🏢 MIT

This paper presents a novel watermarking scheme for language models that is both undetectable and robust to a constant fraction of adversarial edits (insertions, deletions, substitutions).

This research introduces novel algorithms to dynamically price third-party platform service fees under strategic buyer behavior, achieving optimal revenue with a theoretically proven regret bound.

Boosting Human-AI teamwork via interactive, explainable AI!

Federated data acquisition via experimental design (DAVED) achieves lower prediction error without labeled validation data, optimizing cost-effectively for test-set predictions in decentralized market…

CONTEXTCITE pinpoints which parts of a given context led a language model to generate a specific statement, improving model verification and response quality.

Researchers developed a novel method using mechanistic interpretability to create compact formal proofs for AI model performance, improving AI safety and reliability.

Researchers strengthen understanding of transformer self-attention by proving asymptotic convergence to single clusters under causal masking, linking it to the Rényi parking problem.

BrainBits reveals that surprisingly little brain information is needed for high-fidelity image & text reconstruction, highlighting the dominance of generative model priors over neural signal extractio…

BitDelta drastically shrinks fine-tuned LLMs by quantizing their weight deltas to just one bit, achieving 10x memory reduction and latency improvements without sacrificing performance.

BEND-VLM: A novel, efficient test-time debiasing method for vision-language models, resolving bias without retraining.

Evidential deep learning’s uncertainty quantification is unreliable; this paper reveals its limitations, proposes model uncertainty incorporation for improved performance.

ARCHITECT: Generating realistic 3D scenes using hierarchical 2D inpainting!

ALPS: An optimization-based framework achieves state-of-the-art one-shot LLM pruning, significantly reducing test perplexity and improving zero-shot performance.

Randomly initialized transformers, with only embedding layers optimized, surprisingly excel at various algorithmic tasks, revealing inherent capabilities even before training.

Cert-LSVI-UCB achieves constant regret in RL with linear function approximation, even under model misspecification, using a novel certified estimator.

LLMs dynamically adjust restless multi-armed bandit (RMAB) resource allocation policies in public health via human-language commands.