Posters

This paper unveils the information-theoretic limits of online classification with noisy labels, showing that the minimax risk is tightly characterized by the Hellinger gap of noisy label distributions…

New theoretical analysis reveals optimal binning strategies for minimizing bias in expected calibration error (ECE), improving machine learning model calibration evaluation.

InfoRE: A novel method improving large language models’ reasoning by reorganizing information to highlight logical relationships, resulting in a 4% average accuracy boost across various tasks.

InfoRM tackles reward hacking in RLHF using an information-theoretic approach, enhancing generalizability and enabling overoptimization detection.

InfLLM: Training-free long-context extrapolation for LLMs via efficient context memory.

Calibrated Bayesian inference achieved via novel diffusion models uniquely mapping high-dimensional data to lower-dimensional Gaussian distributions.

InfiNet achieves state-of-the-art results by enabling feature interaction in an infinite-dimensional space using RBF kernels, surpassing models limited to finite-dimensional interactions.

Researchers reveal how the training dynamics of transformer models behave at infinite width, depth, and head count, providing key insights for scaling up these models.

Researchers developed a method using variational sequential Monte Carlo to fit stochastic low-rank recurrent neural networks to neural data, enabling efficient analysis and generation of realistic neu…

This research presents LocalN2NM, a novel method for inferring neural signed distance functions (SDF) from single, noisy point clouds by finetuning data-driven priors, achieving faster inference and b…

Contrastive learning enables efficient probabilistic inference in high-dimensional time series by creating Gaussian representations that form a Gauss-Markov chain, allowing for closed-form solutions t…

SRNNs reveal behaviorally-relevant neural dynamics switches!

This paper proves that inexact ALMs applied to SDPs achieve linear convergence for both primal and dual iterates, contingent solely on strict complementarity and a bounded solution set, thus resol…

Injecting watermarks into LLM outputs while speeding up generation is impossible; this paper proves this trade-off and offers methods prioritizing either watermark strength or speed.

Multi-task learning and finetuning show surprising feature reuse biases, including a novel ’nested feature selection’ regime where finetuning prioritizes a sparse subset of pretrained features, signif…

INDICT, a novel framework, empowers LLMs with internal dialogues of critiques to enhance code generation, prioritizing both safety and helpfulness, resulting in +10% absolute improvement across variou…

IsCiL: a novel adapter-based continual imitation learning framework that efficiently adapts to new tasks by incrementally learning and retrieving reusable skills.

Meta-learning enhances human mesh recovery by unifying training and test-time objectives, significantly improving accuracy and generalization.

IGNITE improves offline optimization by incorporating surrogate gradient norm to reduce model sharpness, boosting performance up to 9.6%

Cost-robust contracts, inspired by statistical hypothesis tests, incentivize quality in LLM text generation, overcoming the moral hazard of pay-per-token models.