A new framework named ATLAS aims to resolve a fundamental tension in visual reasoning systems by merging two distinct approaches: agentic reasoning and latent reasoning. Developed by researchers and detailed in a paper on arXiv, ATLAS uses a single discrete "word," termed a functional token, to serve as both an agentic operation and a latent visual reasoning unit. This innovation seeks to improve efficiency and task generalization in visual reasoning, areas where prior methods have faced limitations.
Traditional visual reasoning often involves intermediate visual states, and a direct approach of generating images during reasoning is computationally expensive and complex to implement. Alternatives have emerged, including agentic reasoning, which uses code or tool calls, and latent reasoning, which relies on learned hidden embeddings. However, agentic methods introduce latency due to context switching between external executions, while latent methods can struggle with task generalization and are difficult to train using autoregressive parallelization.
The ATLAS framework addresses these shortcomings by internalizing visual operations within these functional tokens. Each token is associated with a visual operation but does not require explicit visual supervision, functioning as a standard token within the tokenizer's vocabulary that can be generated through next-token prediction. This design bypasses the need for generating verbose intermediate visual content and maintains compatibility with standard scalable supervised fine-tuning (SFT) and reinforcement learning (RL) training methods without requiring architectural or methodological changes.
To further enhance the training process, particularly in RL where functional tokens can be sparse, the researchers introduced Latent-Anchored GRPO (LA-GRPO). This method stabilizes training by anchoring functional tokens with a statically weighted auxiliary objective, which provides stronger gradient updates.
Extensive experiments and analyses reported in the paper indicate that ATLAS achieves superior performance on challenging benchmarks while maintaining clear interpretability. The researchers suggest that ATLAS could offer a new paradigm for future visual reasoning research. The framework's ability to potentially collapse the architectural trade-off between speed and flexibility could significantly impact how large-scale multimodal AI systems are developed.
The research originates from institutions including Meta AI and CUHK. The paper's contribution lies in proposing a learned token as a routing mechanism, allowing a single token to direct the reasoning process rather than forcing a choice between code execution or embedding-based reasoning. This approach aims to sidestep the limitations of each method by offering a unified solution.