Perception‑R1: RL Gives Visual Insight Without Chain‑of‑Thought, Beats Four Tasks

The authors observe that recent large language models (LLMs) have shifted from non‑reasoning foundations (e.g., GPT‑4/4o, DeepSeek‑V3) to strong reasoning models (e.g., OpenAI o1/o3, DeepSeek‑R1, Kimi‑1.5). DeepSeek‑R1 introduced a simple rule‑based reinforcement learning (RL) method that produces emergent reasoning without Monte‑Carlo tree search or separate reward models. Building on this, the paper explores applying RL to multimodal large language models (MLLMs) for visual perception tasks.

Problem Statement

Current RL research for LLMs focuses on pure language or language‑centric multimodal tasks, while visual perception tasks differ fundamentally: they have explicit physical ground truth (points, lines, bounding boxes) and are often single‑step predictions, offering limited structured search space for RL. The authors therefore ask whether RL can improve perception strategies and under what conditions.

Method: Perception‑R1

Perception‑R1 applies the rule‑based RL algorithm GRPO [55] during the post‑training phase of an MLLM. The baseline model is Qwen2‑VL‑2B‑Instruct [61]; for detection the authors also use Qwen2.5‑VL‑3B‑Instruct [3]. The framework introduces two key innovations:

First systematic use of RL for visual perception strategy learning , filling a gap where prior work only applied RL to language reasoning.

GRPO‑based Perception‑R1 with task‑specific reward design and multi‑topic reward matching.

Rule‑Based Reward Modeling

Two reward types are defined:

Format Reward : enforces the output format required by each visual task (e.g., bounding‑box coordinates must follow [x1,y1,x2,y2]).

Answer Reward : measures correctness against the physical ground truth. For localization it uses IoU, for counting it uses Euclidean distance between predicted and true points, for OCR it uses edit distance, and for detection it combines F1‑score, IoU, and a binary penalty.

Multi‑Topic Reward Matching

Because visual scenes contain multiple objects, the authors formulate reward assignment as a bipartite‑graph matching problem. Predictions form one node set, ground‑truth objects the other; edge weights are given by the task‑specific reward function. The optimal assignment maximizes total reward and is solved efficiently with the Hungarian algorithm [27]. This matching is crucial for counting and detection tasks where many objects coexist.

Experiments

Perception‑R1 is evaluated on four representative perception benchmarks:

RefCOCO+ : +4.2% improvement.

PixMo‑Count : +17.9% improvement.

PageOCR : F1‑score +4.2%.

COCO2017 validation : first pure MLLM to reach 31.9 mAP, surpassing expert models.

These gains are reported in Tables 1‑4 of the paper. Additional cross‑task benefits are observed on general multimodal understanding benchmarks (Table 4).

Ablation Studies

Key findings from Table 5 include:

Replacing bipartite matching with sequential matching degrades counting and detection performance, confirming the importance of the matching mechanism.

Introducing an explicit chain‑of‑thought (CoT) process harms all four perception tasks, indicating that CoT is unnecessary for current visual perception problems.

Perception confusion—a measure of task difficulty—predicts when RL outperforms supervised fine‑tuning (SFT). High‑confusion tasks (counting, detection) benefit most from RL, while low‑confusion tasks (localization, OCR) do not.

Further Analyses

Reward‑design experiments (Table 6) show that progressively adding finer‑grained rewards (format → answer) continuously improves detection performance, underscoring reward design as a critical factor.

Rollout scalability experiments (Figure 2) reveal that increasing the number of rollouts improves reward optimization and final accuracy on localization and counting, demonstrating that Perception‑R1 scales well with more exploration.

Limitations

The authors note four main limitations:

Many existing visual perception tasks are overly simple, restricting RL’s exploration space.

Lack of more complex meta‑tasks hampers comprehensive evaluation of RL’s potential.

The current study focuses on localization, counting, OCR, and detection; broader perception tasks such as action recognition remain unexplored.

Rule‑based reward designs may have limited generalization and need further validation on unseen domains.

Conclusion

Perception‑R1 demonstrates that rule‑based RL can unlock “visual insight” in multimodal LLMs without the need for chain‑of‑thought reasoning, achieving state‑of‑the‑art results on four major perception benchmarks. The work highlights perception confusion and carefully crafted rewards as decisive factors for RL effectiveness, and opens avenues for more challenging perception meta‑tasks.