We introduce Precise Information Control (PIC), a framework that reframes faithful hallucination in long-form generation as a claim-level control problem, and present a benchmark, initial models, and practical demonstrations.

We develop a variance-aware gap-dependent regret bound with better $H$ dependence for tabular MDPs.

The token embeddings for LLMs are unlikely to be manifolds with low curvature, based upon a novel rigorous statistical hypothesis test.

We propose CellCLIP, a multimodal contrastive learning framework for Cell Painting data. It aligns image profiles and perturbation descriptions using pretrained encoders and a novel channel-wise scheme.

The volume transmission of small molecules implements a context-factorized hypernetwork capable of online credit assignment and compositional learning reminiscent of neurobiology.

We present G-Vendi, a data diversity measure that strongly correlates with LLM reasoning generalization in OOD benchmarks; we use this insight to diverse synthetic reasoning data, which leads to SOTA distilled models in NLI and math reasoning.

Minimax rate of robust estimation when different samples have different but known rates of corruption.

We provide a computational efficient algorithm to achieve $O(H)$ deployment cost with polynomial sample complexity.

Our work the first to incorporate human driving cognition to enhance end-to-end autonomous driving models, yielding significant findings.

A framework and benchmark to evaluate language models' reasoning on imperfect tabular data

Measuring and improving the signal-to-noise ratio in language model benchmarks.

We use physics-guided semi-supervised learning to align human-labeled wildfire smoke annotations with GOES satellite imagery, creating SmokeViz, a large-scale dataset for smoke plume segmentation.

A framework that uses Monte Carlo Tree Search for guiding an LLM to ask information-seeking questions, learning from past successful strategies to solve problems more efficiently and accurately.

To achieve personalization in LLMs, we leverage the user model to incorporate a curiosity-based intrinsic reward into multi-turn RLHF.

A systematic review of benchmarks for large language models focusing on issues of construct validity, which results in a set of recommendations for best practices.

Fair Universe HiggsML was a NeurIPS 2024 competition where participants dealt with a biased dataset and estimated a confidence interval on a parameter of interest in fundamental physics. This paper is about the dataset and the competition findings.

Chatbot Arena has become a leading platform for ranking AI models. Our extensive study uncovers hidden dynamics that distort rankings and provides concrete steps to enhance fairness and transparency in evaluation of models on Chatbot Arena.

URDF-Anything is an end-to-end MLLM-based framework that jointly optimizes geometric segmentation and kinematic prediction for articulated object reconstruction.

Debate between AI experts outperforms single-advisor consultancy in helping humans make more accurate factual judgments, especially benefiting those with mainstream beliefs.

Language models are surprisingly robust to non-canonical tokenizations of the input, which can even lead to improved performance

We solve noisy linear inverse problems 10-50x faster than existing diffusion methods and with exact recovery of measurement constraints

We design a scheme to embed upto 24576 fingerprints into an LLM for better security

We meta-learn a transformer based in-context learning fMRI visual cortex encoder which can adapt to new human subjects without any fine-tuning

This paper tackles the problem of zero-shot inference on text-attributed graphs, and proposes a novel method that uses bundles to query LLMs and supervise GNNs.

We argue that there are important gaps between what unlearning methods can do and claims about/hopes for what these methods can do for broader impact.

A new minimalist example to understand the Edge of Stability and Progressive Sharpening phenomenon

Meta-RL with self-supervised predictive coding modules can learn interpretable, task-relevant representations that better approximate Bayes-optimal belief states than black-box meta-RL models across diverse partially observable environments.

We perform large-scale strong MIAs on pre-trained LLMs to clarify the extent of actual privacy risk MIAs pose in this setting.

We study how different changes to the auto-encoder rate-distortion trade-off affect discrete auto-regressive generation performance through scaling laws.

A cognition grounded interpretability method

We build a benchmark on attribute-focused text-to-image retrieval and propose a pipeline of using promptable image embeddings for solving it, leading to performance gain.

We efficetively address the resolution scalability challenge caused by exponential growth in token sequences in VAR models, while preserving pixel-level fidelity.

ReAgent-V enables reward-driven, multi-agent video understanding with dynamic reflection and frame selection.

How do models generalize to unseen constraints for precise instruction following?

We introduce a visual knowledge dataset aiming to evaluate current models' latest visual knowledge seeking and updating capabilities.

We only need one example for RLVR on LLMs to achieve significant improvement on math tasks

We introduce a framework for evaluating & improving LLM consistency in simulated human dialogue. Our metrics correlate with human judgments and when used with multi-turn RL, reduce inconsistency across chit-chat, teaching and mental health dialogue.

Using simple DNNs with dendrite-inspired architectures, we show that single neurons exhibit a morphology-driven phase transition in learnability, with robustness–adaptability trade-offs that suggest new inductive biases for NeuroAI model design.

We theoretically analyze the benefit of filtering a noisy training dataset on model performance in multimodal contrastive learning, and identify two regimes with different amounts of gain.

A multi-metric Common Task Framework that standardizes characterization and comparison of modeling methods for dynamic systems.