MIT CSAIL

A simple visual harness transforms a general multimodal model into a perfect reasoner on ARC-AGI-3, completing all games with drastically fewer actions than humans.

Heterogeneous visual agents can boost downstream performance by up to 4.9 points across diverse tasks simply by reading from a shared, self-organizing 2B memory tree—without updating a single host model weight.

It is shown that the workspace token can be used as a drop-in replacement for observations during deployment, enabling policies to solve memory-intensive tasks without the need for VLM reasoning in-the-loop.

ModaLens, a paired image-swap audit, measures how report availability changes image sensitivity and limits conclusions about visual correctness; the direction replicates in two further model lineages.

Achieving a 40% reduction in character error rate, this syllable-level UASR framework unlocks new potential for low-resource language recognition without costly phoneme resources.

RoboEdit transforms abundant human manipulation videos into high-fidelity training data for robots, unlocking new avenues for scalable robot learning.

FiGuRO reveals that effective intrinsic dimension estimation can emerge as a byproduct of optimizing low-rank projections, transforming how we approach multi-modal representation learning.

Dropping echocardiogram data nearly doubles error rates in cardiac AI models, revealing critical modality dependencies that could impact clinical decisions.

SciFigAlign achieves a remarkable 59% reduction in error over traditional LLM-based scoring methods by grounding figure assessments in manuscript context.

Ms.Forcing achieves a 39.6% speedup in streaming video generation while significantly enhancing quality by intelligently adapting spatial granularity to noise levels.

Richer visual representations don't always lead to better human alignment; sometimes, simpler methods like temporally averaged images outperform full video in identifying engaging moments.

Sinusoidal filters in the first layer of CNNs can unlock optimal representations for contrastive learning, even with complex augmentations.

A groundbreaking dataset reveals that 3D particle models significantly outperform 2D video models in capturing the complexities of deformable object dynamics.

Relying on text-based rationales for dementia classification can actually degrade performance, revealing the need for more sophisticated approaches like DeTAiL.

FracEvent achieves superior event timing and downstream performance by accurately modeling pixel dynamics, outperforming traditional simulators.

MixTIME reveals that integrating diverse image modalities can significantly enhance the precision of immune biomarker predictions in oncology, outperforming traditional single-modality approaches.

Retaining visual figures in skill artifacts boosts CUA performance by over 23 points, proving that seeing is believing in agent training.

Action-chunking policies can lead to premature robot assistance, but a novel steering method effectively mitigates this issue, enhancing collaboration efficiency.

Achieving long-range consistency in video generation is now feasible with a hierarchical token approach that balances detail and computational efficiency.

USAD 2.0 achieves state-of-the-art audio understanding by seamlessly integrating self-supervised and supervised learning techniques, scaling to one billion parameters.

Text-to-image models may only need basic word meanings and order, not complex contextual embeddings, to produce high-quality images.

Multimodal pretraining doesn't guarantee better alignment with human reading patterns, suggesting that language-internal representations are still king when modeling how humans process text.

Quantum kernels unlock signal in medical image embeddings where classical methods fail, suggesting a new path for extracting value from medical foundation models.

Multi-event video generation gets a 33% quality boost with TS-Attn, a training-free attention mechanism that dynamically aligns video content with complex temporal prompts.

Imagine a medical "Google" where you can search for similar patients using text, images, and medical history, and predict future health risks years in advance – Apollo brings this closer to reality.

Even the best LLMs still stumble on Olympiad-level math, and retrieval quality is the bottleneck for retrieval-augmented problem solving, according to the new MathNet benchmark.

Users feel more creative and in control when building images step-by-step from sketches, rather than wrestling with a one-shot text-to-image generator's fully-formed (and often unwanted) details.

Multi-agent systems can find 5x more real-world events in satellite imagery than traditional methods, unlocking a wealth of training data for multi-temporal change detection.

Track unseen objects through total occlusion without CAD models, using just a handful of 2D points.

Gaze-tracking unlocks a new level of personalized AI assistance, enabling LLMs to infer user cognitive states and boost recall performance.

Robots can now "see" hidden objects and understand articulation by learning from human egocentric video, even if they can't physically explore those areas themselves.

VLMs hide a surprisingly compact set of recurrent hub neurons that, when perturbed, can drastically alter model outputs, suggesting a new path for targeted interventions.

By dynamically adjusting contrastive learning temperatures based on data density, MM-TS achieves state-of-the-art results on multimodal long-tail datasets.

Forget hand-engineered features: this approach learns symbolic representations for robotic planning directly from pixels using VLMs, enabling impressive zero-shot generalization to new environments and goals.

Forget simulated manipulation—ManipulationNet offers a global infrastructure for benchmarking robots in the real world, complete with standardized hardware and software, to finally measure progress toward general manipulation.

Injecting spatial transcriptomics data into existing pathology foundation models unlocks significant performance gains across a range of downstream tasks, including molecular status prediction and gene-to-image retrieval.

Achieving 80% accuracy on VQA v2.0 proves that combining Visual BERT, ViLT, and memory-augmented attention can significantly outperform traditional VQA models.