All papers

The paper proposes a self-distillation method weighting loss by judged reliability and student predictive uncertainty on open-ended reasoning tasks.

The paper presents a framework to steer language models using multimodal context to expand abbreviated text inputs while preserving user intent.

The paper introduces a diagnostic pipeline to identify low-confidence patient subgroups and detect harmful covariate shifts in clinical prediction models.

The paper uses intermediate transformer hidden states instead of semantic embeddings to improve classifier-based prompt routing across unseen tasks.

The paper derives statistical confidence bounds on coverage regret for selecting budget-constrained document subsets in retrieval systems under adaptive auditing.

The paper presents a method to retrofit attention models with hierarchical test-time neural memory modules to improve long-context inference efficiency.

The paper provides theoretical conditions and algorithms for preserving causal graph identifiability when working with missing data distributions.

The paper introduces a framework using psychometric dimensions and player personas to benchmark how language models evaluate and refine video games.

The paper analyzes whether syntactic attention heads in transformers play a direct causal role in syntactic language predictions.

The paper develops an efficient two-pass scheduling method to update nearest-neighbor graphs in SE(3)-equivariant structure prediction models.

The paper proposes adding an explicit copy-paste prediction head to language models to improve in-context retrieval and copying efficiency.

The paper introduces a framework to audit cultural, geographic, and user-representation biases embedded within popular machine learning benchmarks.

The paper evaluates recurrent neural networks and feature attribution methods for tree-species classification using satellite time-series imagery.

The paper establishes theoretical bounds and sample complexity laws for models generalizing across varying input sizes in structured data.

The paper proposes training language models on hierarchically clustered semantic identifiers instead of standard ICD codes for clinical coding.

The paper provides a theoretical generalization and stability analysis of biased stochastic gradient methods for minimax optimization problems.

The paper details the development, training data criteria, and performance of the open Japanese sovereign models, LLM-jp-4.

The paper evaluates the reliability, score variance, and consistency of automated peer reviews generated by large language models.

The paper mathematically analyzes long-sequence generation in transformers, establishing temperature thresholds required for internal cache convergence.

The paper introduces a test showing that steering vectors designed for competing ethical frameworks behave interchangeably in open-weight models.

The paper analyzes Best-of-N sampling as an extreme-value competition between tails of correct and incorrect responses to explain reward hacking.

The paper introduces an internal probing method to detect whether safety-aligned language models still retain accessible representations of hazardous concepts.

The paper evaluates whether language models genuinely benefit from compressed memory by comparing them against memory-free baseline models.

The paper introduces a speech dataset spanning 675 underrepresented languages to support speech recognition and text-to-speech technologies.

The paper introduces a benchmark to evaluate how well language models recognize, infer, and coin newly emerging Korean terms.

The paper presents a systematic framework for selecting optimal parallelism strategies to balance latency and throughput during distributed language model serving.

The paper proposes a Bayesian cue-combination model to explain how vision-language models integrate conflicting visual and textual evidence.

The paper proposes a hybrid projection and simulation method to automatically configure distributed training parallelisms across GPU clusters.

The paper introduces a benchmark to evaluate how well language models understand legal philosophy and apply it to legal reasoning.

The paper introduces a multilingual benchmark to evaluate how language models identify financial concepts and extract supporting numerical evidence.

The paper compares physics simulations and chain-of-thought language reasoning for physical problem solving in multimodal language models.

The paper proposes a graph-based method to generate explainable recommendations by identifying influential user peers in graph neural networks.

The paper analyzes five internal representation reading methods to examine how they detect prompt injections in language models.

The paper proposes marginal inverse probability scaling to reduce mode collapse in reinforcement learning with large outcome spaces.

The paper introduces a multi-task benchmark and a masked autoencoder to evaluate transfer learning for wrist-worn inertial measurement units.

The paper combines camera video and insole sensor signals to improve real-time monocular human pose estimation under visual occlusion.

The paper introduces a document-level data augmentation method using discourse structure parsing to preserve logical flow during text generation.

The paper learns 3D shape representations by mapping text hierarchies into hyperbolic space without relying on synthetic 2D renderings.

The paper compares typed probabilistic decision models with large language models on tabular data classification and format robustness.

The paper presents an automated framework to predict suspected drug overdose deaths from investigation reports using expert-guided rationales.

The paper presents a contrastive learning method for time series that unifies morphological similarity and temporal proximity.

The paper introduces a training-free controller that uses token entropy to target and regenerate faulty reasoning steps at test time.

The paper introduces a spatial reasoning benchmark to evaluate how models handle varying reference frames and altered visual evidence.

The paper proposes divergence-routed on-policy distillation to target critical token positions where students diverge from teacher language models.

The paper studies transferring learned predictive structure between continuous-time dynamic graphs and temporal knowledge graphs across distinct representations.

The paper proposes a likelihood-based attribution method to identify the source model of machine-generated text under distribution shifts.

The paper proposes a Bayesian formulation of self-attention to address non-stationary temporal dependencies in long-term time series forecasting.

The paper evaluates an artificial intelligence decision model across cognitive reflection, classical bias benchmarks, and pricing tasks.

The paper proposes a Transformer architecture that plans dynamic cross-layer connections globally across the model based on input tokens.

The paper combines a vision-language model with a Transformer-VAE to generate expressive, human-like gaze dynamics for humanoid robots.

The paper uses sparse autoencoders to establish a shared coordinate system for matching symbols across heterogeneous neuro-symbolic systems.

The paper presents a human-in-the-loop framework providing role-bounded explanations of cyber-physical system resilience graphs using Formal Concept Analysis.

The paper simulates student behavior by modeling individual learner history and constraining simulated actions via a third-person prediction harness.

The paper introduces a benchmark for medical knowledge poisoning and a routing controller that selectively avoids unreliable retrieved evidence during clinical reasoning.

The paper introduces a single-policy optimization framework combining privileged adaptation with stored historical next-token distributions for language model reasoning.

The paper proves computational complexity results for contiguous envy-free cake-cutting and indivisible goods allocations under additive valuations.

The paper presents a benchmark evaluating how reliably language models correct and maintain formatting requirements across changing instructions.

The paper proposes an algorithm using betting-based sequential tests to monitor supermartingale certificate violations during runtime execution of stochastic systems.

The paper proposes a 3D-anchored latent space representation for simultaneous scene reconstruction and view generation.

The paper introduces an uncertainty-driven self-play method where exploring agents target policy weaknesses to learn robust autonomous driving policies.