Research map / Vision-language and multimodal / Multimodal LLMs

MILR: Improving Multimodal Image Generation via Test-Time Latent Reasoning

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal 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 MILR, a test-time method that optimizes continuous latent vectors of text and image tokens in a unified multimodal understanding and generation model via REINFORCE guided by a reward model, without updating parameters.

Problem

Improving text-guided image generation through reasoning without curated reasoning data or fine-tuning, while reasoning jointly over image and text in a shared space.

Approach

Instead of searching over discrete image/text tokens, run REINFORCE-style policy gradient updates on the intermediate Transformer output latents (text and image prefix tokens) of a Janus-Pro MUG model at test time, using a benchmark evaluator or off-the-shelf critic as reward, with early stopping.

Key results

• Janus-Pro-7B+MILR reaches 0.95 overall on GenEval, matching Flow-GRPO (0.95) and beating the best test-time baseline ReflectionFlow/PARM/Best-of-N (0.91).
• On WISE, Janus-Pro-7B+MILR scores 0.63 versus 0.35 for base Janus-Pro-7B (+80%).
• On T2I-CompBench overall, Janus-Pro-7B+MILR scores 0.5325 versus 0.3921 for the base model and slightly above T2I-R1 (0.5281).

Surprising findings

• Optimizing only the first 2% of image tokens gives peak GenEval performance, and random-subset image optimization is worse than prefix optimization.
• Text-only latent optimization performs nearly as well as joint optimization on all benchmarks.

Limitations

• Relies on a reward model; the best results use the benchmark's oracle evaluator, and non-oracle critics lag behind.
• Failure modes include textual and visual reasoning collapse and reward hacking; an L2 latent regularizer is needed to mitigate reward hacking.
• Evaluated only on autoregressive MUG (Janus-Pro); transfer to diffusion-based MUG is untested.

Experimental setup

Datasets: GenEval T2I-CompBench WISE

Benchmarks: GenEval T2I-CompBench WISE

Metrics: GenEval overall score T2I-CompBench category scores WISE overall score

Baselines: Flow-GRPO T2I-R1 GoT-R1 ReflectionFlow PARM Best-of-N Reflect-DiT Janus-Pro-7B+GRPO Janus-Pro-7B+DPO BAGEL GPT-4o SD3-Medium

Builds on: Janus-Pro REINFORCE GRPO Best-of-N

Closest papers by meaning

• ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.90)
• RMIR: A Benchmark Dataset for Reasoning-Intensive Multimodal Image Retrieval (CVPR 2026 · similarity 0.89)
• DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.89)
• GIR-Bench: Versatile Benchmark for Generating Images with Reasoning (ICLR 2026 · similarity 0.89)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.89)
• MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning (ICLR 2026 · similarity 0.89)