Research map / Vision-language and multimodal / Multimodal LLMs

VICO: Visual Environments Co-Evolving for Vision-Language Model Reasoning

NeurIPS 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.

Proposes VICO, a co-evolution framework where a VLM actor and an EnvRewriter jointly train, with the EnvRewriter editing verifiable image structures (CLEVR scene graphs, ChartQA tables, ZwZ masks) to keep tasks near the actor's 50% pass-rate frontier.

Problem

RLVR for VLMs uses a static training environment, so as the actor improves tasks become trivial or unsolvable and the GRPO learning signal collapses.

Approach

Use the actor's rollout pass rate as a shared signal: an EnvRewriter edits verifiable image-side structures and is rewarded by a bell-shaped alpha-curriculum reward peaked at pass rate 0.5, while the actor is trained with GRPO on the edited, label-valid samples.

Key results

• VICO-8B (Chart) reaches 64.9% OOD average excluding ChartQA, +5.0% over Qwen3-VL-8B base, and +4.3% over the strongest self-evolution baseline (MM-Zero).
• On ChartQA, VICO-Qwen3-VL-8B (Chart) reaches 89.1% vs 76.9% base, exceeding offline Nano Banana augmentation (81.9%) and Direct GRPO (83.8%).
• VICO achieves comparable chart results to chart-specialized RLVR methods using 16x-160x fewer labeled samples (2000 seed inputs).

Surprising findings

• Text-only perturbation (V-Zero-style) yields only +2.4% overall gain versus +6.1% for visual editing under the same CLEVR seed.
• Static offline augmentation with stronger generators (GPT-5, Nano Banana) plateaus below 82% on ChartQA, below the co-evolution result.

Limitations

• Requires a verifiable image-side structure (scene graph, table, or region mask), limiting applicability to domains without such structure.
• Extra compute from repeated rollouts and joint updates of two models per round.
• Evaluation restricted to verifiable-answer visual reasoning tasks; open-ended generation not covered.

Experimental setup

Datasets: CLEVR ChartQA ZwZ

Benchmarks: MathVista MathVision ChartQA MathVerse WeMath LogicVista MMMU MMVP BLINK

Metrics: accuracy

Baselines: Direct GRPO Supervised GRPO MM-Zero VisionZero V-Zero R1-OneVision-7B MM-Eureka OpenVLThinker-7B GPT-5 Nano Banana Gemini-2.5-Pro

Builds on: GRPO RLVR VisionZero V-Zero MM-Zero

Closest papers by meaning

• VisPlay: Self-Evolving Vision-Language Models (CVPR 2026 · similarity 0.90)
• Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning (ICML 2026 · similarity 0.90)
• ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs (NeurIPS 2025 · similarity 0.90)
• Bridging Environments and Language with Rendering Functions and Vision-Language Models (ICML 2024 · similarity 0.90)
• Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models (ICML 2024 · similarity 0.89)
• SSL4RL: Revisiting Self-supervised Learning as Intrinsic Reward for Visual-Language Reasoning (ICML 2026 · similarity 0.89)