Research map / Vision-language and multimodal / Multimodal LLMs

ProxyThinker: Test-Time Guidance through Small Visual Reasoners

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 ProxyThinker, a training-free decoding method that adds the logit difference between a small RFT visual reasoner and its base model to a large base VLM, transferring slow-thinking reasoning and speeding inference via vLLM multi-model scheduling.

Problem

RFT of large vision-language models with RLVR is computationally expensive, making it hard to scale visual reasoning to larger model sizes.

Approach

Steer a large base LVLM at decoding time by adding alpha times the logit difference between a small RFT expert and a small amateur base model (same family), so the large model inherits slow-thinking reasoning behaviors without any training; implemented with multi-model vLLM scheduling for speed.

Key results

• On MathVision, Qwen2.5-VL-32B improves from 38.4% to 40.8% using OpenVLThinker-7B as expert, surpassing VL-Rethinker-32B (40.5%).
• ProxyThinker-72B with VL-Rethinker-7B reaches 54.4% on R1-Bench vs. 50.4% base and 57.9% for VL-Rethinker-72B; average gain +2.7% over base at 72B.
• Implementation with vLLM and optimized tensor parallelism gives about 38x speedup over the Huggingface implementation (19133s to 501s) on MathVision testmini.

Surprising findings

• A 3B RFT expert (VLAA-Thinker-3B) degrades MathVista from 61.0 to 51.5 when used as expert, showing negative transfer from poorly aligned small experts.
• ProxyThinker can exceed the full-scale RFT 32B reasoner on MathVision despite the 7B expert scoring only 25.3% on that benchmark.

Limitations

• Requires an RFT expert and amateur sharing the same vocabulary and ideally architecture, limiting applicability.
• Experiments focus on visual reasoning benchmarks; natural-language-only or embodied tasks are untested.
• Underperforming or undersized experts can cause negative transfer; no statistically significant gains on knowledge-intensive MMMU.

Experimental setup

Datasets: MathVerse MathVista MathVision MMMU-Pro R1-Onevision-Bench EMMA

Benchmarks: MathVista MathVerse MathVision MMMU-Pro R1-Onevision-Bench EMMA MMMU

Metrics: Pass@1 accuracy pass@k runtime (s)

Baselines: Qwen2.5-VL-32B-Instruct Qwen2.5-VL-72B-Instruct VL-Rethinker-32B VL-Rethinker-72B OpenVLThinker-7B ThinkLite-VL-7B VL-Rethinker-7B contrastive decoding VLAA-Thinker-3B

Builds on: Contrastive Decoding DExperts ProxyTuning GRPO vLLM

Closest papers by meaning

• VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
• Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.89)
• OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025 · similarity 0.89)
• SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.89)
• OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.89)