All papers

This paper audits whether decision models and generative language models can serve as valid synthetic respondents for psychometric surveys.

This paper addresses failure modes in model predictive task sampling for reinforcement learning by using in-context modeling.

This paper studies how language model behavioral propensities generalize out-of-distribution across different elicitation methods.

This paper proposes visual counterfactual on-policy distillation to prioritize visually grounded tokens during vision-language reasoning training.

This paper introduces a probabilistic relaxation method with learnable categorical priors for bilevel multi-source data selection.

This paper proposes a causal contrastive reinforcement learning method using counterfactual rollouts to handle unobserved confounding in offline demonstrations.

This paper proposes a steering framework that dynamically personalizes language model responses using low-overhead activation vectors.

This paper introduces a two-layer masked language model approach to generate constrained Thai poetry in a single decoding pass.

This paper investigates failure modes where recurrent algorithmic reasoning models jump ahead prematurely when generalizing to larger graphs.

This paper studies experimentally how the timing of generative AI access impacts team ideation and creative selection.

This paper provides theoretical convergence guarantees for average-reward reinforcement learning algorithms via mixing policy regularization.

This paper proposes an adaptive selection method that identifies the top candidates under uncertainty using minimal scenario evaluations.

This paper introduces parameter-tied recurrent blocks to reduce model size in tabular foundation models without sacrificing performance.

This paper models preference model alignment as a Stackelberg game to identify comparisons where evaluators are prone to error.

This paper maps the computational complexity landscape of robust decision-making in multi-environment partially observable environments.

This paper analyzes pretraining factors such as masking ratios and token budgets to optimize bidirectional transformer encoders.

This paper fine-tunes a quantized decision model to classify diseases from patient symptom descriptions without free-form text generation.

This paper introduces a benchmark to evaluate continual learning in semi-supervised visual industrial anomaly detection across components.

This paper proposes a tree search method using belief embeddings for planning in partially observable environments.

This paper fine-tunes language models to selectively correct speech recognition transcripts using curated candidate pools.

This paper compares four deep learning segmentation models for detecting structures in dental radiographs under controlled experimental settings.

This paper demonstrates an API and network framework to support split AI inference over 5G mobile networks.

This paper examines how recurrent models generalize and transfer rule-based reasoning across discrete puzzle environments.

This paper proposes a watermarking method embedded during speech synthesis to trace and verify generated audio.

This paper presents a neuro-symbolic framework using language models and formal argumentation semantics to detect persuasiveness in online discussions.

Introduces a variable-length coding method for post-training neural network quantization that maintains accuracy without retraining.

Evaluates a compact bidirectional Mamba state-space autoencoder for classifying heart failure disease states from single-nucleus RNA sequencing data.

Introduces an elastic transformer pretraining architecture enabling simultaneous hyperparameter transfer and subnetwork execution within a single model.

Presents an enhanced convolutional neural network architecture for real-time traffic sign classification in autonomous driving applications.

Proposes an online learning algorithm that decides whether language models should answer queries directly or defer to congested human experts.

Introduces a benchmark evaluating how and when multimodal language models perceive human social cues from partial audio and video snippets.

Critiques trust calibration in human-AI decision making for open-ended problems and outlines dimensions for higher-order system alignment.

Trains chemical retrosynthesis models via reinforcement learning using forward reaction model consistency to credit plausible alternative pathways.

Evaluates commercial language models against Universal Declaration of Human Rights principles, revealing misalignments driven by efficiency arguments.

Combines historical violation decay with language-model triage of online consumer reviews to dynamically schedule restaurant health inspections.

Fine-tunes a coding model using synthetic tutoring dialogues to provide step-by-step guidance without revealing program fixes prematurely.

Studies how depth-based early-exit inference affects language model safety, finding shallow exits resist jailbreaks better than deep exits.

Shows that running fewer layer loops in recurrent-depth models significantly reduces safety refusals while preserving most reasoning performance.

Probes attention heads in unified vision-language models to study condition processing and introduces an inference-time guidance tuning method.

Applies an educational measurement framework to prompt engineering to improve the validity and reproducibility of automated essay scoring systems.

Presents a two-stage post-training strategy that improves language model handling of tabular numerical records using auxiliary pretraining and token matching.

Proposes a data-sharing ecosystem and framework to make angle-resolved photoemission spectroscopy data organized and accessible for AI research.

Introduces a medical benchmark of real de-identified patient edge cases pairing imaging with clinical vignettes to test multimodal language models.

The paper presents RECAP, a post-training quantization method that calibrates reasoning models using dynamic prefixes to reduce error accumulation.

The paper proposes LeHami, incorporating symplectic Hamiltonian dynamics into latent predictive architectures to prevent representation collapse over long rollouts.

The paper presents GeoBridge, an evolutionary framework that searches for reusable, transferable decoder topologies across geometric vision tasks.

The paper presents LENS, a benchmark evaluating language models on North Korean linguistic nuances, sociocultural perspectives, and domain knowledge.

The paper proposes using subordinate color channel intensities in fluorescence microscopy to improve protein crystallization segmentation and classification.

The paper presents Co-DSL, enabling users and language models to co-design reusable spatial concepts to guide layout generation solvers.

The paper introduces a post-training quantization framework for autoregressive visual generation models to retain image quality under low precision.

The paper introduces CHEAR, an activity recognition framework that pre-organizes confusion hierarchies offline to reduce inference costs.

The paper proposes an artificial neural network framework designed to reject uncertain predictions to prevent misclassification errors.

The paper proposes a multi-layer detector ensemble to identify deceptive model outputs while calibrating score uncertainty and false-positive rates.

The paper introduces SuperAI, a framework to merge complementary capabilities across heterogeneous foundation models using low-dimensional subspace alignment.

The paper introduces a rule-based curriculum training system that improves reinforcement learning performance on Craftax without language models.

The paper introduces a zeroth-order influence function framework to enable memory-efficient training data selection using only forward passes.

The paper introduces MultiFlow, a generative framework that models distributions of fMRI neural responses using stimulus-informed flow matching.

The paper presents StimFlow, using conditional flow matching to synthesize fMRI brain activity directly in voxel space from visual stimuli.

The paper develops an attack framework to generate physical adversarial textures that deceive video models across varying trajectories and viewpoints.

The paper presents CodeWalk, a method to create challenging coding benchmarks at low cost by programmatically composing verified atomic problems.