# Visual Reasoning

Multimodal large language models can read text in an image accurately yet still fail to reason over the same problem as well as they do over plain text, a shortfall the authors call a modality reasoning-access gap. TIGER renders text-only math problems as images, mines cases the model solves from text but fails from the image, and post-trains on those image inputs with GRPO and related objectives. Trained only on synthetic rendered-text images, Qwen3-VL-8B rises from 54.5 to 65.8 on MathVerse, while Qwen2.5-VL-7B rises from 43.9 to 47.5 on MathVerse and from 22.2 to 25.7 on EMMA.

Multimodal large language models can read text in an image accurately yet still fail to reason over the same problem as well as they do over plain text, a shortfall the authors call a modality reasoning-access gap. TIGER renders text-only math problems as images, mines cases the model solves from text but fails from the image, and post-trains on those image inputs with GRPO and related objectives. Trained only on synthetic rendered-text images, Qwen3-VL-8B rises from 54.5 to 65.8 on MathVerse, while Qwen2.5-VL-7B rises from 43.9 to 47.5 on MathVerse and from 22.2 to 25.7 on EMMA.

VisualCoding trains reusable visual programs into family-level solvers and returns their valid in-scope answer on applicable inputs, beating the strongest evaluated baseline by 4.45 percentage points in mean accuracy across four benchmarks and 69 task families.

VisualCoding trains reusable visual programs into family-level solvers and returns their valid in-scope answer on applicable inputs, beating the strongest evaluated baseline by 4.45 percentage points in mean accuracy across four benchmarks and 69 task families.

Elorian's visual reasoning model Glimpse is now available on AI Gateway, accepting images and video to follow objects through motion, interpret spatial relationships, and connect an action to its physical consequences. While working, it can return to the visual evidence itself by zooming into a detail, revisiting a short interval of video, or tracking the same feature across frames, supporting uses such as scientific chart analysis, video captioning, visual counting, and robotics data annotation.

Elorian's visual reasoning model Glimpse is now available on AI Gateway, accepting images and video to follow objects through motion, interpret spatial relationships, and connect an action to its physical consequences. While working, it can return to the visual evidence itself by zooming into a detail, revisiting a short interval of video, or tracking the same feature across frames, supporting uses such as scientific chart analysis, video captioning, visual counting, and robotics data annotation.

TiCO replaces coordinate regression in thinking-with-images with a probe token that samples evidence over visual tokens, reaching 78.0% overall accuracy on HR-Bench 4K while generating only 6.2 tokens per query.

TiCO replaces coordinate regression in thinking-with-images with a probe token that samples evidence over visual tokens, reaching 78.0% overall accuracy on HR-Bench 4K while generating only 6.2 tokens per query.

Isolating the reinforcement-learning-stage parameter update of a language model, keeping only the single dominant direction of each matrix rescaled to the original magnitude, and injecting it into the language modules of a vision-language model improves matched full-update interpolation in 12 of 15 backbone-benchmark comparisons across three model families and five visual-reasoning benchmarks, including an 8.55 percentage-point MathVision gain on the Qwen recipient. Controls show that matching update magnitude or using arbitrary low-rank directions does not reproduce these gains, indicating that what is acquired during post-training and what remains transferable across models are not the same thing.

Isolating the reinforcement-learning-stage parameter update of a language model, keeping only the single dominant direction of each matrix rescaled to the original magnitude, and injecting it into the language modules of a vision-language model improves matched full-update interpolation in 12 of 15 backbone-benchmark comparisons across three model families and five visual-reasoning benchmarks, including an 8.55 percentage-point MathVision gain on the Qwen recipient. Controls show that matching update magnitude or using arbitrary low-rank directions does not reproduce these gains, indicating that what is acquired during post-training and what remains transferable across models are not the same thing.

The work introduces Flow-of-Thought (FoT), a framework that trains coordinate-, time-, and action-conditioned spatial U-Nets with flow matching on rotation orbits and cumulative shortest-path prefixes, then freezes the learned dynamics; at inference, the source and its horizontal reflection serve as competing hypotheses selected by foreground-weighted reconstruction energy. On locked tests FoT reaches 100.0% on Tetris and 99.0% on colored shapes, with maze endpoint IoU 97.5% and 100% goal reach; under frozen transfer to 133 BLINK Multi-view public validation pairs, the orbit-trained flow reaches 72.2% versus 63.9% for its endpoint-only control.

The work introduces Flow-of-Thought (FoT), a framework that trains coordinate-, time-, and action-conditioned spatial U-Nets with flow matching on rotation orbits and cumulative shortest-path prefixes, then freezes the learned dynamics; at inference, the source and its horizontal reflection serve as competing hypotheses selected by foreground-weighted reconstruction energy. On locked tests FoT reaches 100.0% on Tetris and 99.0% on colored shapes, with maze endpoint IoU 97.5% and 100% goal reach; under frozen transfer to 133 BLINK Multi-view public validation pairs, the orbit-trained flow reaches 72.2% versus 63.9% for its endpoint-only control.

The work proposes FastRL, a plug-and-play reinforcement learning framework that selectively preserves high-advantage trajectories while maximizing inter-trajectory gradient diversity through an advantage-aware pruning strategy, and dynamically adjusts sampling scale across training stages based on historical pruning distributions via an adaptive rollout sampling mechanism, achieving an average 2.07x training speedup on Geometry3K and GeoQA8K-R1V along with approximately 1.64% improvement in average accuracy on visual reasoning benchmarks, and integrating seamlessly into GRPO, DAPO, and GSPO variants.

The work proposes FastRL, a plug-and-play reinforcement learning framework that selectively preserves high-advantage trajectories while maximizing inter-trajectory gradient diversity through an advantage-aware pruning strategy, and dynamically adjusts sampling scale across training stages based on historical pruning distributions via an adaptive rollout sampling mechanism, achieving an average 2.07x training speedup on Geometry3K and GeoQA8K-R1V along with approximately 1.64% improvement in average accuracy on visual reasoning benchmarks, and integrating seamlessly into GRPO, DAPO, and GSPO variants.