Research map / Vision-language and multimodal / Multimodal LLMs

Thyme: Think Beyond Images

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal 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.

Thyme trains Qwen2.5-VL-7B to write and execute Python code for image manipulation and computation via SFT on ~500K samples followed by RL with a novel GRPO-ATS that uses different sampling temperatures for text and code.

Problem

Existing "thinking with images" MLLMs are limited to cropping or generated auxiliary images, lacking the broad image operations and code-based computation offered by proprietary models like OpenAI o3.

Approach

Let the MLLM autonomously generate executable Python code in a sandbox for diverse image operations (crop, rotate, contrast, computation), trained with SFT plus RL using GRPO-ATS, which samples at temperature 1.0 for text and 0.0 for code.

Key results

• Thyme-VL-7B reaches 77.0 overall on HRBench-4K vs 68.8 for Qwen2.5-VL-7B (+8.2) and 72.0 on HRBench-8K (+6.7).
• On MME-RealWorld-Lite, Thyme-VL-7B improves overall accuracy from 44.1 to 55.2 (+11.1).
• On LogicVista, Thyme-VL-7B reaches 49.0 vs 39.8 for the Qwen2.5-VL-7B baseline (+9.2), though Qwen2.5-VL-32B scores 54.4.

Surprising findings

• Naive SFT on the full mixed data gave little gain and produced chaotic output formats, with the model predicting sandbox outputs itself; masking sandbox content was necessary.
• Adding a process reward (LLM-scored reasoning quality) or a code-execution reward reduced average performance relative to the consistency reward variant.

Limitations

• Performance is bounded by the 7B base model's localization and code generation ability, producing incorrect crops or unexecutable code.
• Existing benchmarks underrepresent rotation and low-contrast cases, so these capabilities are hard to evaluate numerically.
• Model sometimes writes unnecessary code for simple questions, and some bad cases stem from answering without cropping on high-resolution images.

Experimental setup

Datasets: LLaVA-OV-Image MM-RLHF SMR V* MM-Eureka arXivQA Retool ThinkLite-VL

Benchmarks: HRBench-4K HRBench-8K MME-RealWorld MME-RealWorld-CN V* RealWorld QA MathVision MathVista MathVerse LogicVista WeMath VisuLogic

Metrics: accuracy

Baselines: Qwen2.5-VL-7B InternVL3-8B Qwen2.5-VL-32B GPT-4o

Builds on: GRPO Qwen2.5-VL DeepSeek-R1 VLMEvalKit

Closest papers by meaning

• Thinking with Programming Vision: Towards a Unified View for Thinking with Images (CVPR 2026 · similarity 0.88)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.87)
• Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.87)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.87)
• When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought (CVPR 2026 · similarity 0.87)
• What, Whether and How? Unveiling Process Reward Models for Thinking with Images Reasoning (AAAI 2026 · similarity 0.87)