Research map / Vision-language and multimodal / Multimodal LLMs

Unleashing Perception-Time Scaling to Multimodal Reasoning Models

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning

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.

Introduces the DisTANCE visual-estimation benchmark and Perception-Time Scaling (PTS), which uses symbolic distance tokens and step-wise reference-segment decomposition with SFT then GRPO to raise Qwen2.5-VL high-precision estimation accuracy.

Problem

LVLMs give imprecise visual estimates of geometric quantities (length, perimeter, area), and inference-time scaling via RL on reasoning gives only marginal perception gains because models use a one-shot "fast perception" output.

Approach

Reformulate perception as an explicit, token-rich process: encode distances as symbolic tick tokens (<==========>, 0.1 unit per "=") and decompose estimation into accumulating a reference segment, then train with SFT cold start and GRPO using an exponential relative-error reward so RL can refine intermediate perception steps.

Key results

• PTS with SFT+GRPO raises Qwen2.5-VL-7B DisTANCE RA_avg from 21.5% to 88.3% and high-precision RA_0.1 from 8.0% to 64.7%.
• On out-of-domain Geoperception LHC, Qwen2.5-VL-7B PTS reaches 78.7% vs 59.8% base.
• Geometry3K plus PTS data improves Qwen2.5-VL-7B on CV-Bench (73.36 to 75.68) and BLINK, while Geometry3K alone yields smaller or negative gains on perception tasks.

Surprising findings

• Inference-time scaling via reasoning-oriented RL yields only marginal perception gains over base LVLMs (e.g., Vision-R1-7B 22.7% vs Qwen2.5-VL-7B 21.5% RA_avg).
• Purely synthetic PTS geometry data combined with math data improves real-world perception benchmarks such as CV-Bench and BLINK.

Limitations

• Benchmark is small (300 image-question pairs) and synthetic geometric images.
• PTS relies on GPT-4o-generated cold-start reasoning chains and a hand-designed symbolic encoding; finer granularity gave slightly worse results.
• Perception-token ratio is measured by keyword filtering on 50 samples per model.

Experimental setup

Datasets: DisTANCE synthetic training images Geometry3K

Benchmarks: DisTANCE DisTANCE_ood Geoperception LHC LEGO-Puzzles Height MathVision MMBench MMVet HallusionBench CV-Bench BLINK

Metrics: RA_0.1 RA_avg accuracy

Baselines: GPT-4o Gemini-2.5-Flash Claude-3.5-Sonnet Qwen2.5-VL-7B InternVL3-78B Vision-R1-7B NoisyRollout-7B Spatial-R1-7B SpaceThinker-3B Visual Sketchpad DetToolChain

Builds on: GRPO RLVR Qwen2.5-VL EasyR1

Closest papers by meaning

• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.92)
• SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.91)
• VideoChat-R1.5: Visual Test-Time Scaling to Reinforce Multimodal Reasoning by Iterative Perception (NeurIPS 2025 · similarity 0.91)
• Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning (ICML 2026 · similarity 0.91)
• Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.91)
• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.90)