🏢 Carnegie Mellon University

VLMs learn to generate their own memories by abstracting experiences from noisy demonstrations and human feedback, significantly boosting in-context learning performance.

CONBIAS tackles dataset bias by representing visual data as concept graphs, diagnosing imbalances via clique analysis, and debiasing through targeted data augmentation for improved model generalizatio…

Diffusion models generate unrealistic images by smoothly interpolating between data modes; this paper identifies this ‘mode interpolation’ failure and proposes a metric to detect and reduce it.

MOHAWK: Distilling Transformers’ quadratic knowledge into faster subquadratic SSMs, achieving state-of-the-art performance with <1% of training data!

This work unveils a causal lens on extrapolation, offering theoretical guarantees for accurate predictions on out-of-support data, even with limited target samples.

Federated Q-learning achieves optimal sample & communication complexities simultaneously via Fed-DVR-Q, a novel algorithm.

Hybrid Preference Optimization (HyPO) outperforms existing offline methods for fine-tuning LLMs by leveraging both offline and online data, achieving better performance and efficiency.

Test-time adaptation strengthens the linear correlation between in- and out-of-distribution accuracy, enabling precise OOD performance prediction and hyperparameter optimization without labeled OOD da…

Tactile DreamFusion: High-resolution tactile sensing enhances 3D generation, creating realistic geometric details previously unattainable.

Synatra synthesizes high-quality digital agent training data from online tutorials and web pages, significantly improving agent performance on complex web-based tasks at a fraction of the cost of huma…

SparseAGS: High-fidelity 3D reconstruction & camera pose estimation from sparse views via generative synthesis.

Slightly corrupting pre-training data significantly improves diffusion models’ image generation quality, diversity, and fidelity.

SIRIUS: A novel correction mechanism boosts the efficiency of contextually sparse LLMs for complex reasoning tasks, achieving significant latency reduction.

SEQUOIA: A novel algorithm boosts Large Language Model (LLM) inference speed by up to 9.5x using a scalable and robust speculative decoding approach!

First finite-sample analysis of interventional causal representation learning shows that surprisingly few samples suffice for accurate graph and latent variable recovery.

S2FT: Structured Sparse Fine-Tuning achieves state-of-the-art LLM fine-tuning performance, training efficiency, and inference scalability by selecting sparsely and computing densely.

Leveraging model-generated synthetic data for LLM finetuning significantly improves efficiency when using both positive and strategically constructed negative examples, resulting in an eight-fold incr…

Researchers propose Adversarial Compression Ratio (ACR) to assess LLM memorization, offering an adversarial, flexible, and computationally efficient method for monitoring data misuse and compliance.

This research shows how to improve Stackelberg game strategies by considering side information, achieving no-regret learning in online settings with stochastic contexts or followers.

RISE: Recursive Introspection teaches LLMs to iteratively improve their responses, enabling self-correction and enhanced performance on challenging reasoning tasks.