Posters

This paper reveals how a one-layer transformer’s training converges for next-token prediction, showing sub-linear convergence for both layers and shedding light on its surprising generalization abilit…

This paper rigorously analyzes biased adaptive stochastic gradient descent (SGD), proving convergence to critical points for non-convex functions even with biased gradient estimations. The analysis c…

NoMAD-Attention achieves up to 2x speedup in 4-bit quantized LLaMA inference on CPUs by replacing computationally expensive multiply-add operations with ultra-low-latency in-register lookups.

Near-optimal algorithm for private resource allocation is introduced, achieving improved accuracy and privacy guarantees.

NoiseGPT uses multi-modal LLMs to detect & fix noisy image labels by identifying probability curvature differences between clean and noisy examples.

Meta-learning and differential privacy combine to enable accurate, well-calibrated private regression, even with limited data, via the novel DPConvCNP model.

This paper introduces InfoNCA and NCA, novel frameworks for language model alignment using noise contrastive estimation, enabling direct optimization from both explicit rewards and pairwise preference…

Noether’s Razor learns conserved quantities and symmetries directly from data via Bayesian model selection, improving dynamical systems modeling accuracy and generalizability.

This paper reveals efficient stochastic bandit algorithms for maximizing M-concave functions and proves NP-hardness for adversarial full-information settings.

This paper solves the open problem of achieving no-regret learning in online multi-agent Nash social welfare maximization.

A novel stochastic bandit algorithm using soft tree ensemble models achieves lower cumulative regret than existing ReLU-based neural bandit algorithms, offering a constrained yet effective hypothesis …

Self-supervised gradients boost frozen deep learning model performance!

Deep RL agents trained under non-stationarity suffer performance collapse due to representation degradation; this work reveals this in PPO and introduces Proximal Feature Optimization (PFO) to mitigat…

AI agents learn better with well-designed training environments. This paper reveals flaws in current environment-selection methods and introduces Sampling for Learnability (SFL), a new approach that …

No free lunch for adversarial optimization: This paper proves that no single algorithm universally outperforms others when finding Nash Equilibrium, introducing black-box complexity analysis to estab…

LLM watermarking faces inherent trade-offs; this paper reveals simple attacks exploiting common design choices, proposing guidelines and defenses for more secure systems.

Passive data collection in complex social systems invalidates standard AI model validation; new methods are needed.

Contrastive vision-language models (VLMs) trained only on English data significantly underperform on culturally diverse benchmarks. This paper reveals this bias, proposes novel evaluation metrics, and…

Multimodal models’ impressive ‘zero-shot’ performance hinges on the frequency of concepts in their training data, not inherent generalization ability; exponentially more data is needed for linear impr…

Nimbus achieves 2.7-4.7x speedup in BERT base inference using novel two-party computation techniques for efficient matrix multiplication and non-linear layer approximation.