280

Papers

11/11

Roadblocks Active

4

Connections

Standard text-to-image models fail on complex prompts because they plan in language without checking whether the generated image actually matches the plan. This paper introduces CLVR, a framework that couples language-level planning with pixel-level diffusion generation in a closed loop, using two judge AI models to verify each reasoning step and filter training trajectories. The result approaches proprietary commercial model quality on multiple benchmarks, and the mechanism — verify-then-train rather than generate-and-hope — offers a reusable template for grounding any generative model to its own outputs.

██████████ 0.9 reasoning-reliability Preprint

ReadSaveConnections

MemLens tests 34 vision-language systems — 27 large models and 7 memory-augmented agents — on 789 questions requiring visual evidence across conversation lengths up to 256,000 tokens. The key finding is a clean trade-off: large models with extended context windows are accurate on short conversations but degrade fast, while memory agents stay stable but lose visual fidelity through compression. Removing evidence images drops frontier model accuracy below 2%, confirming that visual grounding, not just text recall, is the true bottleneck.

██████████ 0.9 long-context Preprint

ReadSaveConnections

Building AI agents that can operate graphical interfaces requires enormous labeled data, which is expensive to collect manually. Video2GUI scrapes 500 million Internet tutorial video records and automatically extracts 12 million labeled GUI interaction trajectories across 1,500+ applications — no human annotation needed. Pre-training standard vision-language models on this dataset yields 5–20% accuracy gains on GUI benchmarks, demonstrating that scale of weakly-supervised, automatically-mined demonstrations can substitute for hand-labeled data in agent training.

██████████ 0.9 multimodal-understanding Preprint

ReadSaveConnections

Medical AI agents currently hallucinate clinical trends because they try to reason about time-series patient data in plain language rather than computing statistics explicitly. COTCAgent fixes this by having the model write executable code to analyze longitudinal electronic health records, then match the computed trends against a knowledge base of 9,948 diseases using weighted scoring. This separation of computation from language generation achieves 70.41% on the public HealthBench benchmark, outperforming frontier models including o4-mini, by eliminating a systematic source of confabulation.

██████████ 0.9 hallucination-grounding Preprint

ReadSaveConnections

When the correct answer is deliberately removed from a medical multiple-choice question, frontier AI models still commit to a wrong answer 53–82% of the time instead of saying 'I don't know' or asking for clarification — a behavior the authors call premature closure. This is measured across five leading models (including GPT, Claude, Gemini, and DeepSeek variants) on MedQA and AfriMed-QA benchmarks. The result matters because confident wrong answers in clinical settings are more dangerous than expressed uncertainty, and the high failure rates suggest this is a structural property of current training, not an edge case.

██████████ 0.9 alignment-safety Preprint

ReadSaveConnections

Video-Zero teaches a model to improve its own video understanding without any human-labeled data, by having it generate hard questions about videos and then answer them — but with a critical twist: questions must be grounded to specific temporal evidence segments, not just any frame. This evidence-grounding requirement is the key finding: models that generate questions without temporal anchoring learn mostly from static visual cues and fail to improve on genuinely temporal tasks. Validated across 13 benchmarks and three model families, the framework shows self-supervised training can work for video if the supervision signal enforces temporal specificity.

██████████ 0.9 multimodal-understanding Preprint

ReadSaveConnections

Vision-language models systematically misinterpret historical artifacts by applying modern or culturally mismatched concepts — for example, describing a prehistoric tool using contemporary manufacturing terminology. The authors test ten state-of-the-art systems on a benchmark of 600 questions about 1,600 Indian cultural artifacts spanning prehistoric to modern periods, and find even the best model scores only 58.7%. Crucially, performance does not improve with model scale, suggesting this is a structural limitation from training data bias rather than a capacity problem that more parameters will fix.

██████████ 0.9 multimodal-understanding Preprint

ReadSaveConnections

As AI systems increasingly rely on multiple AI agents working together, errors compound and become hard to trace back to their source. This survey maps the landscape of multi-agent AI research through a four-stage framework — build individual capability, integrate agents, find where failures originate, then evolve the system — and identifies that the failure attribution stage is severely underdeveloped compared to collaboration research. The practical implication is that current multi-agent deployments lack principled debugging tools, making them fragile in production settings where errors propagate silently.

██████████ 0.8 agent-tool-use Preprint

ReadSaveConnections

Most memory benchmarks for AI agents can be answered from text captions alone, masking whether an agent truly preserves and reasons over visual evidence. MemEye constructs 742 question-answer pairs designed to require pixel-level visual detail and multi-step temporal reasoning, with validation checks to ensure caption-only shortcuts do not work. Testing 13 memory methods reveals three failure bottlenecks: routing evidence to the right memory slot, tracking how stored state changes over time, and extracting fine-grained visual details at retrieval — each requiring different fixes.

██████████ 0.8 multimodal-understanding Preprint

ReadSaveConnections

Visual reasoning in AI typically requires either calling external tools (slow, brittle) or generating intermediate images (expensive). ATLAS introduces five types of special vocabulary tokens — for lines, text, manipulation, arrows, and shapes — that a model can emit during normal text generation to trigger visual reasoning operations without switching modes or incurring context-switching latency. Trained on 178,000 curated examples with a reinforcement learning objective that rewards valid token use alongside answer correctness, the approach outperforms prior methods on challenging visual reasoning benchmarks while keeping the model architecture standard.

██████████ 0.8 reasoning-reliability Preprint

ReadSaveConnections