All papers

Investigates whether continuous latent reasoning tokens genuinely compute answers or merely settle early on predetermined conclusions.

Presents a test-time method that extracts structured schemas from corpora to guide generative models toward more diverse outputs.

Proposes a lossy compression framework for federated learning that balances user privacy against model utility on resource-constrained devices.

Introduces a condition-contrastive distillation approach to prevent student language models from learning erroneous teacher continuations.

Examines how extended reasoning steps lead language models to interpret everyday logical terms strictly literally instead of inferring pragmatic meaning.

Proposes conformal calibration to weight criterion-level uncertainty when using rubric-based language model judges for reinforcement learning.

Introduces a two-stage model merging method to enable dynamic multi-objective recommendation trade-offs at inference time without retraining.

Analyzes how prompt variations redistribute internal feature activations in sparse autoencoders and traces their impact on downstream model decisions.

Audits how persistent conversational agents remain with therapeutic crisis referrals after users repeatedly refuse help for suicidal thoughts.

Proposes a fractional power-of-two quantization grid to enable efficient multiplier-free integer inference for large language models.

Introduces a metric to measure how non-linearly features learned by sparse autoencoders interact during downstream model behaviors.

The paper explores whether measuring information density across reasoning steps can predict and guide answer correctness in large language models.

The paper applies heterogeneous graph neural networks directly to bank relational schemas to detect money-mule accounts prior to flag dates.

The paper examines how increasing visual token budgets and cropping key evidence improves document retrieval-augmented generation in vision-language models.

The paper evaluates a constrained multi-agent reinforcement learning method that learns policies inside world models to reduce environment interaction costs.

The paper introduces a method that accelerates on-policy knowledge distillation by truncating student rollouts and applying inverse-probability weighting.

The paper compares quantum and classical machine learning models on multi-cohort fMRI datasets to test their utility for neuropsychiatric classification.

The paper introduces a hierarchical graph convolutional network and dataset to detect internal inconsistencies and headline mismatches in news documents.

The paper introduces an in-context memory method storing latent transitions between visual observations to assist robot manipulation without extra annotation.

The paper presents a Korean dataset extracted from corporate sustainability reports to support information extraction on environmental, social, and governance topics.

The paper introduces a system that retrieves document information into a fixed set of prompt embeddings for frozen language models.

The paper introduces a reinforcement learning method that uses internal certainty profiles and selective filtering to train models without external verifiers.

The paper presents a self-guided reasoning framework where language models generate code to detect and diagnose errors in vehicle routing problem instances.

The paper studies combining classical and quantum representations to determine when multi-view fusion improves machine learning predictions over single representations.

The paper introduces a benchmark to evaluate how language models perform counterfactual reasoning regarding AI data center infrastructure and siting commitments.

The paper analyzes multi-reward reinforcement learning failures where auxiliary objectives degrade primary performance and introduces a trust-anchored gating method.

Presents methods for query-guided video grounding across object localization, relationship detection, and temporal moment identification.

Introduces a distillation method that identifies critical sub-spans in student agent trajectories for teacher revision and continual learning.

Provides theoretical convergence guarantees and finite-sample bounds for advantage-based value sharing in reinforcement learning algorithms.

Presents a full-duplex conversational model that generates real-time speech alongside synchronized facial expressions and body gestures.

Investigates whether training multimodal models on visual generation tasks transfers accessible factual knowledge into text question answering.

Combines domain embeddings with language models to detect and repair incorrect subsumption axioms in industrial knowledge ontologies.

Introduces a benchmark and evaluation to test whether language models can successfully generate coherent nonlinear narratives.

Surveys the application of artificial intelligence methods to humanitarian disaster response and displacement monitoring.

Extracts attention statistics directly during fast attention computation to build lightweight hallucination detectors for language models.

Decomposes the alignment between language model activations and human brain scans using neurocognitively grounded semantic property representations.

Introduces a framework using language models to construct context-sensitive semantic representations that align with human brain activity during language comprehension.

Studies a hierarchical conditioning architecture that predicts acoustic scenes directly from intermediate audio event probabilities.

Introduces an optimization framework that localizes functional subnetworks in language models to enable continual learning without catastrophic forgetting.

Reviews and analyzes reinforcement learning methods used to guide and improve diffusion models across vision, language, and robotics.

The paper investigates the mathematical conditions under which the Muon optimizer's polar update acts as an optimal step.

The authors formulate classification and uncertainty quantification methods for pairwise difference learning frameworks.

The paper compares the Muon optimizer against AdamW across controlled data, wall-clock time, and tuning budgets.

The paper demonstrates vulnerabilities in unilateral weight obfuscation schemes used for on-device secure language model inference.

The paper proposes a search framework to recover executable editing programs between before-and-after sewing pattern pairs.

The authors propose an indoor navigation framework that uses user-provided floor plans and topological memory without prior robot site maps.

The paper introduces an open-source dataset and standardized benchmark suite for machine learning methods in analog integrated circuit design.

The paper introduces a speculative decoding framework that uses early recurrent states to draft tokens in looped transformers.

The study evaluates using language models to directly reconstruct mobile network schemas versus generating validation code tools.

The paper studies whether predicting actions from consecutive latent states helps world models retain essential environmental dynamics.

The authors introduce an audio model with a non-autoregressive decision head for selection, judgment, and scoring tasks.

The paper evaluates feature encoders and pair selection strategies for pairwise difference learning in low-data classification regimes.

The paper proposes filtering high-drift responses to mitigate off-policy degradation in mixture-of-experts reinforcement learning.

The authors evaluate how easily item identities and user profiles can be recovered from language embeddings in recommender systems.

The authors introduce a Fisher-informed adapter merging method to maintain retained task performance during federated language model unlearning.

The paper analyzes why search-augmented language models produce outdated answers even when relevant evidence is retrieved.

Introduces an in-context learning approach to map specified user preferences directly to Pareto-optimal solutions using offline data.

Investigates how partial interventions in concept-based models can unexpectedly reduce downstream task performance due to information leakage.

Introduces a two-stage coarse-to-fine speech separation architecture leveraging visual streams to isolate target speaker audio in noisy environments.

Combines multi-layer vision transformer representations with a fusion decoder to localize pixel-level generative image inpainting manipulations.