Reframe aligns visual agents on tool observations during distillation
GD-GRPO and observation-gated multi-teacher distillation beat Mix-RL and vanilla MOPD on 14 benchmarks.
Stories per day · 30 days
3 RSS
Relatedmultimodal5benchmarks3tool-use3elorian2glimpse2model-release2agentic-reasoning1agentic-rl1agent-memory1api1charts1computer-use1data-agents1decoycount1flow-matching1generative-models1gpt1grpo1
24 stories, newest first · one card per story
Stories tagged #visual-reasoning
Elorian launches Glimpse and the DecoyCount visual benchmark
The paper puts frontier models under 50% accuracy; CEO Dai says Glimpse scored 50.48% on one run.
6
OpenAI releases GPT-6.1 Terra for visual reasoning and computer use
OpenAI calls it a step up from GPT-5.6 Terra for proposals, technical drawings, forms and video editing.
Charts used as internal steps improve agent data analysis
A 200-task probe shows agents read self-made charts as well as numbers but often skip plotting.
OpenVLHarness adds tools and memory for visual agent reasoning
On 22 benchmarks, average scores rise 9.7–12.6 points over matched no-tool baselines across eight backbones.
Visual-Attribution Gating filters contaminated corrections in visual RL
On six visual reasoning benchmarks, VAG beats GRPO and on-policy self-distillation on average.
OpenVLHarness adds tool use and memory for visual agent reasoning
On 22 benchmarks it lifts matched no-tool baselines by 9.7 to 12.6 points across eight backbones.
Elorian’s first AI model wants to reason with images, not just about them
Elorian released Glimpse, a research-preview model that revisits images mid-reasoning through a visual scratch space.
Self-Conditioning Enables Deep Sequential Reasoning in Flow Matching
Researchers show self-conditioning lets flow-matching models gain deeper visual reasoning from additional sampling steps.
Game2Real-Bench links toy-game scores to real visual reasoning
Across 25 vision-language models, game and real scores correlate at 0.935; constraint reasoning rose in one-shot transfer.
Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space
Humanity’s Sixth Sense finds GPT-6 Astra at 53.6 pass@1 versus a 93.1 human baseline on intuitive visual reasoning.
9
ThinkDrag: Semantic Drag-Based Image Editing with Visual Reasoning
ThinkDrag uses visual reasoning so drag-based image edits follow semantic changes, not only geometry.
WOVEN: Weaving Visual World Modeling into Multimodal LLMs
WOVEN trains multimodal LLMs on visual transitions, lifting 22 of 26 external benchmarks by up to 27.3 points.
Humanity's Sixth Sense: Benchmarking Intuitive Visual Reasoning in Multimodal Models
Researchers introduce HSS, where humans score 93.1% and GPT-6-astra only 53.6% on intuitive visual reasoning.
Think Before You Paint: Recursive Latent Reasoning for Diffusion Models
Painter-Thinker adds recursive latent reasoning to diffusion, solving 92.5% of hard MNIST Sudoku puzzles.
ProAR: Learning Prospective Reasoning with Autoregressive Video Models
ProAR makes autoregressive video generation goal-oriented via predicted goal frames, improving visual reasoning using only 25% of standard training steps.
Reinforcing Multimodal Reasoning via Token-Level Perception-Grounded Advantage Estimation
TPAE uses token-level visual-dependency and entropy signals to improve multimodal RL reasoning over sequence-level rewards.
NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning
NeuronEye activates query-relevant visual concepts in frozen VLMs, lifting Qwen2.5-VL-7B CV-Bench by 3.1 points.
LoopVL: Recurrent Visual Intelligence
LoopVL extends loop transformers to vision-language models and beats similarly sized non-recurrent baselines.
Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence
ReaLVR adds visual-evidence supervision to latent reasoning, hitting 63.7% on Qwen2.5-VL-7B and scaling to 235B.
SolveEdit: Benchmarking Visual Problem Solving in Generative Models
SolveEdit benchmarks goal-driven visual edits; SolveEdit-PLAN lifts GPT-Image-2 from 57.0% to 71.6%.
OpenAI's GPT-6 Astra can now tell you exactly where you screwed up your IKEA shelf
OpenAI's GPT-6 Astra scores 80% on Epoch AI's furniture-assembly benchmark, up from 28% ten months earlier.
Reasoning with Image Generation
ReImaGin uses image generators as flexible visual reasoning tools for multimodal LLMs, beating text-only and specialist baselines.
Reason Through the Latent! Making Latent Visual Reasoning Necessary
Researchers introduce CVRR, forcing multimodal models to rely on recurrent latent computation rather than accessible image tokens, validated via causal interventions and benchmarks.