EVAThink trains video models to reason when frame evidence is sparse
Using 6.7K curated samples, it improves the base model under both 32- and 64-frame budgets.
Stories per day · 30 days
3 RSS
Relatedbenchmark4benchmarks43d-reasoning3agents2embodied-ai2qwen2reinforcement-learning2spatial-reasoning2act3d1activation-steering1adaptive-reasoning1adversarial-patches1agent-evaluation1agentic-frameworks1agentvqa1bedbench1cell-tracking1computer-using-agents1
47 stories, newest first · one card per story
Stories tagged #vision-language-models
Spatial Harness grounds VLM spatial reasoning in physics-engine states
On MindCube it raises Qwen3.5-27B from 52.70% to 71.14%, and to 98.25% when poses are annotated.
EviSeek Framework Improves Embodied Question Answering
EviSeek makes evidence demands explicit to boost accuracy on 3,436 ESI-Bench questions.
GUILD guided data collection improves imitation learning sample efficiency
GUILD alternates vision-language model task proposals with predictive data attribution evaluation to close the collection loop.
PlantVLM-Edge enables quantized vision-language models for on-device crop diagnosis
4-bit quantization retains 98.6% answer quality while reducing memory by 49%, running in 0.12s for first token on edge devices.
TriPhysBench evaluates physical intelligence coordination in vision-language models
Benchmark shows top vision-language models achieve only 50.38% accuracy on integrated physical tasks.
VeriCUA verifies computer-using agents by gathering GUI and code evidence
The framework builds a task rubric, then runs GUI or code checks and rejects unsupported completion claims.
BedBench reveals weaknesses in vision-language 3D reasoning
Best vision-language models achieve only 7.2% on per-color counting in BedBench compared to humans at 93.3%.
EditAgent expands tools and verification for agentic image editing
Experts preferred parallel listwise judging over sequential refinement in 78.5% of decided comparisons.
StreamVibe-VL infers evolving user intent for proactive video agents
The benchmark has 4.1K evaluation and 85K training samples; timing and answer quality rose 10.7% over baselines.
VLM safety steering differs by layer depth, study finds
On Qwen3-VL-4B, early vision-layer steering moved unsafe rejection between 6% and 96% without retraining.
AgentVQA benchmarks vision-language models on agentic visual tasks
The best of 22 models scores 58.5% on 1,000 examples drawn from 23 datasets across five agent domains.
CellTrack-Bench compares vision-language models with specialized cell trackers
The benchmark uses 15 datasets, four imaging modalities, and 1,537 problems covering 9,041 cell-linking decisions.
Edge vision-language models match detectors on physical patch evasion
A unified metric finds compact grounding VLMs can fail with confident wrong boxes, not only missed people.
Study finds inverse scaling in vision-language models on some subtasks
Across 19 models and 558,334 evaluations, 4,342 queries were answered by a smaller model and missed by a larger one.
ACT3D uses reinforcement learning so VLMs explore 3D from one view
Average accuracy reached 53.2% on six 3D benchmarks, 22.0 points above base Qwen and above Gemini 3.6 Flash.
RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment
RoTA-Bench shows vision-language models struggle to recover target 3D projections through successive rotations.
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.
SIMIT Self-Improves Vision-Language Models
SIMIT generates query-specific practice examples at test-time yielding +7.20% mean gain over BAGEL-7B.
When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors
A new benchmark finds vision-language models drop spatial relations from the representations used downstream.
Studying Comics at Scale with Artificial Intelligence: The PAGE Framework and Three Studies of American Comics and Japanese Manga
Researchers propose PAGE, an AI framework that parses and annotates thousands of American comics and Japanese manga.
SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
SpatialASM trains vision-language models to plan viewpoints and imagine scenes, improving accuracy on seven spatial benchmarks.
RealityBench tests VLMs on documenting physical work
Across 26 models, documentation is strong for in-frame text but weak for measurements and small spatial relations.
When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models
MOTIVE uses multi-view self-verification so vision-language models accept reliable answers and rethink uncertain ones.