Skip to content

DeepEyesV2: Toward Agentic Multimodal Model

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; DeepEyesV2 integrates code execution and web search into agentic reasoning

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

DeepEyesV2 is an agentic multimodal model that integrates code execution and web search directly into its iterative reasoning loop. Direct RL alone fails to induce stable tool use; the authors show that a cold-start SFT stage is necessary before RL can reliably strengthen tool invocation. The resulting 7B model surpasses larger (32B) baselines on several benchmarks and achieves state-of-the-art on search-intensive tasks.

Problem

Existing MLLMs are passive: they interpret images and text but cannot autonomously invoke external tools (code execution, web search) for computation or up-to-date knowledge retrieval. Prior "Think with Image" approaches either use a very limited tool set (crop only) or cannot handle knowledge-intensive queries. Crucially, applying RL directly to a base MLLM fails to produce reliable tool use—models either abandon code generation or degenerate to reward hacking (emitting placeholder code blocks).

Method

Two-stage training on Qwen2.5-VL-7B:

  1. Cold-start SFT: A curated dataset spans three domains (perception, reasoning, search) plus long chain-of-thought reasoning data. Filtering is two-pass: (i) difficulty filtering—retain only questions the base model answers correctly ≤2/8 rollouts; (ii) tool-benefit classification—keep only examples where tool invocation actually improves accuracy. Trajectories are synthesized by prompting frontier models (Gemini 2.5 Pro, GPT-4o, Claude Sonnet 4) in an interactive loop with real code execution, retaining only correct, error-free traces. SFT trains for 3 epochs (lr \(= 10^{-5}\), batch 128).

  2. Agentic RL: Uses DAPO with a sparse outcome-driven reward \(R = R_{\text{acc}} + R_{\text{format}}\), KL coefficient 0, 16 rollouts per prompt, max 16,384 token responses. No complex reward shaping; RL encourages dynamic multi-tool combinations that were absent in cold-start data (e.g., webpage scraping via requests/BeautifulSoup emerges spontaneously).

At inference, the model operates in an iterative loop: generate reasoning → emit Python code or SerpAPI search query → receive observation → continue reasoning, until a final answer is produced. Code runs sandboxed; image search returns top-5 matched webpages; text search returns top-5 snippets.

Key Contributions

  • Empirical demonstration that direct RL fails to bootstrap tool use in MLLMs (reward hacking and degeneration to text-only CoT), motivating cold-start necessity.
  • A principled data curation pipeline with difficulty filtering and tool-benefit classification, producing complementary SFT and RL subsets.
  • DeepEyesV2: first agentic MLLM unifying code execution and web search (both text and image queries) in a single reasoning trajectory, built on a 7B backbone.
  • Quantitative analysis of task-adaptive tool invocation: image ops dominate for perception, numerical code for reasoning, search for knowledge-intensive tasks.
  • Evidence that RL enables emergent complex tool combinations (multi-step web scraping) not present in cold-start demonstrations.

Results

  • Real-world understanding: +3.3 on HRBench-4K, +6.3 on HRBench-8K, +5.9 on MME-RealWorld, +7.6 on V*Bench vs. Qwen2.5-VL-7B; surpasses Qwen2.5-VL-32B on multiple benchmarks.
  • Math reasoning: MathVista 71.9% (+3.6), MathVerse 52.7% (+7.1), MathVision 28.9% (+3.3), WeMath 38.1% (+3.5), DynaMath 57.2% (+3.9) vs. Qwen2.5-VL-7B.
  • Search-intensive: MMSearch 63.7% vs. MMSearch-R1 53.8% (+9.9); InfoSeek 51.1% (+2.1 vs. Qwen search baseline); SimpleVQA 59.4% (+7.8).
  • Chart understanding: CharXiv descriptive 78.6% (+5.9), CharXiv reasoning 48.9% (+8.7), ChartQA 88.4% (+2.2).
  • Ablation confirms all three cold-start data categories (perception, reasoning, long CoT) contribute; removing any category degrades performance.

Limitations

  • Evaluation is confined to 7B scale; it is unclear whether the cold-start necessity and RL dynamics transfer to larger or differently architected models.
  • RL data for search is required to unlock search capability; without it, search performance degrades substantially (MMSearch drops from 63.7% to 55.0% in ablation).
  • Tool reliability depends on external services (SerpAPI, sandboxed executor); latency and availability are not analyzed.
  • Cold-start trajectories are distilled from proprietary frontier models, introducing dependency on their capabilities and potential ceiling effects.
  • No analysis of failure modes, hallucinated tool calls, or safety implications of unrestricted web access.

Relevance to Agentic AI / LLM Agents

DeepEyesV2 directly addresses a core challenge in agentic LLM research: how to reliably bootstrap tool-use behavior when the base model lacks it. The finding that RL alone leads to reward hacking—and that cold-start SFT is a necessary precursor—is an important empirical result for anyone designing tool-augmented agent training pipelines. The paper also demonstrates emergent multi-tool chaining from RL (web scraping absent from SFT data appears spontaneously), which connects to debates about whether agentic capabilities can be induced purely through outcome-based optimization. For the broader agentic AI community, the data curation methodology (difficulty + tool-benefit filtering, trajectory synthesis with real execution) offers a replicable recipe for building training sets that go beyond static instruction tuning toward interactive, environment-grounded learning.