LARC Achieves 0.88 GenEval Score with Continuous Reasoning

A new reasoning framework called LARC (Latent Reasoning in Continuous Space) has achieved 0.88 overall on GenEval when applied to the BAGEL base model, representing a seven-percentage-point improvement and marking the first reasoning-augmented unified multimodal model to demonstrate monotonic test-time scaling. The research, published November 2, 2026, addresses a core challenge in multimodal AI: reasoning methods that work for language models have failed to improve image generation because they operate in incompatible mathematical spaces.
The Modality Mismatch Problem
According to the LARC paper on Hugging Face, previous reasoning-augmented approaches failed to help image generation models scale at test time because of a fundamental incompatibility. Text-based reasoning operates in discrete token space, while image generation requires continuous mathematical representations. "We attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space," the authors write. LARC solves this by performing reasoning directly over continuous hidden states rather than converting everything to text tokens, interleaving language-mode and latent-mode processing steps throughout the generation pipeline.
Training Architecture and Methods
The framework employs a two-stage training process. First, a curriculum supervised fine-tuning phase gradually converts token-level reasoning into latent reasoning. Second, an information-gain reinforcement learning stage enables the model to self-evolve latent reasoning patterns without requiring ground-truth supervision labels. This second phase uses GRPO (Group Relative Policy Optimization) with two custom reward signals: an information-gain reward that encourages useful reasoning steps, and a collapse-prevention term that maintains diversity in the generated outputs. The training approach allows the model to learn which continuous reasoning patterns actually improve final image quality.
Benchmark Performance Gains
LARC demonstrates substantial improvements across multiple evaluation metrics when applied to BAGEL:
- GenEval overall score: 0.88, up from 0.81 baseline (7-point gain)
- Attribute binding: 15-point improvement, the largest category gain
- Spatial positioning: 13-point improvement
- DPG-Bench prompt following: 90.04, up from 85.07
- MIRA visual reasoning: 11.7, up from 7.5
- OneIG-Bench: 0.866, representing a 10.2-point improvement
The attribute binding and spatial positioning gains suggest LARC particularly excels at understanding and rendering complex relationships between objects in generated images, areas where baseline diffusion models often struggle.
Test-Time Scaling Behavior
The headline finding is clear test-time scaling: doubling the latent compute budget yields monotonic performance improvements, and LARC outperforms best-of-N sampling strategies when compute budgets are matched. This represents a breakthrough for unified multimodal models, which until now have not shown the reliable inference-time scaling behavior that language models achieve through chain-of-thought reasoning. The paper demonstrates that both parallel and sequential test-time scaling strategies work, giving practitioners flexibility in how they deploy additional compute during image generation.
Release Date and Availability
The LARC framework was officially released on November 2, 2026, through a paper published on Hugging Face. The research comes from a team documenting their methods for continuous-space reasoning in unified multimodal architectures. As of the release date, no independent third-party reproductions have appeared, and the results are based on experiments with a single base model architecture. The timing places LARC amid an active period of multimodal model releases throughout 2026.
What This Means
LARC demonstrates that reasoning can successfully scale image generation when the reasoning process matches the mathematical structure of the generation task. By working in continuous hidden states rather than discrete tokens, the framework achieves the test-time scaling benefits that have made language models increasingly capable as they think longer. The approach opens a path for future multimodal systems to inherit scaling properties from both language and vision domains. However, broader validation across different base models and independent reproductions will determine whether these gains generalize beyond the initial BAGEL experiments.
on Emergent today





