test-Time Scaling for Resource-Constrained Visual Question Answering
Electronics · 2026-11-13 · 15 (22) · 5291
DOI: 10.3390/electronics15225291 ↗At a glance
Journal article in Electronics, published online on 2026-11-13. Written by Rayyan Najam and 2 co-authors. The full text is freely available (open access, CC BY 4.0). Funded by: Deanship of Scientific Research (DSR), King Abdulaziz University.
Abstract (original language)
State-of-the-art vision–language models (VLMs) show emerging reasoning but demand prohibitive computation, limiting deployment in constrained environments. This research addresses this via a dual-phase framework: trace distillation for structured reasoning priors, reinforcement learning for policy alignment, and test-time scaling. On ScienceQA-IMG, it decouples visual context from textual reasoning via dense visual metadata from Gemini 3.7 Flash (long context) or Qwen2.5-VL-7B-Instruct (short context). Qwen-3-8B undergoes LoRA supervised fine-tuning on only 5214 verified teacher reasoning traces (40.97% of the full training set), achieving 95.14% accuracy. GRPO with custom rewards yields stable convergence (policy loss: 2.49 × 10−6; KL: 1.25 × 10−3). Distillation generalizes across model sizes (LFM2.5-2.6B, Qwen3-4B, Gemma4-12B, Qwen3-14B, Qwen3.5-9B) and visual context variations (Long/Short) with consistent gains. Further, we propose a three-phase test-time scaling pipeline: (1) Gaussian Calibration-Weighted Voting (five rollouts), (2) Continuous Uncertainty Error Detection to flag high-risk samples, and (3) Sequential Remediation with a dynamic logit-calibrated critic followed by targeted N = 15 CWV with dampened weighting. This elevates the long-context model (DRLTSL-Qwen3-8B-Stage3-Phase3) to 95.79% IMG accuracy (seven-seed mean: 95.66 ± 0.12%), surpassing the fine-tuned LLM-based methods compared in the ScienceQA IMG category (accuracies as reported on the test split), while using 59.03% fewer training examples than the full ScienceQA training split and 16.15% fewer than the ScienceQA-IMG training set, conditioned on teacher-correct supervision. The short-context model (DRLTSS-Qwen3-8B-Stage3-Phase3) achieves 92.51%. It shows that careful design and efficient distillation/scaling can achieve performance competitive with the best reported IMG results with reduced compute and data, enabling consumer-grade GPU deployment of the text-only reasoning model with 32 GB VRAM or more, given visual descriptions from an upstream extractor.
Funding
- Deanship of Scientific Research (DSR), King Abdulaziz University
Cite
APA
Najam, R., Alsolami, F., & Saleh, M. (2026). Structured Reasoning Distillation with Policy Optimization and Test-Time Scaling for Resource-Constrained Visual Question Answering. Electronics, 15(22), 5291. https://doi.org/10.3390/electronics15225291
Vancouver
Najam R, Alsolami F, Saleh M. Structured Reasoning Distillation with Policy Optimization and Test-Time Scaling for Resource-Constrained Visual Question Answering. Electronics. 2026;15(22):5291. doi:10.3390/electronics15225291
BibTeX
@article{Najam2026,
author = {Najam, Rayyan and Alsolami, Fawaz and Saleh, Mostafa},
title = {Structured Reasoning Distillation with Policy Optimization and Test-Time Scaling for Resource-Constrained Visual Question Answering},
journal = {Electronics},
year = {2026},
volume = {15},
number = {22},
pages = {5291},
doi = {10.3390/electronics15225291}
}
EUVIMED